text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
author:
- |
James Bannon[^1]\
Wenbo Song\
Brad Windsor\
Tao Li\
bibliography:
- 'ref.bib'
title: |
Causality and Batch Reinforcement Learning:\
Complementary Approaches To Planning In Unknown Domains
---
[^1]: Equal contribution
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In supersymmetric theories of nature the Higgsino fermionic superpartner of the Higgs boson can arise as the lightest standard model superpartner depending on the couplings between the Higgs and supersymmetry breaking sectors. In this letter the production and decay of Higgsino pairs to the Goldstone fermion of supersymmetry breaking and the Higgs boson, $h$, or gauge bosons, $Z$ or $\gamma$ are considered. Relatively clean di-boson final states, $hh$, $h \gamma$, $hZ$, $Z \gamma$, or $ZZ$, with a large amount of missing energy result. The latter channels provide novel discovery modes for supersymmetry at high energy colliders since events with $Z$ bosons are generally rejected in supersymmetry searches. In addition, final states with real Higgs bosons can potentially provide efficient channels to discover and study a Higgs signal at the Fermilab Tevatron Run II.'
author:
- |
\
[Konstantin T. Matchev]{}\
[Theoretical Physics Department]{}\
[Fermi National Accelerator Laboratory]{}\
[Batavia, IL 60510]{}\
\
[Scott Thomas]{}\
[Physics Department]{}\
[Stanford University]{}\
[Stanford, CA 94305]{}\
title: |
\
\
\
Higgs and $Z$-boson Signatures of Supersymmetry
---
\#1\#2\#3\#4\]\#5
\#1 \#2 \#3 \#4 [Nucl. Phys. [**B\#1**]{}, \#2 (\#3)\#4 ]{} \#1 \#2 \#3 \#4 [Phys. Lett. [**B\#1**]{}, \#2 (\#3)\#4 ]{} \#1 \#2 \#3 \#4 [Phys. Rev. [**D\#1**]{}, \#2 (\#3)\#4 ]{} \#1 \#2 \#3 \#4 [Phys. Rev. Lett. [**\#1**]{}, \#2 (\#3)\#4 ]{} \#1 \#2 \#3 \#4 [Phys. Rept. [**\#1**]{}, \#2 (\#3)\#4 ]{} \#1 \#2 \#3 \#4 [Mod. Phys. Lett. [**A\#1**]{}, \#2 (\#3)\#4 ]{} \#1 \#2 \#3 \#4 [Z. Phys. [**C\#1**]{}, \#2 (\#3)\#4 ]{}
Introduction
============
Supersymmetry (SUSY) provides perhaps the best motivated extension of the Standard Model. Spontaneous SUSY breaking leads naturally to radiative electroweak symmetry breaking with masses of order the electroweak scale for the superpartners of the Standard Model (SM) particles. If the messenger interactions which couple the SM superpartners to the SUSY breaking sector are stronger than gravity, the lightest supersymmetric particle (LSP) is the Goldstone fermion of supersymmetry breaking, the Goldstino $\tilde G$. The next to lightest supersymmetric particle (NLSP) is generally the lightest SM superpartner. If the intrinsic scale of supersymmetry breaking is below $\sim 10^3$ TeV the NLSP can decay to its SM partner and the Goldstino on laboratory length scales [@goldstino]. This has an important impact on experimental SUSY signatures at high energy colliders. Since superpartners are generally produced in pairs, these decays give rise to final states with two hard partons and missing energy (${\rlap{\,/}E_T}$) carried by the Goldstino pair, and with possibly other partons in the final state from cascade decays to the NLSP [@goldstino; @supp; @review].
The identity of the NLSP determines the type of final states which arise from decay to the Goldstino [@review]. A neutralino NLSP, $\tilde \chi_1^0$, which is gaugino-like, can decay by $\tilde \chi_1^0 \rightarrow \gamma \tilde G$, leading to final states with $\gamma \gamma {\rlap{\,/}E_T}$. A slepton NLSP, $\tilde \ell$, can decay by $\tilde \ell \rightarrow \ell \tilde G$, giving $\ell\ell {\rlap{\,/}E_T}$ final states. In this letter we consider in detail the possibility of a fermionic Higgsino-like neutralino NLSP. Because it is the superpartner of the Higgs boson, $h$, a Higgsino NLSP can decay by $\tilde \chi_1^0 \rightarrow h \tilde G$. In addition, since the longitudinal component of the $Z$ boson mixes with the Goldstone mode of the Higgs field, $\tilde \chi_1^0 \rightarrow Z \tilde G $ can also result. Because of a strong phase space suppression of the $h$ and $Z$ final states near threshold, decay to a photon can also be important for Higgsinos not too much heavier than the $Z$ boson. Pair production of Higgsinos which decay to Goldstinos can then give rise to the di-boson final states $(hh,h \gamma, hZ, Z \gamma, ZZ) {\rlap{\,/}E_T}$ [@talks].
Di-boson signatures which include Higgs and $Z$ bosons and ${\rlap{\,/}E_T}$ are quite novel discovery modes for supersymmetry in the mass range accessible to the current generation of high energy collider experiments. In conventional SUSY signatures, in which the lightest neutralino, $\tilde \chi_1^0$, is assumed to escape the detector without decay to the Goldstino, the mass splittings between supersymmetric particles required in order for $h$ or $Z$ to arise in a cascade decay, typically imply the superpartners are too heavy to be produced in sufficient numbers at present colliders. For this reason events with reconstructed $Z$ bosons are in fact generally rejected in present SUSY searches. However, since the Goldstino is essentially massless, sufficient phase space is available for the $h {\tilde{G}}$ and $Z {\tilde{G}}$ modes for a Higgsino somewhat heavier than $h$ or $Z$. And this mass range will be accessible at the upcoming Run II at the Fermilab Tevatron. The Higgs final states also present the exciting possibility of discovering and studying the Higgs boson in association with supersymmetry.
If supersymmetry is broken at a low scale, as required for the di-boson sigatures discussed here, it is very likely that the SM gauge interactions play some role in coupling the SUSY breaking sector to the SM superpartners [@supp]. However, such gauge-mediated SUSY breaking requires additional interactions between the Higgs and SUSY breaking sectors in order to break certain Higgs sector global symmetries and obtain acceptable electroweak symmetry breaking [@review]. These interactions can modify the Higgsino mass from minimal expectations, and allow for a Higgsino NLSP. So searches for di-boson signatures of a Higgsino NLSP within theories of low scale gauge-mediated SUSY breaking are very well motivated as possible indirect probes for the existence of these additional couplings.
Higgsino decays and production
==============================
The Higgsinos $\tilde H_u$ and $\tilde H_d$ are fermionic superpartners of the Higgs boson fields $H_u$ and $H_d$. The neutral Higgsinos mix with the gaugino superpartners of the $\gamma$ and $Z$ gauge bosons, while the charged Higgsino mixes with the gaugino superpartner of the $W$ gauge boson. In the limit relevant here, in which the gauginos are heavier than the Higgsinos, the two lightest neutralinos and lightest chargino, $\tilde \chi_1^0$, $\tilde \chi_2^0$, $\tilde \chi_1^{\pm}$, are predominantly Higgsino and approximately degenerate. The splitting between these states is on the order of 10-15 GeV for masses in the range 120-250 GeV discussed below. If the $U(1)_Y$ and $SU(2)_L$ gaugino mass parameters, $M_1$ and $M_2$, have the same sign, ${\rm sgn}(M_1 M_2)=+$ then $\tilde \chi_1^0$ is the NLSP. For ${\rm sgn}(M_1 M_2) = -$ it is however possible in certain regions of parameter space that $\tilde \chi_1^{\pm}$ is the NLSP. In this letter only a $\tilde \chi_1^0$ NLSP, which leads to the interesting di-boson signatures, will be considered.
The branching ratios ${\rm Br}(\tilde \chi^0_1 \rightarrow \tilde G + (\gamma,h,Z))$ are determined by the Higgsino and gaugino content of $\tilde \chi_1^0$ [@review; @BR-NLSP].
=3.5in
This is illustrated in Fig. \[br\] as a function of the neutralino mixing angle $\tan^{-1}(\mu/M_1)$ for fixed $\tilde \chi_1^0$ mass, where $\mu$ is the Higgsino mass parameter, and $\tan \beta = v_u / v_d$ is the ratio of Higgs expectation values. For definiteness the Higgs decoupling limit in which decays to the heavy scalar and pseudoscalar Higgs bosons, $H$ and $A$, are kinematically blocked is employed throughout. For gaugino-like $\tilde \chi_1^0$ the $\gamma$ mode dominates, but for Higgsino-like $\tilde \chi_1^0$ the $h$ and $Z$ modes become important. The dependence on ${\rm sgn}(\mu)$ and $\tan \beta$ apparent in Fig. \[br\] can be understood in terms of the $\tilde \chi_1^0$ quantum numbers and couplings and will be presented elsewhere.
The branching ratios also depend on the $\tilde \chi_1^0$ mass through the phase space available to the $h$ and $Z$ modes which suffer a $\beta^4$ velocity suppression near threshold [@review; @BR-NLSP]. So even a Higgsino-like $\tilde \chi_1^0$ decays predominantly by $\tilde \chi_1^0 \rightarrow \gamma \tilde G$ for masses not too far above the $h$ and $Z$ masses.
=3.5in
The mass dependence of the branching ratios is illustrated in Fig. \[sigbr\] in which the $p \bar{p}$ signal cross section times branching ratio into the di-boson final states is given as a function of the $\tilde \chi_1^0$ mass for fixed Higgsino-neutralino mixing. With $\tilde \chi_1^0$ Higgsino-like the $hh$, $ZZ$, or $hZ$ modes dominate for very large masses, while the $\gamma \gamma$ mode dominates for smaller masses. However, because of the strong phase space suppression near threshold there is a transition region which extends over a significant range of mass between these limits in which the mixed final states $\gamma h$ and/or $\gamma Z$ (depending on ${\rm sgn(\mu)}$ and $\tan \beta)$ are important. These final states are particularly useful for masses in the transition region since the photon is quite hard.
The total cross section $\sigma_{tot}(p \bar{p} \rightarrow \tilde{\chi}_i^0 \tilde{\chi}_j^0,
\tilde{\chi}^{+}_1 \tilde{\chi}^{-}_1)$ for $i,j=1,2$ in Fig. \[sigbr\], summed over all the Higgsino-like states, is the relevant signal cross section since these states are approximately degenerate, and can all be produced at similar rates. The heavier states cascade decay to $\tilde \chi_1^0$ through neutral and charged current interactions. The partons from these cascade decays are relatively soft and probably not particularly useful at the trigger level.
$Z$ boson final states
======================
The final states with a $Z$ boson can be significant for large $\tan \beta$, or at small $\tan \beta$ with $\mu > 0$. The $Z$ boson can decay invisibly, leptonically, or hadronically, $Z \rightarrow \nu \nu, \ell\ell, jj$, leading to many possible signatures. The $ee$ and $\mu \mu$ leptonic decays allow the possibility of precise reconstruction of the $Z$ invariant mass, but suffer from small branching ratio, ${\rm Br}(Z \rightarrow ee,\mu \mu) \simeq 6.7 \%$. In contrast, the invisible and hadronic decay modes can be useful because of larger branching ratios, ${\rm Br}(Z \rightarrow \nu \nu) \simeq 20 \%$, and ${\rm Br}(Z \rightarrow jj) \simeq 70 \%$.
The $\gamma Z {\rlap{\,/}E_T}$ di-boson mode dominates the total cross section in the transition region of masses as shown in Fig. 2(b). Leptonic decay of the $Z$ provides the cleanest final state, $\gamma \ell^+ \ell^- {\rlap{\,/}E_T}$, which is similar to existing SM $Z\gamma$ studies without ${\rlap{\,/}E_T}$ [@Zgamma_CDF; @Zgamma_D0]. For a Higgsino search, however, an additional large ${\rlap{\,/}E_T}$ cut, as well as a more stringent photon $E_T$ cut should reduce the backgrounds to a negligible level. Our Monte Carlo estimates indicate that this channel is practically background free, but is limited by the small leptonic branching ratio of the $Z$ boson. The Tevatron Run IIa with $2$ fb$^{-1}$ of integrated luminosity will have a reach at the 3$\sigma$ discovery level for $\tilde \chi_1^0$ masses up to 155 GeV for the parameters of Fig. 2(b), while the reach in Run IIb with $30$ fb$^{-1}$ should approach 220 GeV.
Invisible decay of the $Z$ gives rise to the signature $\gamma {\rlap{\,/}E_T}$. This channel has been studied in Run I as a probe for anomalous $\gamma Z$ couplings [@photon_D0; @diboson_review_D0]. Backgrounds include $\gamma j$ and $jj$ with one jet faking a photon [*and*]{} in each case the remaining jet energy mismeasured to be below the minimum pedastool. The largest background in Run I was from single $W$ production with $W \rightarrow e \nu$ and the electron misidentified as a photon. This background can be substantially reduced by raising the photon $E_T$ and ${\rlap{\,/}E_T}$ cuts above 50 GeV, beyond the Jacobian peak for $W\rightarrow \ell\nu$ [@Greg]. This also reduces the hadronic background. The 3$\sigma$ discovery reach in $\chi_1^0$ mass should then approach 150 (185) GeV in Run IIa (IIb) for the parameters of Fig. \[sigbr\](b).
Hadronic decay of the $Z$ in the $\gamma Z{\rlap{\,/}E_T}$ mode gives rise to the signature $\gamma jj {\rlap{\,/}E_T}$. Backgrounds are similar to those of the $\gamma {\rlap{\,/}E_T}$ channel. The $\gamma jj {\rlap{\,/}E_T}$ channel has been studied in Run I in order to place limits on squark and gluino masses in very specific supersymmetric models [@photon_jets_D0]. Further background suppressions not included in the Run I study are possible with acoplanarity, sphericity and invariant dijet mass cuts to reconstruct the $Z$ boson, and a lepton veto. In any case, the total background is expected to be smaller than for the $\gamma{\rlap{\,/}E_T}$ channel, due to the presence of two additional hard partons. Given the significant $Z$ hadronic branching ratio, the $\gamma jj {\rlap{\,/}E_T}$ channel should provide somewhat better reach than the $\gamma \ell^+\ell^- {\rlap{\,/}E_T}$ or $\gamma {\rlap{\,/}E_T}$ channels in Run II.
The $ZZ {\rlap{\,/}E_T}$ di-boson mode dominates at larger $\tilde \chi_1^0$ mass as shown in Fig. 2(b). Leptonic decay of each $Z$ boson gives rise to the spectacular signature $\ell^+\ell^-\ell^{\prime +} \ell^{\prime -} {\rlap{\,/}E_T}$, with the lepton pairs reconstructing the $Z$ mass (in one choice of pairing for $\ell=\ell^{\prime}$). This channel is expected to be essentially background free, but suffers from small leptonic branching ratio. Because of this Run IIb will not be sensitive to this channel for the paramters of Fig. \[sigbr\]. But for $\mu / M_1 = 1/3$ and $\tan \beta = 3$ with larger ${\rm Br}(\tilde \chi_1^0 \rightarrow Z \tilde G)$ (c.f. Fig. \[br\]), the 3$\sigma$ discovery reach in Run IIb for the $\tilde \chi_1^0$ mass is 170 GeV. At the LHC $\ell^+\ell^-\ell^{\prime +} \ell^{\prime -} {\rlap{\,/}E_T}$ would represent the gold plated channel for the $ZZ{\rlap{\,/}E_T}$ di-boson mode from Higgsino decay.
Hadronic decay of one of the $Z$ bosons gives the signature $\ell^+\ell^-jj {\rlap{\,/}E_T}$. An important background in this channel comes from $t\bar{t}$ production with $t \rightarrow Wb$ and $W \rightarrow \ell \nu$ with the $\ell^+\ell^-$ pair reconstructing the $Z$ mass, and each $b$-jet not identified as a heavy flavor. Other backgrounds arise from $ZZ$ and $WZ$ in association with jets. In Run IIb the 3$\sigma$ discovery reach in $\tilde \chi_1^0$ mass should approach 195 GeV for $\mu / M_1 = 1/3$ and $\tan \beta =3$. Rejecting backgrounds for the other decay channels of the $ZZ{\rlap{\,/}E_T}$ di-boson mode presents more serious challenges.
Higgs boson final states
========================
The decay of Higgsinos to real Higgs bosons gives perhaps the most interesting di-boson final states because of the opportunity to study both supersymmetry and the Higgs sector. Higgs boson final states are important for small $\tan \beta$ and $\mu <0$ or for large $\tan \beta$ with sufficiently large $\tilde \chi_1^0$ mass, as shown in Figs. \[br\] and \[sigbr\].
In the transition region of $\tilde \chi_1^0$ mass, $\gamma h {\rlap{\,/}E_T}$ is the most important di-boson mode. With the dominant decay $h \rightarrow bb$ this leads to the signature $\gamma bb {\rlap{\,/}E_T}$. Backgrounds include $Z \gamma j$ and $Z jj$ with $Z \rightarrow bb$ and $bb \gamma j$ and $bbjj$ with one jet misidentified as a photon [*and*]{} in each case the remaining jet energy mismeasured to be below the minimum pedastool. Based on the work presented here [@talks] it has been estimated [@report] that with a single $b$-tag the 3$\sigma$ discovery reach in $\tilde \chi_1^0$ mass should approach 210 (250) GeV in Run IIa (IIb) for the parameters of Fig. \[sigbr\](a).
For larger $\tilde \chi_1^0$ masses the $hZ{\rlap{\,/}E_T}$ and/or $hh{\rlap{\,/}E_T}$ modes can become important, as shown in Fig. \[sigbr\]. The $hh {\rlap{\,/}E_T}$ di-boson final state gives rise to the signature $bbbb {\rlap{\,/}E_T}$. The sizeable QCD and electroweak backgrounds to this final state can be significantly reduced by requiring at least 3 tagged $b$-jets with large invariant mass for two $b$-jet pairs [@talks], as verified by Monte Carlo simulation [@bbbb]. Remaining backgrounds include $ZZj$ with each $Z \rightarrow bb$, $bbjj$ with one jet misidentified as a $b$-jet, and $bbbbj$ with in each case the jet energy mismeasured to be below the minimum pedastool, and $t \bar{t}$ production with $t \rightarrow Wb$ and one hadronic decay $W \rightarrow jj$ with one jet misidentified as a $b$-jet, and one leptonic decay $W \rightarrow \ell \nu$ with $\ell$ not identified. Accounting for the $t\bar{t}$ background [@bbbb], the 3$\sigma$ discovery reach in $\tilde \chi_1^0$ mass at Run IIb should approach 240 GeV for the parameters of Fig. \[sigbr\](a).
The $hZ{\rlap{\,/}E_T}$ di-boson mode arising from Higgsino decay is similar to direct $hZ$ production. Invisible decay of the $Z$ gives the signature $bb {\rlap{\,/}E_T}$, and would contribute slightly to searches for the SM Higgs boson in this channel. Leptonic decay of the $Z$ gives the signature $\ell^+\ell^-bb{\rlap{\,/}E_T}$. Unfortunately, the dominant background from $t\bar{t}$ production with $t\rightarrow Wb$ and $W \rightarrow \ell \nu$ with the $\ell^+\ell^-$ pair reconstructing the $Z$ mass, is very similar to the signal. Because of this, Run II is not expected to be sensitive to this channel. Hadronic decay with $Z \rightarrow bb$ gives the signature $bbbb {\rlap{\,/}E_T}$, similar to the $hh{\rlap{\,/}E_T}$ mode. However, because of the smaller branching ratio, ${\rm Br}(Z \rightarrow bb) / {\rm Br}(h \rightarrow bb) \simeq
20\%$, Run II will just marginally not be sensitive to the $hZ{\rlap{\,/}E_T}$ mode in this channel for the parameters of Fig. \[sigbr\].
The Higgs boson final states of Higgsino decay discussed above present the possibility of collecting a relatively clean sample of events which contain real Higgs bosons. It is therefore interesting to consider the reach as a general function of both Higgsino and Higgs masses. The total cross section times branching ratio contours for the $\gamma bb {\rlap{\,/}E_T}$ and $bbbb {\rlap{\,/}E_T}$ channels as a function of the $h$ and $\tilde \chi_1^0$ masses are shown in Fig. 3.
=3.5in
These contours include ${\rm Br}(\chi_1^0 \to (\gamma, h) \tilde G)$ for $\tan\beta=3$ and $\mu/M_1=-3/4$ and SM values for ${\rm Br}(h \to bb)$. The Run IIa 3$\sigma$ dicovery reach quoted above for the $\gamma bb {\rlap{\,/}E_T}$ channel corresponds to a signal times branching ratio cross section of 5 fb. For the parameters of Fig. 3 this corresponds to a Higgs mass of up to at least 120 GeV for $\tilde \chi_1^0$ masses in the range 135-200 GeV, with a maximum reach in Higgs mass of just over 130 GeV. This is to be contrasted with the search for the SM Higgs from direct $Wh$ and $Zh$ production. These SM channels are background limited, and no sensitivity to a Higgs mass beyond current limits is expected in Run IIa [@higgsreport]. So the $\gamma bb {\rlap{\,/}E_T}$ channel presents the interesting possibility for Run IIa of a SUSY signal which contains real Higgs bosons. The Run IIb 3$\sigma$ dicovery reaches quoted above for the $\gamma bb {\rlap{\,/}E_T}$ and $bbbb {\rlap{\,/}E_T}$ channels correspond to signal times branching ratio cross sections of 1 fb and 4 fb respectively. For the parameters of Fig. 3 the maximum reach in Higgs mass then corresponds to just over 145 GeV and 115 GeV respectively.
In order to identify the Higgs boson directly in a sample of events arising from Higgsino decays it is necessary to observe a peak in the $bb$ invariant mass. The identifiable di-boson final states and large ${\rlap{\,/}E_T}$ carried by the Goldstinos render the supersymmetric Higgs boson final states discussed here relatively clean. Reconstructing the Higgs mass peak should be relatively straightforward compared to SM $Wh$ and $Zh$ production modes which suffer from much larger continuum $bb$ backgrounds.
All the new signatures presented here involve hard photons, leptons, and/or $b$-jets, in association with significant missing energy. New triggers are therefore not required, but final state specific off line analysis should be implemented in order to search for supersymmetry and/or the Higgs boson in these interesting channels.
Finally, Higgsino decay with a measurable macroscopic decay length to the Goldstino would render all the di-boson final states discussed here essentially background free. A search for such final states requires a special analysis for displaced $\ell^+ \ell^-$, $jj$, or $bb$ with large invariant mass and approximately uniform angular distribution with respect to the beam axis [@mesino].
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank R. Demina, G. Landsberg, and J. Wells for invaluable discussions. The work of K. M. was supported by Fermilab which is operated under DOE contract DE-AC02-76CH03000, and that of S. T. by the US National Science Foundation under grant PHY98-70115, the Alfred P. Sloan Foundation, and Stanford University through the Frederick E. Terman Fellowship.
[99]{}
S. Dimopoulos, M. Dine, S. Raby, and S. Thomas, [hep-ph/9601367]{}, Phys. Rev. Lett. [**76**]{} 3494 (1996).
S. Dimopoulos, M. Dine, S. Raby, S. Thomas, and J. Wells, [hep-ph/9607450]{}, Nucl. Phys. Proc. Suppl. A [**52**]{} 38 (1997).
For a review of hadron collider signatures of low scale supersymmetry breaking see S. Dimopoulos, S. Thomas, and J. D. Wells, [hep-ph/9609434]{}, Nucl. Phys. B [**488**]{} 39 (1997).
K. Matchev, talk given at SUSY’98, July 11-17 1998, Oxford, England, [http://hepnts1.rl.ac.uk/SUSY98]{};\
S. Thomas, talks presented at the SUSY/Higgs Run II workshops, Batavia, IL, May 14-16 1998, September 3-4 1998.
S. Ambrosanio, G. L. Kane, G. D. Kribs, S. P. Martin, and S. Mrenna, [hep-ph/9605398]{}, Phys. Rev. D [**54**]{} 5396 (1996);\
J. Bagger, K. Matchev, D. Pierce and R.-J. Zhang, [hep-ph/9609444]{}, Phys. Rev. D [**55**]{} 3188 (1997).
F. Abe [*et al.*]{}, 74 1941 1995 .
B. Abbott [*et al.*]{}, [hep-ex/9710031]{}, Phys. Rev. D [**57**]{} R3817 (1998).
S. Abachi [*et al.*]{}, [hep-ex/9702011]{}, 78 3640 1997 .
S. Abachi [*et al.*]{}, 56 6742 1997 .
G. Landsberg, private communication.
B. Abbott [*et al.*]{}, [hep-ex/9808010]{}. Phys. Rev. Lett. [**82**]{} 29 (1999).
J. Qian, [hep-ph/9903548]{}, FERMILAB-FN-680; Report of the Low-Scale Supersymmetry Breaking working group, conveners R. Culbertson, S. Martin, J. Qian, and S. Thomas, [http://fnth37.fnal.gov/gm/gmrun2.html]{}.
H. Baer, P. Mercadante, X. Tata, and Y. Wang, [hep-ph/9903333]{}, Phys. Rev. D [**60**]{} 055001 (1999).
Report of the Higgs Working Group, conveners M. Carena, J. Conway, H. Haber, and J. Hobbs, [http://fnth37.fnal.gov/higgs.html]{}.
U. Sarid and S. Thomas, SU-ITP-98-70.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We develop a representation suitable for the unconstrained recognition of words in natural images, where unconstrained means that there is no fixed lexicon and words have unknown length.
To this end we propose a convolutional neural network (CNN) based architecture which incorporates a Conditional Random Field (CRF) graphical model, taking the whole word image as a single input. The unaries of the CRF are provided by a CNN that predicts characters at each position of the output, while higher order terms are provided by another CNN that detects the presence of N-grams. We show that this entire model (CRF, character predictor, N-gram predictor) can be jointly optimised by back-propagating the structured output loss, essentially requiring the system to perform multi-task learning, and training requires only synthetically generated data. The resulting model is a more accurate system on standard real-world text recognition benchmarks than character prediction alone, setting a benchmark for systems that have not been trained on a particular lexicon. In addition, our model achieves state-of-the-art accuracy in lexicon-constrained scenarios, without being specifically modelled for constrained recognition. To test the generalisation of our model, we also perform experiments with random alpha-numeric strings to evaluate the method when no visual language model is applicable.
author:
- |
Max Jaderberg[^1], Karen Simonyan^\*^, Andrea Vedaldi & Andrew Zisserman^+^\
Visual Geometry Group, Department of Engineering Science, University of Oxford\
`{max,karen,vedaldi,az}@robots.ox.ac.uk`\
title: |
Deep Structured Output Learning for\
Unconstrained Text Recognition
---
Introduction
============
In this work we tackle the problem of *unconstrained text recognition* – recognising text in natural images without restricting the words to a fixed lexicon or dictionary. Usually this problem is decomposed into a word detection stage followed by a word recognition stage. The word detection stage generates bounding boxes around words in an image, while the word recognition stage takes the content of these bounding boxes and recognises the text within.
This paper focuses on the text recognition stage, developing a model based on deep convolutional neural networks (CNNs) ([@Lecun98]). Previous methods using CNNs for word recognition (discussed in more detail in section Section \[sec:related\]) has either constrained ([@Jaderberg14d]) or heavily weighted ([@Bissacco13]) the recognition results to be from a dictionary of known words. This works very well when training and testing are limited to a fixed vocabulary, but does not generalise to where previously unseen or non-language based text must be recognised – for example for generic alpha-numeric strings such as number plates or phone numbers.
The shift of focus towards a model which performs accurately without a fixed dictionary increases the complexity of the text recognition problem. To solve this, we propose a novel CNN architecture (Figure \[fig:path\]) employing a *Conditional Random Field* (CRF) whose unary terms are outputs of a CNN character predictor, which are position-dependent, and whose higher order terms are outputs of a CNN N-gram predictor, which are position-independent. The recognition result is then obtained by finding the character sequence that maximises the CRF score, enforcing the consistency of the individual predictions.
The CRF model builds on our previous work where we explored dictionary-based recognition ([@Jaderberg14c]) for two scenarios: the first was to train a different CNN character classifier for each position in the word being recognised, using the whole image of the word as input to each classifier (an idea also expored by [@Goodfellow13]); the second was to construct a CNN predictor to detect the N-grams contained in the word, effectively encoding the text as a bag-of-N-grams.
The dictionary-free joint model proposed here is trained by defining a structured output learning problem, and back-propagating the corresponding *structured output loss*. This formulation results in multi-task learning of both the character and N-gram predictors, and additionally learns how to combine their representations in the CRF, resulting in more accurate text recognition.
The result is a highly flexible text recognition system that achieves excellent unconstrained text recognition performance as well as state-of-the-art recognition performance when using standard dictionary constraints. While performance is measured on real images as contained in standard text recognition benchmarks, all results are obtained by training the model *purely on synthetic data*. The model is evaluated on this synthetic data as well in order to study its performance under different scenarios.
Section \[sec:related\] outlines work related to ours. Section \[sec:characters\] reviews the character sequence model and Section \[sec:ngrams\] the bag-of-N-grams model. Section \[sec:joint\] shows how these predictors can be combined to form a joint CRF model and formulates the training of the latter as structured-output learning. Section \[sec:eval\] evaluates these models extensively and Section \[sec:conclusion\] summarises our findings.
Related Work {#sec:related}
============
We concentrate here on text recognition methods, recognising from a cropped image of a single word, rather than the text detection stages of scene text recognition (‘text spotting’) that generate the word detections.
Traditional text recognition methods are based on sequential character classification, finding characters by sliding window methods ([@Wang11; @Wang12; @Jaderberg14a], after which a word prediction is made by integrating character classifier predictions in a left-to-right manner. The character classifiers include random ferns ([@Ozuysal07]) in [@Wang11], and CNNs in [@Wang12; @Jaderberg14a]. Both [@Wang11] and [@Wang12] use a small fixed lexicon as a language model to constrain word recognition.
More recent works such as [@Bissacco13; @Alsharif13] make use of over-segmentation methods, guided by a supervised classifier, to generate candidate character proposals in a single-word image, which are subsequently classified as true or false positives. For example, PhotoOCR ([@Bissacco13]) uses binarization and a sliding window classifier to generate candidate character regions, with words recognised through a beam search driven by classifier scores and static N-gram language model, followed by a re-ranking using a dictionary of 100k words. [@Jaderberg14a] uses the convolutional nature of CNNs to generate response maps for characters and bigrams which are integrated to score lexicon words.
In contrast to these approaches based on character classification, the work by [@Almazan14; @Gordo14; @Goel13; @Rodriguez13; @Novikova12; @Mishra12] instead uses the notion of holistic word recognition. [@Mishra12; @Novikova12] still rely on explicit character classifiers, but construct a graph to infer the word, pooling together the full word evidence. [@Rodriguez13] use aggregated Fisher Vectors ([@Perronnin10]) and a Structured SVM framework to create a joint word-image and text embedding. [@Almazan14] and more recently [@Gordo14] also formluate joint embedding spaces, achieving impressive results with minimal training data. [@Goel13] use whole word-image features to recognize words by comparing to simple black-and-white font-renderings of lexicon words. In our own previous work ([@Jaderberg14c; @Jaderberg14d]) we use large CNNs acting on the full word image region to perform 90k-way classification to a dictionary word.
It should be noted that all the methods make use of strong static language models, either relying on a constrained dictionary or re-ranking mechanism. [@Goodfellow13] had great success using a CNN with multiple position-sensitive character classifier outputs (closely related to the character sequence model in Section \[sec:characters\]) to perform street number recognition. This model was extended to CAPTCHA sequences (up to 8 characters long) where they demonstrated impressive performance using synthetic training data for a synthetic problem (where the generative model is known), but we show that synthetic training data can be used for a real-world data problem (where the generative model is unknown).
There have been previous uses of graphical models with back-propagated loss functions for neural networks, such as the early text recognition work of [@Lecun98] to combine character classifier results on image segmentations. Another example is the recent work of [@Tompson14] for human pose estimation, where an MRF-like model over the distribution of spatial locations for each body part is constructed, incorporating a single round of message-passing.
CNN Text Recognition Models
===========================
We now review the component CNN models, originally presented in our tech report [@Jaderberg14c], that form the basis of our joint model in Section \[sec:joint\].
Character Sequence Model Review {#sec:characters}
-------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![(a) The character sequence model. A word image is recognised by predicting the character at each position in the output, spelling out the text character by character. Each positional classifier is learnt independently but shares a jointly optimised set of features. (b) The N-gram encoding model. The recognised text is represented by its bag-of-N-grams. This can be thought of as 10k independently trained binary classifiers using a shared set of jointly learnt features, trained to detect the presence of a particular N-gram.[]{data-label="fig:basenets"}](charnet.png "fig:"){width="0.5\linewidth"} ![(a) The character sequence model. A word image is recognised by predicting the character at each position in the output, spelling out the text character by character. Each positional classifier is learnt independently but shares a jointly optimised set of features. (b) The N-gram encoding model. The recognised text is represented by its bag-of-N-grams. This can be thought of as 10k independently trained binary classifiers using a shared set of jointly learnt features, trained to detect the presence of a particular N-gram.[]{data-label="fig:basenets"}](ngramnet.png "fig:"){width="0.36\linewidth"}
(a) (b)
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
In this section we describe our character sequence model. This model encodes the character at each position in the word and so predicts the sequence of characters in an image region (hereafter we simply refer to the image region as an image). Each position in the word is modelled by an independent classifier acting on a shared set of features from a single CNN. By construction, this model makes no assumptions about the underlying language and allows completely unconstrained recognition.
A word $w$ of length $N$ is modelled as a sequence of characters such that $w = (c_1, c_2, \dots, c_N)$ where each $c_i \in {\mathcal{C}}= \{1,2,\dots,36\}$ represents a character at position $i$ in the word, from the set of 10 digits and 26 letters. Each $c_i$ can be predicted with a single classifier, one for each character in the word. However, since words have variable length $N$ which is unknown at test time, we fix the number of characters to $N_\text{max}$ (here set to 23), the maximum length of a word in the training set, and introduce a null character class. Therefore a word is represented by a string $w \in ({\mathcal{C}}\cup \{\phi\})^{N_\text{max}}$.
For a given input image $x$, we want to return the estimated word $w^*$ which maximises $P(w^*|x)$. Since we seek an unconstrained recognition system with this model, we assume independence between characters leading to $$w^* = \arg\max_{w} P(w|x) = \arg\max_{c_1,c_2,\dots,c_{N_\text{max}}} \prod_{i=1}^{N_\text{max}} P(c_i|\Phi(x))
\label{eqn:wordprob}$$ where $P(c_i|\Phi(x))$ is given by the classifier for the $i$-th position acting on a single set of shared CNN features $\Phi(x)$. The word $w^*$ can be computed by taking the most probable character at each position $
c_i^* = \arg\max_{c_i \in {\mathcal{C}}\cup \{\phi\}} P(c_i|\Phi(x)).
$ The CNN (Figure \[fig:basenets\] (a)) takes the whole word image $x$ as input. Word images can be of different sizes, in particular due to the variable number of characters in the image. However, our CNN requires a fixed size input for all input images. This problem is overcome by simply resampling the original word image to a canonical height and width, without regard to preserving the aspect ratio, producing a fixed size input $x$.
The base CNN has a number of convolutional layers followed by a series of fully connected layers, giving $\Phi(x)$. The full details of the network architecture are given in Section \[sec:implementation\]. $\Phi(x)$ is fed to $N_\text{max}$ separate fully connected layers with 37 neurons each, one for each character class including the null character. These fully connected layers are independently softmax normalised and can be interpreted as the probabilities $P(c_i|\Phi(x))$ of the width-resized input image $x$.
The CNN is trained with multinomial logistic regression loss, back-propagation, and stochastic gradient descent (SGD) with dropout regularisation similar to [@Hinton12].
Bag-of-N-grams Model Review {#sec:ngrams}
---------------------------
This section describes our second word recognition model, which exploits compositionality to represent words. In contrast to the sequential character encoding of Section \[sec:characters\], words can be seen as a composition of an unordered set of character N-grams, a *bag-of-N-grams*. In the following, if $s\in {\mathcal{C}}^N$ and $w\in{\mathcal{C}}^M$ are two strings, the symbol $s \subset w$ indicates that $s$ is a substring of $w$. An $N$-gram of word $w$ is a substring $s \subset w$ of length $|s|=N$. We will denote with $G_N(w) = \{ s : s\subset w \,\wedge\,|s|\leq N\}$ the set of all N-grams of word $w$ of length up to $N$ and with $G_N
=\cup_{w\in{\mathcal{W}}} G_N(w)$ the set of all such grams in the language. For example, $G_3(\texttt{spires})=\{\texttt{s},\texttt{p},\texttt{i},\texttt{r},\texttt{e},\texttt{sp},\texttt{pi},\texttt{ir},\texttt{re},\texttt{es},\texttt{spi},\texttt{pir},\texttt{ire},\texttt{res}\}$. This method of encoding variable length sequences is similar to the *Wickelphone* phoneme-encoding methods ([@Wickelgran69]).
Even for small values of $N$, $G_N(w)$ encodes each word $w\in\mathcal{W}$ nearly uniquely. For example, with $N=4$, this map has only 7 collisions out of a dictionary of 90k words. The encoding $G_N(w)$ can be represented as a $|G_N|$-dimensional binary vector of N-gram occurrences. This vector is very sparse, as on average $|G_N(w)|\approx 22$ whereas $|G_N|=10k$.
Using a CNN we can predict $G_N(w)$ for a word $w$ depicted in the input image $x$. We can use the same architecture as in Section \[sec:characters\], but now have a final fully connected layer with $G_N$ neurons to represent the encoding vector. The scores from the fully connected layer can be interpreted as probabilities of an N-gram being present in the image by applying the logistic function to each neuron. The CNN is therefore learning to recognise the presence of each N-gram somewhere within the input image, so is an N-gram detector.
With the applied logistic function, the training problem becomes that of $|G_N|$ separate binary classification tasks, and so we back-propagate the logistic regression loss with respect to each N-gram class independently. To jointly train a whole range of N-grams, some of which occur very frequently and some barely at all, we have to scale the gradients for each N-gram class by the inverse frequency of their appearance in the training word corpus. We also experimented with hinge loss and simple regression to train but found frequency weighted binary logistic regression was superior. As with the other model, we use dropout and SGD.
In this model we exploit the statistics of our underlying language in choosing a subset of $|G_N|$ N-grams from the space of all possible N-grams to be modelled. This can be seen as using a language model to compress the representation space of the encoding, but is not restraining the predictive capability for unconstrained recognition. While the encoding $G_N(w)$ is almost always unique for words from natural language, non-language words often contain much fewer N-grams from the modelled set $G_N$ leading to more ambiguous and non-unique encodings.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![An illustration of the construction of the path score $S(\texttt{camel},x)$ for the word $\texttt{camel}$. The unary and edge terms used for the score are selected by the path through the graph of character positions shown in the upper right corner. The values of these terms, $S_c(c_i,x)$ and $S_e(s,x)$, where $s \subset w$, are given by the outputs of the character sequence model CNN (CHAR CNN) and the N-gram encoding CNN (NGRAM CNN).[]{data-label="fig:path"}](pathscore.png "fig:"){width="\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
Joint Model {#sec:joint}
===========
In Section \[sec:characters\], maximising the posterior probability of a character sequence is equivalent to maximising the log-score $
\log P(w|x) = S(w,x) = \sum_{i=1}^{N_\text{max}} S_c^i(c_i,x)\label{eqn:wordscore}
$ where $S_c^i(c_i,x)=\log P(c_i|\Phi(x))$ is the logarithm of the posterior probability of the character at position $i$ in the sequence. The graph associated with this function is a set of nodes, one for each unary term $S_c^i(c_i,x)$, and does not contain any edges. Hence maximising the function reduces to maximising each term individually.
The model can now be extended to incorporate the N-gram predictors of Section \[sec:ngrams\], encoding the presence of N-grams in the word image $x$. The N-gram scoring function $S_e(s,x)$ assigns a score to each string $s$ of length $|s| \leq N$, where $N$ is the maximum order of N-gram modelled. Note that, differently from the functions $S_c^i$ defined before, the function $S_e$ is position-independent. However, it is applied repeatedly at each position $i$ in the word: $$S(w,x) =
\sum_{i=1}^{N_\text{max}} S_c^i(c_i,x)
+ \sum_{i=1}^{|w|} \sum_{n=1}^{\min(N,|w|-i+1)}
S_e(c_ic_{i+1}\dots c_{i+n-1},x).
\label{eqn:jointscore}$$ As illustrated in Figure \[fig:path\], the scores $S_c^i(c_i,x)$ are obtained from the CNN character predictors of Section \[sec:characters\] whereas the score $S_e(s,x)$ is obtained from the CNN N-gram predictor of Section \[sec:ngrams\]; note that the N-gram scoring function is only defined for the subset $G_N$ of N-grams modelled in the CNN; if $s\not\in G_N$, the score $S_e(s,x)=0$ is defined to be zero.
The graph associated with the function has cliques of order $N$; hence, when $N$ is even moderately large, we resort to beam search ([@Russel94]) to maximise and find the predicted word $w^*$. Also, the score can be interpreted as a potential function defining a word posterior probability as before; however, evaluating this probability would require computing a normalisation factor, which is non-trivial. Instead, the function is trained discriminatively, as explained in the next section.
#### Structured Output Loss.
The unary and edge score functions $S_c^i(c_i,x)$ and $S_e(s,x)$, should incorporate the outputs of the character sequence model and N-gram encoding model respectively. A simple way to do this is to apply a weighting to the output of the CNNs after removing the softmax normalisation and the logistic loss: $$S(w,x) =
\sum_{i=1}^{N_\text{max}} \alpha_{c_i}^i f_{c_i}^i(x)
+ \sum_{i=1}^{|w|} \sum_{n=1}^{\min(N,|w|-i+1)}
\beta_{c_ic_{i+1}\dots c_{i+n-1}} g_{c_ic_{i+1}\dots c_{i+n-1}}(x),
\label{eqn:weighted}$$ where $f_{c_i}^i(x)$ is the output of the character sequence CNN for character $c_i$ at position $i$ and $g_{s}(x)$ is the output of the N-gram encoding CNN for the N-gram $s$. If desired, the character weights $\alpha = \{\alpha_{c_i}^i\}$ and edge weights $\beta = \{\beta_s\}$ can be constrained to be shared across different characters, character positions, different N-grams of the same order, or across all N-grams.
The sets of weights $\alpha$ and $\beta$ in Equation \[eqn:weighted\], or any weight-constrained variant of Equation \[eqn:weighted\], can be learnt in a structured output learning framework, encouraging the score of the ground-truth word $w_{\text{gt}}$ to be greater than or equal to the highest scoring *incorrect* word prediction plus a margin, $S(w_{\text{gt}},x) \geq \mu + S(w^*,x)$ where $S(w^*,x) = \max_{w\not= w_{\text{gt}}} S(w,x)$. Enforcing this as a soft-constraint results in the convex loss $$L(x_i,w_{\text{gt},i},S)
=\max_{w\not= w_{\text{gt},i}} \max(0, \mu + S(w,x) - S(w_{\text{gt},i},x_i))$$ and averaging over $M$ example pairs $(x_i,w_{\text{gt},i})$ results in the regularised empirical risk objective $$E(S) =
\frac{\lambda_\alpha}{2}\|\alpha\|^2 +
\frac{\lambda_\beta }{2}\|\beta\|^2 +
\frac{1}{M}\sum_{i=1}^M
L(x_i,w_{\text{gt},i},S).
\label{eqn:cost}$$
However, in the general scenario of Equation \[eqn:weighted\], the weights can be incorporated into the CNN functions $f$ and $g$, resulting in the score $$S(w,x) =
\sum_{i=1}^{N_\text{max}} f_{c_i}^i(x)
+ \sum_{i=1}^{|w|} \sum_{n=1}^{\min(N,|w|-i+1)}
g_{c_ic_{i+1}\dots c_{i+n-1}}(x),
\label{eqn:cnnscore}$$ The functions $f$ and $g$ are defined by CNNs and so we can optimise the parameters of them to reduce the cost in Equation \[eqn:cost\]. This can be done through standard back-propagation and SGD. Differentiating the loss $L$ with respect to $S$ gives $$\frac{\partial L(x,w_{\text{gt}},S)}{\partial S(w^*,x)} =
\begin{cases}
1 & \mbox{if } z > 0 \\
0 & \mbox{otherwise}
\end{cases}
\qquad
\frac{\partial L(x,w_{\text{gt}},S)}{\partial S(w_{\text{gt}},x)} =
\begin{cases}
-1 & \mbox{if } z > 0 \\
0 & \mbox{otherwise}
\end{cases}$$ where $z=\max_{w\not= w_{\text{gt},i}} \mu + S(w,x) - S(w_{\text{gt}},x)$. Differentiating the score function of Equation \[eqn:cnnscore\] with respect to the character sequence model and N-gram encoding model outputs $f_{c_i}^i$ and $g_s$ gives $$\frac{\partial S(w,x)}{\partial f_{c}^i} =
\begin{cases}
1 & \mbox{if } c_i = c \\
0 & \mbox{otherwise}
\end{cases},
\qquad
\frac{\partial S(w,x)}{\partial g_{s}} =
\sum_{i=1}^{|w|-|s|+1}
\mathbf{1}_{\{ c_ic_{i+1}\dots c_{i+|s|-1} = s\}}$$ This allows errors to be back-propagated to the entire network. Intuitively, the errors are back-propagated through the CNN outputs which are responsible for margin violations, since they contributed to form an incorrect score.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![The architecture for training the joint model, comprising of the character sequence model (CHAR) and and the N-gram encoding model (NGRAM) with structured output loss. The Path Select Layer generates the score $S(w_{\text{gt}},x)$ by summing the inputs of the groundtruth word. The Beam Search Layer uses beam search to try to select the path with the largest score $S(w^*,x)$ from the inputs. The hinge loss implements a ranking loss, constraining the highest scoring path to be the groundtruth path, and can be back-propagated through the entire network to jointly learn all the parameters.[]{data-label="fig:training"}](learningnet.png "fig:"){width="0.8\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
Using this structured output loss allows the parameters of the entire model to be jointly optimised within the structure imposed by Equation \[eqn:cnnscore\]. Figure \[fig:training\] shows the training architecture used. Due to the presence of high order scores in Equation \[eqn:cnnscore\], it is too expensive to exhaustively search the space of all possible paths to find $w^*$, even with dynamic programming, so instead we use beam search to find the approximate highest scoring path.
The structured output loss described in this section bares resemblance to the discriminative Viterbi training introduced by [@Lecun98]. However, our model includes higher-order terms, terms of a different nature (N-grams), and uses a structured-output formulation. Furthermore, our method incorporates only a very weak language model, limited to assigning a score of 0 to all N-grams outside a target set $G_N$. Note that this does not mean that these N-grams cannot be recognised (this would require assigning to them a score of $-\infty$); instead, it is a smoothing technique that assigns a nominal score to infrequent N-grams.
Evaluation {#sec:eval}
==========
In this section we evaluate the three models introduced in the previous sections. The datasets used for training and testing are described in Section \[sec:datasets\], the implementation details given in Section \[sec:implementation\], and the results of experiments reported in Section \[sec:experiments\].
Datasets {#sec:datasets}
--------
We evaluate our models on a number of standard datasets – ICDAR 2003, ICDAR 2013, Street View Text, and IIIT5k, whereas for training, as well as testing across a larger vocabulary, we turn to the synthetic Synth90k and SynthRand datasets.
[**ICDAR 2003**]{} ([@ICDAR03]) is a scene text recognition dataset, with the test set containing 251 full scene images and 860 groundtruth cropped images of the words contained with the full images. We follow the standard evaluation protocol defined by [@Wang11] and perform recognition on the words containing only alphanumeric characters and at least three characters. The test set of 860 cropped word images is referred to as IC03. The lexicon of all test words is IC03-Full (563 words), and the per-image 50 word lexicons defined by [@Wang11] and used in a number of works ([@Wang11; @Wang12; @Alsharif13]) are referred to as IC03-50.
[**ICDAR 2013**]{} ([@ICDAR2013]) test dataset contains 1015 groundtruth cropped word images from scene text. Much of the data is inherited from the ICDAR 2003 datasets. We refer to the 1015 groundtruth cropped words as IC13.
[**Street View Text**]{} ([@Wang11]) is a more challenging scene text dataset than the ICDAR datasets. It contains 250 full scene test images downloaded from Google Street View. The test set of 647 groundtruth cropped word images is referred to as SVT. The lexicon of all test words is SVT-Full (4282 words), and the smaller per-image 50 word lexicons defined by [@Wang11] and used in previous works ([@Wang11; @Wang12; @Alsharif13; @Bissacco13]) are referred to as SVT-50.
[**IIIT 5k-word**]{} ([@Mishra12]) test dataset contains 3000 cropped word images of scene text downloaded from Google image search. Each image has an associated 50 word lexicon (IIIT5k-50) and 1k word lexicon (IIIT5k-1k).
[**Synth90k[^2]**]{} ([@Jaderberg14c; @Jaderberg14d]) is a dataset of 9 million cropped word images that have been synthetically generated. The synthetic data is highly realistic and can be used to train on and as a challenging test benchmark. The dataset covers 90k different English words, and there are predefined training and test splits with approximately 8 million training images and 900k test images. In addition, we use the same synthetic text engine from [@Jaderberg14c; @Jaderberg14d] to generate word images with completely random strings of up to 10 uniformly sampled alphanumeric characters. We refer to this dataset as [**SynthRand**]{}. The training set consists of 8 million training images and the test set of 900k images. In this corpus there are very few word repetitions (in addition to the random rendering variations). There is a wide range of difficulty in this dataset, from perfectly readable text to almost impossible to read samples.
Implementation Details {#sec:implementation}
----------------------
In the following, the character sequence model is referred to as CHAR, the N-gram encoding model as NGRAM, and the joint model as JOINT.
The CHAR and NGRAM models both have the same base CNN architecture. The base CNN has five convolutional layers and two fully connected layers. The input is a $32 \times 100$ greyscale image obtained by resizing the word image (ignoring its aspect ratio) and then subtracting its mean and dividing by its standard deviation. Rectified linear units are used throughout after each weight layer except for the last one. In forward order, the convolutional layers have 64, 128, 256, 512, and 512 square filters with an edge size of 5, 5, 3, 3, and 3. Convolutions are performed with stride 1 and there is input feature map padding to preserve spatial dimensionality. $2 \times 2$ max-pooling follows the first, second and third convolutional layers. The fully connected layers have 4096 units.
On top of this base CNN, the CHAR model has 23 independent fully connected layers with 37 units, allowing recognition of words of up to $N_\text{max}=23$ characters long. The NGRAM model operates on a selection of 10k frequent N-grams of order $N\leq 4$ (identified as the ones that occur at least 10 times in the Synth90k word corpus, resulting in 36 1-grams, 522 2-grams, 3965 3-grams, and 5477 4-grams). This requires a final fully connected layer on top of the base CNN with 10k units. Therefore, the graph of function has cliques of sizes at most 4. Beam search uses a width of 5 during training and of 10 during testing. If a lexicon is used to constrain the output, instead of performing beam search, the paths associated with the lexicon words are scored with Equation \[eqn:cnnscore\], and the word with the maximum score is selected as the final result.
The three models are all trained with SGD and dropout regularisation. The learning rates are dynamically decreased as training progresses. The JOINT model is initialised with the pre-trained CHAR and NGRAM network weights and the convolutional layers’ weights are frozen during training.
Experiments {#sec:experiments}
-----------
We evaluate our models on a combination of real-world test data and synthetic data to highlight different operating characteristics.
#### N-gram Encoding Results.
The NGRAM model predicts the N-grams contained in input word image. Due to the highly unbalanced nature of this problem (where only 10-20 N-grams are contained in any given image), results are reported as the maximum achieved F-score, computed as the harmonic mean of precision and recall. The latter are computed by sweeping the threshold probability for an N-gram to be classified as present in the word. The maximum achieved F-score on Synth90k is 87.0% and on IC03 is 87.1%. This demonstrates that, while not perfect, the NGRAM model accurately models the presence of N-grams in word images.
-- --------------- ------ --------------------------------------------
Synth90k-test 87.3 **91.0\
& IC03 & 85.9 & **89.6\
& SVT & 68.0 & **71.7\
& IC13 & 79.5 & **81.8\
Synth1-72k & Synth72k-90k & 82.4 & **89.7\
Synth1-45k & Synth45k-90k & 80.3 & **89.1\
SynthRand & SynthRand & **80.7 & 79.5\
**************
-- --------------- ------ --------------------------------------------
: *Left:* The accuracy (%) of the character sequence model, CHAR, and the joint model, JOINT. Different combinations of training and test data are evaluated. Synth$x$-$y$ refers to a subset of the Synth90k that only contains words in the label interval $[x,y]$ (word label indices are in random, non-alphabetical order). Training and testing on completely distinct words demonstrates the power of a general, unconstrained recognition model. *Right:* Some results of the CHAR model on the SynthRand test dataset. Letters in red have been predicted incorrectly with the groundtruth (GT) shown below. Notice the range in difficulty of the SynthRand data.[]{data-label="table:internal"}
#### Character Sequence and Joint Model Results.
The CHAR and JOINT models are evaluated on standard as well as synthetic benchmarks (Table \[table:internal\]), but both models are trained on Synth90k. While the CHAR model achieves good performance, it is consistently outperformed by the JOINT model; the accuracy improvement is as much as +4% on IC03 and SVT, despite the difficulty of the latter. Figure \[fig:results\] shows some example results using the JOINT model.
Next, we evaluate the ability of our model to generalise by recognising words unseen during training. This effectively amounts to zero-shot learning and is a key contribution compared to [@Jaderberg14c; @Jaderberg14d]. In order to do so, the training vocabulary is split into two parts, with one part (50% or 80%) used for training and the other one for evaluation (50% or 20%). In this case the CHAR model is significantly penalised, but the JOINT model can recover most of the performance. For instance, on the 50/50 split, the JOINT model accuracy is 89.1%, only -2% compared to the 91.0% obtained when the training and testing vocabularies are equal.
The final test pushes generalisation by training and testing on completely random strings from SynthRand. As this dataset is a lot less regular than a natural language, the performance of the CHAR model suffers, dropping to 80.7% accuracy. Furthermore, as could be expected form the absence of common N-grams in the random language, the JOINT model performs *slightly worse* at 79.5% accuracy. However this drop is very small because N-grams are not used as hard constraints on the predicted words, but rather to nudge the word scores based on further visual cues.
#### Comparison to the state-of-the-art.
Table \[table:comparison\] compares the accuracy of CHAR and JOINT to previous works. Whereas these works make use of strong language models, our models make minimal assumptions about the language. In the constrained lexicon cases (the starred columns of Table \[table:comparison\]), both CHAR and JOINT are very close to the state-of-the-art DICT model of [@Jaderberg14c; @Jaderberg14d]. Furthermore, if the same 90k dictionary used in by the DICT model is used to constrain the output of the JOINT model, the performance is identical at 93.1% accuracy on IC03. While in the constrained lexicon experiments the lexicon is limited at test time, these results are still remarkable because, differently from DICT, CHAR and JOINT are not trained on a specific dictionary. In particular, DICT would not be able to operate on random strings.
The recognition results without a lexicon are still behind that of some constrained models, however the JOINT model provides competitive performance and is far more flexible to recognise unseen words than previous works, while still achieving state-of-the-art performance if a lexicon is then applied as a constraint at test time. Figure \[fig:results\] shows some example results where the CHAR model does not recognise the word correctly but the JOINT model succeeds.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Results where the unary terms of the JOINT model cannot solely recognise the word correctly, but the addition of the edge scores result in correct recognition, from SVT (a,b) and IC13 (c). The input image is shown in the top left corner. The unary scores for characters (rows) at each position (columns, up to 12 out of 23 characters) are shown, with the selected path using only the unary score term $S^i_c$ (orange) and when edge scores $S_e$ are incorporated (cyan). The bars show the NGRAM strengths, with lighter colour representing larger values.[]{data-label="fig:results"}](res4.png "fig:"){width="0.3\linewidth"} ![Results where the unary terms of the JOINT model cannot solely recognise the word correctly, but the addition of the edge scores result in correct recognition, from SVT (a,b) and IC13 (c). The input image is shown in the top left corner. The unary scores for characters (rows) at each position (columns, up to 12 out of 23 characters) are shown, with the selected path using only the unary score term $S^i_c$ (orange) and when edge scores $S_e$ are incorporated (cyan). The bars show the NGRAM strengths, with lighter colour representing larger values.[]{data-label="fig:results"}](res5.png "fig:"){width="0.3\linewidth"} ![Results where the unary terms of the JOINT model cannot solely recognise the word correctly, but the addition of the edge scores result in correct recognition, from SVT (a,b) and IC13 (c). The input image is shown in the top left corner. The unary scores for characters (rows) at each position (columns, up to 12 out of 23 characters) are shown, with the selected path using only the unary score term $S^i_c$ (orange) and when edge scores $S_e$ are incorporated (cyan). The bars show the NGRAM strengths, with lighter colour representing larger values.[]{data-label="fig:results"}](res6.png "fig:"){width="0.32\linewidth"}
(a) (b) (c)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------- ------------------------------------------------------------------------------- ------ ------ --- ------ ------ ------ ------ ------
*Baseline ABBYY* ([@Wang11]) 56.0 55.0 - - 35.0 - - 24.3 -
[@Wang11] 76.0 62.0 - - 57.0 - - - -
[@Mishra12] 81.8 67.8 - - 73.2 - - - -
[@Novikova12] 82.8 - - - 72.9 - - 64.1 57.5
[@Wang12] 90.0 84.0 - - 70.0 - - - -
[@Goel13] 89.7 - - - 77.3 - - - -
[@Bissacco13] - - - - 90.4 78.0 87.6 - -
[@Alsharif13] 93.1 88.6 85.1 - 74.3 - - - -
[@Almazan14] - - - - 89.2 - - 91.2 82.1
[@Yao14] 88.5 80.3 - - 75.9 - - 80.2 69.3
[@Jaderberg14a] 96.2 91.5 - - 86.1 - - - -
[@Gordo14] - - - - 90.7 - - 93.3 86.6
DICT [@Jaderberg14c; @Jaderberg14d] **98.7 & **98.6 & 93.3 & **93.1 & **95.4 & **80.7 & **90.8 & **97.1 & **92.7\
CHAR & 98.5 & 96.7 & 92.3 & 85.9 & 93.5 & 68.0 & 79.5 & 95.0 & 89.3\
JOINT & 97.8 & 97.0 & **93.4 & 89.6 & 93.2 & 71.7 & 81.8 & 95.5 & 89.6\
******************
------------------------------------- ------------------------------------------------------------------------------- ------ ------ --- ------ ------ ------ ------ ------
: Comparison to previous methods. The baseline method is from a commercially available OCR system. Note that the training data for DICT includes the lexicons of the test sets, so it has the capacity to recognise all test words. ^\*^Results are constrained to the lexicons described in Section \[sec:datasets\].[]{data-label="table:comparison"}
Conclusion {#sec:conclusion}
==========
In this paper we have introduced a new formulation for word recognition, designed to be used identically in language and non-language scenarios. By modelling character positions and the presence of common N-grams, we can define a joint graphical model. This can be trained effectively by back propagating structured output loss, and results in a more accurate word recognition system than predicting characters alone. We show impressive results for unconstrained text recognition with the ability to generalise recognition to previously unseen words, and match state-of-the-art accuracy when comparing in lexicon constrained scenarios.
#### Acknowledgments.
This work was supported by the EPSRC and ERC grant VisRec no. 228180. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the GPUs used for this research.
[^1]: Current affiliation Google DeepMind. ^+^Current affiliation University of Oxford and Google DeepMind.
[^2]: <http://www.robots.ox.ac.uk/~vgg/data/text/>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A question of the time the system spends in the specified state, when the final state of the system is given, is raised. The model of weak measurements is used to obtain the expression for the time. The conditions for determination of such a time are obtained.'
address: |
Institute of Theoretical Physics and Astronomy,\
A. Goštauto 12, 2600 Vilnius, Lithuania
author:
- 'J. Ruseckas'
- 'B. Kaulakys'
title: Time problem in quantum mechanics and weak measurements
---
Quantum measurement ,Time measurement 03.65.Bz ,03.65.Sq
Introduction
============
Time plays a special role in quantum mechanics. Unlike other observables, time remains a classical variable. It cannot be simply quantized because, as it is well known, the self-adjoint operator of time does not exist for bounded Hamiltonians. The problems with the time rise from the fact that in quantum mechanics many quantities cannot have definite values simultaneously. The absence of the time operator makes this problem even more complicated. However, in practice the time often is important for an experimenter. If quantum mechanics can correctly describe the outcomes of the experiments, it must also give the method for the calculation of the time the particle spends in some region.
The most-known problem of time in quantum mechanics is the so-called ”tunneling time problem”. There had been many attempts to define physical time for quantum mechanical tunneling processes, since the question was raised by MacColl [@maccol] in 1932. This question is still the subject of much controversy, since numerous theories contradict each other in their predictions for ”the tunneling time” [@haugestovneng; @olkhovskyrecami; @landauermartin]. We can raise another, more general, question about the time. Let us consider a system which evolves with time. Let $\chi$ is one of the observables of the system. During the evolution the value of $\chi$ changes. We are considering a subset $\Gamma$ of possible values of $\chi$. The question is [*how much time the values of $\chi$ belong to this subset?*]{}.
There is another version of the question. If we know the final state of the system, we may ask how much time the values of $\chi$ belong to the subset under consideration when the system evolves from the initial to the definite final state. The question about the tunneling time belongs to such class of the problems. Really, in the tunneling time problem we ask about the duration the particle spends in a specified region of the space and we know that the particle has tunneled, i.e., it is on the other side of the barrier. We can expect that such a question may not always be answered. Here our goal is to obtain the conditions under which it is possible to answer such a question.
One of the possibilities to solve the problem of time is to answer what exactly the word “time” means. The meaning of every quantity is determined by the procedure of measurement. Therefore, we have to construct a scheme of an experiment (this can be a [*gedanken*]{} experiment) to measure the quantity with the properties corresponding to the classical time.
The experiment for the measurement of time must obey certain conditions. The time in classical mechanics describes not a single state of the system but the process of the evolution. This property is an essential concept of the time. We speak about the time belonging to a certain evolution of the system. If the measurement of the time disturbs the evolution we cannot attribute this measured duration to the undisturbed evolution. Therefore, we should require that the measurement of the time does not disturb the motion of the system. This means that the interaction of the system with the measuring device must be weak. In quantum mechanics this means that we cannot use the strong measurements described by the von-Neumann’s projection postulate. We have to use the weak measurements of Aharonov, Albert and Vaidman [@AAV1; @AAV; @WM2; @AV; @AV2; @AV3], instead.
We proceed as follows: In Secs. \[sec:mod\] and \[sec:mes\], we present the model and the procedure of the time measurement, respectively. In Sec. \[sec:condprob\] we modify the proposed procedure of the time measurement to make the distinction between different final states of the system. This procedure allows us to determine when the concept of the time with the known final state is correct in quantum mechanics. In Sec. \[sec:exampl\] an example of application of our formalism for the two-level system is presented. Section \[sec:concl\] summarizes our findings.
The model of the time measurement {#sec:mod}
=================================
We consider a system evolving with time. One of the quantities describing the system is $\chi$. Operator $\hat{\chi}$ corresponds to this quantity. For simplicity we assume that the operator $\hat{\chi}$ has a continuous spectrum. The case with discrete spectrum will be considered later.
The measuring device interacts with the system only if $\chi$ is near some point $\chi_{D}$, depending only on the detector. If we want to measure the time the system is in a large region of $\chi$, we have to use many detectors. In the case of tunneling a similar model had been introduced by A. M. Steinberg [@steinberg] and developed in our paper [@JR]. The strong limit of such a model for analysis of the measurement effect for the quantum jumps has been used in Ref. [@JR2].
In order the weak measurements can provide the meaningful information, the measurements have to be performed on an ensemble of identical systems. Each system with its own detector is prepared in the same initial state. After time $t$ the readings of the detectors are collected and averaged.
Our model consists of the system [**S**]{} under consideration and of the detector [**D**]{}. The Hamiltonian is $$\hat{H}=\hat{H}_{\mathrm{S}}+\hat{H}_{\mathrm{D}}+\hat{H}_{\mathrm{I}}$$ where $\hat{H}_{\mathrm{S}}$ and $\hat{H}_{\mathrm{D}}$ are the Hamiltonians of the system and of the detector, respectively, and the operator $$\hat{H}_{\mathrm{I}}=\gamma \hat{q}\hat{D}\left(\chi_{\mathrm{D}}\right)
\label{interact}$$ represents the interaction between the system and the detector. The interaction term (\[interact\]) only slightly differs from the one used by Aharonov, Albert and Vaidman [@AAV]. The operator $\hat{q}$ acts in the Hilbert space of the detector. We require a continuous spectrum of the operator $\hat{q}$. For simplicity, we can consider this operator as the coordinate of the detector. The operator $\hat{D}\left(\chi_{\mathrm{D}}\right) $ acts in the Hilbert space of the system. In an ideal case the operator $\hat{D}\left(
\chi_{\mathrm{D}}\right) $ can be expressed as $\delta $ function $$\hat{D}\left(\chi_{\mathrm{D}}\right)=\left| \chi_{\mathrm{D}}\right\rangle
\left\langle \chi_{\mathrm{D}}\right| =\delta \left(
\hat{\chi}-\chi_{\mathrm{D}}\right) . \label{delta}$$ Parameter $\gamma $ in Eq. (\[interact\]) characterizes the strength of the interaction. A very small parameter $\gamma $ ensures the undisturbance of the system’s evolution.
Hamiltonian (\[interact\]) with $\hat{D}$ given by (\[delta\]) represents the constant force acting on the detector [**D**]{} when the quantity $\chi$ is very close to the value $\chi_{\mathrm{D}}$. This force induces the change of the detector’s momentum. From the classical point of view, the change of the momentum is proportional to the time the particle spends in the region around $\chi_{\mathrm{D}}$ and the coefficient of proportionality equals to the force acting on the detector. We assume that the change of the mean momentum of the detector is proportional to the time the constant force acts on the detector and that the time the particle spends in the detector’s region is the same as the time the force acts on the detector.
We can replace the $\delta$ function by the narrow rectangle of the height $1/L$ and of the width $L$ in the $\chi$ space. From Eq. (\[interact\]) it follows that the force acting on the detector when the particle is in the region around $\chi_{\mathrm{D}}$ is $F=-\gamma/L$. The time the particle spends until time moment $t$ in the unit length region is $$\tau\left(t\right)=-\frac{1}{\gamma}\left(\left\langle p_{\mathrm{q}}\left(
t\right)\right\rangle -\left\langle p_{\mathrm{q}}\right\rangle \right)
\label{timedef}$$ where $\left\langle p_{\mathrm{q}}\right\rangle$ and $\left\langle
p_{\mathrm{q}}\left(t\right) \right\rangle$ are the mean initial momentum and momentum after time $t$, respectively. If we want to find the time the system spends in the region of the finite width, we have to add many times (\[timedef\]).
When the operator $\hat{\chi}$ has a discrete spectrum, we may ask how long the quantity $\chi$ has the value $\chi_{\mathrm{D}}$. To answer this question the detector must interact with the system only when $\chi=\chi_{\mathrm{D}}$. In such a case the operator $\hat{D}\left(\chi_{\mathrm{D}}\right)$ takes the form $$\hat{D}\left(\chi_{\mathrm{D}}\right)=\left|\chi_{\mathrm{D}}\right\rangle
\left\langle \chi_{\mathrm{D}}\right| .$$ The force, acting on the detector in this case equals to $-\gamma$. The time the quantity $\chi$ has the value $\chi_{\mathrm{D}}$ is given by Eq.(\[timedef\]), too. Further formulae do not depend on the spectrum of the operator $\hat{\chi}$.
In the time moment $t=0$ the density matrix of the whole system is $\hat{\rho}
\left(0\right)=\hat{\rho}_{\mathrm{S}}\left(0\right)\otimes
\hat{\rho}_{\mathrm{D}}\left(0\right)$, where $\hat{\rho}_{\mathrm{S}}\left(
0\right)$ is the density matrix of the system and $\hat{\rho}_{\mathrm{D}}
\left(0\right)=\left|\Phi\right\rangle\left\langle\Phi\right|$ is the density matrix of the detector with $\left|\Phi\right\rangle$ being the normalized vector in the Hilbert space of the detector. After the interaction the density matrix of the detector is $\hat{\rho}_{\mathrm{D}}\left(t\right)
={\mathop{\mathrm{Tr}}}_{\mathrm{S}}\left\{\hat{U}\left(t\right)\left(\hat{\rho}_{\mathrm{S}}\left(
0\right)\otimes\left|\Phi\right\rangle\left\langle\Phi\right|\right)
\hat{U}^{\dag}\left( t\right) \right\}$ where $\hat{U}\left( t\right) $ is the evolution operator.
Later on, for simplicity we will neglect the Hamiltonian of the detector. Then, the evolution operator approximately equals to the operator $\hat{U}\left(
t,\gamma\hat{q}\right)$ with $\hat{U}\left(t,\alpha\right) $ being the solution of the equation $$\mathrm{i}\hbar\frac{\partial}{\partial t}\hat{U}\left( t,\alpha\right) =\left(
\hat{H}_{\mathrm{S}}+\alpha\hat{D}\left(\chi_{\mathrm{D}}\right)\right)
\hat{U}\left( t,\alpha \right) . \label{aproxevol}$$ After such assumptions we can obtain the dwell time for our model explicitly.
The dwell time {#sec:mes}
==============
We can expand the operator $\hat{U}\left( t,\gamma \hat{q}\right) $ into the series of the parameter $\gamma $, assuming that $\gamma $ is small. Introducing the operator $\hat{D}\left(\chi_{\mathrm{D}}\right)$ in the interaction representation $$\tilde{D}\left(\chi_{\mathrm{D}},t\right)=\hat{U}^{\dag}_{\mathrm{S}}\left(
t\right)\hat{D}\left(\chi_{\mathrm{D}}\right)\hat{U}_{\mathrm{S}}\left(
t\right)$$ where $\hat{U}_{\mathrm{S}}\left(t\right)$ is the evolution operator of unperturbed system we obtain the first-order approximation for the operator $\hat{U}\left( t,\gamma \hat{q}\right)$, $$\hat{U}\left(t,\gamma\hat{q}\right)\approx\hat{U}_{\mathrm{S}}\left(t\right)
\left(1+\frac{\gamma\hat{q}}{\mathrm{i}\hbar}\int_{0}^{t}\d t_{1}
\tilde{D}\left(\chi_{\mathrm{D}},t_{1}\right)\right) .$$ For shortening of the notation we introduce the operator $$\hat{F}\left(\chi_{\mathrm{D}},t\right)=\int_{0}^{t}\d t_{1}
\tilde{D}\left(\chi_{\mathrm{D}},t_{1}\right). \label{opf}$$
From the density matrix of the detector after the measurement in the first-order approximation we find that the average change of the momentum of the detector during the time $t$ is $-\gamma\left\langle\hat{F}\left(\chi_{\mathrm{D}},t\right)\right\rangle$. From Eq. (\[timedef\]) we obtain the dwell time until time moment $t$, $$\tau\left(\chi,t\right)=\left\langle\hat{F}\left(\chi,t\right)\right\rangle .
\label{fulltime}$$ The time spent in the region $\Gamma$ is $$t\left(\Gamma;t\right)=\int_{\Gamma}\d \chi\tau\left(\chi,t\right)= \int_0^t\d
t'\int_{\Gamma}\d \chi P\left(\chi,t'\right), \label{eq:dw}$$ where $P\left(\chi,t'\right)=\left\langle\tilde{D}\left(\chi,t\right)
\right\rangle$ is the probability for the system to have the value $\chi$ at time moment $t'$.
In the case when $\chi$ is the coordinate of the particle Eq. (\[eq:dw\]) yields the well-known expression for the dwell time [@olkhovskyrecami; @JR]. This time is the average over the entire ensemble of the systems, regardless of their final states.
The time on condition that the system is in the given final state {#sec:condprob}
=================================================================
Further we will consider the case when the final state of the system is known. We ask [*how much time the values of $\chi$ belong to the subset under consideration, $\Gamma$, on condition that the system evolves to the definite final state $f$*]{}. More generally, we might know that the final state of the system belongs to the certain subspace ${\mathcal H}_{\mathrm{f}}$ of system’s Hilbert space.
The projection operator which projects the vectors from the Hilbert space of the system into the subspace ${\mathcal H}_{\mathrm{f}}$ of the final states is $\hat{P}_{\mathrm{f}}$. As far as our model gives the correct result for the time averaged over the entire ensemble of the systems, we can try to take the average only over the subensemble of the systems with the given final states. We measure the momenta $p_{\mathrm{q}}$ of each measuring device after the interaction with the system. Subsequently we perform the final, postselection measurement on the systems of our ensemble. Then we collect the outcomes $p_{q}$ only of the systems the final state of which turns out to belong to the subspace ${\mathcal H}_{\mathrm{f}}$.
The joint probability that the state of the system belongs to ${\mathcal
H}_{\mathrm{f}}$ [*and*]{} the detector has the momentum $p_{\mathrm{q}}(t)$ at the time moment $t$ is $W\left( P_{\mathrm{f}},p_{\mathrm{q}};t\right)
={\mathop{\mathrm{Tr}}}\left\{\hat{P}_{\mathrm{f}}\left|p_{\mathrm{q}}\right\rangle \left\langle
p_{\mathrm{q}}\right|\hat{\rho}\left(t\right)\right\}$, where $\left|
p_{\mathrm{q}}\right\rangle $ is the eigenfunction of the momentum operator $\hat{p}_{\mathrm{q}}$. In quantum mechanics the probability that two quantities simultaneously have definite values does not always exist. If the joint probability does not exist then the concept of the conditional probability is meaningless. However, in our case operators $\hat{P}_{\mathrm{f}}$ and $\left| p_{\mathrm{q}}\right\rangle \left\langle
p_{\mathrm{q}}\right| $ act in different spaces and commute, therefore, the probability $W\left( P_{\mathrm{f}},p_{\mathrm{q}};t\right) $ exists.
Let us define the conditional probability, i.e., the probability that the momentum of the detector is $p_{\mathrm{q}}$ provided that the state of the system belongs to ${\mathcal H}_{\mathrm{f}}$. This probability is given according to the Bayes’s theorem $$W\left( p_{\mathrm{q}};t\left|P_{\mathrm{f}}\right.\right)=\frac{W\left(
P_{\mathrm{f}},p_{\mathrm{q}};t\right)}{W\left(P_{\mathrm{f}};t\right) }
\label{condprob}$$ where $W\left(P_{\mathrm{f}};t\right) ={\mathop{\mathrm{Tr}}}\left\{\hat{P}_{\mathrm{f}}\hat{\rho}\left( t\right) \right\} $ is the probability that the state of the system belongs to the subspace ${\mathcal H}_{\mathrm{f}}$. The average momentum of the detector on condition that the state of the system belongs to the subspace ${\mathcal H}_{\mathrm{f}}$ is $$\left\langle p_{\mathrm{q}}\left(t\right)\right\rangle=\int p_{\mathrm{q}}\d
p_{\mathrm{q}}W\left(p_{\mathrm{q}};t\left| P_{\mathrm{f}}\right. \right) .
\label{condave}$$
From Eqs. (\[timedef\]) and (\[condave\]), in the first-order approximation we obtain the duration on condition that the final state of the system belongs to the subspace ${\mathcal H}_{\mathrm{f}}$ $$\begin{aligned}
\tau_{\mathrm{f}}\left( \chi,t\right) &=& \frac{1}{2\left\langle
\tilde{P}_{\mathrm{f}}\left(t\right)\right\rangle}\left\langle
\tilde{P}_{\mathrm{f}}\left(t\right)\hat{F}\left(\chi,t\right)+\hat{F}\left(
\chi,t\right)\tilde{P}_{\mathrm{f}}\left( t\right)
\right\rangle \nonumber \\
&& +\frac{1}{\mathrm{i}\hbar \left\langle \tilde{P}_{\mathrm{f}}\left(t\right)
\right\rangle }\left( \left\langle q\right\rangle \left\langle
p_{\mathrm{q}}\right\rangle-{\mathop{\mathrm{Re}}}\left\langle
\hat{q}\hat{p}_{\mathrm{q}}\right\rangle \right) \left\langle \left[
\tilde{P}_{\mathrm{f}}\left( t\right) ,\hat{F}\left( \chi,t\right) \right]
\right\rangle . \label{condtime}\end{aligned}$$
Eq. (\[condtime\]) consists of two terms and we can introduce two expressions with the dimension of time $$\begin{aligned}
\tau_{\mathrm{f}}^{(1)}\left(\chi,t\right) &=&\frac{1}{2\left\langle
\tilde{P}_{\mathrm{f}}\left(t\right)\right\rangle}\left\langle
\tilde{P}_{\mathrm{f}}\left(t\right)\hat{F}\left(\chi,t\right)+\hat{F}\left(
\chi,t\right)
\tilde{P}_{\mathrm{f}}\left(t\right)\right\rangle , \label{timere} \\
\tau_{\mathrm{f}}^{(2)}\left(\chi,t\right) &=&\frac{1}{2\mathrm{i}\left\langle
\tilde{P}_{\mathrm{f}}\left(t\right)\right\rangle}\left\langle \left[
\tilde{P}_{\mathrm{f}}\left(t\right),\hat{F}\left(\chi,t\right) \right]
\right\rangle . \label{timeim}\end{aligned}$$ Then the time the system spends in the subset $\Gamma$ on condition that the final state of the system belongs to the subspace ${\mathcal
H}_{\mathrm{f}}$ can be rewritten in the form $$\tau_{\mathrm{f}}\left(\chi,t\right)=\tau_{\mathrm{f}}^{(1)}\left(\chi,t\right)+
\frac{2}{\hbar}\left(\left\langle q\right\rangle\left\langle p_{\mathrm{q}}
\right\rangle-{\mathop{\mathrm{Re}}}\left\langle\hat{q}\hat{p}_{\mathrm{q}}\right\rangle\right)
\tau_{\mathrm{f}}^{(2)}\left(\chi,t\right). \label{condtime2}$$ The quantities $\tau_{\mathrm{f}}^{(1)}(\chi,t)$ and $\tau_{\mathrm{f}}^{(2)}(\chi,t)$ are related to the real and imaginary parts of the complex time, introduced by D. Sokolovski [*et. al*]{} [@sokolovskibaskin]. In our model the quantity $\tau_{\mathrm{f}}(\chi,t)$ is real, contrary to the complex-time approach. The components of time $\tau_{\mathrm{f}}^{(1)}$ and $\tau_{\mathrm{f}}^{(2)}$ are real, too. Therefore, this time can be interpreted as the duration of an event.
If the commutator $\left[\tilde{P}_{\mathrm{f}}\left( t\right) ,\hat{F}\left(
\chi,t \right) \right] $ in Eq. (\[condtime\]) is not zero then, even in the limit of the very weak measurement, the measured value depends on the particular detector. This fact means that in such a case we cannot obtain the [*definite*]{} value for the conditional time. Moreover, the coefficient $(\langle q\rangle\langle p_{\mathrm{q}}\rangle-{\mathop{\mathrm{Re}}}\langle\hat{q}
\hat{p}_{\mathrm{q}}\rangle)$ may be zero for the specific initial state of the detector, e.g., for the Gaussian distribution of the coordinate $q$ and momentum $p_{\mathrm{q}}$.
The conditions of the possibility to determine the time uniquely in a case when the final state of the system is known takes the form $$\left[\tilde{P}_{\mathrm{f}}\left(t\right),\hat{F}\left(\chi,t\right)\right]=0
. \label{eq:poscond}$$ This result can be understood basing on general principles of quantum mechanics, too. We ask [*how much time the values of $\chi$ belong to the certain subset when the system evolves to the given final state*]{}. We know with certainty the final state of the system. In addition, we want to have some information about the values of the quantity $\chi$. However, if we know the final state with certainty, we may not know the values of $\chi$ in the past and, vice versa, if we know something about $\chi$, we may not definitely determine the final state. Therefore, in such a case the question about the time when the system evolves to the given final state cannot be answered definitely and the conditional time has no reasonable meaning.
The quantity $\tau_{\mathrm{f}}\left(t\right)$ according to Eqs.(\[condtime\]) and (\[timere\]) has many properties of the classical time. So, if the final states $\left\{f\right\}$ constitute the full set, then the corresponding projection operators obey the equality of completeness $\sum_f
\hat{P}_{\mathrm{f}}=1$. Then from Eq. (\[condtime\]) we obtain the expression $$\sum_f \left\langle \tilde{P}_{\mathrm{f}}\left( t\right) \right\rangle
\tau_{\mathrm{f}}\left( \chi,t\right) =\tau\left( \chi,t\right) .
\label{clasprop}$$ The quantity $\left\langle \tilde{P}_{\mathrm{f}}\left( t\right) \right\rangle$ is the probability that the system at the time $t$ is in the state $f$. Eq.(\[clasprop\]) shows that the full duration equals to the average over all possible final states, as it is a case in the classical physics. From Eq.(\[clasprop\]) and Eqs. (\[timere\]), (\[timeim\]) it follows $$\begin{aligned}
\sum_f \left\langle \tilde{P}_{\mathrm{f}}\left( t\right) \right\rangle
\tau_{\mathrm{f}}^{(1)}\left( \chi,t\right) &=& \tau\left( \chi,t\right) ,\\
\sum_f \left\langle \tilde{P}_{\mathrm{f}}\left( t\right) \right\rangle
\tau_{\mathrm{f}}^{(2)}\left( \chi,t\right) &=& 0 .\end{aligned}$$ We suppose that quantities $\tau_{\mathrm{f}}^{(1)}\left(\chi,t\right)$ and $\tau_{\mathrm{f}}^{(2)}\left(\chi,t\right)$ can be useful even in the case when the time has no definite value, since in the tunneling time problem the quantities (\[timere\]) and (\[timeim\]) correspond to the real and imaginary parts of the complex time, respectively [@JR].
The eigenfunctions of the operator $\hat{\chi}$ constitute the full set $\int\left|\chi\right\rangle\left\langle\chi\right|\d \chi=1$ where the integral must be replaced by the sum for the discrete spectrum of the operator $\hat{\chi}$. From Eqs. (\[delta\]), (\[opf\]), (\[condtime\]) we obtain the equality $$\int \tau_{\mathrm{f}}\left(\chi,t\right)\d \chi=t . \label{eq:eqx}$$ Eq. (\[eq:eqx\]) shows that the time during which the quantity $\chi$ has any value equals to $t$, as it is in the classical physics.
Example {#sec:exampl}
=======
The obtained formalism can be applied to the tunneling time problem [@JR]. In this paper, however, we will consider a simpler system than the tunneling particle, i.e., a two-level system. The system is forced by the perturbation $V$ which causes the jumps from one state to another. We will determine the time the system is in the given state.
The Hamiltonian of this system is $$\hat{H}=\hat{H}_0+\hat{V}$$ where $\hat{H}_0=\hbar\omega\hat{\sigma}_3/2$ is the Hamiltonian of the unperturbed system and $\hat{V}=v\hat{\sigma}_{+}+v^{*}\hat{\sigma}_{-}$ is the perturbation. Here $\sigma_1,\sigma_2,\sigma_3$ are Pauli matrices and $\sigma_{\pm}=\frac{1}{2}\left( \sigma_1 \pm i\sigma_2\right)$. The Hamiltonian $\hat{H}_0$ has two eigenfunctions $\left|0\right\rangle$ and $\left|1\right\rangle$ with the eigenvalues $-\hbar\omega/2$ and $\hbar\omega/2$, respectively. The initial state of the system is $\left|0\right\rangle$.
From Eq. (\[fulltime\]) we obtain the times the system spends in the energy levels $0$ and $1$, respectively, $$\begin{aligned}
\tau\left(0,t\right) &=&\frac{1}{2}\left(1+\frac{\omega^2}{\Omega^2}\right)
t+\frac{1}{2\Omega}\sin\left(\Omega t\right) \left(1-
\frac{\omega^2}{\Omega^2}\right) , \label{eq:31}\\
\tau\left(1,t\right)&=&\frac{1}{2}\left(1- \frac{\omega^2}{\Omega^2}\right)
t-\frac{1}{2\Omega}\sin\left( \Omega t\right)
\left(1-\frac{\omega^2}{\Omega^2}\right) \label{eq:32}\end{aligned}$$ where $\Omega=\sqrt{\omega^{2}+4\frac{\left|v\right|^2}{\hbar^2}}$. From Eqs. (\[timere\]) and (\[timeim\]) we can obtain the conditional time. The components $\tau^{(1)}$ (\[timere\]) and $\tau^{(2)}$ (\[timeim\]) of the time the system spends in the level $0$ on condition that the final state is $\left|1\right\rangle$ are $$\begin{aligned}
\tau_1^{(1)}\left( 0,t\right) &=& \frac{t}{2},\label{eq:34}\\
\tau_1^{(2)}\left( 0,t\right) &=& \frac{\omega}{2\Omega}\left( 1-t\cot \left(
\frac{\Omega}{2} t\right) \right) .\end{aligned}$$ When $\Omega t=2\pi n$, $n\in \Zset$, the quantity $\tau_1^{(2)}\left(
0,t\right)$ tends to infinity. This is because at these time moments the system is in state $|1\rangle$ with the probability $0$ and we cannot consider the interaction with the detector as very weak.
![The times the system spends in the energy levels $0$, $\tau\left(
0,t\right)$ (dashed line) and level $1$, $\tau\left( 1,t\right)$ (dotted line), according to Eqs. (\[eq:31\]) and (\[eq:32\]), respectively. The quantity $\tau_1^{(1)}\left( 0,t\right) $, Eq. (\[eq:34\]) is shown as solid straight line. The quantities $\tau^{(1)}_0(0,t)$ (1) and $\tau^{(1)}_0(1,t)$ (2) are calculated according to Eqs. (\[eq:28\]) and (\[eq:30\]), respectively. The parameters are $\omega=2$ and $\Omega=4$.[]{data-label="dwt"}](pav1.eps){width=".6\hsize"}
The components of the time (\[timere\]) and (\[timeim\]) the system spends in level $0$ on condition that the final state is $\left|0\right\rangle$ are $$\begin{aligned}
\tau_0^{(1)}\left(0,t\right)&=&\frac{\left(1+3\frac{\omega^2}{\Omega^2}\right)
t+\left(1-\frac{\omega^2}{\Omega^2}\right)\left(\frac{2}{\Omega}\sin\left(
\Omega t\right)+t\cos\left(\Omega t\right)\right)}{2\left( \left(
1+\frac{\omega^2}{\Omega^2}\right)+\left(1-\frac{\omega^2}{\Omega^2}\right)
\cos\left(\Omega t\right)\right)} ,\label{eq:28} \\
\tau_0^{(2)}\left(0,t\right)&=&\frac{\frac{\omega}{\Omega}\left(1-\frac{\omega
^2}{\Omega^2}\right)\sin\left(\frac{\Omega}{2}t\right)\left(t\cos\left(
\frac{\Omega}{2}t\right)-\frac{2}{\Omega}\sin\left(\frac{\Omega}{2}t\right)
\right)}{2\left( \left( 1+ \frac{\omega^2}{\Omega^2}\right)+
\left(1-\frac{\omega^2}{\Omega^2}\right) \cos\left(\Omega t\right)\right)} .
\label{eq:29}\end{aligned}$$
![The quantity $\tau_0^{(2)}\left( 0,t\right)$, Eq. (\[eq:29\]). The parameters are the same as in Fig. \[dwt\].[]{data-label="ti00"}](fig4.eps){width=".4\hsize"}
The time the system spends in level $1$ on condition that the final state is $\left|0\right\rangle$ may be expressed as $$\tau_0^{(1)}\left( 1,t\right) = \frac{\left( 1- \frac{\omega ^2}{\Omega
^2}\right) \left( t+t\cos\left(\Omega t\right)- \frac{2}{\Omega}\sin\left(
\Omega t\right) \right)}{2\left(\left(1+\frac{\omega^2}{\Omega^2}\right)
+\left(1-\frac{\omega^2}{\Omega^2}\right) \cos\left(\Omega t\right)\right)} .
\label{eq:30}$$
The quantities $\tau\left(0,t\right)$, $\tau\left(1,t\right)$, $\tau_1^{(1)}\left( 0,t\right)$, $\tau_0^{(1)}\left( 0,t\right)$ and $\tau_0^{(1)}\left(1,t \right)$ are shown in Fig. \[dwt\]. The quantity $\tau_0^{(2)}\left( 0,t\right)$ is shown in Fig. \[ti00\]. Note that the duration with the given final state is not necessarily monotonic as it is with the full duration, because this final state at different time moments can be reached by different ways. We can interpret the quantity $\tau_0^{(1)}\left(
0,t\right)$ as the time the system spends in the level $0$ on condition that the final state is $\left|0\right\rangle$, but at certain time moments this quantity is greater than $t$. The quantity $\tau_0^{(1)}\left( 1,t\right)$ becomes negative at certain time moments. This is the consequence of the fact that for the system under consideration the condition (\[eq:poscond\]) is not fulfilled. The peculiarities of the behavior of the conditional times show that it is impossible to decompose the unconditional time into two components having all classical properties of the time.
Conclusion {#sec:concl}
==========
We consider the problem of the time in quantum mechanics. The tunneling time problem is a part of this more general problem. The problem of time is solved adapting the weak measurement theory to the measurement of time. In this model expression (\[fulltime\]) for the duration when the arbitray observable $\chi$ has the certain value is obtained. This result is in agreement with the known results for the dwell time in the tunneling time problem.
Further we consider the problem of the duration when the observable $\chi$ has the certain value on condition that the system is in the given final state. Our model of measurement allows us to obtain the expression (\[condtime\]) of this duration as well. This expression has many properties of the corresponding classical time. However, such a duration not always has the reasonable meaning. It is possible to obtain the duration the quantity $\chi$ has the certain value on condition that the system is in a given final state only when the condition (\[eq:poscond\]) is fulfilled. In the opposite case, there is a dependence in the outcome of the measurements on particular detector even in an ideal case and, therefore, it is impossible to obtain the definite value of the duration. When the condition (\[eq:poscond\]) is not fulfilled, we introduce two quantities (\[timere\]) and (\[timeim\]), characterizing the conditional time. These quantities are useful in the case of tunneling and we suppose that they can be useful also for other problems.
[00]{} L. A. MacColl, Phys. Rev. 40 (1932) 621. E. H. Hauge and J. A. Støvneng, Rev. Mod. Phys. 61 (1989) 917. V. S. Olkhovsky, and E. Recami, Phys. Rep. 214 (1992) 339. R. Landauer and Th. Martin, Rev. Mod. Phys. 66 (1994) 217. Y. Aharonov, D. Albert, A. Casher, L. Vaidman, Phys. Lett. A 124 (1987) 199. Y. Aharonov, D. Z. Albert, L. Vaidman, Phys. Rev. Lett. 60 (1988) 1351. I. M. Duck, P. M. Stevenson, E. C. G. Sudarshan, Phys. Rev. D 40 (1989) 2112. Y. Aharonov, L. Vaidman, Phys. Rev. A 41 (1990) 11. Y. Aharonov, L. Vaidman, J. Phys. A 24 (1991) 2315. Y. Aharonov, L. Vaidman, Phys. Scr. T 76 (1998) 85. A. M. Steinberg, Phys. Rev. A 52 (1995) 32. J. Ruseckas, Phys. Rev. A 63 (2001) 052107; quant-ph/0101136. J.Ruseckas and B. Kaulakys, Phys. Rev. A 63 (2001) 062103. D. Sokolovski and L. M. Baskin, Phys. Rev. A 36 (1987) 4604.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the Casimir interaction between two spheres in $(D+1)$-dimensional Minkowski spacetime due to the vacuum fluctuations of scalar fields. We consider combinations of Dirichlet and Neumann boundary conditions. The TGTG formula of the Casimir interaction energy is derived. The computations of the T matrices of the two spheres are straightforward. To compute the two G matrices, known as translation matrices, which relate the hyper-spherical waves in two spherical coordinate frames differ by a translation, we generalize the operator approach employed in \[IEEE Trans. Antennas Propag. **36**, 1078 (1988)\]. The result is expressed in terms of an integral over Gegenbauer polynomials. In contrast to the $D=3$ case, we do not re-express the integral in terms of $3j$-symbols and hyper-spherical waves, which in principle, can be done but does not simplify the formula. Using our expression for the Casimir interaction energy, we derive the large separation and small separation asymptotic expansions of the Casimir interaction energy. In the large separation regime, we find that the Casimir interaction energy is of order $L^{-2D+3}$, $L^{-2D+1}$ and $L^{-2D-1}$ respectively for Dirichlet-Dirichlet, Dirichlet-Neumann and Neumann-Neumann boundary conditions, where $L$ is the center-to-center distance of the two spheres. In the small separation regime, we confirm that the leading term of the Casimir interaction agrees with the proximity force approximation, which is of order $d^{-\frac{D+1}{2}}$, where $d$ is the distance between the two spheres. Another main result of this work is the analytic computations of the next-to-leading order term in the small separation asymptotic expansion. This term is computed using careful order analysis as well as perturbation method. In the case the radius of one of the sphere goes to infinity, we find that the results agree with the one we derive for sphere-plate configuration. When $D=3$, we also recover previously known results. We find that when $D$ is large, the ratio of the next-to-leading order term to the leading order term is linear in $D$, indicating a larger correction at higher dimensions. The methodologies employed in this work and the results obtained can be used to study the one-loop effective action of the system of two spherical objects in the universe.'
author:
- 'L. P. Teo'
title: ' Casimir interaction between spheres in $\boldsymbol{(D+1)}$-dimensional Minkowski spacetime'
---
Introduction
============
Motivated by the advent of string theory and the endeavor to solve fundamental problems in physics such as dark energy and cosmological constant problem, studying physics in higher dimensional spacetime has become a norm rather than exception.
Casimir effect [@8] which was proposed more than 60 years ago plays an important role in high energy physics since it is intimately related to the one-loop effective action of a quantum field [@9]. Although most of the works in Casimir effect were focused in (3+1)-dimensional spacetime, there have been quite a considerable amount of work on Casimir effect in higher dimensional spacetime. One of the pioneering works is the work [@10] where the Casimir effect of a $D$-dimensional rectangular cavity is studied. Subsequently, Casimir effect of a $D$-dimensional spherical cavity were also considered [@11; @12; @13]. Nonetheless, in contrast to the Casimir interaction between two parallel plates considered by Casimir in his pioneering work [@8], the Casimir energy of a $D$-dimensional rectangular or spherical cavity is a self-energy rather than interaction energy.
In the end of last century, it has gradually been realized that the Casimir interaction between two objects should play a more important role since this is physically observable. Partly also due to the need to compare to Casimir experiments, researchers have started to research on the method to compute the Casimir interactions between two objects, in particular between a sphere and a plate. In the early phase of this research, most of the methods proposed were numerical [@14; @15; @16; @17; @18; @19; @20]. About eight years ago, new light has been shed on this problem by two new developments. First, Gies and collaborators [@21; @22; @23; @24; @25] used worldline representation to compute the Casimir interaction between two objects imposed with Dirichlet boundary conditions. Later a few groups of researchers [@55; @26; @27; @28; @43; @6; @44; @45; @32; @33; @34; @36; @37; @38; @39] independently developed some methods to compute functional representations for the Casimir interaction between specific pairs of objects which all used the idea of multiple scattering in some form. In a time span of three years, a scheme has been worked out which in principle can be used to compute the Casimir interactions between any two objects [@6; @32]. Motivated by the mode summation approach proposed for the cylinder-cylinder configuration [@40; @41; @42], we managed to interpret the scheme proposed in [@6; @32] from the point of view of mode summation approach [@7]. Despite that we only considered examples in $(3+1)$-dimensional spacetime, it is quite easy to see that the mathematical scheme we deployed in [@7] is not limited to $(3+1)$-dimensional spacetime.
There have been quite a number of works on the Casimir effect between two parallel plates in higher dimensional spacetime. The case of $(D+1)$-dimensional Minkowski spacetime has been treated in [@10] as limiting case of rectangular cavity. As a first step to consider Casimir interaction between nontrivial objects in higher dimensional spacetime, we have considered the Casimir interaction between a sphere and a plate in $(D+1)$-dimensional Minkowski spacetime in [@1]. In this work, we consider two spheres with Dirichlet or Neumann boundary conditions. This scenario is more interesting since it can be used to model two spherical objects in the universe. The mathematical scheme we developed in [@7] is used to compute the Casimir interaction energy. In hyper-spherical coordinates, it is easy to write down the scattering matrices of the two spheres. The most technical part of the problem is to derive the translation matrices which relate the hyper-spherical coordinate system centered at the two spheres. For this, we generalize the operator approach developed for the $D=3$ case [@31; @7]. A major difference is that we do not use $3j$-symbols to rewrite the translation matrices as linear combinations of spherical waves but leave it as integrals over Gegenbauer polynomials.
After deriving the functional representation for the Casimir interaction energy, a natural question to ask is what we can infer from the formula. One of the most important things we want to know is the strength of the interaction. To have an idea of this, we need to compute the asymptotic expansions of the Casimir interaction energy at small and large separations. The computation of the large separation asymptotic behavior is quite straightforward since it only involves a few terms which can be computed explicitly. For the small separation asymptotic expansion, the computations are more complicated and a careful order counting is needed. For $D=3$ case, such analysis have been carried out in [@33; @34; @46; @47; @48; @2; @50; @51; @5] for different configurations. Our current scenario is closest to [@5] where two spheres in $(3+1)$-dimensional spacetime is considered. However, a major difference is that we do not have $3j$-symbols in the formula for the Casimir interaction energy. So instead of the integral representation for the $3j$-symbols, we have to use an integral representation for the Gegenbauer polynomials. It turns out that this is in fact not any more complicated.
This work will shed some light on how to compute the Casimir effect between two objects with nontrivial geometry in higher dimensional spacetime. It will also be interesting for researchers wanted to study quantum system of two spherical objects in higher dimensional spacetime.
The Casimir interaction between a sphere and a plate
====================================================
In this section, we consider the Casimir interaction energy between two spheres in $(D+1)$-dimensional Minkowski spacetime equipped with the standard metric $$ds^2=dt^2-dx_1^2-\ldots-dx_D^2.$$ The equation of motion of a scalar field $\varphi$ is $$\label{eq12_3_2}
\left(\frac{{\partial}^2}{{\partial}x_1^2}+\ldots+\frac{{\partial}^2}{{\partial}x_D^2}\right)\varphi = -\frac{\omega^2}{c^2}\varphi.$$
Assume that the radii of the spheres are $R_1$ and $R_2$ respectively, and the centers are at $O_1=(0, 0, \ldots, 0)$ and $O_2=(L, 0, \ldots, 0)$ respectively. We use the hyper-spherical coordinate system: $$\begin{aligned}
x_1=&r\cos\theta_1\\
x_2=&r\sin\theta_1\cos\theta_2\\
&\vdots\\
x_{D-1}=&r\sin\theta_1\ldots\sin\theta_{D-2}\cos\theta_{D-1}\\
x_D=&r\sin\theta_1\ldots\sin\theta_{D-2}\sin\theta_{D-1}\end{aligned}$$ When $\boldsymbol{r}=(x_1,\ldots,x_D)$ ranges over $\mathbb{R}^D$, $r$ ranges from 0 to $\infty$, whereas $$\begin{aligned}
0\leq \theta_i\leq\pi, \quad i=1,2,\ldots,D-2,\end{aligned}$$ and $$-\pi\leq\theta_{D-1}\leq\pi.$$In the following, we will denote by $S^{D-1}$ the region $$\begin{aligned}
0\leq \theta_i\leq\pi, \quad i=1,2,\ldots,D-2;\quad -\pi\leq\theta_{D-1}\leq\pi.\end{aligned}$$ The volume element $\displaystyle\prod_{i=1}^D dx_i$ is equal to $$r^{D-1}\prod_{i=1}^{D-1}\sin^{D-i-1}\theta_id\theta_i$$ in spherical coordinates. Denote by $d\Omega_{D-1}$ the measure $$\prod_{i=1}^{D-1}\sin^{D-i-1}\theta_id\theta_i.$$ Then $$\int_{S^{D-1}}d\Omega_{D-1}=\frac{2\pi^{\frac{D}{2}}}{\Gamma\left(\frac{D}{2}\right)}$$ is the volume of the unit sphere $x_1^2+x_2^2+\ldots+x_{D}^2=1$.
In spherical coordinates, the equation of motion becomes $$\left(\frac{{\partial}^2}{{\partial}r^2}+\frac{D-1}{r}\frac{{\partial}}{{\partial}r}+\frac{1}{r^2}\sum_{i=1}^{D-1}\frac{1}{\prod_{j=1}^{i-1}\sin^2\theta_j}\left(\frac{{\partial}^2}{{\partial}\theta_i^2}
+(D-i-1)\frac{\cos\theta_i}{\sin\theta_i}\frac{{\partial}}{{\partial}\theta_i}\right)\right)\varphi=-\frac{\omega^2}{c^2}\varphi.$$ The solutions of this differential equation are parametrized by $\boldsymbol{m}=(m_1,\ldots,m_{D-1})$, with $$l=m_1\geq m_2\geq \ldots\geq m_{D-2}\geq |m_{D-1}|.$$The regular and outgoing spherical waves are [@29; @30]: $$\label{eq12_3_6}
\begin{split}
\varphi_{\boldsymbol{m}}^{\text{reg}}(\boldsymbol{x},k)=\mathcal{C}_l^{\text{reg}}C_{\boldsymbol{m}}j_l( k r)\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}),\\
\varphi_{\boldsymbol{m}}^{\text{out}}(\boldsymbol{x},k)=\mathcal{C}_l^{\text{out}}C_{\boldsymbol{m}}h_l^{(1)}( k r)\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}),
\end{split}$$ where $$\begin{aligned}
j_l(z)=z^{-\frac{D-2}{2}}J_{l+\frac{D-2}{2}}(z),\hspace{1cm}h_l^{(1)}(z)=z^{-\frac{D-2}{2}}H^{(1)}_{l+\frac{D-2}{2}}(z),\end{aligned}$$ $$\begin{aligned}
\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta})=e^{im_{D-1}\theta_{D-1}}\prod_{j=1}^{D-2}\sin^{|m_{j+1}|}\theta_{j}C_{m_j-|m_{j+1}|}^{|m_{j+1}|+\frac{D-j-1}{2}}(\cos\theta_j),\end{aligned}$$$J_{l+\frac{D-2}{2}}(z)$ and $H^{(1)}_{l+\frac{D-2}{2}}(z)$ are Bessel functions, and $C_n^{\nu}(z)$ is a Gegenbauer polynomial defined by $$\begin{aligned}
(1-2zt+t^2)^{-\nu}=\sum_{n=0}^{\infty} C_n^{\nu}(z)t^n.\end{aligned}$$ The Gegenbauer polynomials satisfy the orthogonality relation $$\begin{aligned}
\label{eq12_3_11}
\int_{-1}^1dx C_n^{\nu}(x)C_m^{\nu}(x)(1-x^2)^{\nu-\frac{1}{2}}=\frac{\pi 2^{1-2\nu}\Gamma(n+2\nu)}{n!(n+\nu)\Gamma(\nu)^2}\delta_{n,m}.\end{aligned}$$ Hence, the hyperspherical harmonics $\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta})$ satisfy the orthogonality condition $$\int_{S^{D-1}} d\Omega_{D-1}\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}) \boldsymbol{Y}_{\boldsymbol{m}'}(\boldsymbol{\theta})^* =\frac{1}{C_{\boldsymbol{m}}^2}\delta_{\boldsymbol{m},\boldsymbol{m}'},$$ where $$\begin{aligned}
C_{\boldsymbol{m}}=\sqrt{\frac{1}{2\pi^{D-1}}\prod_{j=1}^{D-2}\frac{2^{2|m_{j+1}|+D-j-2}\Gamma\left(|m_{j+1}|+\frac{D-j-1}{2}\right)^2\left(m_j+\frac{D-j-1}{2}\right)(m_j-|m_{j+1}|)!}
{\Gamma\left(m_j+|m_{j+1}|+D-j-1\right)}}.\end{aligned}$$ The constants $\mathcal{C}_l^{\text{reg}}$ and $\mathcal{C}_l^{\text{out}}$ are defined by $$\begin{aligned}
\mathcal{C}_l^{\text{reg}}=i^{-l},\quad \mathcal{C}_l^{\text{out}}=\frac{\pi}{2}i^{l+D-1},\end{aligned}$$ so that $$\begin{aligned}
\mathcal{C}_l^{\text{reg}}j_l(iz)=z^{-\frac{D-2}{2}}I_{l+\frac{D-2}{2}}(z),\hspace{1cm}\mathcal{C}_l^{\text{out}}h^{(1)}_l(iz)=z^{-\frac{D-2}{2}}K_{l+\frac{D-2}{2}}(z).\end{aligned}$$
In [@7], we have discussed the mathematical formalism underlying the TGTG formula for the Casimir interaction energy $$\label{eq12_3_1}
E_{\text{Cas}}=\frac{\hbar}{2\pi}\int_0^{\infty} d\xi \ln\det\left(1-\mathbb{T}^1\mathbb{G}^{1}\mathbb{T}^2\mathbb{G}^{2}\right),$$ between two objects. It is easy to see that this formalism does not depend on the dimension of spacetime and the type of quantum field involved. It can be applied for Casimir interaction in $(D+1)$-dimensional spacetime. The $\mathbb{T}^1$ and $\mathbb{T}^2$ in this formula are the Lippmann-Schwinger T-operators of the two spheres, which are related to the scattering matrices of the spheres. As in [@7], it is easy to find that for Dirichlet (D) and Neumann (N) boundary conditions, they are diagonal in $\boldsymbol{m}$ with diagonal elements given by $$\begin{split}
&T_{\boldsymbol{m}}^{i, \text{D}}(\kappa)=T_{l}^{i, \text{D}}(\kappa)=\frac{I_{l+\frac{D-2}{2}}(\kappa R_i)}{K_{l+\frac{D-2}{2}}(\kappa R_i)},\\
&T_{\boldsymbol{m}}^{i, \text{N}}(\kappa)=T_{l}^{i, \text{N}}(\kappa)=\frac{-\frac{D-2}{2}I_{l+\frac{D-2}{2}}(\kappa R_i)+\kappa R_iI_{l+\frac{D-2}{2}}'(\kappa R_i)}
{-\frac{D-2}{2}K_{l+\frac{D-2}{2}}(\kappa R_i)+\kappa R_i K_{l+\frac{D-2}{2}}'(\kappa R_i)}.
\end{split}$$Here $\kappa=\xi/c$ and $k=i\kappa$.
The translation matrices $\mathbb{G}^{1} $ and $\mathbb{G}^{2}$ in are defined by $$\label{eq12_3_5}
\varphi_{\boldsymbol{m}'}^{\text{out}}(\boldsymbol{x}-\boldsymbol{L},k)=\sum_{\boldsymbol{m}}G^1_{\boldsymbol{m},\boldsymbol{m}'}\varphi_{\boldsymbol{m}}^{\text{reg}}(\boldsymbol{x},k),$$ $$\label{eq3_4_1}
\varphi_{\boldsymbol{m}}^{\text{out}}(\boldsymbol{x}'+\boldsymbol{L},k)=\sum_{\boldsymbol{m}'}G^2_{\boldsymbol{m}',\boldsymbol{m}}\varphi_{\boldsymbol{m}'}^{\text{reg}}(\boldsymbol{x}',k),$$where the summation over $\boldsymbol{m}$ is $$\begin{aligned}
\sum_{\boldsymbol{m}}=\sum_{l=0}^{\infty}\sum_{m_2=0}^{l}\sum_{m_3=0}^{m_2}\ldots\sum_{m_{D-2}=0}^{m_{D-3}}\sum_{m_{D-1}=-m_{D-2}}^{m_{D-2}}.\end{aligned}$$In the following, we will derive the explicit expressions for $G^1_{\boldsymbol{m},\boldsymbol{m}'}$ and $G^2_{\boldsymbol{m}',\boldsymbol{m}}$.
Express $\boldsymbol{k}=(k_1,k_2,\ldots,k_D)$ in hyper-spherical coordinates: $$\label{eq12_3_8}\begin{split}
k_1=&k\cos\theta^k_1,\\
k_2=&k\sin\theta^k_1\cos\theta^k_2,\\
&\vdots\\
k_{D-1}=&k\sin\theta^k_1\ldots\sin\theta^k_{D-2}\cos\theta^k_{D-1},\\
k_{D}=&k\sin\theta^k_1\ldots\sin\theta^k_{D-2}\sin\theta^k_{D-1},\end{split}$$ and let $S^{D-1}_k$ be the region $$\begin{gathered}
0\leq \theta^k_j\leq \pi,\quad 1\leq j\leq D-2,\quad
-\pi\leq \theta^k_{D-1}\leq\pi.\end{gathered}$$ As is shown in [@30], $$\begin{aligned}
\label{eq12_3_4}
\int_{S^{D-1}_k}d\Omega_{D-1}^k\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)e^{i\boldsymbol{k}\cdot\boldsymbol{r}}=(2\pi)^{\frac{D}{2}} i^{-l}j_l(kr)\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}).\end{aligned}$$ This is equivalent to $$\begin{aligned}
e^{i\boldsymbol{k}\cdot\boldsymbol{r}}= (2\pi)^{\frac{D}{2}} \sum_{\boldsymbol{m}}
i^{-l} C_{\boldsymbol{m}}^2j_l(kr)\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta})\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)^*.\end{aligned}$$Notice that $$\begin{aligned}
j_0(kr)=a_D\int_{S_k^{D-1}} d\Omega_{D-1}^ke^{i\boldsymbol{k}\cdot\boldsymbol{r}},\end{aligned}$$ where $$a_D=\frac{1}{(2\pi)^{\frac{D}{2}}}.$$ A counterpart for the outgoing wave is $$\begin{aligned}
h_0^{(1)}(kr)=2a_D\int_{\mathbb{R}^{D-1}}d\boldsymbol{k}_{\perp}\frac{e^{i\boldsymbol{k}\cdot\boldsymbol{r}}}{ k^{D-2}\sqrt{k^2-k_{\perp}^2}},\end{aligned}$$with $k_1=\pm \sqrt{k^2-k_{\perp}^2}$, where the sign $\pm$ is the same as the sign of $x_1$. Now we will use the method in [@31; @7]. Using the fact that the normalized hyper-spherical harmonics $C_{\boldsymbol{m}}\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta})$ can be written as $$\begin{aligned}
C_{\boldsymbol{m}}\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta})=H_{\boldsymbol{m}}\left(\frac{x_1}{r},\frac{x_2}{r},\ldots,\frac{x_D}{r}\right)\end{aligned}$$for some homogeneous polynomial $H_{\boldsymbol{m}}\left(x_1,\ldots,x_D\right)$ of degree $m_1=l$, we can define an operator $$\mathcal{H}_{\boldsymbol{m}}(\boldsymbol{{\partial}})=H_{\boldsymbol{m}}\left(\frac{{\partial}_{x_1}}{ik},\ldots,\frac{{\partial}_{x_D}}{ik}\right)$$which generalizes the operator $\mathcal{P}_{lm}$ defined in [@31]. It follows from definition that $$\begin{aligned}
\label{eq3_3_3}
\mathcal{H}_{\boldsymbol{m}}(\boldsymbol{{\partial}})e^{i\boldsymbol{k}\cdot\boldsymbol{r}}
=C_{\boldsymbol{m}}\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)e^{i\boldsymbol{k}\cdot\boldsymbol{r}}.\end{aligned}$$ Hence, can be written as $$\label{eq3_3_2}\begin{split}
\varphi_{\boldsymbol{m}}^{\text{reg}}(\boldsymbol{x},k)=&\mathcal{C}_l^{\text{reg}}C_{\boldsymbol{m}}a_Di^l\int_{S^{D-1}_k}d\Omega_{D-1}^k\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)
e^{i\boldsymbol{k}\cdot\boldsymbol{r}}\\
=&\mathcal{C}_l^{\text{reg}}a_Di^l\mathcal{H}_{\boldsymbol{m}}(\boldsymbol{{\partial}})\int_{S^{D-1}_k} d\Omega_{D-1}^ke^{i\boldsymbol{k}\cdot\boldsymbol{r}}\\
=&i^l\mathcal{C}_l^{\text{reg}}\mathcal{H}_{\boldsymbol{m}}(\boldsymbol{{\partial}})j_0(kr),
\end{split}$$ which says that $\varphi_{\boldsymbol{m}}^{\text{reg}}(\boldsymbol{x},k)$ can be obtained by applying the operator $\mathcal{H}_{\boldsymbol{m}}(\boldsymbol{{\partial}})$ to $j_0(kr)$. Since $j_{\nu}(z)$ and $h_{\nu}^{(1)}(z)$ satisfies the same differential equation, it follows that $$\begin{aligned}
\varphi_{\boldsymbol{m}}^{\text{out}}(\boldsymbol{x},k)=&i^l\mathcal{C}_l^{\text{out}}\mathcal{H}_{\boldsymbol{m}}(\boldsymbol{{\partial}})h_0^{(1)}(kr).\end{aligned}$$ Namely, $$\begin{aligned}
\varphi_{\boldsymbol{m}}^{\text{out}}(\boldsymbol{x},k)=&2i^l \mathcal{C}_l^{\text{out}} a_DC_{\boldsymbol{m}}\int_{\mathbb{R}^{D-1}}d\boldsymbol{k}_{\perp}\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)\frac{e^{i\boldsymbol{k}\cdot\boldsymbol{r}}}{ k^{D-2}\sqrt{k^2-k_{\perp}^2}}.\end{aligned}$$
Applying the operator $\mathcal{H}_{\boldsymbol{m}''}(\boldsymbol{{\partial}})$ to $\varphi_{\boldsymbol{m}}^{\text{reg}}(\boldsymbol{x},k)$ and set $\boldsymbol{x}$ equal to $\mathbf{0}$, and imply that $$\label{eq3_3_4}\begin{split}
\left[\mathcal{H}_{\boldsymbol{m}''}(\boldsymbol{{\partial}})\varphi_{\boldsymbol{m}}^{\text{reg}}\right](\boldsymbol{0},k)=&\mathcal{C}_l^{\text{reg}}C_{\boldsymbol{m}}
C_{\boldsymbol{m}''}a_Di^l\int_{S^{D-1}_k}d\Omega_{D-1}^k\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)
\boldsymbol{Y}_{\boldsymbol{m}''}(\boldsymbol{\theta}_k)\\=&\mathcal{C}_l^{\text{reg}}a_Di^l\delta_{\boldsymbol{m}^*,\boldsymbol{m}''}.
\end{split}$$ For $\boldsymbol{m}=(m_1,\ldots,m_{D-1})$, we define $\boldsymbol{m}^*=(m_1,\ldots,m_{D-2},-m_{D-1})$, so that $$\boldsymbol{Y}_{\boldsymbol{m}^*}(\boldsymbol{\theta}_k)=\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)^*.$$
Hence, by applying $\mathcal{H}_{\boldsymbol{m}^*}(\boldsymbol{{\partial}})$ to and setting $\boldsymbol{x}=\mathbf{0}$, we find that $$\label{eq3_4_2}
\begin{split}
G^1_{\boldsymbol{m},\boldsymbol{m}'}=&\frac{1}{C_l^{\text{reg}}a_Di^l}\left[\mathcal{H}_{\boldsymbol{m}^*}(\boldsymbol{{\partial}})\varphi_{\boldsymbol{m}'}^{\text{out}}\right](-\boldsymbol{L},k)\\
=&\pi(-1)^{l'}i^{D-1}C_{\boldsymbol{m}}C_{\boldsymbol{m}'}\int_{\mathbb{R}^{D-1}}d\boldsymbol{k}_{\perp}\boldsymbol{Y}_{\boldsymbol{m}^*}(\boldsymbol{\theta}_k)
\boldsymbol{Y}_{\boldsymbol{m}'}(\boldsymbol{\theta}_k)\frac{e^{-i\boldsymbol{k}\cdot\boldsymbol{L}}}{ k^{D-2}\sqrt{k^2-k_{\perp}^2}}.
\end{split}$$ In principle, one can express $\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)
\boldsymbol{Y}_{\boldsymbol{m}'}(\boldsymbol{\theta}_k)$ as a linear combinations of $\boldsymbol{Y}_{\boldsymbol{m}''}(\boldsymbol{\theta}_k)$: $$\begin{aligned}
\boldsymbol{Y}_{\boldsymbol{m}}(\boldsymbol{\theta}_k)
\boldsymbol{Y}_{\boldsymbol{m}'}(\boldsymbol{\theta}_k)=\sum_{\boldsymbol{m}''}H_{\boldsymbol{m},\boldsymbol{m}'}^{\boldsymbol{m}''}\boldsymbol{Y}_{\boldsymbol{m}''}(\boldsymbol{\theta}_k).\end{aligned}$$When $D=3$, the constants $H_{\boldsymbol{m},\boldsymbol{m}'}^{\boldsymbol{m}''}$ are well-known and can be expressed as $3j$-symbols. However, the computations of these constants are not simple tasks. Therefore, we will use an alternative approach.
As in [@1], we express the integration over $\boldsymbol{k}_{\perp}$ in polar coordinates $$\begin{split}
k_2=&k_{\perp}\cos\theta_2^k,\\
k_3=&k_{\perp}\sin\theta_2^k\cos\theta_3^k,\\
&\vdots\\
k_{D-1}=&k_{\perp}\sin\theta_2^k\ldots\sin\theta_{D-2}^k\cos\theta_{D-1}^k,\\
k_{D}=&k_{\perp}\sin\theta_2^k\ldots\sin\theta_{D-2}^k\sin\theta_{D-1}^k.
\end{split}$$Replacing $k$ with $i\kappa$, we have $$\begin{split}
G^1_{\boldsymbol{m},\boldsymbol{m}'}=&\pi(-1)^{l'}i^{D-1}C_{\boldsymbol{m}}C_{\boldsymbol{m}'}\int_{0}^{\infty}d\boldsymbol{k}_{\perp} \boldsymbol{Y}_{\boldsymbol{m}^*}(\boldsymbol{\theta}_k)
\boldsymbol{Y}_{\boldsymbol{m}'}(\boldsymbol{\theta}_k)\frac{e^{-i\boldsymbol{k}\cdot\boldsymbol{L}}}{ k^{D-2}\sqrt{k^2-k_{\perp}^2}}\\
=&\pi(-1)^{l} i^{-m_2-m_2'} C_{\boldsymbol{m}}^2 \int_{0}^{\infty}dk_{\perp}k_{\perp}^{D-2}\int_0^{\pi}d\theta_2^k\sin^{D-3}\theta_2^k\ldots\int_0^{\pi}d\theta_{D-2}^k\sin\theta_{D-2}^k\int_{-\pi}^{\pi}d\theta_{D-1}^k
e^{-im_{D-1}\theta_{D-1}^k+im_{D-1}'\theta_{D-1}^k}\\
&\times \frac{e^{-L\sqrt{\kappa^2+k_{\perp}^2}}}{\kappa^{D-2}\sqrt{\kappa^2+k_{\perp}^2}}\left(\frac{k_{\perp}}{\kappa}\right)^{m_2+m_2'}
C_{l-m_2}^{m_{2}+\frac{D-2}{2}}\left(\frac{\sqrt{\kappa^2+k_{\perp}^2}}{\kappa}\right)C_{l'-m_2'}^{m_{2}'+\frac{D-2}{2}}\left(\frac{\sqrt{\kappa^2+k_{\perp}^2}}{\kappa}\right)
\\&\times\prod_{j=2}^{D-2}\sin^{|m_{j+1}|}\theta_j^k C_{m_j-|m_{j+1}|}^{|m_{j+1}|+\frac{D-j-1}{2}}(\cos\theta_j^k)\prod_{j'=2}^{D-2}\sin^{|m_{j'+1}'|}\theta_{j'}^k C_{m_{j'}'-|m_{j'+1}'|}^{|m_{j'+1}'|+\frac{D-j'-1}{2}}(\cos\theta_{j'}^k).
\end{split}$$Integrating over $\theta_2^k,\ldots,\theta_{D-1}^k$ using the orthogonality relation gives $$\begin{split}G^1_{\boldsymbol{m},\boldsymbol{m}'}=&(-1)^{l+m_2} \delta_{\boldsymbol{m}_{\perp},\boldsymbol{m}_{\perp}'}2^{2m_2+D-3} \Gamma\left(m_2+\frac{D-2}{2}\right)^2
\sqrt{\frac{\left(l+\frac{D-2}{2}\right)\left(l'+\frac{D-2}{2}\right)(l-m_2)!(l'-m_2)!}{(l+m_2+D-3)!(l'+m_2+D-3)!}}
\\&\times\int_{0}^{\infty}dk_{\perp}\,
\frac{e^{-L\sqrt{\kappa^2+k_{\perp}^2}}}{ \sqrt{\kappa^2+k_{\perp}^2}}\left(\frac{k_{\perp}}{\kappa}\right)^{2m_2 +D-2}
C_{l-m_2}^{m_{2}+\frac{D-2}{2}}\left(\frac{\sqrt{\kappa^2+k_{\perp}^2}}{\kappa}\right)C_{l'-m_2}^{m_{2}+\frac{D-2}{2}}\left(\frac{\sqrt{\kappa^2+k_{\perp}^2}}{\kappa}\right).
\end{split}$$Here $\boldsymbol{m}_{\perp}=(m_2,\ldots,m_{D-1})$. Making a change of variables $k_{\perp}=\kappa\sinh\theta$ gives $$\begin{split}G^1_{\boldsymbol{m},\boldsymbol{m}'}=&(-1)^{l+m_2}\delta_{\boldsymbol{m}_{\perp},\boldsymbol{m}_{\perp}'} 2^{2m_2+D-3} \Gamma\left(m_2+\frac{D-2}{2}\right)^2
\sqrt{\frac{\left(l+\frac{D-2}{2}\right)\left(l'+\frac{D-2}{2}\right)(l-m_2)!(l'-m_2)!}{(l+m_2+D-3)!(l'+m_2+D-3)!}}
\\&\times\int_{0}^{\infty}d\theta
\left(\sinh\theta\right)^{2m_2 +D-2}
C_{l-m_2}^{m_{2}+\frac{D-2}{2}}\left(\cosh\theta\right)C_{l'-m_2}^{m_{2}+\frac{D-2}{2}}\left(\cosh\theta\right)e^{-\kappa L\cosh\theta}.
\end{split}$$ In the same way, we find that $$\begin{split}G^2_{\boldsymbol{m}',\boldsymbol{m}}=&(-1)^{l+m_2}\delta_{\boldsymbol{m}_{\perp},\boldsymbol{m}_{\perp}'} 2^{2m_2+D-3} \Gamma\left(m_2+\frac{D-2}{2}\right)^2
\sqrt{\frac{\left(l+\frac{D-2}{2}\right)\left(l'+\frac{D-2}{2}\right)(l-m_2)!(l'-m_2)!}{(l+m_2+D-3)!(l'+m_2+D-3)!}}
\\&\times\int_{0}^{\infty}d\theta
\left(\sinh\theta\right)^{2m_2 +D-2}
C_{l-m_2}^{m_{2}+\frac{D-2}{2}}\left(\cosh\theta\right)C_{l'-m_2}^{m_{2}+\frac{D-2}{2}}\left(\cosh\theta\right)e^{-\kappa L\cosh\theta}.
\end{split}$$ Finally, the Casimir interaction energy between the two spheres are given by $$\begin{aligned}
\label{eq12_3_13}
E_{\text{Cas}}=\frac{\hbar c}{2\pi}\int_0^{\infty} d \kappa \text{Tr}\,\ln\left(1-\mathbb{M}(\kappa)\right),\end{aligned}$$ where the $(\boldsymbol{m},\boldsymbol{m}')$ element of $\mathbb{M}$ is given by $$\begin{aligned}
M_{\boldsymbol{m},\boldsymbol{m}'}=&T_{\boldsymbol{m}}^1\sum_{\boldsymbol{m}''}G^1_{\boldsymbol{m},\boldsymbol{m}''}T_{\boldsymbol{m}''}^2G_{\boldsymbol{m}'',\boldsymbol{m}'}^2.\end{aligned}$$
Since $\mathbb{M}$ is diagonal in $\boldsymbol{m}_{\perp}=(m_2,\ldots,m_{D-1})$, we can simplify the trace in as follows. When $D\geq 5$, $$\begin{aligned}
\label{eq12_4_1}
\sum_{m_3=0}^{m_2}\ldots\sum_{m_{D-2}=0}^{m_{D-3}}\sum_{m_{D-1}=-m_{D-2}}^{m_{D-2}}1= \frac{(2m_2+D-3)(m_2+D-4)!}{(D-3)!m_2!}.\end{aligned}$$When $D=4$, $$\begin{aligned}
\sum_{m_3=-m_2}^{m_2}1=2m_2+1,\end{aligned}$$ which is equal to the right hand side of when $D=4$. Hence, when $D\geq 4$, the Casimir interaction energy can be rewritten as $$\begin{aligned}
\label{eq12_3_14}
E_{\text{Cas}}=\frac{\hbar c}{2\pi}\int_0^{\infty} d \kappa \sum_{m=0}^{\infty}\frac{(2m+D-3)(m+D-4)!}{(D-3)!m!}\text{Tr}\,\ln\left(1-\mathbb{M}_{m}(\kappa)\right),\end{aligned}$$ where $m=m_2$ and the elements $M_{m;l,l'}$ of $\mathbb{M}_{m}$ is $$\begin{aligned}
M_{m;l,l'}=&T_{l}^1 \sum_{l''=m}^{\infty}G^1_{m;l,l''}T_{l''}^2G_{m;l'',l'}^2.\end{aligned}$$ For fixed $m$, $l,l'$ ranges from $m$ to $\infty$.
When $D=3$, we can also represent the Casimir interaction energy by provided that the summation $\sum_{m=0}^{\infty}$ is replaced by the summation $\sum_{m=0}^{\infty}\!'$, where the prime $\prime$ indicates that the term $m=0$ is summed with weight $1/2$.
Large separation asymptotic behavior
====================================
In this section, we consider the large separation asymptotic behavior of the Casimir interaction energy. Expanding the logarithm in , we have $$\begin{aligned}
\label{eq12_3_14_2}
E_{\text{Cas}}=-\frac{\hbar c}{2\pi}\sum_{s=0}^{\infty}\frac{1}{s+1}\int_0^{\infty} d \kappa \sum_{m=0}^{\infty}\frac{(2m+D-3)(m+D-4)!}{(D-3)!m!}\sum_{l_0=m}^{\infty}
\sum_{l_1=m}^{\infty}\ldots\sum_{l_s=m}^{\infty}\prod_{j=0}^sM_{m; l_jl_{j+1}}.\end{aligned}$$ Making the change of variables $$\kappa\mapsto\frac{\kappa}{L},$$ we find from the definition of $M_{m;l,l'}$ that in order to obtain the leading asymptotic behavior of the Casimir interaction energy when $L\gg 1$, we need the following asymptotic behaviors when $z\rightarrow 0$: $$\begin{aligned}
I_{\nu}(z)\sim & \frac{1}{\Gamma(\nu+1)}\left(\frac{z}{2}\right)^{\nu},\\
K_{\nu}(z)\sim &\frac{\Gamma(\nu)}{2}\left(\frac{z}{2}\right)^{-\nu}.\end{aligned}$$ It follows that for $i=1, 2$, $$\label{eq12_4_7}\begin{split}
T_{l}^{i, \text{D}}\left(\frac{\kappa}{L}\right)\sim & \frac{1}{2^{2l+D-3}\Gamma\left(l+\frac{D-2}{2}\right)\Gamma\left(l+\frac{D}{2}\right)}\left(\frac{\kappa R_i}{L}\right)^{2l+D-2},\\
T_{l}^{i, \text{N}}\left(\frac{\kappa}{L}\right)\sim & -\frac{l}{l+D-2}\frac{1}{2^{2l+D-3}\Gamma\left(l+\frac{D-2}{2}\right)\Gamma\left(l+\frac{D}{2}\right)}\left(\frac{\kappa R_i}{L}\right)^{2l+D-2},\quad\text{if}\;l\neq 0,\\
T_{0}^{i,\text{N}}\left(\frac{\kappa}{L}\right)\sim &-\frac{1}{2^{D-1}\Gamma\left(\frac{D}{2}\right)\Gamma\left(\frac{D+2}{2}\right)}\left(\frac{\kappa R_i}{L}\right)^{D}.
\end{split}$$ From these asymptotic behaviors, we find that the leading contribution to the large separation asymptotic behavior of the Casimir interaction energy comes from lower $l$ (and hence lower $m$) as well as smallest possible $s$, i.e., $s=0$. For Dirichlet boundary conditions, we only take $l=0$ (and hence $m=0$) for the leading term of large separation asymptotic expansion. However, for Neumann boundary conditions, both the $l=0$ and $l=1$ terms of $T_{l}^{i, \text{N}}\left(\kappa/L\right)$ have the same order in $L$, so we have to take both $l=0$ and $l=1$ terms to compute the leading term of the large separation asymptotic expansion. We have $$\begin{split}
E_{\text{Cas}}^{\text{DD}}\sim &-\frac{\hbar c}{2\pi L}\int_0^{\infty}d\kappa T_0^{1,\text{D}}G_{0;00}^1T_{0}^{2,\text{D}}G_{0;00}^2,
\\
E_{\text{Cas}}^{\text{DN}}\sim &-\frac{\hbar c}{2\pi L}\int_0^{\infty}d\kappa \left(T_0^{1,\text{D}}G_{0;00}^1T_{0}^{2,\text{N}}G_{0;00}^2+T_0^{1,\text{D}}G_{0;01}^1T_{1}^{2,\text{N}}G_{0;10}^2\right),\\
E_{\text{Cas}}^{\text{NN}}\sim & -\frac{\hbar c}{2\pi L}\int_0^{\infty}d\kappa \left(T_0^{1,\text{N}}G_{0;00}^1T_{0}^{2,\text{N}}G_{0;00}^2+T_0^{1,\text{N}}G_{0;01}^1T_{1}^{2,\text{N}}G_{0;10}^2+T_1^{1,\text{N}}G_{0;10}^1T_{0}^{2,\text{N}}G_{0;01}^2
\right.\\&\hspace{4cm}\left.+T_1^{1,\text{N}}G_{0;11}^1T_{1}^{2,\text{N}}G_{0;11}^2+(D-1)T_1^{1,\text{N}}G_{1;11}^1T_{1}^{2,\text{N}}G_{1;11}^2\right).
\end{split}$$ Using $$\begin{aligned}
C_0^{\nu}(z)=1,\hspace{1cm}C_1^{\nu}(z)=2\nu z,\end{aligned}$$we find that $$\begin{split}
G_{0;00}^1=G_{0;00}^2=&\sqrt{\pi}\frac{\Gamma\left(\frac{D}{2}\right)}{\Gamma\left(\frac{D-1}{2}\right)}\int_0^{\infty}d\theta\,\sinh^{D-2}\theta e^{-\kappa\cosh\theta}\\=& \left(\frac{2}{\kappa}\right)^{\frac{D-2}{2}}
\Gamma\left(\frac{D}{2}\right)K_{\frac{D-2}{2}}\left(\kappa\right),\end{split}$$ $$\begin{split}
G_{0;01}^1=G_{0;10}^2=-G_{0;10}^1=-G_{0;01}^2=&\sqrt{\pi D}\frac{\Gamma\left(\frac{D}{2}\right)}{\Gamma\left(\frac{D-1}{2}\right)}\int_0^{\infty}d\theta\,\sinh^{D-2}\theta\cosh\theta e^{-\kappa\cosh\theta}\\=&\sqrt{D} \left(\frac{2}{\kappa}\right)^{\frac{D-2}{2}}\Gamma\left(\frac{D}{2}\right)K_{\frac{D}{2}}\left(\kappa\right),
\end{split}$$ $$\begin{split}
G_{0;11}^1=G_{0;11}^2=& -2\sqrt{\pi }\frac{\Gamma\left(\frac{D+2}{2}\right)}{\Gamma\left(\frac{D-1}{2}\right)}\int_0^{\infty}d\theta\,\sinh^{D-2}\theta\cosh^2\theta e^{-\kappa\cosh\theta}\\
=&-2\Gamma\left(\frac{D+2}{2}\right)\left(\left(\frac{2}{\kappa}\right)^{\frac{D-2}{2}}
K_{\frac{D-2}{2}}\left(\kappa\right)+\frac{D-1}{2}\left(\frac{2}{\kappa}\right)^{\frac{D}{2}}
K_{\frac{D}{2}}\left(\kappa\right)\right),
\end{split}$$ $$\begin{split}
G_{1;11}^1=G_{1;11}^2=&\sqrt{\pi}\frac{\Gamma\left(\frac{D+2}{2}\right)}{\Gamma\left(\frac{D+1}{2}\right)}\int_0^{\infty}d\theta\,\sinh^{D}\theta e^{-\kappa\cosh\theta}\\=& \left(\frac{2}{\kappa}\right)^{\frac{D}{2}}
\Gamma\left(\frac{D+2}{2}\right)K_{\frac{D}{2}}\left(\kappa\right).\end{split}$$ It follows that the leading term of the large separation asymptotic expansion of the Casimir interaction energy is given by $$\begin{split}
E_{\text{Cas}}^{\text{DD}}\sim & -\frac{\hbar c}{2\sqrt{\pi}}\frac{\Gamma\left(\frac{D-1}{2}\right)^2\Gamma\left(D-\frac{3}{2}\right)}{\Gamma\left(\frac{D-2}{2}\right)^2(D-2)!}
\frac{(R_1R_2)^{D-2}}{L^{2D-3}},\\
E_{\text{Cas}}^{\text{DN}}\sim & \frac{\hbar c}{4\sqrt{\pi}}\frac{\Gamma\left(\frac{D+1}{2}\right)^2\Gamma\left(D-\frac{1}{2}\right)}{\Gamma\left(\frac{D-2}{2}\right)\Gamma\left(\frac{D+2}{2}\right)D!}\frac{2D^2-1}{D-1}
\frac{R_1^{D-2}R_2^{D}}{L^{2D-1}},\\
E_{\text{Cas}}^{\text{ND}}\sim & \frac{\hbar c}{4\sqrt{\pi}}\frac{\Gamma\left(\frac{D+1}{2}\right)^2\Gamma\left(D-\frac{1}{2}\right)}{\Gamma\left(\frac{D-2}{2}\right)\Gamma\left(\frac{D+2}{2}\right)D!}\frac{2D^2-1}{D-1}
\frac{R_1^{D}R_2^{D-2}}{L^{2D-1}},\\
E_{\text{Cas}}^{\text{NN}}\sim & -\frac{\hbar c}{8\sqrt{\pi}}\frac{\Gamma\left(\frac{D+3}{2}\right)^2\Gamma\left(D+\frac{1}{2}\right)}{\Gamma\left(\frac{D+2}{2}\right)^2 (D+2)!}\frac{(2D^2+2D-1)(2D^2+2D-3)}{(D-1)^2}
\frac{(R_1R_2)^{D}}{L^{2D+1}}.
\end{split}$$ In other words, the large separation leading terms of the Casimir interaction energies are of order $L^{-2D+3}$, $L^{-2D+1}$ and $L^{-2D-1}$ respectively for DD, DN/ND and NN boundary conditions. Hence, when the separation between the spheres is large, the interaction is strongest for Dirichlet-Dirchlet boundary conditions, and weakest for Neumann-Neumann boundary conditions. Moreover, the interaction gets weaker for higher dimensions.
When $D=3$, we find that $$\begin{aligned}
E_{\text{Cas}}^{\text{DD}}\sim &-\frac{\hbar cR_1R_2}{4\pi L^3},\\
E_{\text{Cas}}^{\text{DN}}\sim &\frac{17\hbar c R_1R_2^3}{48\pi L^5},\\
E_{\text{Cas}}^{\text{NN}}\sim &\frac{17 \hbar cR_1^3R_2}{48\pi L^5},\\
E_{\text{Cas}}^{\text{NN}}\sim &-\frac{161\hbar c R_1^3R_2^3}{96\pi L^7},\end{aligned}$$which agree with the results derived in [@6].
Proximity force approximation and small separation asymptotic behavior
======================================================================
The proximity force approximation approximates the Casimir interaction energy between two objects by summing the local Casimir energy density between two planes over the surfaces. For two planes both with Dirichlet (D) or Neumann (N) boundary conditions in $(D+1)$-dimensional Minkowski spacetime, the Casimir energy density is $$\begin{aligned}
\mathcal{E}_{\text{Cas}}^{\parallel,\text{DD/NN}} (d)=-\hbar c\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{D+1}\pi^{\frac{D+1}{2}}}\frac{1}{d^{D}}=\frac{b_D^{\text{DD/NN}}}{d^{D}};\end{aligned}$$whereas if one plane is Dirichlet and one is Neumann, the Casimir energy density is $$\begin{aligned}
\mathcal{E}_{\text{Cas}}^{\parallel,\text{DN/ND}} (d)=\hbar c\left(1-2^{-D}\right)\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{D+1}\pi^{\frac{D+1}{2}}}\frac{1}{d^{D}}=\frac{b_D^{\text{DN/ND}}}{d^{D}}.\end{aligned}$$Here $\zeta(z)=\sum_{n=1}^{\infty}n^{-z}$ is the Riemann zeta function.
Let $(R_1,\theta_1,\ldots,\theta_{D-1})$ be a point on the sphere with radius $R_1$ in hyper-spherical coordinates, the distance from this point to the sphere with radius $R_2$ is $$d(\boldsymbol{\theta})=\sqrt{L^2-2R_1L\cos\theta_1+R_1^2}-R_2.$$Notice that this only depends on $\theta_1$.
The proximity force approximation to the Casimir interaction energy between the two spheres is given by $$\begin{split}
E_{\text{Cas}}^{\text{PFA}}=&R_1^{D-1}\int_0^{\pi}d\theta_1\sin^{D-2}\theta_1\int_0^{\pi}d\theta_2\sin^{D-3}\theta_2\ldots\int_0^{\pi}d\theta_{D-2}\sin\theta_{D-2}\int_{-\pi}^{\pi}
d\theta_{D-1} \mathcal{E}_{\text{Cas}}^{\parallel}\left(d\left(\boldsymbol{\theta}\right)\right)\\
=&\frac{2\pi^{\frac{D-1}{2}}}{\Gamma\left(\frac{D-1}{2}\right)}b_DR_1^{D-1}\int_0^{\pi}d\theta_1\frac{\sin^{D-2}\theta_1}{\left(\sqrt{L^2-2R_1L\cos\theta_1+R_1^2}-R_2\right)^D}.
\end{split}$$ Let $$u=\sqrt{L^2-2R_1L\cos\theta_1+R_1^2}-R_2.$$ Then $$\begin{split}
E_{\text{Cas}}^{\text{PFA}}=&\frac{2\pi^{\frac{D-1}{2}}}{\Gamma\left(\frac{D-1}{2}\right)}b_D\frac{R_1 }{2^{D-3} L^{D-2}}\int_d^{L+R_1-R_2}du\left(u+R_2\right)
u^{-D}\\&\hspace{4cm}\times\left(\left[u-d\right]\left[u+R_2+L-R_1\right]\left[u+L+R_1+R_2\right]\left[L+R_1-R_2-u\right]\right)^{\frac{D-3}{2}}\\
\sim & \frac{(2\pi)^{\frac{D-1}{2}}}{\Gamma\left(\frac{D-1}{2}\right)} \left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}}\frac{b^D}{d^{\frac{D+1}{2}}}\int_1^{\infty} dv v^{-D}(v-1)^{\frac{D-3}{2}}\\
=&\frac{(2\pi)^{\frac{D-1}{2}}\Gamma\left(\frac{D+1}{2}\right)}{\Gamma(D)} \left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}}\frac{b^D}{d^{\frac{D+1}{2}}}\\
=&\frac{\pi^{\frac{D}{2}}}{2^{\frac{D-1}{2}}\Gamma\left(\frac{D}{2}\right)} \left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}}\frac{b^D}{d^{\frac{D+1}{2}}}.
\end{split}$$ Here $d=L-R_1-R_2$ is the distance between the two spheres. Hence, we find that the proximity force approximation to the Casimir interaction energy between two spheres is $$\label{eq12_5_10}\begin{split}
E_{\text{Cas}}^{\text{PFA, DD/NN}}\sim &-\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}} \frac{\hbar c}{d^{\frac{D+1}{2}}},\\
E_{\text{Cas}}^{\text{PFA, DN/ND}}\sim &(1-2^{-D})\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}} \frac{\hbar c}{d^{\frac{D+1}{2}}}.
\end{split}$$
Next, we derive the small separation asymptotic behavior of the Casimir interaction energy from the TGTG formula. Set $$\tilde{m}=m+\frac{D-3}{2},$$and make a change of variables $$\kappa =\frac{\omega}{R_1+R_2},$$we obtain from that $$\begin{aligned}
\label{eq3_6_1}
E_{\text{Cas}}=-\frac{\hbar c}{\pi(R_1+R_2)}\frac{1}{(D-3)!}\sum_{s=0}^{\infty}\frac{1}{s+1}\int_0^{\infty} d \omega \sum_{m=0}^{\infty}\frac{\tilde{m}\left(\tilde{m}+\frac{D-5}{2}\right)!}{ \left(\tilde{m}-\frac{D-3}{2}\right)!} \sum_{l_0=\tilde{m}-\frac{D-3}{2}}^{\infty}\sum_{l_1=\tilde{m}-\frac{D-3}{2}}^{\infty}\ldots \sum_{l_s=\tilde{m}-\frac{D-3}{2}}^{\infty}
\prod_{j=0}^s M_{l_j,l_{j+1}},\end{aligned}$$ where $$\begin{aligned}
\label{eq3_10_1}
M_{l_j,l_{j+1}}=T_{l_j}^1\sum_{l_j'=\tilde{m}}^{\infty}U^1_{l_j,l_j'}T_{l_j'}^2U^2_{l_j',l_{j+1}},\end{aligned}$$ $$\begin{split}
&T_{l_j}^{1, \text{D}} =\frac{I_{l_j+\frac{D-2}{2}}(\omega a_1)}{K_{l_j+\frac{D-2}{2}}(\omega a_1)},\\
& T_{l_j}^{1, \text{N}} =\frac{-\frac{D-2}{2}I_{l_j+\frac{D-2}{2}}(\omega a_1)+\omega a_1I_{l_j+\frac{D-2}{2}}'(\omega a_1)}
{-\frac{D-2}{2}K_{l_j+\frac{D-2}{2}}(\omega a_1)+\omega a_1 K_{l_j+\frac{D-2}{2}}'(\omega a_1)},\\
&T_{l_j'}^{2, \text{D}} =\frac{I_{l_j'+\frac{D-2}{2}}(\omega a_2)}{K_{l_j'+\frac{D-2}{2}}(\omega a_2)},\\
& T_{l_j'}^{2, \text{N}} =\frac{-\frac{D-2}{2}I_{l_j'+\frac{D-2}{2}}(\omega a_2)+\omega a_2I_{l_j'+\frac{D-2}{2}}'(\omega a_2)}
{-\frac{D-2}{2}K_{l_j'+\frac{D-2}{2}}(\omega a_2)+\omega a_2 K_{l_j'+\frac{D-2}{2}}'(\omega a_2)},
\end{split}$$ $$\begin{split}U^1_{l_j,l_j'}=& 2^{2\tilde{m}} \Gamma\left(\tilde{m}+\frac{1}{2}\right)^2
\sqrt{\frac{\left(l_j+\frac{D-2}{2}\right)\left(l_j'+\frac{D-2}{2}\right)\Gamma\left(l_j+\frac{D-1}{2}-\tilde{m}\right)\Gamma\left(l_j'+\frac{D-1}{2}-\tilde{m}\right)}
{\Gamma\left(l_j+\frac{D-1}{2}+\tilde{m}\right)\Gamma\left(l_j'+\frac{D-1}{2}+\tilde{m}\right)}}
\\&\times\int_{0}^{\infty}d\theta\sinh\theta
\left(\sinh\theta\right)^{2\tilde{m}}
C_{l_j+\frac{D-3}{2}-\tilde{m}}^{\tilde{m}+\frac{1}{2}}\left(\cosh\theta\right)C_{l_j'+\frac{D-3}{2}-\tilde{m}}^{\tilde{m}+\frac{1}{2}}
\left(\cosh\theta\right)e^{-\omega\left(1+{\varepsilon}\right)\cosh\theta},\\
U^2_{l_{j+1},l_j'}=& 2^{2\tilde{m}} \Gamma\left(\tilde{m}+\frac{1}{2}\right)^2
\sqrt{\frac{\left(l_{j+1}+\frac{D-2}{2}\right)\left(l_j'+\frac{D-2}{2}\right)\Gamma\left(l_{j+1}+\frac{D-1}{2}-\tilde{m}\right)\Gamma\left(l_j'+\frac{D-1}{2}-\tilde{m}\right)}
{\Gamma\left(l_{j+1}+\frac{D-1}{2}+\tilde{m}\right)\Gamma\left(l_j'+\frac{D-1}{2}+\tilde{m}\right)}}
\\&\times\int_{0}^{\infty}d\theta\sinh\theta
\left(\sinh\theta\right)^{2\tilde{m}}
C_{l_{j+1}+\frac{D-3}{2}-\tilde{m}}^{\tilde{m}+\frac{1}{2}}\left(\cosh\theta\right)C_{l_j'+\frac{D-3}{2}-\tilde{m}}^{\tilde{m}+\frac{1}{2}}
\left(\cosh\theta\right)e^{-\omega\left(1+{\varepsilon}\right)\cosh\theta}.
\end{split}$$Here $$\begin{aligned}
a_i=\frac{R_i}{R_1+R_2},\hspace{1cm}{\varepsilon}=\frac{d}{R_1+R_2}.\end{aligned}$$Notice that $U^2_{l_{j+1},l_j'}$ is obtained from $U^1_{l_j,l_j'}$ by replacing $l_j$ with $l_{j+1}$.
Next, we introduce new variables $n_1, \ldots, n_s$, $q_0, q_1, \ldots, q_s$ and $\tau$ such that $$\begin{aligned}
l_0=&l,\hspace{1cm} l_j=l+n_j,\quad 1\leq j\leq s, \\
l_j'=&\frac{a_2}{2a_1}\left(l_j+l_{j+1}\right)+q_j=\frac{a_2}{a_1}l+\frac{a_2}{2a_1}\left(n_j+n_{j+1}\right)+q_j,\quad 0\leq j\leq s,\\
\omega=&\frac{l\sqrt{1-\tau^2}}{a_1\tau}.\end{aligned}$$When ${\varepsilon}$ is small, the leading contributions to the Casimir interaction energy come from terms with $l\sim {\varepsilon}^{-1}$, $n_i, q_i, \tilde{m} \sim {\varepsilon}^{-1/2}$ and $\tau\sim 1$.
As explained in [@1], we have $$\label{eq3_12_1}\begin{split}
&\sinh^{\tilde{m}}\theta C_{l_j+\frac{D-3}{2}-\tilde{m}}^{\tilde{m}+\frac{1}{2}}(\cosh\theta)\\=&\frac{\Gamma\left(l_j+\frac{D-1}{2}+\tilde{m}\right)}{\Gamma\left(\tilde{m}+\frac{1}{2}\right)
\Gamma\left(l_j+\frac{D-1}{2}\right)}
\frac{1}{ 2^{\tilde{m}}\sqrt{\pi} }\int_{-\frac{\pi}{2}}^{\frac{\pi}{2}} d\varphi\left(\cosh\theta+\sinh\theta\cos2\varphi\right)^{l_j+\frac{D-3}{2}}e^{2i\tilde{m}\varphi}\\
=&\frac{\Gamma\left(l_j+\frac{D-1}{2}+\tilde{m}\right)}{\Gamma\left(\tilde{m}+\frac{1}{2}\right)}\frac{1}{ 2^{\tilde{m}}\sqrt{\pi} }
\sum_{k=0}^{\infty} \frac{1}{k!\Gamma\left(l_j+\frac{D-1}{2}-k\right)}e^{\left(l_j+\frac{D-3}{2}-2k\right)\theta}\int_{-\frac{\pi}{2}}^{\frac{\pi}{2}}d\varphi
\left(\cos\varphi\right)^{2l_j+D-3-2k}\left(\sin\varphi\right)^{2k}e^{2i\tilde{m}\varphi}.
\end{split}$$ Hence, $$\begin{aligned}
U^1_{l_j,l_j'}=& \frac{1}{\pi}\sum_{k=0}^{\infty}\sum_{k'=0}^{\infty}\frac{1}{k!k'!}\mathcal{N}_{l_j,l_j';k,k'}\int_{0}^{\infty}d\theta\sinh\theta e^{\left(l_j+l_j'+D-3-2k-2k'\right)\theta}
e^{-\omega\left(1+{\varepsilon}\right)\cosh\theta}\\
&\times \int_{-\frac{\pi}{2}}^{\frac{\pi}{2}}d\varphi
\left(\cos\varphi\right)^{2l_j+D-3-2k}\left(\sin\varphi\right)^{2k}e^{2i\tilde{m}\varphi}\int_{-\frac{\pi}{2}}^{\frac{\pi}{2}}d\varphi'
\left(\cos\varphi'\right)^{2l_j'+D-3-2k'}\left(\sin\varphi'\right)^{2k'}e^{2i\tilde{m}\varphi'},\end{aligned}$$where $$\begin{aligned}
\mathcal{N}_{l_j,l_j';k,k'}=&\sqrt{ \left(l_j+\frac{D-2}{2}\right)\left(l_j'+\frac{D-2}{2}\right)}\\&\times \frac{\sqrt{\Gamma\left(l_j+\frac{D-1}{2}-\tilde{m}\right)\Gamma\left(l_j'+\frac{D-1}{2}-\tilde{m}\right) \Gamma\left(l_j+\frac{D-1}{2}+\tilde{m}\right)\Gamma\left(l_j'+\frac{D-1}{2}+\tilde{m}\right)}}{\Gamma\left(l_j+\frac{D-1}{2}-k\right)\Gamma\left(l_j'+\frac{D-1}{2}-k'\right)}.\end{aligned}$$ As in [@1; @2], we can now find the asymptotic behaviors of $U^1_{l_j,l_j'}$. Using $$\begin{aligned}
\left(\cos\varphi\right)^{2l_j+D-3-2k}=&\exp\left(\left(2l_j+D-3-2k\right)\ln\cos\varphi\right)\\= &\exp\left(-\left(l_j+\frac{D-3}{2}-k\right)\varphi^2-
\frac{\left(2l_j+D-3-2k\right)}{12}\varphi^4+\ldots\right),\\
\left(\sin\varphi\right)^{2k}=&\varphi^{2k}\left(1-\frac{\varphi^2}{6}+\ldots\right)^{2k},\end{aligned}$$we observe that when ${\varepsilon}$ is small, the main contribution to the Casimir interaction energy comes from terms with $\varphi\sim {\varepsilon}^{1/2}$. Making a change of variable $$\varphi\mapsto \frac{\varphi}{\sqrt{l}}$$ so that $\varphi\sim 1$, we have an asymptotic expansion of the form: $$\begin{split}\int_{-\frac{\pi}{2}}^{\frac{\pi}{2}}d\varphi
\left(\cos\varphi\right)^{2l_j+D-3-2k}\left(\sin\varphi\right)^{2k}e^{2i\tilde{m}\varphi}\sim \frac{1}{l^{k+\frac{1}{2}}}\int_{-\infty}^{\infty}
d\varphi \varphi^{2k}\left(1+\mathcal{B}_{j,2}\right)\exp\left(\mathcal{A}_{j,1}+\mathcal{A}_{j,2}\right)\exp\left(-\varphi^2+\frac{2i\tilde{m}}{\sqrt{l}}\varphi\right).
\end{split}$$Here and in the following, for any $\mathcal{X}$, $\mathcal{X}_{j,1}$ and $\mathcal{X}_{j,2}$ represent respectively terms of order $\sqrt{{\varepsilon}}$ and ${\varepsilon}$.
In a similar way, we have an expansion of the form $$\begin{split}\int_{-\frac{\pi}{2}}^{\frac{\pi}{2}}d\varphi'
\left(\cos\varphi'\right)^{2l_j'+D-3-2k'}\left(\sin\varphi'\right)^{2k'}e^{2i\tilde{m}\varphi'}\sim \frac{1}{l^{k'+\frac{1}{2}}}\int_{-\infty}^{\infty}
d\varphi' \varphi^{\prime 2k'}\left(1+\mathcal{D}_{j,2}\right)\exp\left(\mathcal{C}_{j,1}+\mathcal{C}_{j,2}\right)\exp\left(-\frac{a_2}{a_1}\varphi^{\prime 2}+\frac{2i\tilde{m}}{\sqrt{l}}\varphi'\right).
\end{split}$$
Let $\theta_0>0$ be such that $$\begin{aligned}
\sinh\theta_0=\frac{\tau}{\sqrt{1-\tau^2}}.\end{aligned}$$Namely, $$\begin{aligned}
\theta_0=\frac{1}{2}\ln\frac{1+\tau}{1-\tau}.\end{aligned}$$ Then $$\begin{split}
&\int_{0}^{\infty}d\theta\sinh\theta e^{\left(l_j+l_j'+D-3-2k-2k'\right)\theta}
e^{-\omega\left(1+{\varepsilon}\right)\cosh\theta}\\=&\int_{-\theta_0}^{\infty}d\theta\sinh\left(\theta+\theta_0\right) e^{\left(l_j+l_j'+D-3-2k-2k'\right)(\theta+\theta_0)}
e^{-\omega\left(1+{\varepsilon}\right)\cosh\left(\theta+\theta_0\right)}.
\end{split}$$When ${\varepsilon}\ll 1$, the main contribution to the Casimir interaction energy comes from terms with $\theta\sim {\varepsilon}^{1/2}$. Hence up to leading contributions, the integration from $-\theta_0$ to $\infty$ can be replaced by integration from $-\infty$ to $\infty$. Since $$\begin{split}
\sinh\left(\theta+\theta_0\right) =&\frac{\tau}{\sqrt{1-\tau^2}}\left(1+\frac{\theta}{\tau}+\frac{\theta^2}{2}+\ldots\right),\\
\cosh\left(\theta+\theta_0\right) =&\frac{1}{\sqrt{1-\tau^2}}\left(1+ \tau\theta +\frac{\theta^2}{2}+\frac{\tau}{6}\theta^3+\frac{\theta^4}{24}+\ldots\right),
\end{split}$$ we obtain an expansion of the form: $$\begin{split}
&\int_{-\theta_0}^{\infty}d\theta\sinh\left(\theta+\theta_0\right) e^{\left(l_j+l_j'+D-3-2k-2k'\right)(\theta+\theta_0)}
e^{-\omega\left(1+{\varepsilon}\right)\cosh\left(\theta+\theta_0\right)}\\
\sim &\left(\frac{1+\tau}{1-\tau}\right)^{\frac{l}{2a_1}+\frac{(a_1+1)}{4a_1}n_j+\frac{a_2}{4a_1}n_{j+1}+\frac{q_j}{2}+\frac{D-3}{2}-k-k'}\frac{\tau}{\sqrt{1-\tau^2}}
\int_{-\infty}^{\infty} d\theta \left(1+\mathcal{F}_{j,1}+\mathcal{F}_{j,2}\right)\\
&\times\exp\left(\mathcal{E}_{j,1}+\mathcal{E}_{j,2}\right)
\exp\left(-\frac{l}{a_1\tau}-\frac{l\theta^2}{2a_1\tau}+\left(\frac{a_1+1}{2a_1}n_j+\frac{a_2}{2a_1}n_{j+1}+q_j\right)\theta-\frac{l_1{\varepsilon}}{a_1\tau}\right).
\end{split}$$For the term $\mathcal{N}_{l_j,l_j';k,k'}$, $$\begin{aligned}
\sqrt{ \left(l_j+\frac{D-2}{2}\right)\left(l_j'+\frac{D-2}{2}\right)}\sim \sqrt{\frac{a_2}{a_1}}l\left(1+\mathcal{G}_{j,1}+\mathcal{G}_{j,2}\right),\end{aligned}$$and by using the Stirling’s formula $$\ln\Gamma(z)\sim \left(z-\frac{1}{2}\right)\ln z-z +\frac{1}{2}\ln (2\pi)+\frac{1}{12z}+\ldots,$$we have $$\begin{split}
&\frac{\sqrt{\Gamma\left(l_j+\frac{D-1}{2}-\tilde{m}\right)\Gamma\left(l_j'+\frac{D-1}{2}-\tilde{m}\right) \Gamma\left(l_j+\frac{D-1}{2}+\tilde{m}\right)\Gamma\left(l_j'+\frac{D-1}{2}+\tilde{m}\right)}}{\Gamma\left(l_j+\frac{D-1}{2}-k\right)\Gamma\left(l_j'+\frac{D-1}{2}-k'\right)}\\
\sim & l^{k+k'}\left(\frac{a_2}{a_1}\right)^{k_2}\exp\left(\mathcal{H}_{j,1}+\mathcal{H}_{j,2}\right)\exp\left(\frac{\tilde{m}^2}{2a_2l}\right)
\end{split}$$ Hence, $$\begin{split}
U^1_{l_j,l_j'}\sim & \frac{1}{\pi}\sum_{k=0}^{\infty}\sum_{k'=0}^{\infty}\frac{1}{k!k'!}\left(\frac{a_2}{a_1}\right)^{k'+\frac{1}{2}}\left(\frac{1+\tau}{1-\tau}\right)^{\frac{l}{2a_1}+\frac{(a_1+1)}{4a_1}n_j+\frac{a_2}{4a_1}n_{j+1}+\frac{q_j}{2}+\frac{D-3}{2}-k-k'}\frac{\tau}{\sqrt{1-\tau^2}}
\exp\left(\frac{\tilde{m}^2}{2a_2l}\right)\\&\times \int_{-\infty}^{\infty} d\theta
\exp\left(-\frac{l}{a_1\tau}-\frac{l\theta^2}{2a_1\tau}+\left(\frac{a_1+1}{2a_1}n_j+\frac{a_2}{2a_1}n_{j+1}+q_j\right)\theta
-\frac{l_1{\varepsilon}}{a_1\tau}\right)\\&\times \int_{-\infty}^{\infty}d\varphi \varphi^{2k}\exp\left(-\varphi^2+\frac{2i\tilde{m}}{\sqrt{l}}\varphi\right)
\int_{-\infty}^{\infty}d\varphi' \varphi^{\prime 2k'}\exp\left(-\frac{a_2}{a_1}\varphi^{\prime 2}+\frac{2i\tilde{m}}{\sqrt{l}}\varphi'\right)\left(1+\mathcal{J}_{j,1}+\mathcal{J}_{j,2}\right),
\end{split}$$ where $$\begin{aligned}
\mathcal{I}_{j,1}=&\mathcal{A}_{j,1}+\mathcal{C}_{j,1}+\mathcal{E}_{j,1}+\mathcal{H}_{j,1},\\
\mathcal{I}_{j,2}=&\mathcal{A}_{j,2}+\mathcal{C}_{j,2}+\mathcal{E}_{j,2}+\mathcal{H}_{j,2},\\
\mathcal{J}_{j,1}=&\mathcal{F}_{j,1}+\mathcal{G}_{j,1}+\mathcal{I}_{j,1},\\
\mathcal{J}_{j,2}=&\mathcal{F}_{j,1}\mathcal{G}_{j,1}+\mathcal{F}_{j,1}\mathcal{I}_{j,1}+\mathcal{G}_{j,1}\mathcal{I}_{j,1}+\mathcal{B}_{j,2}+\mathcal{D}_{j,2}+\mathcal{F}_{j,2}+\mathcal{G}_{j,2}
+\mathcal{I}_{j,2}+\frac{1}{2}\mathcal{I}_{j,1}^2.\end{aligned}$$ Now we perform the summation over $k$ and $k'$ using the following identities: $$\begin{aligned}
&\sum_{k=0}^{\infty} \frac{v^k}{k!}=e^v,\\
&\sum_{k=0}^{\infty} k\frac{v^k}{k!}=ve^v,\\
&\sum_{k=0}^{\infty} k^2\frac{v^k}{k!}=(v^2+v)e^v.\end{aligned}$$ This gives an expansion of the form $$\label{eq3_10_2}
\begin{split}
U^1_{l_j,l_j'}\sim & \frac{1}{\pi}\sqrt{\frac{a_2}{a_1}}
\left(\frac{1+\tau}{1-\tau}\right)^{\frac{l}{2a_1}+\frac{(a_1+1)}{4a_1}n_j+\frac{a_2}{4a_1}n_{j+1}+\frac{q_j}{2}+\frac{D-3}{2} }\frac{\tau}{\sqrt{1-\tau^2}}\exp\left(\frac{\tilde{m}^2}{2a_2l}\right)
\\&\times \int_{-\infty}^{\infty} d\theta
\exp\left(-\frac{l}{a_1\tau}-\frac{l\theta^2}{2a_1\tau}+\left(\frac{a_1+1}{2a_1}n_j+\frac{a_2}{2a_1}n_{j+1}+q_j\right)\theta
-\frac{l_1{\varepsilon}}{a_1\tau}\right)\\&\times \int_{-\infty}^{\infty}d\varphi \exp\left(-\frac{2\tau}{1+\tau}\varphi^2+\frac{2i\tilde{m}}{\sqrt{l}}\varphi\right)
\int_{-\infty}^{\infty}d\varphi' \exp\left(-\frac{a_2}{a_1}\frac{2\tau}{1+\tau}\varphi^{\prime 2}+\frac{2i\tilde{m}}{\sqrt{l}}\varphi'\right)\left(1+\mathcal{K}_{j,1}+\mathcal{K}_{j,2}\right).
\end{split}$$ The integrations over $\varphi, \varphi'$ and $\theta$ are Gaussian and can be performed straightforwardly to give an expansion of the form $$\label{eq3_10_3}
\begin{split}
U^1_{l_j,l_j'}\sim & \sqrt{\frac{\pi a_1 \tau}{2 l}}
\left(\frac{1+\tau}{1-\tau}\right)^{\frac{l}{2a_1}+\frac{(a_1+1)}{4a_1}n_j+\frac{a_2}{4a_1}n_{j+1}+\frac{q_j}{2}+\frac{D-2}{2} }
\\&\times
\exp\left(-\frac{l}{a_1\tau}+\frac{\tau}{2a_1l}\left(\frac{a_1+1}{2}n_j+\frac{a_2}{2}n_{j+1}+a_1q_j\right)^2
-\frac{l{\varepsilon}}{a_1\tau}-\frac{\tilde{m}^2}{2\tau a_2l}\right) \left(1+\mathcal{M}_{j,1}+\mathcal{M}_{j,2}\right).
\end{split}$$ $U^2_{l_j',l_{j+1}}$ is obtained from $U^1_{l_j,l_j'}$ by interchanging $n_j$ and $n_{j+1}$. Namely, we have an expansion of the form $$\begin{split}
U^2_{l_j',l_{j+1}}\sim & \sqrt{\frac{\pi a_1 \tau}{2 l}}
\left(\frac{1+\tau}{1-\tau}\right)^{\frac{l}{2a_1}+\frac{a_2}{4a_1}n_{j}+\frac{(a_1+1)}{4a_1}n_{j+1}+\frac{q_j}{2}+\frac{D-2}{2} }
\\&\times
\exp\left(-\frac{l}{a_1\tau}+\frac{\tau}{2a_1l}\left(\frac{a_2}{2}n_{j}+\frac{a_1+1}{2}n_{j+1}+a_1q_j\right)^2
-\frac{l{\varepsilon}}{a_1\tau}-\frac{\tilde{m}^2}{2\tau a_2l}\right) \left(1+\mathcal{N}_{j,1}+\mathcal{N}_{j,2}\right),
\end{split}$$where $\mathcal{N}_{j,1}$ and $\mathcal{N}_{j,2}$ are obtained respectively from $\mathcal{M}_{j,1}$ and $\mathcal{M}_{j,2}$ by interchanging $n_j$ and $n_{j+1}$.
Next, we consider the asymptotic expansion of $T^2_{l_j'}$. Debye uniform asymptotic expansions of modified Bessel functions state that [@3; @4]:
$$\begin{aligned}
&I_{\nu}(\nu z) \sim \frac{1}{\sqrt{2\pi \nu}}\frac{e^{\nu\eta(z)}}{(1+z^2)^{1/4}} \left(1+\frac{u_1(t(z))}{\nu}\right),\\
&K_{\nu}(\nu z) \sim \sqrt{\frac{\pi}{2 \nu}}\frac{e^{-\nu\eta(z)}}{(1+z^2)^{1/4}} \left(1-\frac{u_1(t(z))}{\nu}\right),\\
&-\frac{D-2}{2}I_{\nu}(\nu z)+\nu zI_{\nu}'(\nu z)\sim \frac{\sqrt{\nu}e^{\nu\eta(z)}(1+z^2)^{1/4}}{\sqrt{2\pi}}\left(1+ \frac{m_1(t (z))}{\nu }\right),\\
&-\frac{D-2}{2}K_{\nu}(\nu z)+\nu zK_{\nu}'(\nu z)\sim- \sqrt{\frac{\pi\nu}{2}} e^{-\nu\eta(z)}(1+z^2)^{1/4}\left(1- \frac{m_1(t (z))}{\nu }\right),\end{aligned}$$
where $$\begin{aligned}
u_1(t )=&\frac{t }{8}-\frac{5t^3}{24},\hspace{1cm}m_1(t )=-\frac{(4D-5)t}{8}+\frac{7t^3}{24},\\
t (z)=&\frac{1}{\sqrt{1+z^2}},\hspace{1cm} \eta(z)=\sqrt{1+z^2}+\ln\frac{z}{1+\sqrt{1+z^2}}.\end{aligned}$$ Hence, $$\begin{aligned}
T^{2, \text{D}}_{l_j'}\sim & \frac{1}{\pi}e^{2\nu\eta(z)}\left(1+\frac{2u_1(t(z))}{\nu}\right),\\
T^{2, \text{N}}_{l_j'}\sim & -\frac{1}{\pi}e^{2\nu\eta(z)}\left(1+\frac{2m_1(t(z))}{\nu}\right),\end{aligned}$$ where $$\begin{aligned}
\nu=l_j'+\frac{D-2}{2},\hspace{1cm} z=\frac{\omega a_2 }{\nu}.\end{aligned}$$ Hence, we have an expansion of the form $$\label{eq3_10_4}\begin{split}
T_{l_j'}^{2, \text{Y}}\sim & \frac{(-1)^{\alpha_\text{Y}}}{\pi}\left(\frac{1+\tau}{1-\tau}\right)^{-\frac{la_2}{a_1}-\frac{a_2}{2a_1}(n_j+n_{j+1})-q_j-\frac{D-2}{2}}
\exp\left(\frac{2a_2l}{a_1\tau}-\frac{a_2\tau(n_j+n_{j+1})^2}{4a_1l}-\frac{a_1\tau q_j^2}{a_2l}-\frac{\tau q_j(n_j+n_{j+1})}{l}\right)\\&\times\exp\left(\mathcal{O}_{j,1}+\mathcal{O}_{j,2}\right)\left(1+\mathcal{P}^{\text{Y}}\right),
\end{split}$$ where $\alpha_{\text{D}}=0, \alpha_{\text{N}}=1$, $$\begin{aligned}
\mathcal{P}^{\text{D}}=&\frac{a_1}{a_2 l}\left(\frac{\tau }{4}-\frac{5\tau^3}{12}\right),\\
\mathcal{P}^{\text{N}}=&\frac{a_1}{a_2 l}\left(-\frac{(4D-5)\tau }{4}+\frac{7\tau^3}{12}\right).\end{aligned}$$ The summation over $l_j'$ in can be replaced by summation over $q$, which, to leading contributions to the Casimir interaction energy, can be approximated by an integration over $q$ from $-\infty$ to $\infty$. From , and , we have $$\begin{aligned}
\sum_{l_j'=\tilde{m}}^{\infty}U^1_{l_j,l_j'}T_{l_j'}^2U^2_{l_j',l_{j+1}}\sim & \left(\frac{1+\tau}{1-\tau}\right)^{l+\frac{1}{2}(n_j+n_{j+1})+\frac{D-2}{2}}
\exp\left(-\frac{2l}{\tau}+\frac{(1+a_1)\tau\left(n_j^2+n_{j+1}^2\right)}{4l}+\frac{(1-a_1)\tau n_jn_{j+1}}{2l}
-\frac{2l{\varepsilon}}{a_1\tau}-\frac{\tilde{m}^2}{\tau a_2l}\right) \\&\times (-1)^{\alpha_\text{Y}}\frac{a_1\tau}{2l}\left(1+\mathcal{P}^{\text{Y}}\right)\int_{-\infty}^{\infty} dq_j
\exp\left( -\frac{a_1^2\tau q_j^2}{a_2l} \right) \left(1+\mathcal{Q}_{j,1}+\mathcal{Q}_{j,2}\right),\end{aligned}$$ where $$\begin{split}
\mathcal{Q}_{j,1}=&\mathcal{M}_{j,1}+\mathcal{N}_{j,1}+\mathcal{O}_{j,1},\\
\mathcal{Q}_{j,2}=&\mathcal{M}_{j,1}\mathcal{N}_{j,1}+\mathcal{M}_{j,1}\mathcal{O}_{j,1}+\mathcal{N}_{j,1}\mathcal{O}_{j,1}+\mathcal{M}_{j,2}+\mathcal{N}_{j,2}+\mathcal{O}_{j,2}+
\frac{1}{2}\mathcal{O}_{j,1}^2.
\end{split}$$ The integration over $q$ is straightforward and we obtain an expansion of the form $$\begin{aligned}
\sum_{l_j'=\tilde{m}}^{\infty}U^1_{l_j,l_j'}T_{l_j'}^2U^2_{l_j',l_{j+1}}\sim & \left(\frac{1+\tau}{1-\tau}\right)^{l+\frac{1}{2}(n_j+n_{j+1})+\frac{D-2}{2}}
\exp\left(-\frac{2l}{\tau}+\frac{(1+a_1)\tau\left(n_j^2+n_{j+1}^2\right)}{4l}+\frac{ a_2\tau n_jn_{j+1}}{2l}
-\frac{2l{\varepsilon}}{a_1\tau}-\frac{\tilde{m}^2}{\tau a_2l}\right) \\&\times \frac{(-1)^{\alpha_\text{Y}}}{2}\sqrt{\frac{\pi a_2\tau}{l}}\left(1+\mathcal{P}^{\text{Y}}\right) \left(1+\mathcal{R}_{j,1}+\mathcal{R}_{j,2}\right).\end{aligned}$$ Next, we consider the asymptotic expansion of $T_{l_j}^1$. Similar to $T_{l_j'}^2$, we have $$\label{eq3_10_4_2}\begin{split}
T_{l_j}^{1, \text{X}}\sim & \frac{(-1)^{\alpha_\text{X}}}{\pi}C^{n_j-n_{j+1}}\left(\frac{1+\tau}{1-\tau}\right)^{-l-\frac{1}{2 }(n_j+n_{j+1}) -\frac{D-2}{2}}
\exp\left(\frac{2 l}{ \tau}-\frac{ \tau(n_j^2+n_{j+1}^2)}{2 l} \right)\\&\times\exp\left(\mathcal{S}_{j,1}+\mathcal{S}_{j,2}\right)\left(1+\mathcal{T}^{\text{Y}}\right),
\end{split}$$ where $$\begin{aligned}
\mathcal{T}^{\text{D}}=&\frac{1}{ l}\left(\frac{\tau }{4}-\frac{5\tau^3}{12}\right),\\
\mathcal{T}^{\text{N}}=&\frac{1}{ l}\left(-\frac{(4D-5)\tau }{4}+\frac{7\tau^3}{12}\right).\end{aligned}$$ Thence, $M_{l_j,l_{j+1}}$ has an expansion of the form $$\begin{aligned}
\label{eq3_10_6}
M_{l_j,l_{j+1}}^{\text{XY}}\sim \frac{(-1)^{\alpha_{\text{X}}+\alpha_{\text{Y}}}}{2}\sqrt{\frac{ a_2\tau}{\pi l}}
\exp\left( -\frac{a_2\tau\left(n_j-n_{j+1}\right)^2}{4l}
-\frac{2l{\varepsilon}}{a_1\tau}-\frac{\tilde{m}^2}{\tau a_2l}\right) \left(1+\mathcal{T}^{\text{X}}+\mathcal{P}^{\text{Y}}\right) \left(1+\mathcal{U}_{j,1}+\mathcal{U}_{j,2}\right),\end{aligned}$$ where $$\begin{aligned}
\mathcal{U}_{j,1}=&\mathcal{R}_{j,1}+\mathcal{S}_{j,1}\\
\mathcal{U}_{j,2}=&\mathcal{R}_{j,1}\mathcal{S}_{j,1}+\mathcal{R}_{j,2}+\mathcal{S}_{j,2}+\frac{1}{2}\mathcal{S}_{j,1}^2.\end{aligned}$$X denotes the boundary condition on the first sphere and Y denotes the boundary condition on the second sphere.
Substitute into . To obtain the leading contributions to the Casimir interaction energy, we can replace the summation over $\tilde{m}$ and $l_j$, $0\leq j\leq s$ by the corresponding integration over $\tilde{m}$, $l$ and $n_j$, $1\leq j\leq s$. Using $$\begin{aligned}
\frac{\tilde{m}\left(\tilde{m}+\frac{D-5}{2}\right)!}{\left(\tilde{m}-\frac{D-3}{2}\right)!}= \tilde{m}^{D-3}-\frac{(D-3)(D-4)(D-5)}{24}\tilde{m}^{D-5}+\ldots.\end{aligned}$$ we find that $$\begin{split}
E_{\text{Cas}}^{\text{XY}}\sim &-\frac{\hbar c}{\pi(R_1+R_2)}\frac{1}{(D-3)!}\sum_{s=0}^{\infty}\frac{1}{s+1}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{2^{s+1}}
\frac{a_2^{\frac{s+1}{2}}}{\pi^{\frac{s+1}{2}}}\int_0^{1} \frac{\tau^{\frac{s-3}{2}}d\tau}{ \sqrt{1-\tau^2}}\int_0^{\infty} dl \, l^{-\frac{s-1}{2}} \int_{-\infty}^{\infty}dn_1\ldots\int_{-\infty}^{\infty} dn_s\\
& \times \int_0^{\infty}d\tilde{m} \left(\tilde{m}^{D-3}-\frac{(D-3)(D-4)(D-5)}{24}\tilde{m}^{D-5}\right) \exp\left( -\sum_{j=0}^{s}\frac{a_2\tau\left(n_j-n_{j+1}\right)^2}{4l}
-\frac{2l(s+1){\varepsilon}}{a_1\tau}-\frac{(s+1)\tilde{m}^2}{\tau a_2l}\right) \\
&\times \left(1+\sum_{j=0}^s\mathcal{U}_{j,1}+\sum_{i=0}^{s-1}\sum_{j=i+1}^s\mathcal{U}_{i,1}\mathcal{U}_{j,1}+\sum_{j=0}^s\mathcal{U}_{j,2}+(s+1)\mathcal{T}^{\text{X}}+(s+1)\mathcal{P}^{\text{Y}}\right). \end{split}$$Upon integration with respect to $n_j$, $1\leq j\leq s$, the term $\sum_{j=0}^s\mathcal{U}_{j,1}$ of order $\sqrt{{\varepsilon}}$ does not give any contribution since it is odd in one of the $n_j$’s, and we have an expansion of the form $$\begin{split}
E_{\text{Cas}}^{\text{XY}}\sim &-\frac{\hbar c\sqrt{a_2}}{2\pi^{\frac{3}{2}} a_1(R_1+R_2)(D-3)!}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{\frac{3}{2}}}
\int_0^{1} \frac{d\tau}{\tau^{\frac{3}{2}} \sqrt{1-\tau^2}}\int_0^{\infty} dl \, l^{\frac{1}{2}} \\
& \times \int_0^{\infty}d\tilde{m} \left(\tilde{m}^{D-3}-\frac{(D-3)(D-4)(D-5)}{24}\tilde{m}^{D-5}\right) \exp\left(
-\frac{2l(s+1){\varepsilon}}{a_1\tau}-\frac{(s+1)\tilde{m}^2}{\tau a_2l}\right) \\
&\times \left(1+\mathcal{V}+(s+1)\mathcal{T}^{\text{X}}+(s+1)\mathcal{P}^{\text{Y}}\right). \end{split}$$ From this, we find that the leading term of the Casimir interaction energy is $$\begin{aligned}
E_{\text{Cas}}^{0, \text{XY}}= &-\frac{\hbar c\sqrt{a_2}}{2\pi^{\frac{3}{2}} a_1(R_1+R_2)(D-3)!}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{\frac{3}{2}}}
\int_0^{1} \frac{d\tau}{\tau^{\frac{3}{2}} \sqrt{1-\tau^2}}\int_0^{\infty} dl \, l^{\frac{1}{2}} \\
& \times \int_0^{\infty}d\tilde{m} \,\tilde{m}^{D-3} \exp\left(
-\frac{2l(s+1){\varepsilon}}{a_1\tau}-\frac{(s+1)\tilde{m}^2}{\tau a_2l}\right) \\
=& -\frac{\hbar c a_2^{\frac{D-1}{2}}}{4\pi^{\frac{3}{2}}a_1(R_1+R_2) }\frac{\Gamma\left(\frac{D-2}{2}\right)}{\Gamma(D-2)}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{\frac{D+1}{2}}}
\int_0^{1} \frac{d\tau\,\tau^{\frac{D-5}{2}} }{\sqrt{1-\tau^2}}\int_0^{\infty} dl \, l^{\frac{D-1}{2}} \exp\left(
-\frac{2l(s+1){\varepsilon}}{a_1\tau} \right) \\
=& -\frac{\hbar c (a_1a_2)^{\frac{D-1}{2}}}{2^{\frac{3D-1}{2}}\pi (R_1+R_2){\varepsilon}^{\frac{D+1}{2}}}\frac{\Gamma\left(\frac{D+1}{2}\right)}{\Gamma\left(\frac{D-1}{2}\right)}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{ D+1 }}
\int_0^{1} \frac{d\tau\,\tau^{D-2} }{\sqrt{1-\tau^2}}\\
=& -\frac{\hbar c (a_1a_2)^{\frac{D-1}{2}}}{2^{\frac{3D+1}{2}}\sqrt{\pi}(R_1+R_2) {\varepsilon}^{\frac{D+1}{2}}}\frac{\Gamma\left(\frac{D+1}{2}\right)}{\Gamma\left(\frac{D}{2}\right)}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{ D+1 }}.\end{aligned}$$Namely, $$\label{eq3_10_7}\begin{split}
E_{\text{Cas}}^{0,\text{DD/NN}}\sim &-\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}} \frac{\hbar c}{d^{\frac{D+1}{2}}},\\
E_{\text{Cas}}^{0, \text{DN/ND}}\sim &(1-2^{-D})\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}} \frac{\hbar c}{d^{\frac{D+1}{2}}},
\end{split}$$which agree with the proximity force approximation .
The next-to-leading order term of the Casimir interaction energy $E_{\text{Cas}}^{1, \text{XY}}$ can be written as a sum of two terms: $$\begin{aligned}
E_{\text{Cas}}^{1, \text{XY}}=E_{\text{Cas}}^{1a, \text{XY}}+E_{\text{Cas}}^{1b, \text{XY}}.\end{aligned}$$ $E_{\text{Cas}}^{1a, \text{XY}}$ vanishes if $D=3, 4$ or 5. For $D\geq 6$, $$\begin{aligned}
&E_{\text{Cas}}^{1a, \text{XY}}\\= & \frac{(D-3)(D-4)(D-5)}{48}\frac{\hbar c\sqrt{a_2}}{ \pi^{\frac{3}{2}} a_1(R_1+R_2)(D-3)!}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{\frac{3}{2}}}
\int_0^{1} \frac{d\tau}{\tau^{\frac{3}{2}} \sqrt{1-\tau^2}}\int_0^{\infty} dl \, l^{\frac{1}{2}} \\
& \times \int_0^{\infty}d\tilde{m} \,\tilde{m}^{D-5} \exp\left(
-\frac{2l(s+1){\varepsilon}}{a_1\tau}-\frac{(s+1)\tilde{m}^2}{\tau a_2l}\right) \\
=& \frac{(D-3)(D-4)(D-5)}{96}\frac{\hbar c a_2^{\frac{D-3}{2}}}{ \pi^{\frac{3}{2}}a_1(R_1+R_2) }\frac{\Gamma\left(\frac{D-4}{2}\right)}{\Gamma(D-2)}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{\frac{D-1}{2}}}
\int_0^{1} \frac{d\tau\,\tau^{\frac{D-7}{2}} }{\sqrt{1-\tau^2}}\int_0^{\infty} dl \, l^{\frac{D-3}{2}} \exp\left(
-\frac{2l(s+1){\varepsilon}}{a_1\tau} \right) \\
=& \frac{(D-3)(D-4)(D-5)}{6}\frac{\hbar c (a_1a_2)^{\frac{D-3}{2}}}{2^{\frac{3D+1}{2}}\pi (R_1+R_2){\varepsilon}^{\frac{D-1}{2}}}\frac{\Gamma\left(\frac{D-4}{2}\right)}{\Gamma\left(\frac{D-2}{2}\right)}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{ D-1 }}
\int_0^{1} \frac{d\tau\,\tau^{D-4} }{\sqrt{1-\tau^2}}\\
=& \frac{(D-3)(D-4)(D-5)}{6}\frac{\hbar c (a_1a_2)^{\frac{D-3}{2}}}{2^{\frac{3D+1}{2}}\sqrt{\pi}(R_1+R_2) {\varepsilon}^{\frac{D-1}{2}}}\frac{\Gamma\left(\frac{D-3}{2}\right)}{(D-4)\Gamma\left(\frac{D-2}{2}\right)}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{ D-1 }}.\end{aligned}$$ Namely, $$\begin{split}
E_{\text{Cas}}^{1a, \text{DD/NN}}=& \left(1-\delta_{D,3}\right)\left(1-\delta_{D,4}\right)\frac{ (D-5)}{3}\frac{\Gamma\left(\frac{D-1}{2}\right)\zeta(D-1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D-2}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-3}{2}} \frac{\hbar c}{d^{\frac{D-1}{2}}}\\=&-E_{\text{Cas}}^{0, \text{DD/NN}}
\left(1-\delta_{D,3}\right)\left(1-\delta_{D,4}\right)\frac{ (D-5)(D-2)}{3(D-1)}\frac{\zeta(D-1)}{\zeta(D+1)}\left(\frac{d}{R_1}+\frac{d}{R_2}\right),\\
E_{\text{Cas}}^{1a, \text{DN/ND}}=& \left(1-\delta_{D,3}\right)\left(1-\delta_{D,4}\right)\left(1-2^{-D+2}\right)\frac{ (D-5)}{3}\frac{\Gamma\left(\frac{D-1}{2}\right)\zeta(D-1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D-2}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-3}{2}} \frac{\hbar c}{d^{\frac{D-1}{2}}}\\=&-E_{\text{Cas}}^{0, \text{DD/NN}}
\left(1-\delta_{D,3}\right)\left(1-\delta_{D,4}\right)\frac{ (D-5)(D-2)}{3(D-1)}\frac{\zeta(D-1)}{\zeta(D+1)}\frac{2^D-4}{2^D-1}\left(\frac{d}{R_1}+\frac{d}{R_2}\right).
\end{split}$$ The second contribution to the next-to-leading order term is given by $$\begin{split}
E_{\text{Cas}}^{1b,\text{XY}}= &-\frac{\hbar c\sqrt{a_2}}{2\pi^{\frac{3}{2}} a_1(R_1+R_2)(D-3)!}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{\frac{3}{2}}}
\int_0^{1} \frac{d\tau}{\tau^{\frac{3}{2}} \sqrt{1-\tau^2}}\int_0^{\infty} dl \, l^{\frac{1}{2}} \\
& \times \int_0^{\infty}d\tilde{m} \tilde{m}^{D-3} \exp\left(
-\frac{2l(s+1){\varepsilon}}{a_1\tau}-\frac{(s+1)\tilde{m}^2}{\tau a_2l}\right) \left(\mathcal{V}+(s+1)\mathcal{T}^{\text{X}}+(s+1)\mathcal{P}^{\text{Y}}\right)\\
=&-\frac{\hbar c a_2^{\frac{D-1}{2}}}{4\pi^{\frac{3}{2}} a_1 (R_1+R_2) }\frac{\Gamma\left(\frac{D-2}{2}\right)}{\Gamma(D-2)}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{\frac{D+1}{2}}}
\int_0^{1} \frac{d\tau\, \tau^{\frac{D-5}{2}}}{ \sqrt{1-\tau^2}}\int_0^{\infty} dl \, l^{\frac{D-1}{2}} \exp\left(
-\frac{2l(s+1){\varepsilon}}{a_1\tau} \right) \\
& \times \left(\mathcal{W}+(s+1)\mathcal{T}^{\text{X}}+(s+1)\mathcal{P}^{\text{Y}}\right)\\
=& -\frac{\hbar c (a_1a_2)^{\frac{D-1}{2}}}{2^{\frac{3D-1}{2}}\pi (R_1+R_2){\varepsilon}^{\frac{D+1}{2}}}\frac{\Gamma\left(\frac{D+1}{2}\right)}{\Gamma\left(\frac{D-1}{2}\right)}\sum_{s=0}^{\infty}\frac{(-1)^{(\alpha_{\text{X}}+\alpha_{\text{Y}})(s+1)}}{\left(s+1\right)^{ D+1 }}
\int_0^{1} \frac{d\tau\,\tau^{D-2} }{\sqrt{1-\tau^2}}\mathcal{Z}^{\text{XY}}.\end{split}$$ Upon integration with respect to $\tau$, we find that $$\begin{split}
E_{\text{Cas}}^{1b,\text{DD}}=& -\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}} \frac{\hbar c}{d^{\frac{D+1}{2}}} \left(\left[-\frac{(D+1)}{4}\frac{d}{R_1+R_2}+\frac{D+1}{12}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\right]
\sum_{s=0}^{\infty}\frac{1}{\left(s+1\right)^{ D+1 }}\right.\\&\left.\hspace{4cm}+\left[\left(1-\delta_{D,3}\right)\frac{(D-2)(D-5)}{3(D-1)}-\frac{(D-2)(D-3)}{3D}\right]\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\sum_{s=0}^{\infty}\frac{1}{\left(s+1\right)^{ D-1 }}\right)\\
=&E_{\text{Cas}}^{0,\text{DD}}\left( -\frac{(D+1)}{4}\frac{d}{R_1+R_2}+\frac{D+1}{12}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)
\right.\\&\left.\hspace{4cm}+\left[\left(1-\delta_{D,3}\right)\frac{(D-2)(D-5)}{3(D-1)}-\frac{(D-2)(D-3)}{3D}\right]\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\frac{\zeta(D-1)}{\zeta(D+1)}\right),\end{split}$$$$\begin{split}
E_{\text{Cas}}^{1b,\text{NN}}=& -\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}} \frac{\hbar c}{d^{\frac{D+1}{2}}} \left(\left[-\frac{(D+1)}{4}\frac{d}{R_1+R_2}+\frac{D+1}{12}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\right]
\sum_{s=0}^{\infty}\frac{1}{\left(s+1\right)^{ D+1 }}\right.\\&\left.\hspace{4cm}+\left[\left(1-\delta_{D,3}\right)\frac{(D-2)(D-5)}{3(D-1)}-\frac{D^2+7D-6}{3D}\right]\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\sum_{s=0}^{\infty}\frac{1}{\left(s+1\right)^{ D-1 }}\right)\\
=&E_{\text{Cas}}^{0,\text{NN}}\left( -\frac{(D+1)}{4}\frac{d}{R_1+R_2}+\frac{D+1}{12}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)
\right.\\&\left.\hspace{4cm}+\left[\left(1-\delta_{D,3}\right)\frac{(D-2)(D-5)}{3(D-1)}-\frac{D^2+7D-6}{3D}\right]\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\frac{\zeta(D-1)}{\zeta(D+1)}\right),\\
\end{split}$$ $$\begin{split}
E_{\text{Cas}}^{1b,\text{DN}}=& -\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}} \frac{\hbar c}{d^{\frac{D+1}{2}}} \left(\left[-\frac{(D+1)}{4}\frac{d}{R_1+R_2}+\frac{D+1}{12}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\right]
\sum_{s=0}^{\infty}\frac{(-1)^{s+1}}{\left(s+1\right)^{ D+1 }}\right.\\&\left.\hspace{1cm}+\left[\left(1-\delta_{D,3}\right)\frac{(D-2)(D-5)}{3(D-1)}-\frac{(D-2)(D-3)}{3D}\frac{d}{R_1}-\frac{D^2+7D-6}{3D}\frac{d}{R_2}\right]\sum_{s=0}^{\infty}\frac{(-1)^{s+1}}{\left(s+1\right)^{ D-1 }}\right)\\
=&E_{\text{Cas}}^{0,\text{DN}}\left( -\frac{(D+1)}{4}\frac{d}{R_1+R_2}+\frac{D+1}{12}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)
\right.\\&\left.\hspace{1cm}+\left[\left(1-\delta_{D,3}\right)\frac{(D-2)(D-5)}{3(D-1)}-\frac{(D-2)(D-3)}{3D}\frac{d}{R_1}-\frac{D^2+7D-6}{3D}\frac{d}{R_2}\right]\frac{2^D-4}{2^D-1}\frac{\zeta(D-1)}{\zeta(D+1)}\right),\\
\end{split}$$ $$\begin{split}
E_{\text{Cas}}^{1b,\text{ND}}=& -\frac{\Gamma\left(\frac{D+1}{2}\right)\zeta(D+1)}{2^{\frac{3D+1}{2}}\sqrt{\pi}\Gamma\left(\frac{D}{2}\right)}\left(\frac{R_1R_2}{R_1+R_2}\right)^{\frac{D-1}{2}} \frac{\hbar c}{d^{\frac{D+1}{2}}} \left(\left[-\frac{(D+1)}{4}\frac{d}{R_1+R_2}+\frac{D+1}{12}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\right]
\sum_{s=0}^{\infty}\frac{(-1)^{s+1}}{\left(s+1\right)^{ D+1 }}\right.\\&\left.\hspace{1cm}+\left[\left(1-\delta_{D,3}\right)\frac{(D-2)(D-5)}{3(D-1)}-\frac{D^2+7D-6}{3D}\frac{d}{R_1}-\frac{(D-2)(D-3)}{3D}\frac{d}{R_2}\right]\sum_{s=0}^{\infty}\frac{(-1)^{s+1}}{\left(s+1\right)^{ D-1 }}\right)\\
=&E_{\text{Cas}}^{0,\text{DN}}\left( -\frac{(D+1)}{4}\frac{d}{R_1+R_2}+\frac{D+1}{12}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)
\right.\\&\left.\hspace{1cm}+\left[\left(1-\delta_{D,3}\right)\frac{(D-2)(D-5)}{3(D-1)}-\frac{D^2+7D-6}{3D}\frac{d}{R_1}-\frac{(D-2)(D-3)}{3D}\frac{d}{R_2}\right]\frac{2^D-4}{2^D-1}\frac{\zeta(D-1)}{\zeta(D+1)}\right).
\end{split}$$
Adding together the leading and next-to-leading order term, we find that if $D=3$, $$\begin{split}
E_{\text{Cas}}^{\text{DD}}=&E_{\text{Cas}}^{0, \text{DD}}\left(1-\frac{d}{R_1+R_2}+\frac{1}{3}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)+\ldots \right),\\
E_{\text{Cas}}^{\text{NN}}=&E_{\text{Cas}}^{0, \text{NN}}\left(1-\frac{d}{R_1+R_2}+\left(\frac{1}{3}-\frac{40}{\pi^2}\right)\left(\frac{d}{R_1}+\frac{d}{R_2}\right)+\ldots \right),\\
E_{\text{Cas}}^{\text{DN}}=&E_{\text{Cas}}^{0, \text{DN}}\left(1-\frac{d}{R_1+R_2}+\frac{1}{3}\frac{d}{R_1}+\left(\frac{1}{3}-\frac{160}{7\pi^2}\right) \frac{d}{R_2} +\ldots \right),\\
E_{\text{Cas}}^{\text{ND}}=&E_{\text{Cas}}^{0, \text{ND}}\left(1-\frac{d}{R_1+R_2}+\left(\frac{1}{3}-\frac{160}{7\pi^2}\right) \frac{d}{R_1}+\frac{1}{3}\frac{d}{R_2} +\ldots \right),
\end{split}$$which agree with the results we obtain in [@5].
If $D=4$, $$\begin{split}
E_{\text{Cas}}^{\text{DD}}=&E_{\text{Cas}}^{0, \text{DD}}\left(1-\frac{5}{4}\frac{d}{R_1+R_2}+\left(\frac{5}{12}-\frac{7}{18}\frac{\zeta(3)}{\zeta(5)}\right)\left(\frac{d}{R_1}+\frac{d}{R_2}\right)+\ldots \right),\\
E_{\text{Cas}}^{\text{NN}}=&E_{\text{Cas}}^{0, \text{NN}}\left(1-\frac{5}{4}\frac{d}{R_1+R_2}+\left(\frac{5}{12}-\frac{61}{18}\frac{\zeta(3)}{\zeta(5)}\right)\left(\frac{d}{R_1}+\frac{d}{R_2}\right)+\ldots \right),\\
E_{\text{Cas}}^{\text{DN}}=&E_{\text{Cas}}^{0, \text{DN}}\left(1-\frac{5}{4}\frac{d}{R_1+R_2}+\left(\frac{5}{12}-\frac{14}{45}\frac{\zeta(3)}{\zeta(5)}\right)\frac{d}{R_1}+\left(\frac{5}{12}-\frac{122}{45}\frac{\zeta(3)}{\zeta(5)}\right) \frac{d}{R_2} +\ldots \right),\\
E_{\text{Cas}}^{\text{ND}}=&E_{\text{Cas}}^{0, \text{ND}}\left(1-\frac{5}{4}\frac{d}{R_1+R_2}+\left(\frac{5}{12}-\frac{122}{45}\frac{\zeta(3)}{\zeta(5)}\right) \frac{d}{R_1}+\left(\frac{5}{12}-\frac{14}{45}\frac{\zeta(3)}{\zeta(5)}\right)\frac{d}{R_2} +\ldots \right).
\end{split}$$
If $D\geq 5$, we have $$\label{eq3_13_1}
\begin{split}
E_{\text{Cas}}^{\text{DD}}=&E_{\text{Cas}}^{0, \text{DD}}\left\{1-\frac{D+1}{4}\frac{d}{R_1+R_2}+\left(\frac{D+1}{12}-\frac{(D-2)(D-3)}{3D}\frac{\zeta(D-1)}{\zeta(D+1)}\right)\left(\frac{d}{R_1}+\frac{d}{R_2}\right)+\ldots \right\},\\
E_{\text{Cas}}^{\text{NN}}=&E_{\text{Cas}}^{0, \text{NN}}\left\{1-\frac{D+1}{4}\frac{d}{R_1+R_2}+\left(\frac{D+1}{12}-\frac{D^2+7D-6}{3D}\frac{\zeta(D-1)}{\zeta(D+1)}\right)\left(\frac{d}{R_1}+\frac{d}{R_2}\right)+\ldots \right\},\\
E_{\text{Cas}}^{\text{DN}}=&E_{\text{Cas}}^{0, \text{DN}}\left\{1-\frac{D+1}{4}\frac{d}{R_1+R_2}+\left(\frac{D+1}{12}-\frac{(D-2)(D-3)}{3D}
\frac{2^D-4}{2^D-1}\frac{\zeta(D-1)}{\zeta(D+1)}\right)\frac{d}{R_1}\right.\\&\left.\hspace{3cm}+\left(\frac{D+1}{12}-\frac{D^2+7D-6}{3D}\frac{2^D-4}{2^D-1}\frac{\zeta(D-1)}{\zeta(D+1)}\right) \frac{d}{R_2} +\ldots \right\},\\
E_{\text{Cas}}^{\text{ND}}=&E_{\text{Cas}}^{0, \text{ND}}\left\{1-\frac{D+1}{4}\frac{d}{R_1+R_2}+\left(\frac{D+1}{12}-\frac{D^2+7D-6}{3D}\frac{2^D-4}{2^D-1}\frac{\zeta(D-1)}{\zeta(D+1)}\right) \frac{d}{R_1}\right.\\&\left.\hspace{3cm}+\left(\frac{D+1}{12}-\frac{(D-2)(D-3)}{3D}
\frac{2^D-4}{2^D-1}\frac{\zeta(D-1)}{\zeta(D+1)}\right)\frac{d}{R_2} +\ldots \right\}.
\end{split}$$
Notice that if we take the limit the radius of the second sphere is very large, i.e. $R_2\rightarrow\infty$, we recover the results for sphere-plane we obtained in [@1]. In fact, if the Casimir interaction energy for the sphere-plate case is given by $$\begin{split}
E_{\text{Cas}}^{\text{SP, XY}}=E_{\text{Cas}}^{0, \text{SP, XY}}\left(1+\vartheta^{\text{ XY}}_E\frac{d}{R}+\ldots\right),
\end{split}$$where $R$ is the radius of the sphere, X is the boundary condition on the sphere, and Y is the boundary condition on the plate, then the Casimir interaction energy for the sphere-sphere case can be written as $$\begin{split}
E_{\text{Cas}}^{\text{DD}}=&E_{\text{Cas}}^{0, \text{DD}}\left\{1-\frac{D+1}{4}\frac{d}{R_1+R_2}+\vartheta_E^{\text{DD}}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)+\ldots\right\},\\
E_{\text{Cas}}^{\text{NN}}=&E_{\text{Cas}}^{0, \text{DD}}\left\{1-\frac{D+1}{4}\frac{d}{R_1+R_2}+\vartheta_E^{\text{NN}}\left(\frac{d}{R_1}+\frac{d}{R_2}\right)+\ldots\right\},\\
E_{\text{Cas}}^{\text{DN}}=&E_{\text{Cas}}^{0, \text{DD}}\left\{1-\frac{D+1}{4}\frac{d}{R_1+R_2}+\vartheta_E^{\text{DN}} \frac{d}{R_1}+\vartheta_E^{\text{ND}}\frac{d}{R_2} +\ldots\right\},\\
E_{\text{Cas}}^{\text{ND}}=&E_{\text{Cas}}^{0, \text{DD}}\left\{1-\frac{D+1}{4}\frac{d}{R_1+R_2}+\vartheta_E^{\text{ND}} \frac{d}{R_1}+\vartheta_E^{\text{DN}}\frac{d}{R_2} +\ldots\right\}.
\end{split}$$ The values of $\vartheta_E^{\text{XY}}$ have been tabulated in [@1] for $3\leq D\leq 12$.
Comparing the representation of the Casimir interaction energy we derive in this paper with the one used in [@5], the major difference is that for the elements of the translation matrices $G^1_{\boldsymbol{m},\boldsymbol{m}'}$ and $G^2_{\boldsymbol{m}',\boldsymbol{m}}$, we leave them as integrals and do not re-express it as combinations of spherical waves with complicated coefficients. When we find the asymptotic behavior of the Casimir interaction energy, it turns out that this does not incur further complications thanks to the integral formula , rather than having to rely on integral formulas for those complicated coefficients as in [@5].
Notice that since $a_2=R_2/(R_1+R_2)=1-R_1/(R_1+R_2)$, $$\begin{aligned}
\label{eq3_18_1}
\frac{R_1+R_2}{d}\frac{E^1_{\text{Cas}}}{E^0_{\text{Cas}}}= \frac{1}{{\varepsilon}}\frac{E^1_{\text{Cas}}}{E^0_{\text{Cas}}}\end{aligned}$$is a function of dimension $D$ and $a_1=R_1/(R_1+R_2)$ only. In Figs. \[f1\], \[f2\], \[f3\] and \[f4\], we plot the dependence of as a function of $D$ and $a_1=R_1/(R_1+R_2)$.
=0.5
=0.5
=0.5
=0.5
We notice that the value of can become very large for large $D$. In fact, from , we see that when $D$ is large, the ratio of the next-to-leading order term $E_{\text{Cas}}^1$ to the leading order term $E_{\text{Cas}}^0$ behaves as $$\begin{aligned}
\frac{E_{\text{Cas}}^1 }{E_{\text{Cas}}^0 }\sim &-\frac{D}{4}\frac{d}{R_1+R_2}+\left(\frac{D}{12}-\frac{D}{3} \right)\left(\frac{d}{R_1}+\frac{d}{R_2}\right)\\
=&-\frac{D}{4}\left(\frac{d}{R_1+R_2}+\frac{d}{R_1}+\frac{d}{R_2}\right)\\
=&-\frac{{\varepsilon}D}{4}\left(1+\frac{1}{a_1}+\frac{1}{1-a_1}\right),\end{aligned}$$ which is proportional to $D$. Therefore in higher dimensions, the proximity force approximation to the Casimir interaction energy becomes less accurate, and the contribution of the next-to-leading order term becomes more significant.
Conclusion
==========
We have derived the TGTG formula for the Casimir interaction energy between two spheres in $(D+1)$-dimensional Minkowski spacetime. The most difficult part in the derivation is the computation of the translation matrices which relate the spherical waves in two coordinate frames differ by a translation. This has not been derived elsewhere and can be considered as a major byproduct of this work. Unlike the three-dimensional case, we do not rewrite the elements of the translation matrices as linear combinations of spherical waves with coefficients expressed in terms of $3j$-symbols. We content with writing them as integrals over Gegenbauer polynomials, which are orthogonal polynomials generalizing Legendre polynomials.
For practical purpose, we explore the strength of the Casimir interaction in the small and large separation regimes.
In the large separation regime, the leading contribution to the Casimir interaction energy comes from a few terms with the lowest wave number(s). It is found that for Dirichlet-Dirichlet, Dirichlet-Neumann and Neumann-Neumann boundary conditions, the leading contributions are of order $L^{-2D+3}$, $L^{-2D+1}$ and $L^{-2D-1}$ respectively, where $L$ is the center-to-center distance of the spheres. Hence in the large separation regime, the Casimir interaction is strongest in the Dirichlet-Dirichlet case, and weakest in the Neumann-Neumann case. It is also observed that the order of the interaction is weaker in higher dimensions.
In the small separation regime, the magnitude of the Casimir interaction is of order $d^{-\frac{D+1}{2}}$, with $d$ the distance between the spheres, which agrees with what predicted by proximity force approximation. One observes that in contrast to large separation, the order of interaction is stronger when the dimension of spacetime is higher. To study the deviation from proximity force approximation, we compute the next-to-leading order term of the Casimir interaction. It is found that the ratio of the next-to-leading term to the leading order term is proportional to $D$, indicating larger corrections in higher dimensions.
In this work, we demonstrate how to compute the Casimir interactions between two spheres in a spacetime with $(D+1)$ dimensions. We only consider scalar field with Dirichlet or Neumann boundary conditions in Minkowski spacetime. Nonetheless, it is easy to see that one can generalize the approach here to any spacetime and any other fields. This also shed some light on how to compute the quantum interaction between two spherical objects in $(D+1)$-dimensional spacetime.
This work is supported by the Ministry of Higher Education of Malaysia under FRGS grant FRGS/1/2013/ST02/UNIM/02/2.
[10]{} H. B. G. Casimir, Proc. Kon. Nederland. Akad. Wetensch. **B51**, 793–795 (1948).
S. K. Blau and M. Visser, Nucl. Phys. B **310**, 163 (1988).
J. Ambjørn and S. Wolfram, Ann. Phys. **147** (1983), 1.
C. M. Bender and K. A. Milton, Phys. Rev. D **50** (1994), 6547. K. A. Milton, Phys. Rev. D **55** (1996), 4940. G. Cognola, E. Elizalde and K. Kirsten, J. Phys. A **34** (2001), 7311.
M. Schaden and L. Spruch, Phys. Rev. A **58**, 935 (1998).
M. Schaden and L. Spruch, Phys. Rev. Lett. **84**, 459 (2000).
M. Schaden, Phys. Rev. A **73**, 042102 (2006).
R. L. Jaffe and A. Scardicchio, Phys. Rev. Lett. **92**, 070402 (2004).
A. Scardicchio and R. L. Jaffe, Nucl. Phys. B **704**, 552 (2005).
A. Scardicchio and R. L. Jaffe, Nucl. Phys. B **743**, 249 (2006).
R. Balian and B. Duplantier, *Geometry of the Casimir effect*, in: *Recent Developments in Gravitational Physics*, Institute of Physics Conference Series 176, Ed. Ciufiolini et al, 2004.
H. Gies, K. Langfeld, and L. Moyaerts, J. High Energy Phys. **0306**, 018 (2003).
H. Gies, K. Langfeld, and L. Moyaerts, J. Phys. A **39**, 6415 (2006).
H. Gies and K. Klingmüller, Phys. Rev. Lett. **96**, 220401 (2006).
H. Gies and K. Klingmüller, Phys. Rev. Lett. **97**, 220405 (2006).
H. Gies and K. Klingmüller, Phys. Rev. D **74**, 045002 (2006).
A. Lambrecht, P.A. Maia-Neto, and S. Reynaud New Journal of Physics **8**, 243 (2006). A. Bulgac, P. Magierski and A. Wirzba, Phys. Rev. D **73**, 025007 (2006).
T. Emig, R. L. Jaffe, M. Kadar and A. Scardicchio, Phys. Rev. Lett. **96**, 080403 (2006).
S. J. Rahi, T. Emig, R. L. Jaffe and M. Kardar, Phys. Rev. A **78**, 012104 (2008).
T. Emig, N. Graham, R. L. Jaffe and M. Kardar, Phys. Rev. Lett. **99**, 170403 (2007).
T. Emig, N. Graham, R. L. Jaffe and M. Kadar, Phys. Rev. D **77**, 025005 (2008).
T. Emig and R. L. Jaffe, J. Phys. A: Math. Theor. **41**, 164001 (2008).
T. Emig, J. Stat. Mech. **0804**, P04007 (2008).
S. J. Rahi, T. Emig, N. Graham, R. L. Jaffe and M. Kadar, Phys, Rev. D **80**, 085021 (2009).
M. Bordag, Phys. Rev. D **73**, 125018 (2006).
M. Bordag, Phys. Rev. D **75**, 065003 (2007).
O. Kenneth and I. Klich, Phys. Rev. Lett. **97**, 160401 (2006).
O. Kenneth and I. Klich, Phys. Rev. B **78**, 014103 (2008).
K. A. Milton and J. Wagner, J. Phys. A: Math. Theor. **41**, 155402 (2008).
K. Milton and J. Wagner, Phys. Rev. D **77**, 045005 (2008).
D. A. R. Dalvit, F. C. Lombardo, F. D. Mazzitelli and R. Onofrio, Phys. Rev. A **74**, 020101(R) (2006).
F. D. Mazzitelli, D. A. R. Dalvit and F. C. Lombardo, New. J. Phys. **8**, 240 (2006).
F. C. Lombardo, F. D. Mazzitelli, P. I. Villar and D. A. R. Dalvit, Phys. Rev. A **82**, 042509 (2010).
L. P. Teo, Int. J. Mod. Phys. A **27**, 1230021 (2012).
L. P. Teo, preprint arXiv:1312.1768.
R. C. Wittman, IEEE Trans. Antennas Propag. **36**, 1078 (1988).
M. Bordag and V. Nikolaev, J. Phys. A: Math. Theor. **41**, 164002 (2008). M. Bordag and V. Nikolaev, Phys. Rev. D **81**, 065011 (2010). L. P. Teo, M. Bordag and V. Nikolaev, Phys. Rev. D **84**, 125037 (2011). L. P. Teo, Phys. Rev. D **88**, 045019 (2013).
L. P. Teo, Phys. Rev. D **84**, 025022 (2011).
L. P. Teo, Phys. Rev. D **84**, 065027 (2011). L. P. Teo, Phys. Rev. D **85**, 045027 (2012).
A. Erd$\acute{\text{e}}$lyi et al., *Higher transcendental functions*, Vol. 2, McGraw Hill, New York, 1953.
G. Andrews, R. Askey and R. Roy, *Special functions*, Cambridge University Press, Cambridge, 1999.
M. Abramowitz and I. A. Stegun, *Handbook of Mathematical Functions* (Dover, New York, 1972).
F.W. J. Olver, Phil. Trans. R. Soc. A **247**, 328 (1954).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Extending the Feynman-Kleinert variational approach, we calculate the temperature-dependent effective classical potential governing the quantum statistics of a hydrogen atom in a uniform magnetic at all temperatures. The zero-temperature limit yields the binding energy of the electron which is quite accurate for all magnetic field strengths and exhibits, in particular, the correct logarithmic growth at large fields.'
address: 'Institut für Theoretische Physik, Freie Universität Berlin, Arnimallee 14, 14195 Berlin'
author:
- 'M. Bachmann, H. Kleinert, and A. Pelster'
title: |
Variational Approach to Hydrogen Atom in Uniform\
Magnetic Field of Arbitrary Strength
---
Introduction
============
The quantum statistical and quantum mechanical properties of a hydrogen atom in an external magnetic field are not exactly calculable. Perturbative approaches yield good results only for weak uniform fields as discussed in detail by Le Guillou and Zinn-Justin [@zinn-justin], who interpolated with analytic mapping techniques the ground state energy between weak- and strong-field. Other approaches are based on recursive procedures in higher-order perturbation theory [@cizek1; @cizek2; @gani]. Zero-temperature properties were also investigated with the help of an operator optimization method in a second-quantized variational procedure [@feranshuk]. The behaviour at high uniform fields was inferred from treatments of the one-dimensional hydrogen atom [@landau; @loudon; @haines]. Hydrogen in strong magnetic fields is still a problem under investigation, since its solution is necessary to understand the properties of white dwarfs and neutron stars, as emphasized in Refs. [@cohen; @ivanov; @heyl].
A compact and detailed presentation of the bound states and highly accurate numerically values for the energy levels is given in Ref. [@wunner].
Equations for a first-order variational approach to the ground-state energy of hydrogen in a uniform magnetic field based on the Jensen-Peierls inequality were written down a long time ago [@dev0], but never evaluated. Apparently, they merely served as a preparation for attacking the more complicated problem of a polaron in a magnetic field [@dev0; @dev1; @dev2].
In our approach, we calculate the quantum statistical properties of the system by an extension of variational perturbation theory [@PI]. The crucial quantity is the effective classical potential. In the zero-temperature limit, it yields the ground state energy. Our calculations in a magnetic field require an extension of the formalism in Ref. [@PI] which derives the effective classical potential from the phase space representation of the partition function.
Variational perturbation theory has an important advantage over other approaches: The calculation yields a good effective classical potential for [*all*]{} temperatures and coupling strengths. The quantum statistical partition function is obtained from a simple integral over a Boltzmann-factor involving the effective classical potential. The ground state energy is then obtained from its zero-temperature limit. The asymptotic behaviour in the strong-coupling limit is emerging automatically and does not have to be derived from other sources.
Effective Classical Representations for the Quantum Statistical Partition Function {#effrep}
==================================================================================
A point particle in $D$ dimensions with a potential $V({\bf x})$ and a vector potential ${\bf A}({\bf x})$ is described by a Hamiltonian $$\label{ham00}
H({\bf p},{\bf x})=\frac{1}{2M}\left[{\bf p}-\frac{e}{c}{\bf A}({\bf x})\right]^2+V({\bf x}).$$ The quantum statistical partition function is given by the euclidean phase space path integral $$\label{ham01}
Z=\oint {\cal D'}^Dx {\cal D}^Dp\,e^{-{\cal A}[{\bf p},{\bf x}]/\hbar}$$ with an action $$\label{ham02}
{\cal A}[{\bf p},{\bf x}]=\int_0^{\hbar\beta}d\tau\left[-i{\bf p}(\tau)\cdot\dot{\bf x}(\tau)
+H({\bf p}(\tau),{\bf x}(\tau))\right],$$ and the path measure $$\label{ham03}
\oint{\cal D'}^Dx {\cal D}^Dp=\lim_{N\to\infty}\prod\limits_{n=1}^{N+1}
\left[\int\frac{d^Dx_nd^Dp_n}{(2\pi\hbar)^D} \right].$$ The parameter $\beta=1/k_BT$ denotes the usual inverse thermal energy at temperature $T$, where $k_B$ is the Boltzmann constant. From $Z$ we obtain the free energy of the system: $$\label{ham04}
F=-\frac{1}{\beta}{\rm ln}\,Z.$$ In perturbation theory, one treats the external potential $V({\bf x})$ as a small quantity, and expands the partition function into powers of $V({\bf x})$. Such a naive expansion is applicable only for extremely weak couplings, and has a vanishing radius of convergence. Convergence is achieved by variational perturbation theory [@PI], which yields good approximations for all potential strengths, as we shall see in the sequel.
Effective Classical Potential
-----------------------------
All quantum-mechanical systems studied so far in variational perturbation theory were governed by a Hamiltonian of the standard form $$\label{ham05}
H({\bf p},{\bf x})=\frac{{\bf p}^2}{2M}+V({\bf x}).$$ The simple quadratic dependence on the momenta makes the momentum integrals in the path integral (\[ham01\]) trivial. The remaining configuration space representation of the partition function is used to define an effective classical potential $V_{\rm eff}({\bf x}_0)$, from which quantum mechanical partition function is found by a classically looking integral $$\label{ham06}
Z=\int \frac{d^Dx_0}{\lambda_{\rm th}^{D}}\,\exp\left[-\beta V_{\rm eff}({\bf x}_0)\right],$$ where $\lambda_{\rm th}=\sqrt{2\pi\hbar^2\beta/M}$ is the thermal wavelength. The Boltzmann factor plays the role of a [*local partition function*]{} $Z^{{\bf x}_0}$, which is calculated from the restricted path integral $$\label{ham07}
e^{-\beta V_{\rm eff}({\bf x}_0)}\equiv
Z^{{\bf x}_0}=
\lambda_{\rm th}^{D}\oint {\cal D}^Dx\,
\delta({\bf x}_0-\overline{{\bf x}(\tau)})\,e^{-{\cal A}[{\bf x}]/\hbar},$$ with the action $$\label{ham08}
{\cal A}[{\bf x}]=\int_0^{\hbar\beta}d\tau\,\left[\frac{M}{2}\dot{\bf x}^2(\tau)+
V({\bf x}(\tau))
\right],$$ and the path measure $$\label{ham09}
\oint{\cal D}^Dx=\lim_{N\to\infty}\prod\limits_{n=1}^{N+1}
\left\{\int\frac{d^Dx_n}{[2\pi\hbar^2\beta/M(N+1)]^{D/2}} \right\}.$$ The special treatment of the temporal average of the Fourier path $$\label{ham10}
{\bf x}_0=\overline{{\bf x}(\tau)}=\frac{1}{\hbar\beta}\int_0^{\hbar\beta}d\tau\,{\bf x}(\tau)$$ is essential for the quality of the results. It subtracts from the harmonic fluctuation width $\langle{\bf x}^2\rangle^{\rm cl}$ the classical divergence proportional to $T=1/k_B\beta$ of the Dulong-Petit law [@PI; @density]. Such diverging fluctuations cannot be treated perturbatively, and require the final integration in expression (\[ham06\]) to be done numerically.
For the Coulomb potential $V({\bf x})=-e^2/4\pi\varepsilon_0\,|{\bf x}|$ in three dimensions, the effective classical potential in Eq. (\[ham07\]) can be approximated well by variational perturbation theory [@PI; @density; @hatom].
Effective Classical Hamiltonian
-------------------------------
In order to deal with Hamiltonians like (\[ham00\]) which contain a ${\bf p}\cdot{\bf A}({\bf x})$-term, we must generalize the variational procedure. Extending (\[ham07\]), we define an [*effective classical Hamiltonian*]{} by the phase space path integral $$\label{ham12}
e^{-\beta H_{\rm eff}({\bf p}_0,{\bf x}_0)}\equiv
Z^{{\bf p}_0,{\bf x}_0}=
(2\pi\hbar)^D\oint {\cal D'}^Dx{\cal D}^Dp\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\delta({\bf p}_0-\overline{{\bf p}(\tau)})\,e^{-{\cal A}[{\bf p},{\bf x}]/\hbar},$$ with the action (\[ham02\]) and the measure (\[ham03\]). This allows us to express the partition function as the classically looking phase space integral $$\label{ham11}
Z=\int\frac{d^Dx_0d^Dp_0}{(2\pi\hbar)^D}\,\exp\left[-\beta H_{\rm eff}({\bf p}_0,{\bf x}_0)
\right],$$ where ${\bf p}_0$ is the temporal average of the momentum: $$\label{ham13}
{\bf p}_0=\overline{{\bf p}(\tau)}=\frac{1}{\hbar\beta}\int_0^{\hbar\beta}d\tau\,{\bf p}(\tau).$$ The fixing of ${\bf p}_0$ is done for the same reason as that for ${\bf x}_0$, since the classical expectation value $\langle {\bf p}^2\rangle^{\rm cl}$ is diverging linearly with $T$, just as $\langle{\bf x}^2 \rangle^{\rm cl}$.
In the special case of a standard Hamiltonian (\[ham05\]), the effective Hamiltonian in Eq. (\[ham11\]) reduces to the effective classical potential, since the momentum integral in Eq. (\[ham12\]) can then be easily performed, and the resulting restricted partition function becomes $$\label{ham14}
Z^{{\bf p}_0,{\bf x}_0}=\exp\left(-\beta\frac{{\bf p}_0^2}{2M} \right)\,
Z^{{\bf x}_0}$$ with the local partition function $Z^{{\bf x}_0}=\exp[-\beta V_{\rm eff}({\bf x}_0)]$ of Eq. (\[ham07\]). Thus the complete quantum statistical partition function is given by (\[ham11\]), with an effective classical Hamilton function $$\label{ham15}
H_{\rm eff}({\bf p}_0,{\bf x}_0)=\frac{{\bf p}_0^2}{2M}+V_{\rm eff}({\bf x}_0).$$ As a consequence of the purely quadratic momentum dependence of $H({\bf p},{\bf x})$ in (\[ham05\]), the ${\bf p}_0$-integral in (\[ham11\]) can be done, thus expressing the quantum statistical partition function as a pure configuration space integral over the Boltzmann factor involving the effective classical potential $V_{\rm eff}({\bf x}_0)$, as in Eq. (\[ham06\]).
Exact Effective Classical Hamiltonian for an Electron in a Constant Magnetic Field {#exham}
----------------------------------------------------------------------------------
The effective classical Hamiltonian for the electron moving in a constant magnetic field can be calculated exactly. We consider a magnetic field ${\bf B}=B {\bf e}_z$ pointing along the positive $z$-axis. The only nontrivial motion of the electron is in the $x\!\!-\!\!y$-plane. In symmetric gauge the vector potential is given by $$\label{ham16}
{\bf A}({\bf x})=\frac{B}{2}(-y,x,0).$$ The choice of the gauge does not affect the partition function since the periodic path integral (\[ham01\]) is gauge invariant. Ignoring the trivial free particle motion along the $z$-direction, we may restrict our attention to the two-dimensional Hamiltonian $$\label{ham17}
H({\bf p},{\bf x})=\frac{{\bf p}^2}{2M}-\frac{1}{2}\omega_c l_z({\bf p},{\bf x})+
\frac{1}{8}M\omega_c^2{\bf x}^2$$ with ${\bf x}=(x,y)$ and ${\bf p}=(p_x,p_y)$. Here, $\omega_c=eB/Mc$ is the Landau frequency, and $$\label{ham18}
l_z({\bf p},{\bf x})=({\bf x}\times{\bf p})_z=xp_y-yp_x$$ the third component of the orbital angular momentum. The partition function of the problem is given by Eq. (\[ham11\]), with $D=2$. Being interested in an effective classical formulation, we have to calculate the path integral (\[ham12\]). First we express the $\delta$-function for the averaged momentum as a Fourier integral $$\begin{aligned}
\label{ham19}
\delta({\bf p}_0-\overline{{\bf p}(\tau)})&=&
\int \frac{d^2\xi}{(2\pi\hbar)^2}\,\exp\left(-\frac{i}{\hbar}{\mbox{\boldmath$\xi$}}\cdot{\bf p}_0
\right)\exp\left[-\frac{1}{\hbar}\int_0^{\hbar\beta}d\tau\,
{\bf v}_0({\mbox{\boldmath$\xi$}})\cdot {\bf p}(\tau)\right]\end{aligned}$$ involving an auxiliary source $$\label{ham20}
{\bf v}_0({\mbox{\boldmath$\xi$}})=-\frac{i}{\hbar\beta}\,{\mbox{\boldmath$\xi$}}$$ which is constant in time. Substituting the $\delta$-function in Eq. (\[ham12\]) by this source representation, the partition function reads $$\begin{aligned}
\label{ham21}
Z^{{\bf p}_0,{\bf x}_0}&=&\int d^2\xi\,\exp\left(-\frac{i}{\hbar}{\mbox{\boldmath$\xi$}}\cdot{\bf p}_0\right)
\oint {\cal D'}^2x{\cal D}^2p\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})\nonumber\\
&&\times\exp\left\{-\frac{1}{\hbar}\int_0^{\hbar\beta}d\tau\,
\left[-i{\bf p}(\tau)\cdot\dot{\bf x}(\tau)+H({\bf p}(\tau),{\bf x}(\tau))
+{\bf v}_0({\mbox{\boldmath$\xi$}})\cdot{\bf p}(\tau)\right]\right\}.\end{aligned}$$ Evaluating the momentum integrals and utilizing the periodicity property ${\bf x}(0)={\bf x}(\hbar\beta)$, we obtain the configuration space path integral $$\begin{aligned}
\label{ham22}
Z^{{\bf p}_0,{\bf x}_0}=\lim_{\Omega\to 0}
&&\int d^2\xi\,\exp\left(-\frac{i}{\hbar}{\mbox{\boldmath$\xi$}}\cdot{\bf p}_0-\frac{M}{2\hbar^2\beta}
{\mbox{\boldmath$\xi$}}^2\right)
\oint {\cal D}^2x\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})\nonumber\\
&&\times\exp\left\{-\frac{1}{\hbar}\int_0^{\hbar\beta}d\tau\,\left[\frac{M}{2}\dot{\bf x}^2(\tau)
+\frac{1}{2}M\Omega^2{\bf x}^2(\tau)
-\frac{i}{2}M\omega_c({\bf x}(\tau)\times \dot{\bf x}(\tau))_z+
{\bf x}(\tau)\cdot {\bf j}_1
({\mbox{\boldmath$\xi$}})\right]\right\},\end{aligned}$$ where the source ${\bf v}_0$ coupled to the momentum in (\[ham21\]) has turned to a source ${\bf j}_1$ coupled to the path in configuration space [@correlation], with components $$\label{ham23}
{\bf j}_1({\mbox{\boldmath$\xi$}})=\frac{M}{2}\omega_c\left(\,v_{0y}({\mbox{\boldmath$\xi$}}),-v_{0x}({\mbox{\boldmath$\xi$}})
\,\right)=\frac{i\omega_cM}{2\hbar\beta}\left(\,-\xi_y,\xi_x \,\right).$$ We have introduced an additional harmonic oscillator in Eq. (\[ham22\]) which will turn out to be useful at intermediate stages of the development. At the end of the calculation, only the limit $\Omega\to 0$ will be relevant.
Expressing the $\delta$-function in the path integral of Eq. (\[ham22\]) by the Fourier integral $$\label{ham23b}
\delta({\bf x}_0-\overline{{\bf x}(\tau)})=\int \frac{d^2\kappa}{(2\pi)^2}\,
\exp\left(\frac{i}{\hbar}{\mbox{\boldmath$\kappa$}}\cdot {\bf x}_0 \right)
\exp\left[-\frac{1}{\hbar}\int_0^{\hbar\beta}d\tau\,{\bf j}_2({\mbox{\boldmath$\kappa$}})\cdot
{\bf x}(\tau) \right]$$ with the new source $$\label{ham23c}
{\bf j}_2({\mbox{\boldmath$\kappa$}})=\frac{i{\mbox{\boldmath$\kappa$}}}{\beta},$$ the partition function (\[ham22\]) can be written as $$\label{ham23d}
Z^{{\bf p}_0,{\bf x}_0}=\lim_{\Omega\to 0}
\int d^2\xi\,\exp\left(-\frac{i}{\hbar}{\mbox{\boldmath$\xi$}}\cdot{\bf p}_0-\frac{M}{2\hbar^2\beta}
{\mbox{\boldmath$\xi$}}^2\right)\int\frac{d^2\kappa}{(2\pi)^2}\,\exp\left(\frac{i}{\hbar}
{\mbox{\boldmath$\kappa$}}\cdot {\bf x}_0 \right)\,Z_\Omega[{\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})].$$ The functional $Z_\Omega[{\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})]$ is defined as the configuration space path integral $$\label{ham23e}
Z_\Omega[{\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})]=\oint{\cal D}^2x\,
\exp\left[-\frac{1}{2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
{\bf x}(\tau)\,{\bf G}^{-1}(\tau,\tau')\,{\bf x}(\tau')-\frac{1}{\hbar}\int_0^{\hbar\beta}
d\tau\,{\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})\cdot{\bf x}(\tau) \right],$$ where we have introduced the combined source ${\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})={\bf j}_1({\mbox{\boldmath$\xi$}})+{\bf j}_2({\mbox{\boldmath$\kappa$}})$. Formally, the solution reads $$\label{ham23ee}
Z_\Omega[{\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})]=Z_\Omega[0]
\exp\left[-\frac{1}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
{\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})\,{\bf G}(\tau,\tau')\,{\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}}) \right],$$ where ${\bf G}(\tau,\tau')$ is the matrix of Green functions obtained by inverting $$\label{ham23f}
{\bf G}^{-1}(\tau,\tau')=\frac{M}{\hbar}\,\left(\begin{array}{cc}
-\frac{d^2}{d \tau^2}+\Omega^2 & -i\omega_c\frac{d}{d\tau}\\
i\omega_c\frac{d}{d\tau} & -\frac{d^2}{d \tau^2}+\Omega^2
\end{array}\right)\,\delta(\tau-\tau').$$ The inversion is easily done in frequency space after spectrally decomposing the $\delta$-function into the Matsubara frequencies $\omega_m=2\pi m/\hbar\beta$, $$\label{ham23g}
\delta(\tau-\tau')=\frac{1}{\hbar\beta}\sum\limits_{m=-\infty}^\infty\,e^{i\omega_m(\tau-\tau')}.$$ The result is $$\label{ham23h}
\tilde{\bf G}(\omega_m)=\frac{\hbar}{M}\frac{1}{{\rm det}\,\tilde{\bf G}}
\left(\begin{array}{cc}\omega_m^2+\Omega^2 & -\omega_c\omega_m\\
\omega_c\omega_m & \omega_m^2+\Omega^2\end{array}\right).$$ At this point, the additional oscillator in Eq. (\[ham22\]) proves useful: It ensures that the determinant $$\label{ham23i}
{\rm det}\,\tilde{\bf G}(\omega_m)=(\omega_m^2+\Omega^2)^2+\omega_c^2\omega_m^2$$ is nonzero for $m=0$, thus playing the role of an infrared regulator. The Fourier expansion $$\label{ham23j}
{\bf G}(\tau,\tau')=\frac{1}{\hbar\beta}\sum\limits_{m=-\infty}^\infty\,\tilde{\bf G}(\omega_m)
e^{-i\omega_m(\tau-\tau')}$$ yields the matrix of Green functions $$\label{ham23k}
{\bf G}(\tau,\tau')=\left(\begin{array}{cc}G_{xx}(\tau,\tau') & G_{xy}(\tau,\tau')\\
G_{yx}(\tau,\tau') & G_{yy}(\tau,\tau') \end{array}\right)$$ which inherits the symmetry properties from the kernel (\[ham23f\]): $$\label{ham23l}
G_{xx}(\tau,\tau')=G_{yy}(\tau,\tau'),\qquad G_{xy}(\tau,\tau')=
-G_{yx}(\tau,\tau').$$ A more detailed description of these Green functions is given in Apps. \[appgen\] and \[appprop\].
Since the current ${\bf J}$ does not depend on the euclidean time, the expression (\[ham23ee\]) simplifies therefore to $$\label{ham23m}
Z_\Omega[{\bf J}({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})]=Z_\Omega[0]
\exp\left[-\frac{1}{\hbar^2}{\bf J}^2({\mbox{\boldmath$\xi$}},{\mbox{\boldmath$\kappa$}})
\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,G_{xx}(\tau,\tau')\right].$$ The Green function has the Fourier decomposition $$\label{ham23n}
G_{xx}(\tau,\tau')=\frac{1}{M\beta}\sum\limits_{m=-\infty}^\infty\,
\frac{\omega_m^2+\Omega^2}{(\omega_m^2+\Omega_+^2)(\omega_m^2+\Omega_-^2)}
\,e^{-i\omega_m(\tau-\tau')},$$ where $\Omega_\pm$ are the frequencies $$\label{ham23o}
\Omega_\pm=\sqrt{\Omega^2+\frac{1}{2}\omega_c^2\pm\omega_c\sqrt{\Omega^2+\frac{1}{4}\omega_c^2}}.$$ The ratios in the sum of (\[ham23n\]) can be decomposed into two partial fractions, each of them representing a single harmonic oscillator with frequency $\Omega_+$ and $\Omega_-$, respectively. The analytic form of the periodic Green function of a single harmonic oscillator is well known (see Chap. 3 in [@PI]), and we obtain for the present Green function (\[ham23o\]): $$\label{ham23p}
G_{xx}(\tau,\tau')=\frac{1}{M\beta}\left(\frac{\hbar\beta}{2\Omega_+}\frac{\Omega_+^2-\Omega^2}
{\Omega_+^2-\Omega_-^2}\frac{\cosh{\Omega_+(|\tau-\tau'|-\hbar\beta/2)}}
{\sinh{\hbar\beta\Omega_+/2}}- \frac{\hbar\beta}{2\Omega_-}\frac{\Omega_-^2-\Omega^2}
{\Omega_+^2-\Omega_-^2}\frac{\cosh{\Omega_-(|\tau-\tau'|-\hbar\beta/2)}}
{\sinh{\hbar\beta\Omega_-/2}}\right).$$ By writing the determinant (\[ham23i\]) as $$\label{ham23ii}
{\rm det}\,\tilde{\bf G}(\omega_m)=(\omega_m^2+\Omega_+^2)(\omega_m^2+\Omega_-^2)$$ and summing over the logarithms of this, we calculate the partition function as a product of two single harmonic oscillators: $$\label{ham23q}
Z_\Omega=Z_\Omega[0]=
\frac{1}{2\sinh{\hbar\beta\Omega_+/2}}\frac{1}{2\sinh{\hbar\beta\Omega_-/2}}.$$ The results (\[ham23p\]) and (\[ham23q\]) determine the generating functional (\[ham23m\]). The euclidean time integrations are then easily done, and subsequently the ${\mbox{\boldmath$\kappa$}}$- and ${\mbox{\boldmath$\xi$}}$-integrations in (\[ham23d\]). As a result, we obtain the restricted partition function $$\label{ham25}
Z^{{\bf p}_0,{\bf x}_0}=\lim_{\Omega\to 0}\exp\left\{-\beta\left(-\frac{1}{\beta}{\rm ln}\,
\frac{\sinh \hbar\beta\Omega_+/2}{\hbar\beta\Omega_+/2}
\frac{\sinh \hbar\beta\Omega_-/2}{\hbar\beta\Omega_-/2}
+\frac{{\bf p}_0^2}{2M}-\frac{1}{2}\omega_c l_z({\bf p}_0,{\bf x}_0)
+\frac{1}{8}M\omega_c^2{\bf x}_0^2+\frac{1}{2}M\Omega^2{\bf x}_0^2
\right)\right\}.$$ If we now remove the additional oscillator by taking the limit $\Omega\to 0$, we find from (\[ham23o\]): $\Omega_+\rightarrow \omega_c$, $\Omega_-\rightarrow 0$, and therefore $$\label{ham26}
\lim_{\Omega\to 0}\frac{\sinh{\hbar\beta\Omega_+/2}}{\hbar\beta\Omega_+/2}=
\frac{\sinh{\hbar\beta\omega_c/2}}{\hbar\beta\omega_c/2},\quad
\lim_{\Omega\to 0}\frac{\sinh{\hbar\beta\Omega_-/2}}{\hbar\beta\Omega_-/2}=1.$$ Recalling the definition (\[ham12\]), we identify the exact effective classical Hamiltonian for an electron in a magnetic field as $$\label{ham27}
H_{\rm eff}({\bf p}_0,{\bf x}_0)=\frac{1}{\beta}\,{\rm ln}\,\frac{\sinh{\hbar\beta\omega_c/2}}
{\hbar\beta\omega_c/2}+\frac{{\bf p}_0^2}{2M}-\frac{1}{2}\omega_c\,l_z({\bf p}_0,{\bf x}_0)
+\frac{1}{8}M\omega_c^2{\bf x}_0^2.$$ Integrating out the momenta in Eq. (\[ham11\]), the configuration space representation (\[ham06\]) for the partition function contains the effective classical potential for a charged particle in the plane perpendicular to the direction of a uniform magnetic field $$\label{ham28}
V_{\rm eff}({\bf x}_0)=\frac{1}{\beta}\,{\rm ln}\,\frac{\sinh{\hbar\beta\omega_c/2}}
{\hbar\beta\omega_c/2}.$$ Note that this is a constant potential.
Denoting the area $\int d^2x_0$ by $A$, we find the exact quantum statistical partition function $$\label{ham29}
Z=\frac{A}{\lambda_{\rm th}^2}\frac{\hbar\beta\omega_c/2}{\sinh{\hbar\beta\omega_c/2}}.$$ After these preparations, we can turn our attention to the system we want to study in this paper: the hydrogen atom in a uniform magnetic field, where the additional Coulomb interaction prevents us from finding an exact solution for the effective classical Hamilton function.
Hydrogen Atom in Constant Magnetic Field
========================================
The zero-temperature properties of the hydrogen atom without external fiels are exactly known. For the quantum statistics at finite temperatures, an analytic expression exists, but it is hard to evaluate. It is easier to find an accurate approximative result with the help of variational perturbation theory [@hatom]. Similar calculations have been performed for the electron-proton pair distribution function which can be interpreted as the unnormalized density matrix [@density].
Here we extend this method of calculation to the hydrogen atom in a constant magnetic field. This extension is quite nontrivial since the weak- and strong-field limits will turn out to exhibit completely different asymptotic behaviours. Let us first generalize variational perturbation theory to an electron in a constant magnetic field and arbitrary potential.
Generalized Variational Perturbation Theory
-------------------------------------------
We consider once more the effective classical form (\[ham11\]) of the quantum statistical partition function, which requires the path integration (\[ham12\]) in phase space. Fluctuations parallel and vertical to the magnetic field lines are now both nontrivial, and we must deal with the full three-dimensional system and the components of the electron position and momentum are now denoted by ${\bf x}=(x,y,z)$ and ${\bf p}=(p_x,p_y,p_z)$. For the uniform magnetic field pointing along the $z$-axis, the vector potential ${\bf A}({\bf x})$ is used in the gauge (\[ham16\]). Thus the Hamilton function of an electron in a magnetic field and an arbitrary potential $V({\bf x})$ is $$\label{vpt00}
H({\bf p},{\bf x})=\frac{{\bf p}^2}{2M}-\frac{1}{2}\omega_cl_z({\bf p},{\bf x})+\frac{1}{8}
M\omega_c^2{\bf x}^2+V({\bf x}).$$ The orbital angular momentum $l_z({\bf p},{\bf x})$ was introduced in Eq. (\[ham18\]), and the Landau frequency $\omega_c$ below Eq. (\[ham17\]). The importance of the separation of the zero frequency components ${\bf x}_0$ and ${\bf p}_0$ was discussed in Sect. \[effrep\]. Their divergence with the temperature $T$ prevents a perturbative treatment. Thus it is essential to set up the perturbation theory only for the fluctuations around ${\bf x}_0$ and ${\bf p}_0$. For this we rewrite the action functional (\[ham02\]) associated with the Hamiltonian (\[vpt00\]) as $$\label{vpt01}
{\cal A}[{\bf p},{\bf x}]=
{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf p},{\bf x}]
+{\cal A}_{\rm int}[{\bf p},{\bf x}],$$ where we have introduced the fluctuation action $$\begin{aligned}
\label{vpt02}
{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf p},{\bf x}]=
\int_0^{\hbar\beta}d\tau\, \Big\{&&-i[{\bf p}(\tau)-{\bf p}_0]\cdot\dot{\bf x}(\tau)+
\frac{1}{2M}[{\bf p}(\tau)-{\bf p}_0]^2+\frac{1}{2}{\Omega_{\perp 1}}l_z({\bf p}(\tau)-{\bf p}_0,{\bf x}(\tau)-{\bf x}_0)\nonumber\\
&&+\frac{1}{8}M{\Omega_{\perp 2}}^2
\left[{\bf x}^\perp(\tau)-{\bf x}_0^\perp\right]^2+\frac{1}{2}M{\Omega_{\parallel}}^2[z(\tau)-z_0]^2 \Big\},\end{aligned}$$ in which ${\bf x}^\perp=(x,y)$ denotes the transverse part of ${\bf x}$. The interaction is now $$\label{vpt03}
{\cal A}_{\rm int}[{\bf p},{\bf x}]=\int_0^{\hbar\beta}d\tau\,
V_{\rm int}({\bf p}(\tau),{\bf x}(\tau))
={\cal A}[{\bf p},{\bf x}]-
{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf p},{\bf x}]$$ with the interaction potential $$\begin{aligned}
\label{vpt03b}
V_{\rm int}({\bf p}(\tau),{\bf x}(\tau))&=&\frac{1}{2M}\left\{{\bf p}^2(\tau)-
\left[{\bf p}(\tau)-{\bf p}_0 \right]^2 \right\}+\frac{1}{2}\omega_c\,
{\bf p}^\perp(\tau)\times
{\bf x}^\perp(\tau)
\nonumber\\
&&-\frac{1}{2}{\Omega_{\perp 1}}({\bf p}^\perp(\tau)-{\bf p}^\perp_0)\times
({\bf x}^\perp(\tau)-{\bf x}^\perp_0)
+\frac{1}{8}M\omega_c^2{{\bf x}^\perp}^2(\tau)\nonumber\\
&&-\frac{1}{8}M{\Omega_{\perp 2}}^2
\left[{\bf x}^\perp(\tau)-{\bf x}_0^\perp\right]^2
-\frac{1}{2}M{\Omega_{\parallel}}^2[z(\tau)-z_0]^2+V({\bf x}(\tau)),\end{aligned}$$ where ${\bf p}^\perp=(p_x,p_y)$. The frequencies ${\bf \Omega}=({\Omega_{\perp 1}},{\Omega_{\perp 2}},{\Omega_{\parallel}})$ are for the moment arbitrary. The decomposition (\[vpt01\]) forms the basis for the variational approach, where the first term in the action (\[vpt01\]) allows an exact treatment. The transverse part was given in Sec. \[exham\] and the longitudinal part is trivial, since it is harmonic with frequency ${\Omega_{\parallel}}$. The associated partition function is given by the path integral $$\label{vpt04}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}=
\oint{\cal D'}^3x{\cal D}^3p\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\delta({\bf p}_0-\overline{{\bf p}(\tau)})
e^{-{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf p},{\bf x}]/\hbar},$$ which can be performed. Details are given in Appendix \[appgenrp\]. The result is $$\label{vpt04b}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}=
\frac{\hbar\beta\Omega_+/2}{\sinh{\hbar\beta\Omega_+/2}}\,
\frac{\hbar\beta\Omega_-/2}{\sinh{\hbar\beta\Omega_-/2}}\,
\frac{\hbar\beta{\Omega_{\parallel}}/2}{\sinh{\hbar\beta{\Omega_{\parallel}}/2}},$$ where auxiliary frequencies are composed of the frequencies ${\Omega_{\perp 1}},{\Omega_{\perp 2}}$ in the action (\[vpt02\]) as $$\label{vpt04c}
\Omega_{\pm}({\Omega_{\perp 1}},{\Omega_{\perp 2}})=\frac{1}{2}\,|{\Omega_{\perp 1}}\pm{\Omega_{\perp 2}}|.$$ This partition function serves in the subsequent pertubation expansion as trial system which depends explicitly on the frequencies ${\bf \Omega}$. The correlation functions are a straightforward generalization of (\[ham23k\]) to three dimensions: $$\label{hyd05c}
{\bf G}^{{\bf x}_0}(\tau,\tau')=\left(\begin{array}{ccc}
G_{xx}^{{\bf x}_0}(\tau,\tau') & G_{xy}^{{\bf x}_0}(\tau,\tau') & 0\\
G_{yx}^{{\bf x}_0}(\tau,\tau') & G_{yy}^{{\bf x}_0}(\tau,\tau') & 0\\
0 & 0 & G_{zz}^{{\bf x}_0}(\tau,\tau')\end{array}\right),$$ whose explicit form is derived in App. \[appgenrp\].
The ${\bf \Omega}$-dependent action in Eq. (\[vpt01\]) is treated perturbatively. Writing the partition function (\[ham12\]) as $$\label{vpt05}
Z^{{\bf p}_0,{\bf x}_0}=
(2\pi\hbar)^3\oint {\cal D'}^3x{\cal D}^3p\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\delta({\bf p}_0-\overline{{\bf p}(\tau)})\,
\exp\left\{-\frac{1}{\hbar}{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
[{\bf p},{\bf x}]\right\}
\exp\left\{-\frac{1}{\hbar}\int_0^{\hbar\beta}d\tau\,
V_{\rm int}({\bf p}(\tau),{\bf x}(\tau))\right\},$$ the second exponential is expanded into a Taylor series, yielding $$\begin{aligned}
\label{vpt06}
Z^{{\bf p}_0,{\bf x}_0}&=&
(2\pi\hbar)^3\oint {\cal D'}^3x{\cal D}^3p\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\delta({\bf p}_0-\overline{{\bf p}(\tau)})\,
\exp\left\{-\frac{1}{\hbar}{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
[{\bf p},{\bf x}]\right\}\nonumber\\
&&\times\left[1-\frac{1}{\hbar}\int_0^{\hbar\beta}d\tau\,V_{\rm int}({\bf p}(\tau),{\bf x}(\tau))
+\frac{1}{2!\hbar^2}\int_0^{\hbar\beta}d\tau_1\int_0^{\hbar\beta}d\tau_2\,
V_{\rm int}({\bf p}(\tau_1),{\bf x}(\tau_1))V_{\rm int}({\bf p}(\tau_2),{\bf x}(\tau_2))-
\ldots\right].\end{aligned}$$ Defining harmonic expectation values with respect to the restricted path integral as $$\label{vpt07}
{\left\langle\,\ldots\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
\frac{(2\pi\hbar)^3}{Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}\oint{\cal D'}^3x {\cal D}^3p
\,\ldots\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})\delta({\bf p}_0-\overline{{\bf p}(\tau)})
\exp\left\{-\frac{1}{\hbar}{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf p},{\bf x}]
\right\},$$ the perturbation expansion for the partition function (\[vpt06\]) reads $$\label{vpt08}
Z^{{\bf p}_0,{\bf x}_0}=Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}\,\sum\limits_{n=0}^\infty
\frac{(-1)^n}{\hbar^nn!}\,
{\left\langle\,\left(\int_0^{\hbar\beta}d\tau\,V_{\rm int}({\bf p}(\tau),{\bf x}(\tau))\right)^n\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}.$$ This power series expansion can be rewritten in the exponential form $$\label{vpt09}
Z^{{\bf p}_0,{\bf x}_0}=Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}\,\exp\left\{
\sum\limits_{n=0}^\infty
\frac{(-1)^n}{\hbar^nn!}\,
{\left\langle\,\left(\int_0^{\hbar\beta}d\tau\,V_{\rm int}({\bf p}(\tau),{\bf x}(\tau))\right)^n\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{{\bf \Omega},c}}\right\},$$ where the subscript $c$ on the expectation values indicates cumulants. The lowest cumulants are related to the full expectation values as follows: $$\begin{aligned}
\label{vpt10}
{\left\langle\,O_1({\bf p}(\tau_1),{\bf x}(\tau_1))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{{\bf \Omega},c}}&=&{\left\langle\,O_1({\bf p}(\tau_1),{\bf x}(\tau_1))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}},
\nonumber\\
{\left\langle\,O_1({\bf p}(\tau_1),{\bf x}(\tau_1))O_2({\bf p}(\tau_2),{\bf x}(\tau_2))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{{\bf \Omega},c}}
&=&{\left\langle\,O_1({\bf p}(\tau_1),{\bf x}(\tau_1))O_2({\bf p}(\tau_2),{\bf x}(\tau_2))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}\nonumber\\
&&-{\left\langle\,O_1({\bf p}(\tau_1),{\bf x}(\tau_1))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}
{\left\langle\,O_2({\bf p}(\tau_2),{\bf x}(\tau_2))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}},\nonumber\\
&\vdots&\quad,\end{aligned}$$ where $O_i({\bf p}(\tau_j),{\bf x}(\tau_j))$ denotes any observable depending on position and momentum. Recalling the relation (\[ham12\]) between partition function (\[vpt09\]) and effective classical Hamiltonian $H_{\rm eff}({\bf p}_0,{\bf x}_0)$, we obtain from (\[vpt09\]) the effective classical Hamiltonian as a cumulant expansion: $$\label{vpt11}
H_{\rm eff}({\bf p}_0,{\bf x}_0)=-\frac{1}{\beta}{\rm ln}\,Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
+\frac{1}{\beta}\sum\limits_{n=1}^\infty
\frac{(-1)^n}{\hbar^nn!}\,
{\left\langle\,\left(\int_0^{\hbar\beta}d\tau\,V_{\rm int}({\bf p}(\tau),{\bf x}(\tau))\right)^n\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{{\bf \Omega},c}}.$$ Up to now, we did not make any approximation. The expansion on the right-hand side is an exact expression for the effective classical Hamiltonian for any ${\bf \Omega}$.
For systems with a nontrivial interaction, we are capable of calculating only some initial truncated part of the series (\[vpt11\]), say up to the $N$th order, leading to the approximate effective classical Hamiltonian $$\label{vpt12}
{\cal H}_{\bf \Omega}^{(N)}({\bf p}_0,{\bf x}_0)=-
\frac{1}{\beta}{\rm ln}\,Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
+\frac{1}{\beta}\sum\limits_{n=1}^N
\frac{(-1)^n}{\hbar^nn!}\,
{\left\langle\,\left(\int_0^{\hbar\beta}d\tau\,V_{\rm int}({\bf p}(\tau),{\bf x}(\tau))\right)^n\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{{\bf \Omega},c}}.$$ This depends explicitly on the three parameters ${\bf \Omega}$. Since the exact expression (\[vpt11\]) is independent of ${\bf \Omega}$, the best approximation for ${\cal H}_{\bf \Omega}^{(N)}({\bf p}_0,{\bf x}_0)$ should depend on ${\bf \Omega}$ [*minimally*]{}. Thus the optimal solution will be found by determining the parameters from the conditions $$\label{vpt13}
\nabla_{\bf \Omega} {\cal H}_{\bf \Omega}^{(N)}({\bf p}_0,{\bf x}_0)\stackrel{!}{=}0.$$ Let us denote the optimal variational parameters to $N$th order by $$\label{vpt14}
{\bf \Omega}^{(N)}=\left({\Omega_{\perp 1}}^{(N)}({\bf p}_0,{\bf x}_0),
{\Omega_{\perp 2}}^{(N)}({\bf p}_0,{\bf x}_0), {\Omega_{\parallel}}^{(N)}({\bf p}_0,{\bf x}_0)\right).$$ Inserting these into Eq. (\[vpt12\]) yields the optimal effective classical Hamiltonian ${\cal H}^{(N)}({\bf p}_0,{\bf x}_0)$.
First-Order Effective Classical Potential
-----------------------------------------
The first-order approximation of the effective classical Hamiltonian (\[vpt12\]) reads $$\label{fir00}
{\cal H}_{\bf \Omega}^{(1)}({\bf p}_0,{\bf x}_0)=-
\frac{1}{\beta}{\rm ln}\,Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
-{\left\langle\,V_{\rm int}({\bf p},{\bf x})\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}.$$ In writing the last term we have used the fact that, as a consequence of the time translation invariance of the system, the first-order expectation value of $V_{\rm int}({\bf x})$ is independent of the euclidean time $\tau$.
In order to calculate ${\cal H}_{\bf \Omega}^{(1)}({\bf p}_0,{\bf x}_0)$, we use the two-point correlation functions derived in App. \[appgenrp\], and the vanishing of the linear expectations, e.g. $$\label{fir01}
{\left\langle\,p_x(\tau)-{p_0}_x\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=0$$ to find $$\label{fir02}
{\cal H}_{\bf \Omega}^{(1)}({\bf p}_0,{\bf x}_0)=
\frac{{\bf p}_0^2}{2M}-\frac{1}{2}\omega_c
l_z({\bf p}_0,{\bf x}_0)+\frac{1}{8}M\omega_c^2(x_0^2+y_0^2)+
W_{\bf \Omega}^{(1)}({\bf x}_0),$$ where we have collected all terms depending on the variational parameters ${\bf \Omega}$ in the potential $$\label{fir03}
W_{\bf \Omega}^{(1)}({\bf x}_0)=-\frac{1}{\beta}{\rm ln}\,
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
+(\omega_c-{\Omega_{\perp 1}})\,b^2_\perp({\bf x}_0)-\frac{1}{4}\left({\Omega_{\perp 2}}^2-\omega_c^2 \right)
\,a^2_\perp({\bf x}_0)-\frac{1}{2}M{\Omega_{\parallel}}^2a^2_\parallel({\bf x}_0)+
{\left\langle\,V({\bf x})\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}.$$ The quantities $a^2_\perp({\bf x}_0)$ and $a^2_\parallel({\bf x}_0)$ are the transverse and longitudinal fluctuation widths $$\label{fir03b}
a^2_\perp({\bf x}_0)=G^{{\bf p}_0,{\bf x}_0}_{xx}(0),
\quad a^2_\parallel({\bf x}_0)=G^{{\bf p}_0,{\bf x}_0}_{zz}(0),
\quad b^2_\perp({\bf x}_0)=G^{{\bf p}_0,{\bf x}_0}_{xp_y}(0).$$ Note that the potential (\[fir03\]) is independent of ${\bf p}_0$. This means that the approximation (\[fir02\]) to the effective classical Hamiltonian contains no coupling of the momentum ${\bf p}_0$ to a variational parameter ${\bf \Omega}$, such that the optimal ${\bf \Omega}^{(1)}$ determined by minimizing ${\cal H}_{\bf \Omega}^{(1)}({\bf p}_0,{\bf x}_0)$ is independent of ${\bf p}_0$. We may therefore integrate out ${\bf p}_0$ in the phase space representation of the first-order approximation for the partition function $$\label{fir04}
Z^{(1)}=\int\frac{d^3x_0d^3p_0}{(2\pi\hbar^3)}\,
e^{-\beta{\cal H}_{\bf \Omega}^{(1)}({\bf p}_0,{\bf x}_0)}$$ to find the pure configuration space integral $$\label{fir05}
Z^{(1)}=\int\frac{d^3x_0}{\lambda_{\rm th}^3}\,
e^{-\beta W_{\bf \Omega}^{(1)}({\bf x}_0)},$$ in which $W_{\bf \Omega}^{(1)}({\bf x}_0)$ is the first-order approximation to the effective classical potential of an electron in a potential $V({\bf x})$ and a uniform magnetic field.
Application to the Hydrogen Atom in a Magnetic Field
----------------------------------------------------
We now apply the formulas of the preceding section to the Hamiltonian (\[vpt00\]) with an attracting Coulomb potential $$\label{hyd00}
V({\bf x})=-\frac{e^2}{4\pi\varepsilon_0\,|{\bf x}|}\,,$$ where $|{\bf x}|$ is the distance between the electron and the proton. The only nontrivial problem is the calculation of the expectation value ${\left\langle\,V({\bf x}(\tau))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}$ in Eq. (\[fir03\]). This is done using the so-called [*smearing formula*]{}, which is a Gaussian convolution of $V({\bf x})$. This formula was first derived by Feynman and Kleinert [@fk], and exists now also in an extension to arbitrary order [@density; @hatom]. The generalization to position and momentum dependent observables was given in the phase space formulation [@correlation]. We briefly rederive the first-order smearing formula. The expectation value is defined by $$\label{hyd01}
{\left\langle\,V({\bf x}(\tau'))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=\frac{(2\pi\hbar)^3}{Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}
\oint {\cal D'}^3x{\cal D}^3p\,V({\bf x}(\tau'))\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\delta({\bf p}_0-\overline{{\bf p}(\tau)})
e^{-{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf p},{\bf x}]/\hbar},$$ Now we substitute the potential by the expression $$\label{hyd02}
V({\bf x}(\tau'))=\int d^3x\,V({\bf x})\delta({\bf x}-{\bf x}(\tau'))
=\int d^3x\,V({\bf x})\int\,\frac{d^3\kappa}{(2\pi)^3}\,\exp\left[i{\mbox{\boldmath$\kappa$}}\cdot
({\bf x}-{\bf x}_0) \right]
\exp\left\{-\frac{1}{\hbar}
\int_0^{\hbar\beta}d\tau\,{\bf j}(\tau)\cdot[{\bf x}(\tau)-{\bf x}_0]
\right\},$$ where we have introduced the source $$\label{hyd03}
{\bf j}(\tau)=i\hbar{\mbox{\boldmath$\kappa$}}\delta(\tau-\tau').$$ Inserting the expression (\[hyd02\]) into Eq. (\[hyd01\]) we obtain $$\label{hyd04}
{\left\langle\,V({\bf x}(\tau'))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=\frac{1}{Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}
\int d^3x\,V({\bf x})\int\,\frac{d^3\kappa}{(2\pi)^3}\,\exp\left[i{\mbox{\boldmath$\kappa$}}\cdot
({\bf x}-{\bf x}_0) \right]\,Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j}],$$ with the harmonic generating functional $$\label{hyd05}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j}]=
(2\pi\hbar)^3\oint {\cal D'}^3x{\cal D}^3p\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\delta({\bf p}_0-\overline{{\bf p}(\tau)})
\exp\left\{-\frac{1}{\hbar}{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf p},{\bf x}]
-\frac{1}{\hbar}\int_0^{\hbar\beta}d\tau\,{\bf j}(\tau)\cdot[{\bf x}(\tau)-{\bf x}_0]\right\}.$$ The solution is $$\label{hyd05b}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j}]=
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
\exp\left[\frac{1}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
{\bf j}(\tau)\,{\bf G}^{{\bf x}_0}(\tau,\tau')\,{\bf j}(\tau')\right]$$ with the $3\times 3$-matrix of Green functions of Eq. (\[hyd05c\]). The properties of the Green functions are discussed in the Appendices \[appgen\] and \[appprop\]. Expressing the source ${\bf j}(\tau)$ in terms of ${\mbox{\boldmath$\kappa$}}$ via Eq. (\[hyd03\]) and performing the $\tau$-integrations, we arrive at $$\label{hyd06}
{\left\langle\,V({\bf r}(\tau'))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=\int d^3x\,V({\bf x})\int\,\frac{d^3\kappa}{(2\pi)^3}\,
\exp\left\{i{\mbox{\boldmath$\kappa$}}\cdot[{\bf x}-{\bf x}_0] \right\}\,
\exp\left[-\frac{1}{2}{\mbox{\boldmath$\kappa$}}\,{\bf G}^{{\bf x}_0}(0)\,{\mbox{\boldmath$\kappa$}} \right].$$ Recognizing that $G^{{\bf x}_0}_{yx}(0)=G^{{\bf x}_0}_{xy}(0)$ vanish, the ${\mbox{\boldmath$\kappa$}}$-integral is easily calculated and leads to the first-order smearing formula for an arbitrary position dependent potential $$\label{hyd07}
{\left\langle\,V({\bf x}(\tau'))\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=\frac{1}{(2\pi)^{3/2}\,a^2_\perp({\bf x}_0)
\sqrt{a^2_\parallel({\bf x}_0)}}\int d^3x
\,V({\bf x})\,
\exp\left[-\frac{(x-x_0)^2+(y-y_0)^2}{2a^2_\perp({\bf x}_0)}
-\frac{(z-z_0)^2}{2a^2_\parallel({\bf x}_0)} \right],$$ the right-hand side containing the Gaussian fluctuation widths (\[fir03b\]).
For the Coulomb potential (\[hyd00\]) that we are interested in, the integral in the smearing formula (\[hyd07\]) can not be done exactly. An integral representation for a simple numerical treatment is $$\begin{aligned}
\label{hyd08}
{\left\langle\,-\frac{e^2}{4\pi\varepsilon_0\,|{\bf x}|}\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}&=&-\frac{e^2}{4\pi\varepsilon_0}
\sqrt{\frac{2}{\pi}\,a^2_\parallel({\bf x}_0)}\int\limits_0^1\frac{d\xi}
{a^2_\parallel({\bf x}_0)+\xi^2[a^2_\perp({\bf x}_0)-a^2_\parallel({\bf x}_0)]}\nonumber\\
&&\times\exp\left\{-\frac{\xi^2}{2}\left(\frac{x_0^2+y_0^2}{a^2_\parallel({\bf x}_0)+
\xi^2[a^2_\perp({\bf x}_0)-a^2_\parallel({\bf x}_0)]}+\frac{z_0^2}{a^2_\parallel({\bf x}_0)}
\right) \right\}.\end{aligned}$$ With this expression we know the entire first-order effective classical potential (\[fir03\]) for an electron in a Coulomb potential and a uniform magnetic field which has to be optimized in the variational parameters ${\bf \Omega}$.
Results {#res}
=======
We are now going to optimize the effective classical potential by extremizing it in ${\bf \Omega}$ at different temperatures and magnetic field strengths. In the zero-temperature limit this will produce the ground state energy.
Effective Classical Potential for Different Temperatures and Magnetic Field Strengths
-------------------------------------------------------------------------------------
The optimization of $W_{{\bf \Omega}}^{(1)}({\bf x}_0)$ proceeds by minimization in ${\bf \Omega}$ and must be done for each value of ${\bf x}_0$. Reinserting the optimal parameters ${\bf \Omega}^{(1)}({\bf x}_0)$ into the expressions (\[fir03\]) and (\[hyd08\]), we obtain the optimal first-order effective classical potential $W^{(1)}({\bf x}_0)$. The calculations are done numerically, where we used natural units $\hbar=e^2/4\pi\varepsilon_0=k_B=c=M=1$. This means that energies are measured in units of $\epsilon_0=Me^4/(4\pi\varepsilon_0)^2\hbar^2\equiv 2\,{\rm Ryd}\approx 27.21\,{\rm eV}$, temperatures in $\epsilon_0/k_B\approx 3.16\times 10^5\,{\rm K}$, distances in Bohr radii $a_B=(4\pi\varepsilon_0)^2\hbar^2/Me^2\approx
0.53\times 10^{-10}\,{\rm m}$, and magnetic field strengths in $B_0=e^3M^2/\hbar^3(4\pi\varepsilon_0)^2\approx 2.35\times 10^5\,{\rm T}=2.35\times10^9\,
{\rm G}$. Figure \[wofq1\] shows the resulting curves for various magnetic field strengths $B$ and an inverse tempature $\beta=1/T=1$. Examples of the lower temperature behaviour are shown in Fig. \[wofq100\] for $\beta=100$. To see the expected anisotropy of the curves in the magnetic field direction and in the plane perpendicular to it, we plot simultanously the curves for $W^{(1)}({\bf x}_0)$ transversal to the magnetic field as a function of $\rho_0=\sqrt{x_0^2+y_0^2}$ at $z=0$ (solid curves) and parallel as a function of $z_0$ at $\rho_0=0$ (dashed curves). The curves become strongly anisotropic for low temperatures and increasing field strengths (Fig. \[wofq100\]). At a given field strength $B$, the two curves converge for large distances from the origin, where the proton resides, to the same constant depending on $B$. This is due to the decreasing influence of the Coulomb interaction which shows the classical $1/r$-behaviour in each direction. When approaching the classical high-temperature limit, the effect of anisotropy becomes less important since the violent thermal fluctuations do not have a preferred direction (see Fig. \[wofq1\]). For $\rho_0\to\infty$ or $z_0\to\infty$, the expectation value of the Coulomb potential (\[hyd08\]) tends to zero. The remaining effective classical potential $$\label{res01}
W_{\bf \Omega}^{(1)}({\bf x}_0)\longrightarrow -\frac{1}{\beta}{\rm ln}\,
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
+(\omega_c-{\Omega_{\perp 1}})\,b^2_\perp-\frac{1}{4}\left({\Omega_{\perp 2}}^2-\omega_c^2 \right)
\,a^2_\perp-\frac{1}{2}M{\Omega_{\parallel}}^2a^2_\parallel$$ is a constant with regard to the position ${\bf x}_0$, and the optimization yields ${\Omega_{\perp 1}}^{(1)}={\Omega_{\perp 2}}^{(1)}=\omega_c$ and ${\Omega_{\parallel}}^{(1)}=0$, leading to the asymptotic constant value $$\label{res02}
W^{(1)}({\bf x}_0)\longrightarrow -\frac{1}{\beta}{\rm ln}\,
\frac{\hbar\beta\omega_c/2}{\sinh{\hbar\beta\omega_c/2}}.$$ The $B=0$ -curves are of course identical with those obtained from variational perturbation theory for the hydrogen atom [@hatom].
Ground State Energy of the Hydrogen Atom in Uniform Magnetic Field {#vptlow}
------------------------------------------------------------------
In what follows we investigate the zero-temperature behaviour of the theory. Figures \[wofq1\] and \[wofq100\] show that the minimum of each potential curve lies at the origin. This means that the first-order approximation to the ground state energy for a fixed magnitude of the magnetic field $B$ is found by considering the zero-temperature limit of the first-order effective classical potential in the origin $$\label{gr00}
E^{(1)}=\lim_{\beta\to\infty}W^{(1)}(0).$$ Thus we obtain from Eq. (\[fir03\]) the variational expression for the ground state energy: $$\label{gr01}
E^{(1)}_{\bf \Omega}(B)=\frac{\hbar}{4{\Omega_{\perp 2}}}\left({\Omega_{\perp 2}}^2+\omega_c^2\right)+\frac{\hbar{\Omega_{\parallel}}}{4}-
\frac{e^4}{4\pi\varepsilon_0}{\left\langle\,\frac{1}{|{\bf x}|}\,\right\rangle^{\bf 0}_{\bf \Omega}},$$ where the expectation value for the Coulomb potential (\[hyd08\]) can now be calculated exactly since the exponential in the integral simplifies to unity: $$\label{gr02}
{\left\langle\,\frac{1}{|{\bf x}|}\,\right\rangle^{\bf 0}_{\bf \Omega}}=2\sqrt{\frac{M}{\pi\hbar}}\times\left\{
\begin{array}{cl}
{\displaystyle}\sqrt{\frac{{\Omega_{\parallel}}{\Omega_{\perp 2}}}{{\Omega_{\parallel}}-{\Omega_{\perp 2}}}}\,{\rm arctan}\,\sqrt{\frac{2{\Omega_{\parallel}}}{{\Omega_{\perp 2}}}-1}, & \quad 2{\Omega_{\parallel}}> {\Omega_{\perp 2}},\\
{\displaystyle}\sqrt{{\Omega_{\parallel}}}, & \quad 2{\Omega_{\parallel}}= {\Omega_{\perp 2}},\\
{\displaystyle}\frac{1}{2i}\sqrt{\frac{{\Omega_{\parallel}}{\Omega_{\perp 2}}}{{\Omega_{\parallel}}-{\Omega_{\perp 2}}}}\,{\rm ln}\,
\frac{1+i\sqrt{2{\Omega_{\parallel}}/{\Omega_{\perp 2}}-1}}{1-i\sqrt{2{\Omega_{\parallel}}/{\Omega_{\perp 2}}-1}}, & \quad 2{\Omega_{\parallel}}< {\Omega_{\perp 2}}.
\end{array}\right.$$ The equations (\[gr01\]) and (\[gr02\]) are independent of the frequency parameter ${\Omega_{\perp 1}}$ such that the optimization of the first-order expression for the ground state energy (\[gr01\]) requires the satisfying of the equations $$\label{gr03}
\frac{\partial E^{(1)}_{\bf \Omega}(B)}{\partial {\Omega_{\perp 2}}}\stackrel{!}{=}0,\qquad
\frac{\partial E^{(1)}_{\bf \Omega}(B)}{\partial {\Omega_{\parallel}}}\stackrel{!}{=}0.$$ Reinserting the resulting values ${\Omega_{\perp 2}}^{(1)}$ and ${\Omega_{\parallel}}^{(1)}$ into Eq. (\[gr01\]) yields the first-order approximation for the ground state energy $E^{(1)}(B)$. In the absence of the Coulomb interaction the optimization with respect to ${\Omega_{\perp 2}}$ yields ${\Omega_{\perp 2}}^{(1)}=\omega_c$, rendering the ground state energy $E^{(1)}(B)=\omega_c/2$, which is the zeroth Landau level. An optimal value for ${\Omega_{\parallel}}$ does not exist since the dependece of the ground state energy of this parameter is linear in Eq. (\[gr01\]) in this special case. To obtain the lowest energy, this parameter can be set to zero (all optimal frequency parameters used in the optimization procedure turn out to be nonnegative). For a vanishing magnetic field, $B=0$, Eq. (\[gr01\]) exactly reproduces the first-order variational result for the ground state energy of the hydrogen atom, $E^{(1)}(B=0)\approx -0.42\, {\rm [2\,Ryd]}$, obtained in Ref. [@hatom].
To investigate the asymptotics in the strong-field limit $B\to\infty$, it is useful to extract the leading term $\omega_c/2$. Thus we define the binding energy $$\label{gr04}
\varepsilon(B)\equiv\frac{\omega_c}{2}-E(B)$$ which possesses an characteristic strong-field behaviour to be discussed in detail subsequently. The result is shown in Fig. \[bofb\] as a function of the magnitude of the magnetic field $B$, where it is compared with the high-accuracy results of Ref. [@zinn-justin]. As a first-order approximation, this result is satisfactory. It is of the same quality like other first-order results, for example those from the operator optimization method in first order of Ref. [@feranshuk]. The advantage of variational perturbation theory is that it yields good results over the complete range of the coupling strength, here the magnetic field. Moreover, as a consequence of the exponential convergence [@PI Chap. 5], higher orders of variational perturbation theory push the approximative result of any quantity very rapidly towards the exact value.
### The Weak-Field Case
We investigate now the weak-field behaviour of our theory starting from the expression (\[gr04\]) and the expectation value of the Coulomb potential (\[gr02\]) in natural units: $$\label{we00}
\varepsilon^{(1)}_{\eta,\Omega}(B)=
\frac{B}{2}-\frac{\Omega}{4}\left(1+\frac{\eta}{2}\right)-\frac{B^2}{4\Omega}
-\sqrt{\frac{\eta\Omega}{2\pi}}h(\eta)$$ with $$\label{we01}
h(\eta)=\frac{1}{\sqrt{1-\eta}}\,{\rm ln}\,\frac{1-\sqrt{1-\eta}}{1+\sqrt{1-\eta}}.$$ In comparison with Eq. (\[gr01\]) we introduced new variational parameters $$\label{we02}
\eta\equiv\frac{2{\Omega_{\parallel}}}{{\Omega_{\perp 2}}},\qquad \Omega\equiv{\Omega_{\perp 2}}$$ and utilized, as the calculations for the binding energy showed, that always $\eta\le 1$. Performing the derivatives with respect to these variational parameters and setting them zero yields conditional equations which can be written after some manipulations as $$\begin{aligned}
\label{we03}
\frac{\Omega}{8}+\sqrt{\frac{\Omega}{2\pi\eta}}\frac{1}{1-\eta}\left(1+\frac{1}{2}
\frac{1}{\sqrt{1-\eta}}{\rm ln}\,\frac{1-\sqrt{1-\eta}}{1+\sqrt{1-\eta}}\right)&\stackrel{!}{=}&0,
\nonumber\\
\label{we04}
\frac{1}{4}+\frac{\eta}{8}-\frac{B^2}{4\Omega^2}+\frac{1}{2}\sqrt{\frac{\eta}{2\pi\Omega}}
\frac{1}{\sqrt{1-\eta}}{\rm ln}\,\frac{1-\sqrt{1-\eta}}{1+\sqrt{1-\eta}}&\stackrel{!}{=}&0.\end{aligned}$$ Expanding the variational parameters into perturbation series of the square magnetic field $B^2$, $$\label{we05}
\eta(B)=\sum\limits_{n=0}^{\infty}\,\eta_n B^{2n},\qquad
\Omega(B)=\sum\limits_{n=0}^{\infty}\,\Omega_n B^{2n}$$ and inserting these expansions into the self-consistency conditions (\[we03\]) and (\[we04\]) we obtain order by order the coefficients given in Table \[tab1\]. Inserting these values into the expression for the binding energy (\[we00\]) and expand with respect to $B^2$, we obtain the perturbation series $$\label{we06}
\varepsilon^{(1)}(B)=\frac{B}{2}-\sum\limits_{n=0}^\infty\,\varepsilon_n B^{2n}.$$ The first coefficients are also given in Table \[tab1\]. We find thus the important result that the first-order variational perturbation solution possesses a perturbative behaviour with respect to the square magnetic field strength $B^2$ in the weak-field limit thus yielding the correct asymptotics. The coefficients differ in higher order from the exact ones but are improved in higher orders of the variational perturbation theory [@PI Chap. 5].
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$n$ 0 1 2 3
----------------------------------- ------------------------------------------------- -------------------------------------------------------- ----------------------------------------------------------------- -----------------------------------------------------------------------------
$\eta_n$ $1.0$ ${\displaystyle}-\frac{405\pi^2}{7168}\approx -0.5576$ ${\displaystyle}\frac{16828965\pi^4}{1258815488}\approx 1.3023$ ${\displaystyle}-\frac{3886999332075\pi^6}{884272562962432}\approx -4.2260$
\[1mm\] $\Omega_n$ ${\displaystyle}\frac{32}{9\pi}\approx 1.1318$ ${\displaystyle}\frac{99\pi}{224}\approx 1.3885$ ${\displaystyle}-\frac{1293975\pi^3}{19668992}\approx -2.03982$ ${\displaystyle}\frac{524431667187\pi^5}{27633517592576}\approx 5.8077$
\[1mm\] $\varepsilon_n$ ${\displaystyle}-\frac{4}{3\pi}\approx -0.4244$ ${\displaystyle}\frac{9\pi}{128}\approx 0.2209$ ${\displaystyle}-\frac{8019\pi^3}{1835008}\approx -0.1355$ ${\displaystyle}\frac{256449807\pi^5}{322256764928}
\approx 0.2435$
\[1mm\] $\varepsilon_n$ [@cizek1] $-0.5$ $0.25$ ${\displaystyle}-\frac{53}{192}\approx -0.2760$ ${\displaystyle}\frac{5581}{4608}\approx 1.2112$
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[tab1\] Perturbation coefficients up to order $B^6$ for the weak-field expansions of the variational parameters and the binding energy in comparison to the exact ones of Ref. [@cizek1].
### Asymptotical Behaviour in the Strong-Field Regime
In the discussion of the pure magnetic field below Eq. (\[gr03\]) we have mentioned that the variational calculation for the ground state energy which is thus associated with the zeroth Landau level yields a frequency ${\Omega_{\perp 2}}\propto B$ while ${\Omega_{\parallel}}=0$. Therefore we use the assumption (with ${\Omega_{\perp}}\equiv {\Omega_{\perp 2}}$) $$\label{st00}
{\Omega_{\perp}}\gg 2{\Omega_{\parallel}},\qquad {\Omega_{\parallel}}\ll B$$ for the consideration of the ground state energy (\[gr01\]) of the hydrogen atom in a strong magnetic field. In a first step we expand the last expression of the expectation value (\[gr02\]) which corresponds to the condition (\[st00\]) in terms of $2{\Omega_{\parallel}}/{\Omega_{\perp}}$ and reinsert this expansion in the equation of the ground state energy (\[gr01\]). Then we omit all terms proportional to $C/{\Omega_{\perp}}$ where $C$ stands for any expression with a value much smaller than the field strength $B$. In natural units, we thus obtain the strong-field approximation for the first-order binding energy (\[gr04\]) $$\label{st01}
\varepsilon_{{\Omega_{\perp}},{\Omega_{\parallel}}}^{(1)}=\frac{B}{2}-\left(\frac{{\Omega_{\perp}}}{4}+\frac{B^2}{4{\Omega_{\perp}}}+\frac{{\Omega_{\parallel}}}{4}
+\sqrt{\frac{{\Omega_{\parallel}}}{\pi}}\,{\rm ln}\,\frac{{\Omega_{\parallel}}}{2{\Omega_{\perp}}}\right).$$ As usual, we consider the zeros of the derivatives with respect to the variational parameters $$\label{st02}
\frac{\partial \varepsilon_{{\Omega_{\perp}},{\Omega_{\parallel}}}^{(1)}}{\partial {\Omega_{\parallel}}}\stackrel{!}{=}0,\qquad
\frac{\partial \varepsilon_{{\Omega_{\perp}},{\Omega_{\parallel}}}^{(1)}}{\partial {\Omega_{\perp}}}\stackrel{!}{=}0,$$ which lead to the self-consistence equations $$\begin{aligned}
\label{st03}
\sqrt{{\Omega_{\parallel}}}&=&-\frac{2}{\sqrt{\pi}}\left({\rm ln}\,{\Omega_{\parallel}}-{\rm ln}\,{\Omega_{\perp}}+2-{\rm ln}\,2 \right),\\
\label{st04}
{\Omega_{\perp}}&=&2\sqrt{\frac{{\Omega_{\parallel}}}{\pi}}+B\sqrt{1+4\frac{{\Omega_{\parallel}}}{\pi B^2}}\end{aligned}$$ Let us first consider the last equation. Utilizing the second of the conditions (\[st00\]) we expand the second root around unity yielding the expression $$\label{st05}
{\Omega_{\perp}}=B+2\sqrt{\frac{{\Omega_{\parallel}}}{\pi}}+2 \frac{{\Omega_{\parallel}}}{\pi B}-4\frac{{\Omega_{\parallel}}^2}{\pi^2 B^3}+\ldots,$$ where the terms are sorted with regard to their contribution starting with the biggest. Since we are interested in the strong $B$ limit, we can obviously neglect terms suppressed by powers of $1/B$. Thus we only consider the following terms for the moment: $$\label{st06}
{\Omega_{\perp}}\approx B+2\sqrt{\frac{{\Omega_{\parallel}}}{\pi}}.$$ Inserting this into the other condition (\[st03\]), expanding the corresponding logarithm, and, once more, neglecting terms of order $1/B$, we find $$\label{st07}
\sqrt{{\Omega_{\parallel}}}\approx\frac{2}{\sqrt{\pi}}\left({\rm ln}\,B-{\rm ln}\,{\Omega_{\parallel}}^{(1)}+{\rm ln}\,2-2 \right).$$ To obtain a tractable approximation for ${\Omega_{\parallel}}$, we perform some iterations starting from $$\label{st08}
\sqrt{{\Omega_{\parallel}}^{(1)}}=\frac{2}{\sqrt{\pi}}{\rm ln}\,2Be^{-2}$$ Reinserting this on the right-hand side of Eq. (\[st07\]), one obtains the second iteration $\sqrt{{\Omega_{\parallel}}^{(2)}}$. We stop this procedure after an additional reinsertion which yields $$\label{st09}
\sqrt{{\Omega_{\parallel}}^{(3)}}=\frac{2}{\sqrt{\pi}}\left({\rm ln}\,2Be^{-2}-2{\rm ln}\left[\frac{2}{\sqrt{\pi}}
\left\{{\rm ln}\,2Be^{-2}-2{\rm ln}\,\left(\frac{2}{\sqrt{\pi}}{\rm ln}\,2Be^{-2}\right)
\right\}\right] \right).$$ The reader may convince himself that this iteration procedure indeed converges. For a subsequent systematical extraction of terms essentially contributing to the binding energy, the expression (\[st09\]) is not satisfactory. Therefore it is better to separate the leading term in the curly brackets and expand the logarithm of the remainder. Then this proceeding is applied to the expression in the angular brackets and so on. Neglecting terms of order ${\rm ln}^{-3} B$, we obtain $$\label{st10}
\sqrt{{\Omega_{\parallel}}^{(3)}}\approx \frac{2}{\sqrt{\pi}}\left({\rm ln}\,2Be^{-2}+{\rm ln}\frac{\pi}{4}
-2{\rm ln}{\rm ln}\,2Be^{-2}\right).$$ The double-logarithmic term can be expanded in a similar way as described above: $$\label{st11}
{\rm ln}{\rm ln}\,2Be^{-2}={\rm ln}\left[{\rm ln}\,B\left(1+\frac{{\rm ln}\,2-2}{{\rm ln}\,B}
\right) \right]= {\rm ln}{\rm ln}\,B+\frac{{\rm ln}\,2-2}{{\rm ln}\,B}-\frac{1}{2}
\frac{({\rm ln}\,2-2)^2}{{\rm ln}^2B}+{\cal O}({\rm ln}^{-3} B).$$ Thus the expression (\[st10\]) may be rewritten as $$\label{st12}
\sqrt{{\Omega_{\parallel}}^{(3)}}= \frac{2}{\sqrt{\pi}}\left( {\rm ln}\,B -2{\rm ln}{\rm ln}\,B+
\frac{2a}{{\rm ln}\,B}+\frac{a^2}{{\rm ln}^2B}+b \right) +{\cal O}({\rm ln}^{-3} B)$$ with abbreviations $$\label{st13}
a=2-{\rm ln}\,2 \approx 1.307,\qquad b= {\rm ln}\frac{\pi}{2}-2\approx -1.548.$$ The first observation is that the variational parameter ${\Omega_{\parallel}}$ is always much smaller than ${\Omega_{\perp}}$ in the high $B$-field limit. Thus we can further simplify the approximation (\[st06\]) by replacing $$\label{st14}
{\Omega_{\perp}}\approx B\left(1+\frac{2}{B}\sqrt{\frac{{\Omega_{\parallel}}}{\pi}}\right)\longrightarrow B$$ without affecting the following expression for the binding energy. Inserting the solutions (\[st12\]) and (\[st14\]) into the equation for the binding energy (\[st01\]) and expanding the logarithmic term once more as described, we find up to the order ${\rm ln}^{-2}B$: $$\begin{aligned}
\label{st15}
\varepsilon^{(1)}(B)&=&\frac{1}{\pi}\left({\rm ln}^2B -4\,{\rm ln}\,B\; {\rm ln}{\rm ln}\,B
+4\,{\rm ln}^2{\rm ln}\,B -4b\,{\rm ln}{\rm ln}\,B
+2(b+2)\,{\rm ln}\,B
+b^2
-\frac{1}{{\rm ln}\,B}\left[8\,{\rm ln}^2{\rm ln}\,B-8b\,{\rm ln}{\rm ln}\,B+2b^2
\right]\right)\nonumber\\
&&+{\cal O}({\rm ln}^{-2}B)\end{aligned}$$ Note that the prefactor $1/\pi$ of the leading ${\rm ln}^2B$-term differs from a value $1/2$ obtained by Landau and Lifschitz [@landau]. Our different value is a consequence of using a harmonic trial system. The calculation of higher orders in variational perturbation theory would improve the value of the prefactor.
At a magnetic field strength $B=10^5 B_0$, which corresponds to $2.35\times 10^{10}\,{\rm T}=
2.35\times10^{14}\,{\rm G}$, the contribution from the first six terms is $22.87\,[2\,{\rm Ryd}]$. The next three terms suppressed by a factor ${\rm ln}^{-1} B$ contribute $-2.29\,[2\,{\rm Ryd}]$, while an estimate for the ${\rm ln}^{-2} B$-terms yields nearly $-0.3\,[2\,{\rm Ryd}]$. Thus we find $$\label{st16}
\varepsilon^{(1)}(10^5)=20.58\pm 0.3\,[2\,{\rm Ryd}].$$ This is in very good agreement with the value $20.60\,[2\,{\rm Ryd}]$ obtained from the full treatment described in Sec. \[vptlow\].
Table \[asymp\] lists the values of the first six terms of Eq. (\[st15\]). This shows in particular the significance of the second-leading term $-(4/\pi){\rm ln}\,B\;{\rm ln}{\rm ln}\,B$, which is of the same order of the leading term $(1/\pi){\rm ln}^2B$ but with an opposite sign. In Fig. \[bofb\], we have plotted the expression $$\label{st17}
\varepsilon_L(B)=\frac{1}{2}\,{\rm ln}^2B$$ from Landau and Lifschitz [@landau] to illustrate that it gives far too large binding energies even at very large magnetic fields, e.g. at $2000 B_0\propto 10^{12}\,{\rm G}$.
This strength of magnetic field appears on surfaces of neutron stars ($10^{10}-10^{12}\,{\rm G}$). A recently discovered new type of neuton star is the so-called magnetar. In these, charged particles such as protons and electrons produced by decaying neutrons give rise to the giant magnetic field of $10^{15}\,{\rm G}$. Magnetic fields of white dwarfs reach only up to $10^6-10^8\,{\rm G}$. All these magnetic field strengths are far from realization in experiments. The strongest magnetic fields ever produced in a laboratory were only of the order $10^5\,{\rm G}$, an order of magnitude larger than the fields in sun spots which reach about $0.4\times10^4\,{\rm G}$. Recall, for comparison, that the earth’s magnetic field has the small value of $0.6\,{\rm G}$.
As we see in Fig. \[bofb\], the nonleading terms in Eq. (\[st15\]) give important contributions to the asymptotic behaviour even at such large magnetic fields. It is an unusual property of the asymptotic behaviour that the absolute value of the difference between the Landau-expression (\[st17\]) and our approximation (\[st15\]) diverges with increasing magnetic field strengths $B$, only the relative difference decreases.
$(1/\pi){\rm ln}^2B$ $-(4/\pi){\rm ln}\,B\;{\rm ln}{\rm ln}\,B$ $(4/\pi)\,{\rm ln}^2{\rm ln}\,B$ $-(4b/\pi)\,{\rm ln}{\rm ln}\,B$ $[2(b+2)/\pi]\,{\rm ln}\,B$ $b^2/\pi$
---------------------- -------------------------------------------- ---------------------------------- ---------------------------------- ----------------------------- -----------
$42.1912$ $-35.8181$ $7.6019$ $4.8173$ $3.3098$ $0.7632$
: \[asymp\] Example for the competing leading six terms in Eq. (\[st15\]) at $B=10^5B_0\approx 2.35\times 10^{14}\,{\rm G}$.
Summary
=======
We have calculated the effective classical potential for the hydrogen atom in a magnetic field. For this we have generalized variational perturbation theory to make it applicable to physical systems with uniform external magnetic field.
The effective classical potential containing the complete quantum statistical information of the system was determined in first-order variational perturbation theory. For zero-temperature, it gave the energy of the system. Our result consists of a single analytic expression which is quite accurate at all temperatures and magnetic field strengths.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Prof. J. Čížek and Dr. J. Weniger for useful hints and references of a perturbative treatment of the ground state properties of the hydrogen atom in a magnetic field. The authors also thank Dr. J. Ortner and M. Steinberg for discussions of the finite temperature behaviour of this system. For interesting discussions we would also like to thank Prof. J.T. Devreese and Prof. G. Wunner.
One of us (M.B.) is grateful for support by the Studienstiftung des deutschen Volkes.
Generating Functional for Particle in Magnetic Field and Harmonic Oscillator Potential {#appgen}
======================================================================================
For the determination of the correlation functions of a system, we need to know the solution of the two-dimensional generating functional in the presence of an external source ${\bf j}=(j_x,j_y)$: $$\label{app00}
Z^{{\bf x}_0}[{\bf j}]=\lambda_{\rm th}^2\oint {\cal D}^2x\,
\delta({\bf x}_0-\overline{{\bf x}(\tau)})\,
e^{-{\cal A}^{{\bf x}_0}[{\bf x};{\bf j}]/\hbar}.$$ The action of a particle in a magnetic field in $z$-direction and a harmonic oscillator reads $$\label{app01}
{\cal A}^{{\bf x}_0}[{\bf x};{\bf j}]=\int_0^{\hbar\beta}d\tau\,\left[
\frac{M}{2}\dot{\bf x}^2(\tau)-\frac{i}{2}M\omega([{\bf x}(\tau)-{\bf x}_0]
\times\dot{\bf x}(\tau))_z
+\frac{1}{2}M\Omega^2[{\bf x}(\tau)-{\bf x}_0]^2+{\bf j}(\tau)\cdot({\bf x}(\tau)-{\bf x}_0)
\right].$$ The position dependent terms are centered around ${\bf x}_0=(x_0,y_0)$, which is the temporal average of the path ${\bf x}(\tau)$, and thus equal to the zero frequency component of the Fourier path, is $$\label{app02}
{\bf x}(\tau)={\bf x}_0+\sum\limits_{m=1}^\infty\,\left({\bf x}_me^{i\omega_m\tau}+
{\bf x}^\star_me^{-i\omega_m\tau} \right)$$ with the Matsubara frequencies $\omega_m=2\pi m/\hbar\beta$ and complex Fourier coefficients ${\bf x}_m={\bf x}_m^{\rm re}+i{\bf x}_m^{\rm im}$. Introducing a similar Fourier decomposition for the current ${\bf j}(\tau)$ with Fourier components ${\bf j}_m$ and using the orthonormality relation $$\label{app03}
\frac{1}{\hbar\beta}\int_0^{\hbar\beta}d\tau\,e^{i(\omega_m-\omega_n)\tau}=\delta_{m\,n},$$ the generating functional can be written as $$\label{app04}
Z^{{\bf x}_0}[{\bf j}]=\prod\limits_{m=1}^\infty\left[\int\frac{dx_m^{\rm re}dx_m^{\rm im}
dy_m^{\rm re}dy_m^{\rm im}}{(\pi/M\beta\omega_m^2)^2}
e^{-{\cal A}_m({\bf x}_m,{\bf x}^\star_m;{\bf j}_m,{\bf j}^\star_m)/\hbar} \right]$$ with $$\begin{aligned}
\label{app05}
{\cal A}_m({\bf x}_m,{\bf x}^\star_m;{\bf j}_m,{\bf j}^\star_m)&=&
\hbar\beta M(\omega_m^2+\Omega^2)
([x_m^{\rm re}]^2+[x_m^{\rm im}]^2+[y_m^{\rm re}]^2+[y_m^{\rm im}]^2)
+2i\hbar\beta M \omega\omega_m(x_m^{\rm re}y_m^{\rm im}-x_m^{\rm im}y_m^{\rm re})\nonumber\\
&&+2\hbar\beta(x_m^{\rm re}{j_x}_m^{\rm re}+x_m^{\rm im}{j_x}_m^{\rm im}+
y_m^{\rm re}{j_y}_m^{\rm re}+y_m^{\rm im}{j_y}_m^{\rm im}).\end{aligned}$$ Expression (\[app04\]) is equivalent to the path integral (\[app00\]) and we obtain after performing the integrations and retransforming the currents $$\label{app06}
{\bf j}_m=\frac{1}{\hbar\beta}\int_0^{\hbar\beta}d\tau\,{\bf j}(\tau)e^{-i\omega_m\tau}$$ the resulting generating functional $$\label{app07}
Z^{{\bf x}_0}[{\bf j}]=Z^{{\bf x}_0}\exp\left\{\frac{1}{2\hbar^2}
\int_0^{\hbar\beta}d\tau \int_0^{\hbar\beta}d\tau'\,
{\bf j}(\tau){\bf G}^{{\bf x}_0}(\tau,\tau'){\bf j}(\tau') \right\}$$ with the partition function $$\label{app08}
Z^{{\bf x}_0}\equiv Z^{{\bf x}_0}[0]=
\prod\limits_{m=1}^\infty\,\frac{\omega_m^4}{\omega^2\omega_m^2+
(\omega_m^2+\Omega^2)^2}$$ and the $2\times 2$-matrix of Green functions $$\label{app09}
{\bf G}^{{\bf x}_0}(\tau,\tau')=\left(\begin{array}{cc}
G_{xx}^{{\bf x}_0}(\tau,\tau') & G_{xy}^{{\bf x}_0}(\tau,\tau')\\
G_{yx}^{{\bf x}_0}(\tau,\tau') & G_{yy}^{{\bf x}_0}(\tau,\tau') \end{array}\right).$$ The elements of this matrix are position-position correlation functions what can be easily proved by applying two functional derivatives with respect to the desired component of the current to the functional (\[app00\]), for example $$\label{app10}
G_{xx}^{{\bf x}_0}(\tau,\tau')={\left\langle\,(x(\tau)-x_0)\,(x(\tau')-x_0)\,\right\rangle^{{\bf x}_0}}
=\left[\hbar^2\frac{1}{Z^{{\bf x}_0}[{\bf j}]}
\frac{\delta^2}{\delta j_x(\tau)\delta j_x(\tau')}Z^{{\bf x}_0}[{\bf j}]\right]_{{\bf j}=0},$$ where we have defined expectation values by $$\label{app11}
{\left\langle\,\ldots\,\right\rangle^{{\bf x}_0}}=\frac{\lambda_{\rm th}^2}
{Z^{{\bf x}_0}}\oint {\cal D}^2x\,\ldots\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
e^{-{\cal A}^{{\bf x}_0}[{\bf x};0]/\hbar}.$$ From the above calculation we find the following expressions for the Green functions in Fourier space ($0\le\tau,\tau'\le\hbar\beta$): $$\begin{aligned}
\label{app12a}
G_{xx}^{{\bf x}_0}(\tau,\tau')&=&
{\left\langle\,{\tilde{x}}(\tau)\,{\tilde{x}}(\tau')\,\right\rangle^{{\bf x}_0}} = G_{yy}^{{\bf x}_0}(\tau,\tau')={\left\langle\,{\tilde{y}}(\tau)\,{\tilde{y}}(\tau')\,\right\rangle^{{\bf x}_0}}
\nonumber\\
&=&\frac{2}{M\beta}\sum\limits_{m=1}^\infty\,
\frac{\omega_m^2+\Omega^2}{\omega^2\omega_m^2+(\omega_m^2+\Omega^2)^2}\,
e^{-i\omega_m(\tau-\tau')},\\
\label{app12b}
G_{xy}^{{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{x}}(\tau)\,{\tilde{y}}(\tau')\,\right\rangle^{{\bf x}_0}}=
-G_{yx}^{{\bf x}_0}(\tau,\tau')=-{\left\langle\,{\tilde{y}}(\tau)\,{\tilde{x}}(\tau')\,\right\rangle^{{\bf x}_0}}\nonumber\\
&=&\frac{2\omega}{M\beta}\sum\limits_{m=1}^\infty\,
\frac{\omega_m}{\omega^2\omega_m^2+(\omega_m^2+\Omega^2)^2}\,
e^{-i\omega_m(\tau-\tau')},\end{aligned}$$ where, for simplicity, ${\tilde{\bf x}}(\tau)={\bf x}(\tau)-{\bf x}_0$. It is desirable to find analytical expressions for the Green functions and the partition function (\[app08\]). All these quantities possess the same dominator which can be decomposed as $$\label{app13}
\omega^2\omega_m^2+(\omega_m^2+\Omega^2)^2=(\omega_m^2+\Omega_+^2)(\omega_m^2+\Omega_-^2)$$ with frequencies $$\label{app14}
\Omega_\pm(\omega,\Omega)=
\sqrt{\Omega^2+\frac{1}{2}\omega^2\pm\omega\sqrt{\Omega^2+\frac{1}{4}\omega^2}}.$$ Therefore the partition function (\[app08\]) can be split into two products, each of which known from the harmonic oscillator [@PI Chap. 5]: $$\label{app15}
Z^{{\bf x}_0}=\prod\limits_{m=1}^\infty\left[\frac{\omega_m^2}{\omega_m^2+\Omega_+^2}\right]
\prod\limits_{m=1}^\infty\left[\frac{\omega_m^2}{\omega_m^2+\Omega_-^2}\right]
=\frac{\hbar\beta\Omega_+/2}{\sinh{\hbar\beta\Omega_+/2}}\,
\frac{\hbar\beta\Omega_-/2}{\sinh{\hbar\beta\Omega_-/2}}.$$ Now we apply the property (\[app13\]) to decompose the Green functions (\[app12a\]) into partial fractions, yielding $$\label{app16}
G_{xx}^{{\bf x}_0}(\tau,\tau')=G_{yy}^{{\bf x}_0}(\tau,\tau')=\frac{1}{M\beta}
\left(\alpha_1\,\sum\limits_{m=-\infty}^\infty\,\frac{1}{\omega_m^2+\Omega_+^2}\,
e^{-i\omega_m(\tau-\tau')}+\alpha_2\,\sum\limits_{m=-\infty}^\infty\,
\frac{1}{\omega_m^2+\Omega_-^2}\,
e^{-i\omega_m(\tau-\tau')}-\frac{1}{\Omega^2} \right)$$ with coefficients $$\label{app17}
\alpha_1=\frac{\Omega_+^2-\Omega^2}{\Omega_+^2-\Omega_-^2},\qquad
\alpha_2=-\frac{\Omega_-^2-\Omega^2}{\Omega_+^2-\Omega_-^2}.$$ Following Ref. [@PI Chap. 3], sums of the kind occuring in expression (\[app16\]) are spectral decompositions of the correlation function for the harmonic oscillator and can be summed up: $$\label{app18}
\sum\limits_{m=-\infty}^\infty\,\frac{1}{\omega_m^2+\Omega_\pm^2}\,e^{-i\omega_m(\tau-\tau')}=
\frac{\hbar\beta}{2\Omega_\pm}
\frac{\cosh{\Omega_\pm(|\tau-\tau'|-\hbar\beta/2)}}{\sinh{\hbar\beta\Omega_\pm/2}}.$$ Thus, the $xx$- and $yy$-correlation functions can be expressed by $$\begin{aligned}
\label{app19}
&&G_{xx}^{{\bf x}_0}(\tau,\tau')=G_{yy}^{{\bf x}_0}(\tau,\tau')\nonumber\\
&&\hspace*{5mm}=\frac{1}{M\beta}\left(\frac{\hbar\beta}{2\Omega_+}\,
\frac{\Omega_+^2-\Omega^2}{\Omega_+^2-\Omega_-^2}\,
\frac{\cosh{\Omega_+(|\tau-\tau'|-\hbar\beta/2)}}{\sinh{\hbar\beta\Omega_+/2}}
-\frac{\hbar\beta}{2\Omega_-}\,
\frac{\Omega_-^2-\Omega^2}{\Omega_+^2-\Omega_-^2}\,
\frac{\cosh{\Omega_-(|\tau-\tau'|-\hbar\beta/2)}}{\sinh{\hbar\beta\Omega_-/2}}
-\frac{1}{\Omega^2}
\right),\end{aligned}$$ where, from Eq. (\[app14\]), $\Omega_\pm=\Omega_\pm(\omega,\Omega)$ are functions of the original frequencies $\omega$ from the magnetic field and $\Omega$ from the additional harmonic oscillator (\[app01\]). It is obvious that expression (\[app19\]) reduces to the Green function of the harmonic oscillator for $\omega\to 0$: $$\label{app19a}
\lim_{\omega\to 0} G_{ii}^{{\bf x}_0}(\tau,\tau')=\frac{1}{M\beta\Omega^2}\left(
\frac{\hbar\beta\Omega}{2}\,\frac{\cosh{\Omega(|\tau-\tau'|-\hbar\beta/2)}}
{\sinh{\hbar\beta\Omega/2}}-1\right)$$ with $i\in\{x,y\}$. In this limit, the partition function (\[app15\]) turns out to be the usual one [@PI Chap. 5] for such a harmonic oscillator $$\label{app19aa}
\lim_{\omega\to 0} Z^{{\bf x}_0}=\frac{\hbar\beta\Omega/2}{\sinh{\hbar\beta\Omega/2}}.$$ It is worth mentioning that with the last term in Green function (\[app19\]) the classical harmonic fluctuation width $$\label{app19b}
G^{\rm cl}_{xx}=\left\langle x^2\right\rangle^{\rm cl}=\frac{1}{M\beta\Omega^2}$$ is subtracted. This is the consequence of the exclusion of the zero frequency mode of the Fourier path (\[app02\]) in the generating functional (\[app00\]). The necessity to do this has already been discussed in Sect. \[effrep\]. The other terms in Eq. (\[app19\]) are those which we would have obtained [*without*]{} separation of the $x_0$-component. Thus these terms represent the quantum mechanical Green function containing all quantum as well as thermal fluctuations. It is a nice property of all Green functions discussed in this paper that $$\label{app19c}
G^{{\bf x}_0}_{xx}(\tau,\tau')=G^{\rm qm}_{xx}(\tau,\tau')-G^{\rm cl}_{xx}.$$ Such a relation exists for all other Green functions appropriately, including momentum-position correlations which we consider subsequently.
The knowledge of relation (\[app18\]) makes it quite easy to determine the algebraic expression for the mixed $xy$-correlation functions. Rewriting Eq. (\[app12b\]) as $$\label{app20}
G_{xy}^{{\bf x}_0}(\tau,\tau')=-G_{yx}^{{\bf x}_0}(\tau,\tau')
= \frac{i\omega}{M\beta}\,\frac{1}{\Omega_+^2-\Omega_-^2}\frac{\partial}{\partial \tau}
\left(\sum\limits_{m=-\infty}^\infty\,\frac{1}{\omega_m^2+\Omega_+^2}\,e^{-i\omega_m(\tau-\tau')}
+\sum\limits_{m=-\infty}^\infty\,\frac{1}{\omega_m^2+\Omega_-^2}\,e^{-i\omega_m(\tau-\tau')}
\right)$$ and applying the derivative with respect to $\tau$ to relation (\[app18\]), we obtain the following expression for the mixed Green function: $$\begin{aligned}
\label{app21}
G_{xy}^{{\bf x}_0}(\tau,\tau')&=&-G_{yx}^{{\bf x}_0}(\tau,\tau')\nonumber\\
&=&\frac{\hbar\omega}{2iM}\,\frac{1}{\Omega_+^2-\Omega_-^2}\left\{
\Theta(\tau-\tau')[g_{\Omega_+}(\tau,\tau')-g_{\Omega_-}(\tau,\tau')]-
\Theta(\tau'-\tau)[g_{\Omega_+}(\tau',\tau)-g_{\Omega_-}(\tau',\tau)]
\right\},\end{aligned}$$ where we have used the abbreviation $$\label{app22}
g_{\Omega_\pm}(\tau,\tau')=\frac{\sinh{\Omega_\pm(\tau-\tau'-\hbar\beta/2)}}
{\sinh{\hbar\beta\Omega_\pm/2}},\qquad \tau,\tau'\in(0,\hbar\beta).$$ Note that classically $\left\langle xy\right\rangle^{\rm cl}=0$ such that Eq. (\[app19c\]) reduces to $$\label{app22a}
G_{xy}^{{\bf x}_0}(\tau,\tau')=G_{xy}^{\rm qm}(\tau,\tau').$$ The Heaviside function in Eq. (\[app21\]) is defined symmetrically: $$\label{app22b}
\Theta(\tau-\tau')=\left\{\begin{array}{cc} 1 & \quad\tau>\tau',\\ 1/2 & \quad\tau=\tau', \\
0 & \quad\tau<\tau'. \end{array}\right.$$ In the quantum mechanical limit of zero-temperature ($\beta\to \infty$), the Green function (\[app19\]) simplifies to $$\label{app23a}
\lim_{\beta\to\infty} G_{xx}^{{\bf x}_0}(\tau,\tau')=
\lim_{\beta\to\infty} G_{yy}^{{\bf x}_0}(\tau,\tau')=\frac{\hbar}{2M}
\left(\frac{1}{\Omega_+}\,\frac{\Omega_+^2-\Omega^2}{\Omega_+^2-\Omega_-^2}\,
e^{-\Omega_+|\tau-\tau'|}-\frac{1}{\Omega_-}\,\frac{\Omega_-^2-\Omega^2}{\Omega_+^2-\Omega_-^2}\,
e^{-\Omega_-|\tau-\tau'|}\right),$$ while in Eq. (\[app21\]) only $g_{\Omega_\pm}(\tau,\tau')$ changes: $$\label{app23b}
\lim_{\beta\to\infty} g_{\Omega_\pm}(\tau,\tau')=-e^{-\Omega_\pm(\tau-\tau')}.$$
Properties of Green Functions {#appprop}
=============================
In this section we list properties of the Green functions (\[app19\]) and (\[app21\]) which are important for the forthcoming consideration of the generating functional with sources coupling linearily to position or momentum in Appendix \[appgenrp\]. For all relations we suppose that $0\le\tau,\tau'\le\hbar\beta$.
General Properties
------------------
A first observation is the temporal translational invariance of the Green functions: $$\label{prop00}
G_{ij}^{{\bf x}_0}(\tau,\tau')=G_{ij}^{{\bf x}_0}(\tau-\tau'),$$ where each of the indices $i,j$ stands for $x$ or $y$, respectively. For equal times we find $$\label{prop01}
G_{ij}^{{\bf x}_0}(\tau,\tau)=\frac{1}{M\beta}
\left(\frac{\hbar\beta}{2\Omega_+}\,\frac{\Omega_+^2-\Omega^2}{\Omega_+^2-\Omega_-^2}
\coth{\hbar\beta\Omega_+/2}-\frac{\hbar\beta}{2\Omega_-}\,
\frac{\Omega_-^2-\Omega^2}{\Omega_+^2-\Omega_-^2} \coth{\hbar\beta\Omega_-/2}-\frac{1}{\Omega^2}
\right)\times
\left\{\begin{array}{cc}
1 & \quad i=j,\\
0 & \quad i\neq j.
\end{array}\right.$$ Moreover we read off the following symmetries from the expressions (\[app19\]) and (\[app21\]): $$\label{prop02}
G_{ij}^{{\bf x}_0}(\tau,\tau')=G_{ij}^{{\bf x}_0}(\tau',\tau)\times\left\{\begin{array}{cc}
1 & \quad i=j,\\
-1 & \quad i\neq j.
\end{array}\right.$$ Otherwise, $$\label{prop03}
G_{ij}^{{\bf x}_0}(\tau,\tau')=G_{ji}^{{\bf x}_0}(\tau',\tau).$$ Throughout the paper we always use periodic paths. Hence it is obvious that all Green functions are periodic, too: $$\label{prop04}
G_{ij}^{{\bf x}_0}(0,\tau')=G_{ij}^{{\bf x}_0}(\hbar\beta,\tau'),\qquad
G_{ij}^{{\bf x}_0}(\tau,0)=G_{ij}^{{\bf x}_0}(\tau,\hbar\beta).$$
Derivatives of Green Functions
------------------------------
We now proceed with derivatives of the Green functions (\[app19\]) and (\[app21\]), since these are essential for the derivation of the generating functional of position and momentum dependent correlations in the forthcoming Appendix \[appgenrp\].
Before considering the concrete expressions we introduce a new symbol indicating uniquely to which argument the derivative is applied. A dot on the left-hand side means to perform the derivative with respect to the first argument and the dot on the right-hand side indicates that to differentiate with respect to the other argument. Having a dot on both sides the Green function is derived with respect to both arguments: $$\label{der00}
{{^\bullet G}}_{ij}^{{\bf x}_0}(\tau,\tau')=
\frac{\partial G_{ij}^{{\bf x}_0}(\tau,\tau')}{\partial \tau},\quad
{{G^\bullet}}_{ij}^{{\bf x}_0}(\tau,\tau')=
\frac{\partial G_{ij}^{{\bf x}_0}(\tau,\tau')}{\partial \tau'},\quad
{{^\bullet G^\bullet}}_{ij}^{{\bf x}_0}(\tau,\tau')=
\frac{\partial^2 G_{ij}^{{\bf x}_0}(\tau,\tau')}{\partial \tau \partial \tau'}.$$ Applying such derivatives to the Green functions (\[app19\]), we obtain ($i\in \{x,y\}$): $$\label{der01}
{{^\bullet G}}_{ii}^{{\bf x}_0}(\tau,\tau')=\frac{\hbar}{2M}\frac{1}{\Omega_+^2-\Omega_-^2}
\left[\Theta(\tau-\tau')g(\tau,\tau')
-\Theta(\tau'-\tau)g(\tau',\tau)\right]=-{{G^\bullet}}_{ii}^{{\bf x}_0}(\tau,\tau')$$ with $$\label{der02}
g(\tau,\tau')=(\Omega_+^2-\Omega^2)g_{\Omega_+}(\tau,\tau')
-(\Omega_-^2-\Omega^2)g_{\Omega_-}(\tau,\tau'),$$ where $g_\pm(\tau,\tau)$ was defined in Eq. (\[app22\]). Performing the derivatives to both arguments leads to the expression $$\label{der03}
{{^\bullet G^\bullet}}_{ii}^{{\bf x}_0}(\tau,\tau')={{^\bullet \tilde{G}^\bullet}}_{ii}^{{\bf x}_0}(\tau,\tau')+
\frac{\hbar}{M}\delta(\tau-\tau'),$$ where we have introduced the partial function $$\label{der04}
{{^\bullet \tilde{G}^\bullet}}_{ii}^{{\bf x}_0}(\tau,\tau')=-\frac{\hbar}{2M}\left[\Omega_+
\frac{\Omega_+^2-\Omega^2}{\Omega_+^2-\Omega_-^2}
\frac{\sinh{\Omega_+(|\tau-\tau'|-\hbar\beta/2)}}{\sinh{\hbar\beta\Omega_+/2}}-
\Omega_-
\frac{\Omega_-^2-\Omega^2}{\Omega_+^2-\Omega_-^2}
\frac{\sinh{\Omega_-(|\tau-\tau'|-\hbar\beta/2)}}{\sinh{\hbar\beta\Omega_-/2}} \right]$$ which is finite for equal times.
Applying derivatives with respect to the first respective second argument to the mixed correlation function (\[app21\]), we find: $$\label{der05}
{{^\bullet G}}_{xy}^{{\bf x}_0}(\tau,\tau')=\frac{\hbar\omega}{2iM}\frac{1}{\Omega_+^2-\Omega_-^2}
\left[\Omega_+\frac{\cosh{\Omega_+(|\tau-\tau'|-\hbar\beta/2)}}{\sinh{\hbar\beta\Omega_+/2}}
-\Omega_-\frac{\cosh{\Omega_-(|\tau-\tau'|-\hbar\beta/2)}}{\sinh{\hbar\beta\Omega_-/2}}\right]
=-{{G^\bullet}}_{xy}^{{\bf x}_0}(\tau,\tau')$$ and $$\label{der05b}
{{^\bullet G}}_{yx}^{{\bf x}_0}(\tau,\tau')=-{{^\bullet G}}_{xy}^{{\bf x}_0}(\tau,\tau').$$ Differentiating each argument of the mixed Green function results in $$\label{der06}
{{^\bullet G^\bullet}}_{xy}^{{\bf x}_0}(\tau,\tau')=\frac{i\hbar\omega}{2M}\frac{1}{\Omega_+^2-\Omega_-^2}
\left[\Theta(\tau-\tau')h(\tau,\tau')-\Theta(\tau'-\tau)h(\tau',\tau) \right]=
-{{^\bullet G^\bullet}}_{yx}^{{\bf x}_0}(\tau,\tau')$$ with $$\label{der07}
h(\tau,\tau')=\Omega_+^2g_{\Omega_+}(\tau,\tau')-\Omega_-^2g_{\Omega_-}(\tau,\tau').$$ An additional property we read off from Eqs. (\[der01\]) and (\[der05\]) is ($i,j\in \{x,y\}$): $$\begin{aligned}
\label{der08a}
{{^\bullet G}}_{ij}^{{\bf x}_0}(\tau,\tau')={{^\bullet G}}_{ij}^{{\bf x}_0}(\tau',\tau)\times
\left\{\begin{array}{cc}
-1 &\quad i=j,\\
1 & \quad i\neq j,
\end{array}\right.\\
\label{der08b}
{{G^\bullet}}_{ij}^{{\bf x}_0}(\tau,\tau')={{G^\bullet}}_{ij}^{{\bf x}_0}(\tau',\tau)\times
\left\{\begin{array}{cc}
-1 &\quad i=j,\\
1 & \quad i\neq j.
\end{array}\right.\end{aligned}$$ The double-sided derivatives (\[der03\]), (\[der04\]), and (\[der06\]) imply $$\label{der09}
{{^\bullet G^\bullet}}_{ij}^{{\bf x}_0}(\tau,\tau')={{^\bullet G^\bullet}}_{ij}^{{\bf x}_0}(\tau',\tau)\times
\left\{\begin{array}{cc}
1 &\quad i=j,\\
-1 & \quad i\neq j.
\end{array}\right.$$ The derivatives (\[der01\]), (\[der04\]), (\[der05\]), and (\[der06\]) are periodic: $$\begin{aligned}
\label{der10a}
&&{{^\bullet G}}_{ij}^{{\bf x}_0}(\tau,0)={{^\bullet G}}_{ij}^{{\bf x}_0}(\tau,\hbar\beta),\quad
{{^\bullet G}}_{ij}^{{\bf x}_0}(0,\tau')={{^\bullet G}}_{ij}^{{\bf x}_0}(\hbar\beta,\tau'),\\
\label{der10b}
&&{{G^\bullet}}_{ij}^{{\bf x}_0}(\tau,0)={{G^\bullet}}_{ij}^{{\bf x}_0}(\tau,\hbar\beta),\quad
{{G^\bullet}}_{ij}^{{\bf x}_0}(0,\tau')={{G^\bullet}}_{ij}^{{\bf x}_0}(\hbar\beta,\tau'),\\
\label{der10c}
&&{{^\bullet \tilde{G}^\bullet}}_{ii}^{{\bf x}_0}(\tau,0)={{^\bullet \tilde{G}^\bullet}}_{ii}^{{\bf x}_0}(\tau,\hbar\beta),\quad
{{^\bullet \tilde{G}^\bullet}}_{ii}^{{\bf x}_0}(0,\tau')={{^\bullet \tilde{G}^\bullet}}_{ii}^{{\bf x}_0}(\hbar\beta,\tau'),\\
\label{der10d}
&&{{^\bullet G^\bullet}}_{ij}^{{\bf x}_0}(\tau,0)={{^\bullet G^\bullet}}_{ij}^{{\bf x}_0}(\tau,\hbar\beta),\quad
{{^\bullet G^\bullet}}_{ij}^{{\bf x}_0}(0,\tau')={{^\bullet G^\bullet}}_{ij}^{{\bf x}_0}(\hbar\beta,\tau'),\qquad (i\neq j).\end{aligned}$$
Generating Functional for Position- and Momentum-Dependent Correlation Functions {#appgenrp}
================================================================================
With the discussion of the generating functional for position-dependent correlation functions and, in particular, the Green functions in Appendix \[appgen\] and their properties in Appendix \[appprop\], we have layed the foundation to derive the generating functional for correlation functions depending on both, position and momentum. Following the framework presented in an earlier work [@correlation], such a functional involving sources coupled to the momentum can always be reduced to one containing position-coupled sources only.
We start from the three-dimensional effective classical representation for the generating functional $$\label{xp00}
Z_{\bf \Omega}[{\bf j},{\bf v}]=
\int\frac{d^3x_0d^3p_0}{(2\pi\hbar)^3}Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j},{\bf v}]$$ with zero frequency components ${\bf x}_0=(x_0,y_0,z_0)={\rm const.}$ and ${\bf p}_0=({p_x}_0,{p_y}_0,{p_y}_0)={\rm const.}$ of the Fourier path separated. The reduced functional is $$\label{xp01}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j},{\bf v}]=
(2\pi\hbar)^3\oint {\cal D'}^3x{\cal D}^3p\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\delta({\bf p}_0-\overline{{\bf p}(\tau)})\,
\exp\left\{-\frac{1}{\hbar}{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
[{\bf p},{\bf x};{\bf j},{\bf v}] \right\},$$ where the path integral measure is that defined in Eq. (\[ham03\]). Extending the action (\[ham02\]) by source terms, considering a more general Hamilton function than (\[ham16\]), and introducing an additional harmonic oscillator in $z$-direction, the action functional in Eq. (\[xp01\]) shall read $$\begin{aligned}
\label{xp02}
{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf p},{\bf x};{\bf j},{\bf v}]=
\int_0^{\hbar\beta}d\tau\,\Big\{&&-i{\tilde{\bf p}}(\tau)\cdot \dot{\bf x}(\tau)+
\frac{1}{2M}{\tilde{\bf p}}^2(\tau)-\frac{1}{2}{\Omega_{\perp 1}}l_z({\tilde{\bf p}},{\tilde{\bf x}})
+\frac{1}{8}M{\Omega_{\perp 2}}^2\left[
{\tilde{x}}^2(\tau)+{\tilde{y}}^2(\tau)\right]+\frac{1}{2}M{\Omega_{\parallel}}^2{\tilde{z}}^2(\tau)\nonumber\\
&&+{\bf j}(\tau)\cdot{\tilde{\bf x}}(\tau)+{\bf v}(\tau)\cdot{\tilde{\bf p}}(\tau)
\Big\}\end{aligned}$$ with shifted positions and momenta $$\label{xp03}
{\tilde{\bf x}}={\bf x}(\tau)-{\bf x}_0,\qquad {\tilde{\bf p}}={\bf p}(\tau)-{\bf p}_0.$$ The orbital angular momentum $l_z({\bf p},{\bf x})$ is defined in Eq. (\[ham18\]) and is used in Eq. (\[xp02\]) with the shifted phase space coordinates (\[xp03\]). We have introduced three different frequencies in (\[xp02\]), ${\bf \Omega}=({\Omega_{\perp 1}},{\Omega_{\perp 2}},{\Omega_{\parallel}})$, where the first both components are used in regard to the oscillations in the plane perpendicular to the direction of the magnetic field which shall be considered here to point into $z$-direction. The last component, ${\Omega_{\parallel}}$, is the frequency of a trial oscillator parallel to the field lines.
Due to the periodicity of the paths, we suppose that the sources might also be periodic: $$\label{xp03b}
{\bf j}(0)={\bf j}(\hbar\beta),\qquad {\bf v}(0)={\bf v}(\hbar\beta).$$ Since we want to simplify expression (\[xp01\]) such that we can use the results obtained in Appendix \[appgen\], the momentum path integral is solved in the following. In a first step we reexpress the momentum $\delta$-function in (\[xp01\]) by $$\label{xp04}
\delta({\bf p}_0-\overline{{\bf p}(\tau)})=\int\frac{d^3\xi}{(2\pi\hbar)^3}\,
\exp\left\{-\frac{1}{\hbar}\int_0^{\hbar\beta} d\tau \,{\bf v}_0\cdot[{\bf p}(\tau)-{\bf p}_0]
\right\},$$ where $$\label{xp05}
{\bf v}_0({\mbox{\boldmath$\xi$}})=\frac{i}{\hbar\beta}{\mbox{\boldmath$\xi$}}$$ is an additional current which is coupled to the momentum and is constant in time. Defining the sum of all sources coupled to the momentum by $$\label{xp06}
{\bf V}({\mbox{\boldmath$\xi$}},\tau)={\bf v}(\tau)+{\bf v}_0({\mbox{\boldmath$\xi$}}),$$ the functional (\[xp01\]) can be written as $$\begin{aligned}
\label{xp07}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j},{\bf v}]&=&
\int d^3\xi\oint {\cal D'}^3x{\cal D}^3p\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\exp\Bigg\{-\frac{1}{\hbar}\int_0^{\hbar\beta}d\tau\,\Big[-i{\bf p}(\tau)
\cdot\dot{\bf x}(\tau)
+\frac{{\bf p}^2(\tau)}{2M}-\frac{1}{2}{\Omega_{\perp 1}}l_z({\bf p}(\tau),{\tilde{\bf x}}(\tau))\nonumber\\
&&+\frac{1}{8}M{\Omega_{\perp 2}}^2\left\{{\tilde{x}}^2(\tau)+{\tilde{y}}^2(\tau)\right\}+\frac{1}{2}M{\Omega_{\parallel}}^2{\tilde{z}}^2(\tau)
+{\bf j}(\tau)\cdot{\tilde{\bf x}}(\tau)+{\bf V}({\mbox{\boldmath$\xi$}},\tau)\cdot{\bf p}(\tau)
\Big] \Bigg\},\end{aligned}$$ where we have used the translation invariance ${\tilde{\bf p}}\to {\bf p}$ of the path integral. To solve the momentum path integral, it is useful to express it in its discretized form. Performing quadratic completions such that the momentum path integral separates into an infinite product of simple Gaussian integrals which are easily calculated, the remaining functional is reduced to the configuration space path integral $$\begin{aligned}
\label{xp08}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j},{\bf v}]=
\int d^3\xi\,\exp\left[\frac{M}{2\hbar}\int_0^{\hbar\beta}d\tau\,{\bf V}^2({\mbox{\boldmath$\xi$}},\tau)
\right]
\oint{\cal D}^3x\,\delta({\bf x}_0-\overline{{\bf x}(\tau)})
\exp\left\{-\frac{1}{\hbar}{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
[{\bf x};{\bf j},{\bf V}]\right\}\end{aligned}$$ with the measure (\[ham09\]) for $D=3$. The action functional is $$\begin{aligned}
\label{xp09}
{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf x};{\bf j},{\bf V}]&=&
\int_0^{\hbar\beta}d\tau\,\Bigg[\frac{M}{2}\dot{\bf x}^2(\tau)
+\frac{1}{2}iM{\Omega_{\perp 1}}\left\{\dot{x}(\tau){\tilde{y}}(\tau)-\dot{y}(\tau){\tilde{x}}(\tau)\right\}
+\frac{1}{8}M\left({\Omega_{\perp 2}}^2-{\Omega_{\perp 1}}^2\right)\left\{{\tilde{x}}^2(\tau)+{\tilde{y}}^2(\tau)\right\}\nonumber\\
&&\hspace*{-3pt}
+\frac{1}{2}M{\Omega_{\parallel}}^2{\tilde{z}}^2(\tau)+{\tilde{x}}(\tau)\left[j_x(\tau)+\frac{1}{2}M{\Omega_{\perp 1}}V_y({\mbox{\boldmath$\xi$}},\tau)
\right]\nonumber\\
&&\hspace*{-3pt}+{\tilde{y}}(\tau)\left[j_y(\tau)-\frac{1}{2}M{\Omega_{\perp 1}}V_x({\mbox{\boldmath$\xi$}},\tau)\right]+{\tilde{z}}(\tau)j_z(\tau)
\Bigg]
-\frac{iM}{\hbar}\int_0^{\hbar\beta}d\tau\,\dot{\bf x}(\tau)\cdot
{\bf V}({\mbox{\boldmath$\xi$}},\tau),\end{aligned}$$ where the last term simplifies by the following consideration. A partial integration of this term yields $$\label{xp10}
\int_0^{\hbar\beta}d\tau\,\dot{\bf x}(\tau)\cdot{\bf V}({\mbox{\boldmath$\xi$}},\tau)
=-\int_0^{\hbar\beta}d\tau\,({\bf x}(\tau)-{\bf x}_0)\cdot\dot{\bf V}({\mbox{\boldmath$\xi$}},\tau).$$ The surface term vanishes as a consequence of the periodicity of the path and the source. This periodicity is also the reason why we could shift ${\bf x}(\tau)$ by the constant ${\bf x}_0$ on the right-hand side of Eq. (\[xp10\]). Obviously, the importance of this expression lies in the coupling of the time derivative of ${\bf V}({\mbox{\boldmath$\xi$}},\tau)$ to the path ${\bf x}(\tau)$. Thus, $\dot{\bf V}({\mbox{\boldmath$\xi$}},\tau)$ can be handled like a ${\bf j}(\tau)$-current [@correlation] and the action (\[xp09\]) can be written as $$\label{xp11}
{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf x};{\bf j},{\bf V}]=
{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf x};{\bf J},0]=
{\cal A}^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf x};0,0]-\frac{1}{\hbar}
\int_0^{\hbar\beta}d\tau\,{\tilde{\bf x}}(\tau)\cdot{\bf J}({\mbox{\boldmath$\xi$}},\tau)$$ with the new current vector ${\bf J}({\mbox{\boldmath$\xi$}},\tau)$ which has the components $$\begin{aligned}
\label{xp12}
J_x({\mbox{\boldmath$\xi$}},\tau)&=&j_x(\tau)+\frac{1}{2}M{\Omega_{\perp 1}}V_y({\mbox{\boldmath$\xi$}},\tau)-
iM\dot{V}_x({\mbox{\boldmath$\xi$}},\tau),\nonumber\\
J_y({\mbox{\boldmath$\xi$}},\tau)&=&j_y(\tau)-\frac{1}{2}M{\Omega_{\perp 1}}V_x({\mbox{\boldmath$\xi$}},\tau)-
iM\dot{V}_y({\mbox{\boldmath$\xi$}},\tau),\\
J_z({\mbox{\boldmath$\xi$}},\tau)&=&j_z(\tau)-\frac{1}{2}M{\Omega_{\parallel}}V_z({\mbox{\boldmath$\xi$}},\tau)\nonumber.\end{aligned}$$ and couples to the path ${\bf x}(\tau)$ only. With the expression (\[xp08\]) for the generating functional and the action (\[xp11\]), we have derived a representation similar to Eq. (\[app00\]) with the action (\[app01\]), extended by an additional oscillator in $z$-direction. We identify $$\label{xp13}
\omega\equiv {\Omega_{\perp 1}},\quad \Omega\equiv \frac{1}{4}\left({\Omega_{\perp 2}}^2-{\Omega_{\perp 1}}^2\right),
\quad j_x\equiv J_x,\quad
j_y\equiv J_y.$$ Thus the auxiliary frequencies $\Omega_{\pm}$ (\[app14\]) become $$\label{xp13b}
\Omega_{\pm}({\Omega_{\perp 1}},{\Omega_{\perp 2}})=\frac{1}{2}\,|{\Omega_{\perp 1}}\pm{\Omega_{\perp 2}}|.$$ Inserting the substitutions (\[xp13\]) into the solution (\[app07\]) for the generating functional in two dimensions and performing the usual calculation for a harmonic oscillator with external source [@PI Chaps. 3,5] in $z$-direction, we obtain an intermediate result for the generating functional in three dimensions (\[xp01\]): $$\label{xp14}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j},{\bf v}]=
\lambda_{\rm th}^{-3}\,Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
\int d^3\xi\,\exp\left\{\frac{M}{2\hbar}\int_0^{\hbar\beta}d\tau\,{\bf V}^2({\mbox{\boldmath$\xi$}},\tau)
\right\}\,\exp\left\{\frac{1}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
{\bf J}({\mbox{\boldmath$\xi$}},\tau)\,{\bf G}^{{\bf x}_0}(\tau,\tau')\,{\bf J}({\mbox{\boldmath$\xi$}},\tau')\right\}.$$ The partition function follows from Eqs. (\[app15\]) and (\[app19aa\]) $$\label{xp15}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}=Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[0,0]=
\frac{\hbar\beta\Omega_+/2}{\sinh{\hbar\beta\Omega_+/2}}\,
\frac{\hbar\beta\Omega_-/2}{\sinh{\hbar\beta\Omega_-/2}}\,
\frac{\hbar\beta{\Omega_{\parallel}}/2}{\sinh{\hbar\beta{\Omega_{\parallel}}/2}}$$ and ${\bf G}^{{\bf x}_0}(\tau,\tau')$ is the $3\times 3$-matrix of Green functions $$\label{xp16}
{\bf G}^{{\bf x}_0}(\tau,\tau')=\left(\begin{array}{ccc}
G_{xx}^{{\bf x}_0}(\tau,\tau') & G_{xy}^{{\bf x}_0}(\tau,\tau') & 0\\
G_{yx}^{{\bf x}_0}(\tau,\tau') & G_{yy}^{{\bf x}_0}(\tau,\tau') & 0\\
0 & 0 & G_{zz}^{{\bf x}_0}(\tau,\tau')\end{array}\right).$$ Except $G_{zz}^{{\bf x}_0}(\tau,\tau')$, the Green functions are given by the expressions in Eqs. (\[app19\]) and (\[app21\]) with frequencies (\[xp13b\]). The Green function of the pure harmonic oscillator in $z$-direction $$\label{xp17}
G_{zz}^{{\bf x}_0}(\tau,\tau')=\frac{1}{M\beta{\Omega_{\parallel}}^2}\left(
\frac{\hbar\beta{\Omega_{\parallel}}}{2}\,\frac{\cosh{{\Omega_{\parallel}}(|\tau-\tau'|-\hbar\beta/2)}}
{\sinh{\hbar\beta{\Omega_{\parallel}}/2}}-1\right)$$ follows directly from the limit (\[app19a\]). Since the current ${\bf J}$ (\[xp12\]) still depends on time derivatives of ${\bf V}$, we have to perform some partial integrations in the functional (\[xp14\]). This is a very extensive but straightforward work and thus we only present an instructive example. For that we apply the properties and the time derivatives of the Green functions which we presented in Appendix \[appprop\]. Consider the integral $$\label{xp18}
I=-\frac{M^2}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,\dot{V}_i
({\mbox{\boldmath$\xi$}},\tau)\,G_{ii}^{{\bf x}_0}(\tau,\tau')\,\dot{V}_i({\mbox{\boldmath$\xi$}},\tau')$$ occuring in the second exponential of Eq. (\[xp14\]) with $i\in\{x,y,z\}$. A partial integration in the $\tau'$-integral leads to $$\begin{aligned}
\label{xp18a}
I&=&-\frac{M^2}{2\hbar^2}\int_0^{\hbar\beta}d\tau\,\dot{V}_i({\mbox{\boldmath$\xi$}},\tau)
\left(G_{ii}^{{\bf x}_0}(\tau,\tau')\,V_i({\mbox{\boldmath$\xi$}},\tau')\Big|_{\tau'=0}^{\tau'=\hbar\beta}
-\int_0^{\hbar\beta}d\tau'\,\frac{\partial G_{ii}^{{\bf x}_0}(\tau,\tau')}{\partial \tau'}
\,V_i({\mbox{\boldmath$\xi$}},\tau')\right)\nonumber\\
&=&\frac{M^2}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
\dot{V}_i({\mbox{\boldmath$\xi$}},\tau)\,{{G^\bullet}}_{ii}^{{\bf x}_0}(\tau,\tau')\,V_i({\mbox{\boldmath$\xi$}},\tau').\end{aligned}$$ The surface term in the first line vanishes as a consequence of the periodicity of the current (\[xp03b\]) and the Green function (\[prop04\]). A second partial integration, now in the $\tau$-integral, results in $$\begin{aligned}
\label{xp18b}
I&=&-\frac{M^2}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
V_i({\mbox{\boldmath$\xi$}},\tau)\,{{^\bullet G^\bullet}}_{ii}^{{\bf x}_0}(\tau,\tau')\,V_i({\mbox{\boldmath$\xi$}},\tau')\nonumber\\
&=&-\frac{M^2}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
V_i({\mbox{\boldmath$\xi$}},\tau)\,{{^\bullet \tilde{G}^\bullet}}_{ii}^{{\bf x}_0}(\tau,\tau')\,V_i({\mbox{\boldmath$\xi$}},\tau')-
\frac{M}{2\hbar}\int_0^{\hbar}d\tau\,V_i^2({\mbox{\boldmath$\xi$}},\tau).\end{aligned}$$ Here we have applied the periodicity property of the right-hand derivative of the Green function (\[der10b\]), leading to a vanishing surface term in this case, too. In the second line, we have used the decomposition (\[der03\]) of the double-sided differentiated Green function. Note that the last term just cancels the appropriate term in the first exponential of the right-hand side of Eq. (\[xp14\]). Eventually, after performing all such partial integrations, we reexpress Eq. (\[xp14\]) by $$\label{xp19}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j},{\bf v}]=
\lambda_{\rm th}^{-3}\,Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
\int d^3\xi\,\exp\left\{\frac{1}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
\tilde{\bf s}({\mbox{\boldmath$\xi$}},\tau)\,{\bf H}^{{\bf x}_0}(\tau,\tau')\,\tilde{\bf s}({\mbox{\boldmath$\xi$}},\tau')
\right\}$$ with six-dimensional sources $$\label{xp20}
\tilde{\bf s}({\mbox{\boldmath$\xi$}},\tau)=\left(\,{\bf j}(\tau),{\bf V}({\mbox{\boldmath$\xi$}},\tau)\,\right).$$ and the $6\times 6$-matrix ${\bf H}^{{\bf x}_0}(\tau,\tau')$ which has no significance as long as we have not done the ${\mbox{\boldmath$\xi$}}$-integration. We explicitly insert the decomposition (\[xp06\]) into expression (\[xp20\]) of the source vector $\tilde{\bf s}$. Since ${\bf v}_0({\mbox{\boldmath$\xi$}})$ from Eq. (\[xp05\]) is constant in time, some temporal integrals in the exponential of Eq. (\[xp19\]) can be calculated and we obtain $$\begin{aligned}
\label{xp21}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j},{\bf v}]&=&
\lambda_{\rm th}^{-3}\,Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
\exp\left\{\frac{1}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
{\bf s}(\tau)\,{\bf H}^{{\bf x}_0}(\tau,\tau')\,{\bf s}(\tau')\right\}\nonumber\\
&&\times\int d^3\xi\,
\exp\left\{-\frac{M}{2\hbar^2\beta}{\mbox{\boldmath$\xi$}}^2+i\frac{M}{\hbar^2\beta}{\mbox{\boldmath$\xi$}}\cdot
\int_0^{\hbar\beta}d\tau\,{\bf v}(\tau) \right\}\end{aligned}$$ with the new $6$-vector $$\label{xp22}
{\bf s}(\tau)=(\,{\bf j}(\tau),{\bf v}(\tau)\,)$$ consisting of the original sources ${\bf j}$ and ${\bf v}$ only. The Gaussian $\xi$-integral in Eq. (\[xp21\]) can be easily solved and the terms appearing from quadratic completion modify the above matrix ${\bf H}^{{\bf x}_0}(\tau,\tau')$. The final result for the generating functional of all position and momentum dependent correlations is given by $$\label{xp23}
Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}[{\bf j},{\bf v}]=Z^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}
\exp\left\{\frac{1}{2\hbar^2}\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\,
{\bf s}(\tau)\,{\bf G}^{{\bf p}_0,{{\bf x}_0}}(\tau,\tau')\,{\bf s}(\tau')\right\}.$$ The complete $6\times 6$-matrix ${\bf G}^{{\bf p}_0,{{\bf x}_0}}(\tau,\tau')$ contains all possible Green functions describing position-position, position-momentum, and momentum-momentum correlations. As a consequence of separating the fluctuations into those perpendicular and parallel to the direction of the magnetic field, all correlations between $x,y$ on the one and $z$ on the other hand vanish as well as those for the appropriate momenta. The symmetries for the Green functions and their derivatives were investigated in detail in Appendix \[appprop\] and lead to a further reduction of the number of significant matrix elements. It turns out that only 9 elements are independent of each other. Therefore we can write the matrix $$\label{xp24}
{\bf G}^{{\bf x}_0,{{\bf p}_0}}(\tau,\tau')=\left(\begin{array}{cccccc}
G_{xx}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & G_{xy}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & 0 &
G_{xp_x}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & G_{xp_y}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & 0 \\
G_{xy}^{{\bf p}_0,{\bf x}_0}(\tau',\tau) & G_{xx}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & 0 &
-G_{xp_y}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & G_{xp_x}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & 0 \\
0 & 0 & G_{zz}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & 0 & 0 &
G_{zp_z}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')\\
G_{xp_x}^{{\bf p}_0,{\bf x}_0}(\tau',\tau) & -G_{xp_y}^{{\bf p}_0,{\bf x}_0}(\tau',\tau) & 0 &
G_{p_xp_x}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & G_{p_xp_y}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & 0 \\
G_{xp_y}^{{\bf p}_0,{\bf x}_0}(\tau',\tau) & G_{xp_x}^{{\bf p}_0,{\bf x}_0}(\tau',\tau) & 0 &
G_{p_xp_y}^{{\bf p}_0,{\bf x}_0}(\tau',\tau) & G_{p_xp_x}^{{\bf p}_0,{\bf x}_0}(\tau,\tau') & 0 \\
0 & 0 & G_{zp_z}^{{\bf p}_0,{\bf x}_0}(\tau',\tau) & 0 & 0 &
G_{p_zp_z}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')
\end{array}\right).$$ The matrix decays into four $3\times 3$-blocks, each of the which describing another type of correlation: the upper left position-position, the upper right position-momentum (as well as the lower left one), and the lower right momentum-momentum correlations. The different elements of the matrix are $$\begin{aligned}
\label{xp25a}
G_{xx}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{x}}(\tau){\tilde{x}}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
G_{xx}^{{\bf x}_0}(\tau,\tau'),\\
\label{xp25b}
G_{xy}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{x}}(\tau){\tilde{y}}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
G_{xy}^{{\bf x}_0}(\tau,\tau'),\\
\label{xp25c}
G_{zz}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{z}}(\tau){\tilde{z}}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
G_{zz}^{{\bf x}_0}(\tau,\tau'),\\
\label{xp25d}
G_{xp_x}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{x}}(\tau){\tilde{p}_x}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
iM{{G^\bullet}}_{xx}^{{\bf x}_0}(\tau,\tau')-\frac{1}{2}M{\Omega_{\perp 1}}G_{xy}^{{\bf x}_0}(\tau,\tau'),\\
\label{xp25e}
G_{xp_y}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{x}}(\tau){\tilde{p}_y}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
iM{{G^\bullet}}_{xy}^{{\bf x}_0}(\tau,\tau')+\frac{1}{2}M{\Omega_{\perp 1}}G_{xx}^{{\bf x}_0}(\tau,\tau'),\\
\label{xp25f}
G_{zp_z}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{z}}(\tau){\tilde{p}_z}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
iM{{G^\bullet}}_{zz}^{{\bf x}_0}(\tau,\tau'),\\
\label{xp25g}
G_{p_xp_x}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{p}_x}(\tau){\tilde{p}_x}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
-M^2{{^\bullet \tilde{G}^\bullet}}_{xx}^{{\bf x}_0}(\tau,\tau')-iM^2{\Omega_{\perp 1}}{{^\bullet G}}_{xy}^{{\bf x}_0}(\tau,\tau')+\frac{1}{4}
M^2{\Omega_{\perp 1}}^2G_{xx}^{{\bf x}_0}(\tau,\tau')-\frac{M}{\beta},\\
\label{xp25h}
G_{p_xp_y}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{p}_x}(\tau){\tilde{p}_y}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
iM^2{{^\bullet G}}_{xx}^{{\bf x}_0}(\tau,\tau')-M^2{{^\bullet G^\bullet}}_{xy}^{{\bf x}_0}(\tau,\tau')+\frac{1}{4}
M^2{\Omega_{\perp 1}}^2G_{xy}^{{\bf x}_0}(\tau,\tau'),\\
\label{xp25i}
G_{p_zp_z}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')&=&{\left\langle\,{\tilde{p}_z}(\tau){\tilde{p}_z}(\tau')\,\right\rangle^{{\bf p}_0,{\bf x}_0}_{\bf \Omega}}=
-M^2{{^\bullet \tilde{G}^\bullet}}_{zz}^{{\bf x}_0}(\tau,\tau')-\frac{M}{\beta},\end{aligned}$$ where the expectation values are defined by Eq. (\[vpt07\]). Note that all these Green functions are invariant under time translations such that $$\label{xp26}
G_{\mu\nu}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')=G_{\mu\nu}^{{\bf p}_0,{\bf x}_0}(\tau-\tau')$$ with $\mu,\nu\in \{x,y,z,p_x,p_y,p_z\}$.
It is quite instructive to prove that all these Green functions can be decomposed into a quantum statistical and a classical part as we did it in Eq. (\[app19\]). Since we know that the classical correlation functions do not depend on the euclidean time, all derivative terms in Eqs. (\[xp25a\])–(\[xp25i\]) do not contribute a classical term. We can write each Green function $$\label{xp27}
G_{\mu\nu}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')=G_{\mu\nu}^{\rm qm}(\tau,\tau')-
G_{\mu\nu}^{\rm cl},$$ This relation has been already checked for Eqs. (\[xp25a\])-(\[xp25c\]) in Appendix \[appgen\]. The classical contribution is zero in Eqs. (\[xp25d\]), (\[xp25f\]), and (\[xp25h\]) following from the absence of classical terms in derivatives of the Green functions and mixed correlations like (\[app22a\]). It seems surprising that the correlation (\[xp25e\]) contains a classical term while (\[xp25d\]) possesses none. This is, however, a consequence of the cross product of the orbital angular momentum appearing in the action (\[xp02\]) and the explicit classical calculation entails $$\label{xp28}
G_{xp_x}^{\rm cl}=\langle xp_x\rangle^{\rm cl}=0,
\qquad G_{xp_y}^{\rm cl}=\langle xp_y\rangle^{\rm cl}=\frac{2}{\beta}\,\frac{{\Omega_{\perp 1}}}{{\Omega_{\perp 2}}^2-{\Omega_{\perp 1}}^2},$$ where the latter is the subtracted classical term in Eq. (\[app19\]) when considering the first two substitutions in (\[xp13\]). In Eq. (\[xp25i\]), the second term is obviously the classical one since $$\label{xp29}
G_{p_zp_z}^{\rm cl}=\langle p_zp_z\rangle^{\rm cl}=\frac{M}{\beta}.$$ The extraction of the classical terms $$\label{xp30}
G_{p_xp_x}^{\rm cl}=\langle p_xp_x\rangle^{\rm cl}=\frac{M}{\beta}\left(1+
\frac{{\Omega_{\perp 1}}^2}{{\Omega_{\perp 2}}^2-{\Omega_{\perp 1}}^2}\right)$$ in the case of the Green function $G_{p_xp_x}^{{\bf p}_0,{\bf x}_0}(\tau,\tau')$ requires the consideration of the last two terms in Eq. (\[xp25g\]). Thus we have shown that the decomposition (\[xp27\]) holds for each of the Green functions (\[xp25a\])–(\[xp25i\]). Note the necessity of subtracting the classical terms since they all diverge in the classical limit of high temperatures ($\beta\to 0$).
[199]{} J.C. Le Guillou and J. Zinn-Justin, Ann. Phys. (N.Y.) [**147**]{}, 57 (1983). J.E. Avron, B.G. Adams, J. Čížek, M. Clay, M.L. Glasser, P. Otto, J. Paldus, and E. Vrscay, Phys. Rev. Lett. [**43**]{}, 691 (1979). B.G. Adams, J.E. Avron, J. Čížek, P. Otto, J. Paldus, R.K. Moats, and H.J. Silverstone, Phys. Rev. A [**21**]{}, 1914 (1980). V.A. Gani, A.E. Kudryavtsev, and V.M. Weinberg, eprint: physics/9708005 (1997). I.D. Feranshuk and L.I. Komarov, J. Phys. A: Math. Gen. [**17**]{}, 3111 (1984). L.D. Landau and E.M. Lifschitz, [*Quantenmechanik*]{}, Sechste Auflage (Akademie-Verlag Berlin, 1979). R. Loudon, Am. J. Phys. [**27**]{}, 649 (1959). L.K. Haines and D.H. Roberts, Am. J. Phys. [**37**]{}, 1145 (1969). R. Cohen, J. Lodenquai, and M. Ruderman, Phys. Rev. Lett. [**25**]{} (1970). M.V. Ivanov and P. Schmelcher, Phys. Rev. A [**61**]{} (in press 2000). J.S. Heyl and L. Hernquist, Phys. Rev. A [**58**]{}, 3567 (1998). H. Ruder, G. Wunner, H. Herold, and F. Geyer, [*Atoms in Strong Magnetic Fields*]{} (Springer-Verlag, Berlin, 1994). J.T. Devreese and F. Brosens in [*Elementary Excitations in Solids*]{}, p. 283, Eds. J.L. Birman, C. Sébenne, and R.F. Wallis, Elsevier, 1992. J.T. Devreese and F. Brosens, Solid State Comm. [**79**]{}, 819 (1991). J.T. Devreese and F. Brosens, Phys. Rev. B [**45**]{}, 6459 (1992). H. Kleinert, [*Path Integrals in Quantum Mechanics, Statistics, and Polymer Physics*]{}, 2nd ed. (World Scientific, Singapore, 1995). R.P. Feynman and H. Kleinert, Phys. Rev. A [**34**]{}, 5080 (1986); see also Chapter 5 of the textbook [@PI]. M. Bachmann, H. Kleinert, and A. Pelster, Phys. Rev. A [**60**]{}, 3429 (1999), eprint: quant-ph/9812063. H. Kleinert, W. Kürzinger, and A. Pelster, J. Phys. A: Math. Gen. [**31**]{}, 8307 (1998), eprint: quant-ph/9806016. H. Kleinert, A. Pelster, and M. Bachmann, Phys. Rev. E [**60**]{}, 2510 (1999), eprint: quant-ph/9902051.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Existence of dark matter indicates the presence of unknown fundamental laws of nature. Ultralight axion-like particles are well-motivated dark matter candidates, emerging naturally from theories of physics at ultrahigh energies. We report the results of a direct search for the electromagnetic interaction of axion-like dark matter in the mass range that spans three decades from 12 peV to 12 neV. The detection scheme is based on a modification of Maxwell’s equations in the presence of axion-like dark matter, which mixes with a static magnetic field to produce an oscillating magnetic field. The experiment makes use of toroidal magnets with iron-nickel alloy ferromagnetic powder cores, which enhance the static magnetic field by a factor of 24. Using SQUIDs, we achieve a magnetic sensitivity of 150 $\text{aT}/\sqrt{\text{Hz}}$, at the level of the most sensitive magnetic field measurements demonstrated with any broadband sensor. We recorded 41 hours of data and improved the best limits on the magnitude of the axion-like dark matter electromagnetic coupling constant over part of our mass range, at 20 peV reaching $3.3 \times 10^{-11}~\text{GeV}^{-1}$ (95% confidence level). Our measurements are starting to explore the coupling strengths and masses of axion-like particles where mixing with photons could explain the anomalous transparency of the universe to TeV gamma-rays.'
author:
- 'Alexander V. Gramolin'
- Deniz Aybas
- Dorian Johnson
- Janos Adam
- 'Alexander O. Sushkov'
bibliography:
- 'library.bib'
title: 'Search for axion-like dark matter with ferromagnets'
---
Astronomical evidence, built up over eight decades, indicates that only one-sixth of matter in the universe is made up of fundamental particles whose properties we understand [@Spergel2015; @PDG2019]. The existence of dark matter constitutes compelling evidence for new fundamental particles and interactions beyond the Standard Model. Since it has only been observed through its gravitational effects, there is a broad range of dark matter candidates. The weakly-interacting massive particle (WIMP) is a candidate that has inspired a large number of ultra-sensitive experiments of increasing complexity and scale, but so far there has been no unambiguous detection, and fundamental backgrounds (neutrino floor) will soon start to limit the sensitivity of direct WIMP searches [@Liu2017; @PDG2019; @Rajendran2017]. Another well-motivated dark matter candidate is the axion, which also offers a compelling solution to the strong CP problem of quantum chromodynamics [@Preskill1983; @Abbott1983; @Dine1983; @DeMille2017; @Irastorza2018a]. The axion and other light pseudoscalar bosons (axion-like particles, ALPs) emerge naturally from theoretical models of physics at high energies, including string theory, grand unified theories, and models with extra dimensions [@Svrcek2006; @Irastorza2018a]. Astrophysical observations have produced a number of stringent limits on ALPs, but also some tantalizing hints for their possible existence [@PDG2019]. One such hint is the observation that the universe is too transparent to TeV $\gamma$-rays. These are emitted from distant active galactic nuclei and should be significantly attenuated, through pair production, by infrared background radiation during their travel to our galaxy [@Matsuura2017]. The tension with observed TeV gamma-ray source energy spectra can be explained by photon-ALP mixing inducing gamma-ray conversion into axions, which travel unaffected through interstellar space and convert back into gamma-rays in the Milky Way [@Kohri2017].
Experimental searches for the axion and ALPs rely on one of their interactions with Standard Model particles [@Graham2013; @Budker2014; @Arvanitaki2014; @Irastorza2018a]. Most experiments to date have focused on the axion-photon interaction, which can mix photons with axions and ALPs in the presence of a strong magnetic field. The conversion rate scales with magnetic field and volume, so experiments employ large magnetic fields and volumes for better sensitivity [@Battesti2018]. This effect has been used to place stringent limits on the coupling of ALPs produced in the laboratory [@Ehret2010], or in the Sun [@Anastassopoulos2017]. Several searches for axion dark matter in the $\mu$eV mass range have used the technique of resonant conversion into monochromatic microwave photons inside a high-quality-factor cavity permeated by a strong magnetic field [@Graham2015a]. The ADMX experiment has achieved a level of sensitivity sufficient to search for dark matter axions with masses between 2.66 and $2.81~\mu\text{eV}$ and excluded the QCD axion-photon couplings predicted by plausible models for this mass range [@Du2018]. A number of cavity-based axion dark matter searches are in development or already exploring ALP masses up to $\approx 30~\mu\text{eV}$ [@Graham2015a; @Brubaker2017; @Choi2017b]. In order to search for lower-mass axions and ALPs coupled to photons, it is possible to use lumped-element circuits [@Sikivie2014b; @Chaudhuri2015; @Kahn2016; @Chaudhuri2018; @Ouellet2019]. This concept is based on a modification of Maxwell’s equations: in the presence of a large static magnetic field $B$, axion-like dark matter acts as a source of an oscillating magnetic field whose amplitude is proportional to $B$ [@Sikivie1983; @Wilczek1987].
Our approach is to use a toroidally-shaped permeable material to enhance the magnitude of the static magnetic field $B$ and thus improve sensitivity to axion-like dark matter. In the presence of a magnetizable medium, the modified inhomogeneous magnetic Maxwell’s equation takes the form $$\vec{\nabla}\times\vec{H} = \vec{J}_f + \frac{g_{a\gamma}}{\mu_0c}\frac{\partial a}{\partial t}\vec{B},
\label{eq:1}$$
{width="80.00000%"}
where $\vec{H} = \vec{B}/\mu_0 - \vec{M}$ is the auxiliary field, $\vec{M}$ is the magnetization, $\vec{J}_f$ is the macroscopic free electric current density, $\mu_0$ is the permeability of free space, $c$ is the speed of light in vacuum, and $g_{a\gamma}$ is the strength of ALP electromagnetic coupling [@som]. The axion-like dark matter field $a = a_0 \sin{(\omega_a t)}$ oscillates at its Compton frequency $\nu_a=\omega_a/(2\pi)=m_ac^2/h$, where $m_a$ is the ALP mass, and $h$ is the Planck constant. The coherence time of this oscillating field is $\approx 10^6$ periods, set by the kinetic energy of the virialized axionic dark matter [@Sikivie1983; @Graham2013]. We use SI units for electromagnetic fields and natural units for the ALP field so that $g_{a\gamma} a_0$ is unitless, and the ALP field amplitude $a_0$ is given by the dark matter energy density: $m_a^2 a_0^2 / 2 = \rho_{\text{DM}} = 3.6 \times 10^{-42}~\text{GeV}^4$ [@PDG2019; @Graham2013].
Our experiment, the Search for Halo Axions with Ferromagnetic Toroids (SHAFT), contained two independent detection channels. Each channel consisted of two stacked toroids, each of which could be independently magnetized by injecting a current into a superconducting magnetizing coil wound around the toroid, fig. \[fig:1\](a). With azimuthal field $B_0$ inside the magnetized toroid, the second term in the right-hand side of eq. (\[eq:1\]) represents an effective current density, $J_{\rm eff} = \omega_a g_{a\gamma} a_0 \cos{(\omega_a t)} B_0 / (\mu_0 c)$, sourced by the axion-like dark matter field. This loop of current creates an axial magnetic field $B_a$. The pickup coil, wound around the inner circumference of the toroids, was coupled to a Superconducting Quantum Interference Device (SQUID) magnetometer, as shown in fig. \[fig:1\](b), measuring the magnetic flux $\Phi_a$ due to $B_a$. The axion-like dark matter detection signature would be an oscillating SQUID output signal, whose amplitude was correlated with toroid magnetization.
{width="80.00000%"}
The two-channel, four-toroid design enabled a systematic rejection scheme that allowed for discrimination between an axion-like dark mater signal and electromagnetic interference, fig. \[fig:1\](c). For collecting data sensitive to axion-like dark matter, we magnetized the top two toroids counter-clockwise (channel A: $+ +$), and the bottom two toroids clockwise (channel B: $- -$), viewed from the top. An axion-like dark matter signal would then appear in the two detector channels with opposite sign ($180^{\circ}$ out of phase). Electromagnetic interference, uncorrelated with toroid magnetization, would not have such a phase relationship and could be distinguished from an axion-like dark matter signal by forming symmetric and antisymmetric combinations of the two detector channels. The toroids and pickup coils were inside an enclosure formed by two nested coaxial cylinders immersed in liquid helium. Lead foil, affixed to the inner surfaces of the cylinders and their caps, formed a double-layer superconducting magnetic shield, suppressing electromagnetic interference and ambient magnetic field noise.
The sensitivity enhancement in our approach arises from the fact that the magnetic field inside the permeable toroids includes material magnetization in addition to the field created by the free current in the toroidal magnetizing coil: $\vec{B}_0 = \mu_0 (\vec{H}_0 + \vec{M}_0)$. For a linear magnetic material with permeability $\mu$, $B_0 = \mu \mu_0 H_0$ and the magnetic field enhancement is by a factor equal to $\mu$. In practice, magnetic materials are non-linear: the permeability drops with applied field $H_0$ and the magnetization saturates. For our experiment, we chose powdered iron-nickel “High Flux $125\mu$” alloy for its high permeability, high saturation flux density, and electrical insulating properties at liquid helium temperature [@som]. We determined the permeability of the toroid material by injecting currents up to 10 A into the magnetizing coil and measuring its inductance. The initial permeability was $\approx 110$, dropping with increasing coil current as the material saturated, fig. \[fig:2\](a). The material displayed no magnetic hysteresis. We used the permeability measurements to calculate magnetic field $B_0$ inside the toroid at each value of the magnetizing coil current using $B_0(H_0) = \mu_0 \int_0^{H_0} \mu(H) \, dH$, fig. \[fig:2\](b). Before recording data sensitive to ALP dark matter, we set the magnetizing coil current to 6 A, at which point $B_0 = 1.51~\text{T}$ and the material is close to saturation. At this current, the magnetic field inside an air-core toroid would be 0.063 T (dashed red line in fig. \[fig:2\](b)). Therefore, we can quantify the enhancement of $B_0$ due to the magnetic material to be a factor of 24. During ALP-sensitive data collection, we switched each of the magnetizing coils into persistent mode, but still monitored each toroid magnetization by making inductance measurements of separate permeability sensing coils, which had been calibrated in advance [@som].
{width="70.00000%"}
Quantifying the sensitivity of our experiment to an axion-like dark matter signal required measurements of the magnetic field sensitivity of the two detection channels. The experimental design included four independent single-turn calibration loops, mounted at several positions near the toroids [@som]. We performed calibration measurements by injecting sinusoidally-varying current with known frequency and amplitude into each loop and recording the response of both SQUID channels. Measurements with the central calibration loop, which was positioned midway between the upper and the lower toroid pairs, and thus had equal couplings to the two pickup coils, showed that the detector channels were in phase, fig. \[fig:3\](a). We recorded the SQUID output $V_{\rm SQ}$ for a range of calibration current frequencies and used the numerically-computed mutual inductance to calculate the magnetic flux $\Phi$ through each pickup coil. These measurements established the calibration of the flux-to-voltage transfer function $x = V_{\rm SQ}/\Phi$, whose magnitude was consistent with the inductive coupling model shown in fig. \[fig:1\](b), and whose frequency dependence was consistent with a second-order low-pass filter model, fig. \[fig:3\](b). The transfer function measurements for all calibration loops mounted inside the experiment were consistent within a $25\%$ range [@som].
The main factor limiting the sensitivity of our ALP dark matter search was SQUID sensor noise. To quantify magnetic sensitivity, we magnetized the toroids with 6 A injected current, recorded the SQUID output voltage, converted it to magnetic field using the transfer function for each SQUID detector channel, and calculated the Fourier spectra, fig. \[fig:3\](c,d). Below 3 kHz, the spectra were dominated by vibrations of the pickup coil in the magnetic field leaking out of the magnetized toroids, and above 3 MHz, the sensitivity was degraded by the SQUID feedback electronics bandwidth roll-off. Aside from several noise peaks due to residual radiofrequency (RF) interference, between 10 kHz and 1 MHz the noise spectrum was very nearly flat at the level of $150$ aT/$\sqrt{\rm Hz}$. This is at the level of the most sensitive magnetic field measurements demonstrated with broadband SQUID systems [@Storm2017] or atomic magnetometers [@Dang2010], in spite of the direct proximity of ferromagnetic material magnetized to 1.51 T. The absence of detectable magnetization noise from the ferromagnetic toroids was due to the pickup coil geometry, designed to be insensitive to toroid magnetization fluctuations [@Eckel2009; @Sushkov2009]. For our axion-like dark matter search, we chose the frequency range between 3 kHz and 3 MHz so that the noise was within a factor of 3 of the baseline.
{width="70.00000%"}
The experimental run took place on September 12–17, 2019. After performing calibrations described above, we magnetized the four toroids with 6 A injected current in ($+ +$, $- -$) configuration and recorded 41 hours of data that were used to search for axion-like dark matter signals. Afterwards, we magnetized the four toroids with 6 A injected current in ($+ -$, $+ -$) configuration and collected 15 hours of data. Since in each channel the toroid magnetizations were now opposite, the axion-induced flux would vanish and we could use these data to confirm or reject any potential detection extracted from the first data set. We also collected 12 hours of data with no toroid magnetization, for the same purpose. In each data run, output voltages from the two SQUID detector channels were digitized at the sampling rate of 15.625 MS/s and recorded to a hard disk drive. We used the permeability sensing coils to monitor each toroid magnetization, and ensured that all magnetizations were unchanged before and after recording each data set. We note that experimental sensitivity to $g_{a\gamma}$ has a slow $1/4$-power scaling with averaging time, provided it exceeds ALP coherence time [@Budker2014].
Data analysis consisted of several processing, correction, and axion-like dark matter signal search steps. We divided the data into blocks of duration 2199 s ($2^{35}$ samples), chosen to be much longer than the ALP dark matter field coherence time for every frequency in the exclusion region. After removing 20 data blocks where SQUID feedback reset jumps occurred, we performed a discrete Fourier transform on each of the remaining 47 data blocks, used the SQUID detector transfer function $x$ to convert the frequency-domain data from voltage to magnetic flux $\Phi$, and formed symmetric and antisymmetric combinations of the two detector channels: $\Phi_{+}=(\Phi_{A}+\Phi_{B})/2$ and $\Phi_{-}=(\Phi_{A}-\Phi_{B})/2$. We calculated the power spectral density (PSD) for each of the four data streams $\Phi_{(A,B,+,-)}^2$, and averaged each PSD over all the blocks in the experimental run. We modeled each spectrum as the sum of an axion-like dark matter signal, SQUID detector noise, narrowband RF interference, and a broadband spectral baseline due to vibrational pickup of magnetic flux leakage from toroids. Our analysis assumed the boosted Maxwell-Boltzmann ALP signal lineshape predicted by the astrophysically-motivated standard halo model [@Turner1990; @Brubaker2017; @Brubaker2017a; @Du2018]. The linewidth of an axion-like dark matter signal at frequency $\nu_a$ is determined by the parameter $\beta^2 \nu_a = 3v_0^2 \nu_a/(2c^2) \approx 10^{-6} \nu_a$, where $v_0 = 220~\text{km}/\text{s}$ is the local velocity of motion around the center of the galaxy. Following a simplified version of the data analysis approach used by the ADMX and HAYSTAC collaborations, we rejected narrowband RF interference and broadband spectral baseline by using Savitzky-Golay digital filtering to isolate these spectral features in the frequency domain [@Brubaker2017a; @Du2018]. By injecting simulated axion signals into the data, we determined that these corrections attenuated axion spectral signals by 7%, and the final limits were adjusted accordingly [@som].
We performed the search for axion-like signals independently in each of the averaged spectra $\langle \Phi_{(A,B,+,-)}^2\rangle$ in the frequency range from 3 kHz to 3 MHz. In this range, each spectrum contained $6.6\times 10^9$ independent frequency points, which we broke up into 857 frequency bins, with bin width set to $10^4 \beta^2 \nu_b$, where $\nu_b$ is the bin center frequency. We modeled the histogram of PSD values in each bin as the chi-square distribution with 94 degrees of freedom, corresponding to the two quadratures of the Fourier transform averaged over 47 independent data blocks, fig. \[fig:4\](a). We optimally filtered the data within each frequency bin by convolving with the standard halo model axion lineshape. We modeled the histogram of optimally-filtered data points as the normal distribution (verified by a Lilliefors test), and found its standard deviation $\sigma$ by fitting to the Gaussian dependence, fig. \[fig:4\](b). We set the candidate detection threshold to $3.355\sigma$, equivalent to 95% confidence interval for a $5\sigma$ detection [@Brubaker2017a], and flagged all points above the threshold as candidates. In the 3 kHz to 3 MHz frequency range, with our analysis parameters, there were 20043605 possible ALP frequency points. Assuming the optimally-filtered spectra are normally distributed, we would statistically expect 7954 candidate frequencies at which the spectral value is above the $3.355\sigma$ threshold in any given channel. Our analysis procedure produced 16292 candidates in channel A, 35375 candidates in channel B, 27278 candidates in the symmetric combination channel, and 12513 candidates in the antisymmetric combination channel. The extra candidates appeared due to residual electromagnetic interference and lower-frequency vibrational tones that got past our spectral correction procedure.
Resonant experiments, such as ADMX and HAYSTAC, perform re-scans of candidate frequencies to check if they are statistical deviations or a real axion signal. In our analysis, we used the two detector channels to make this decision. Axion-like dark matter detection claim would be made for candidates that are above threshold for the antisymmetric channel and in both channels A and B, but not in the symmetric channel. None of the candidates satisfied all four of these detection conditions. As a check of our data analysis procedure, we injected numerically-generated ALP signals into our data and verified that our analysis recovered the signals with their correct coupling strengths [@som]. We used the conversion $\Phi_a = \omega_a g_{a\gamma} a_0 B_0 V / c$ between the ALP electromagnetic coupling $g_{a\gamma}$ and the induced magnetic flux $\Phi_a$ through the pickup coil, with the the mean magnetic field $B_0 = (1.51 \pm 0.03)~\text{T}$ at 6 A magnetizing current and the effective volume of each channel $V = (10.3 \pm 0.4)~\text{cm}^3$. We calculated the effective volume numerically, taking into account the $\approx 20\%$ suppression due to the Meissner effect in the lead superconducting magnetic shield [@som].
In the absence of a detection, we report for each frequency bin the magnitude of $g_{a\gamma}$ corresponding to the $3.355\sigma$ value of the optimally-filtered magnetic flux PSD in the antisymmetric channel, which represents the 95% confidence interval limit for that bin, fig. \[fig:4\](c). Before our work, the best experimental limit on the ALP-photon coupling for this mass range was set by the CERN CAST helioscope search for solar axions at $0.66\times 10^{-10}$ GeV$^{-1}$ [@Anastassopoulos2017]. Despite the much smaller volume of our apparatus (by a factor of $\approx 300$), our results constitute significant improvements on this limit over a broad range of ALP masses in the scenario that the ALP field is the dominant component of dark matter. Our experiment begins the exploration of the ALP mass and coupling region near $m_a \gtrsim 0.7~\text{neV}$ and $g_{a\gamma}\lesssim 8.8\times 10^{-10}~\text{GeV}^{-1}$, where existence of ALPs may resolve the tension between the observed TeV gamma-ray energy spectrum and the one expected based on the recent cosmic infrared background radiation data [@Kohri2017].
There are several ways to improve experimental sensitivity to axion-like dark matter. Cooling the SQUID magnetic sensors to milli-Kelvin temperature can reduce their noise by a factor of $\approx 10$, and it may be possible to achieve even better sensor performance using quantum upconversion [@Chaudhuri2019b]. For an optimized search, it is necessary to use a scanned single-pole resonator to couple the ALP-induced magnetic flux to an amplifier operating at, or beyond, the Standard Quantum Limit [@Chaudhuri2019]. The ferromagnetic toroidal core allowed us to achieve a 1.51 T static magnetic field with ampere-level injected current, which enabled the two-channel design, but magnetic material saturation prevented further field increase. New magnetic materials, with higher saturation field, could improve sensitivity further, or a high-field air-core toroidal superconducting magnet could be engineered, needing much higher injected current [@Battesti2018]. The most dramatic sensitivity improvement could be achieved by increasing the toroid volume; the Fe-Ni magnetic alloy used in our experiment is inexpensive and readily available commercially. Scaling the volume up to the benchmark $1~\text{m}^3$ would improve the sensitivity by another factor of $10^4$.
The authors acknowledge support from the NSF grant 1806557, the Heising-Simons Foundation grant 2015-039, the Simons Foundation grant 641332, and the Alfred P. Sloan foundation grant FG-2016-6728. The authors thank O. P. Sushkov for valuable comments on the manuscript.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Given a non-hermitean matrix , the structure of its *minimal* polynomial encodes whether is diagonalizable or not. This note will explain how to determine the minimal polynomial of a matrix without going through its *characteristic* polynomial. The approach is applied to a quantum mechanical particle moving in a square well under the influence of a piece-wise constant -symmetric potential. Upon discretizing the configuration space, the system is decribed by a matrix of dimension three. It turns out not to be diagonalizable for a critical strength of the interaction, also indicated by the transition of two real into a pair of complex energy eigenvalues. The systems develops a *three-fold* degenerate eigenvalue, and *two* of the three eigenfunctions disappear at this exceptional point, giving a difference between the *algebraic* and *geometric* multiplicity of the eigenvalue equal to two.'
author:
- |
Stefan Weigert\
$\langle \mbox{Hu} |\mbox{MP} \rangle$ - Hull Mathematical Physics\
Department of Mathematics, University of Hull\
UK-Hull HU6 7RX
title: 'How to Test for Diagonalizability: The Discretized -Invariant Square-Well Potential'
---
Introduction\[sec:Introduction\]
================================
Genuinely -invariant operators may or may not possess a complete set of eigenstates. In other words, -invariance of a matrix $\mathsf{M}$ is compatible with the presence of (non-trivial) Jordan blocks while hermiticity is not. When considering a familiy of -invariant operators depending on a parameter, their spectra often change qualitatively if one passes through an *exceptional* point [@kato84] where diagonalizability breaks down. It is thus important to be able to either check whether a given matrix is diagonalizable, or to locate exceptional points when presented with a continuous family of matrices.
The purpose of this note is to describe a method which allows one to identify exceptional points of finite-dimensional non-hermitean matrices by means of an algorithm. It is different from the method outlined in [@abate97; @weigert05a] as it directly aims at the *minimal* polynomial containing the relevant information about (non-) diagonalizability. While being more transparent in the first place, it also requires no knowledge of the *characteristic* polynomial of the given matrix.
The presentation to follow is problem-based: the algorithm will be developed while studying a specific example, the discretized -invariant square well. This physical system is introduced in Section \[sec: discretized-PT-symmetric square well\], and it is subjected to the test for diagonalizability in the subsequent section. The results will be discussed in Section \[sec:Summary-and-Discussion\] and some open questions will be addressed.
The discretized PT-symmetric square well\[sec: discretized-PT-symmetric square well\]
=====================================================================================
Consider a quantum particle in a one-dimensional box of length $4L$ subjected to a piece-wise constant -symmetric potential,
$$V(x)=\left\{ \begin{array}{cc}
\textrm{$-iZ,$} & -2L<x<0,\\
\quad\:0, & x=0\:,\\
\quad iZ, & \quad0<x<2L,\end{array}\right.\qquad Z\in\mathbb{R}\,,\label{eq: PT-potential}$$
which has proved a useful testbed for the discussion of -symmetric systems. Its eigenvalues are given as the zeros of a transcendental equation [@znojil01plus], and, for each value of $Z$, the lowest two real eigenvalues are known to coalesce and then disappear jointly at critical values of the parameter.
Let us introduce a toy-version of this system by discretizing its configuration space. This strategy has been applied successfully to decribe tunneling phenomena in a driven double-well potential in terms of a three-state model [@schatzer+98]. Effectively, this technique corresponds to turning Feynman’s “derivation” of Schrø"dinger’s equation from a discrete lattice [@feynman65] upside down. Explicitly, the continuous set of points of configuration space with labels $-2L\leq x\leq2L$, are replaced by *five* equidistant points at $0,\pm L$, and $\pm2L$. The wave function is allowed to take nonzero values only at these points, so it will be a vector with five components at most. However, the hard walls of the square well at $\pm L$ force the wave function to vanish there, leaving us with only three non-zero components, $\psi\rightarrow\psi_{k}=\psi(kL),k=0,\pm1$. The potential energy defined in (\[eq: PT-potential\]) turns into a diagonal matrix, $V(x)\rightarrow\mathsf{V=}\mbox{{diag}}(-iZ,0,iZ)$. The operator for the kinetic energy follows from replacing $\partial^{2}\psi(x)/\partial x^{2}\rightarrow(\psi_{k+1}-2\psi_{k}+\psi_{k-1})/L^{2}$. Putting all this together, the Hamiltonian operator of the discrete version of this system reads$$\mathsf{H_{0}}\simeq2\mathsf{E}-\mathsf{H}\:,\quad\mbox{where}\quad\mathsf{H}=\left(\begin{array}{ccc}
i\xi & 1 & 0\\
1 & 0 & 1\\
0 & 1 & -i\xi\end{array}\right),\quad\xi=Z/\eta\:;\label{eq:3by3 hamiltonian}$$ here $\mathsf{E}$ is the $(3\times3)$ identity matrix, and an overall factor $\eta=\hbar^{2}/2mL^{2}$ has been dropped. The matrix $\mathsf{H}_{0}$ inherits $\mathsf{PT}$-invariance from the square well: the matrix $\mathsf{H}$, and hence $\mathsf{H}_{0}$, is invariant under the combined action of parity , represented by a matrix with unit entries equal to one along its minor diagonal and zero elsewhere, and effecting complex conjugation. In the next Section, the diagonalizability of $\mathsf{H}$, the nontrivial part of the Hamiltonian $\mathsf{H}_{0}$, will be studied.
Diagonalizability of the PT-symmetric square well\[sec: Square well diagonalizability\]
=======================================================================================
The *minimal* polynomial **$m_{\mathsf{M}}$ of a matrix $\mathsf{M}$ is defined (see [@lancaster+85], for example) as the polynomial of least degree in $\mathsf{M}$ which annihilates $\mathsf{M}$, that is, $m_{\mathsf{M}}(\mathsf{M})=0$. This polynomial is unique if the coefficient of its highest power is taken to be one: the minimal polynomial is *monic.* Since any matrix $\mathsf{M}$ of size $N$, say, is annihilated by its own characteristic *polynomial,* $p_{\mathsf{M}}(\mathsf{M})=0$*,* the degree of the minimal polynomial does not exceed $N$. Once $m_{\mathsf{M}}$ has been found, one needs to determine whether it has only *single* roots, i.e. whether$$m_{\mathsf{M}}(\lambda)=\prod_{\nu=1}^{\nu_{0}(\leq N)}(\lambda-M_{\nu})\,,\quad\mbox{all }\, M_{\nu}\mbox{ distinct}\,.\label{genminpolydiag}$$ holds. If it does, the matrix $\mathsf{M}$ *is* diagonalizable - otherwise, it is *not* diagonalizable since multiple roots of $m_{\mathsf{M}}$ indicate the presence of Jordan blocks larger than one.
The procedure to determine the minimal polynomial of $\mathsf{M}$, as outlined in [@weigert05a], invokes the characteristic polynomial of $\mathsf{M}$ and repeated applications of the Euclidean algorithm generalized to polynomials. The method presented below, taken from [@horn+85], aims at directly constructing the minimal polynomial. The fundamental observation is that matrices of size $N$ constitute a vector space of dimension $N^{2}$. This is seen immediatly by setting up a one-to-one correspondence between matrices of size 3 and vectors of length $9\equiv3^{2}$, for example, simply by rearranging the elements $\mathsf{M}_{jk}$ of each $\mathsf{M}$ systematically according to$$\mathsf{M}\Leftrightarrow(\mathsf{M}_{11},\mathsf{M}_{12},\mathsf{M}_{13};\mathsf{M}_{21},\mathsf{M}_{22},\mathsf{M}_{23};\mathsf{M}_{31},\mathsf{M}_{32},\mathsf{M}_{33})^{T}.\label{eq: correspondence}$$ In view of this correspondence, the theorem by Cayley-Hamilton - every matrix satisfies its own characteristic equation, $p_{\mathsf{M}}(\mathsf{M})=0$, - turns into a statement about linear *dependence* of the *$(N+1)$ vectors* $\mathsf{E\equiv M}^{0},\mathsf{M},\mathsf{M}^{2},\ldots,\mathsf{M}^{N}$. In order to find the minimal polynomial of a matrix $\mathsf{M}$ one thus simply calculates $\mathsf{M}^{n},n=1\ldots N,$ and then determines successively whether the vectors $\mathsf{E}$ and $\mathsf{M}$ are linearly independent; if not, one adds $\mathsf{M}^{2}$ and asks the same question; etc. Proceeding in this way, one is obviously able to identify linear dependence among the vectors $\mathsf{M}^{n}$ containing only the smallest powers necessary. This, however, comes down to the definition of the minimal polynomial of . Gram-Schmidt orthonormalization effectively provides a systematic test for linear dependence among the first $k$ elements of $\mathsf{M}^{n},n=0\ldots N$.
Applying these ideas explicitly to the matrix $\mathsf{H}$ in (\[eq:3by3 hamiltonian\]) leads to $$\begin{aligned}
\mathsf{E} \, \, & \Leftrightarrow & (1,0,0;0,1;0;0,0,1)\:,\\
\mathsf{H} \, \, & \Leftrightarrow & (i\xi,1,0;1,0,1;0,1,-i\xi)\:,\\
\mathsf{H}^{2} & \Leftrightarrow & (1-\xi^{2},-i\xi,1;-i\xi,2,i\xi;1,i\xi,1-\xi^{2})\:,\\
\mathsf{H}^{3} & \Leftrightarrow & (2-\xi^{2})\:(i\xi,1,0;1,0,1;0,1,-i\xi)\:\equiv(2-\xi^{2})\mathsf{H\:}.\end{aligned}$$ It is easy to see that neither the first two nor the first three vectors in this sequence are linearly dependent. Consequently, there must be a relation expressing $\mathsf{H}^{3}$ in terms of the others, and indeed, the minimal polynomial follows immediately from $$(2-\xi^{2})\mathsf{H}-\mathsf{H}^{3}=0\quad\Rightarrow\quad m_{\mathsf{H}}(\lambda)=\lambda^{3}+(\xi^{2}-2)\lambda\:.\label{eq: minipolyH}$$ In addition, the characteristic polynomial of $\mathsf{H}$ must coincide with $m_{\mathsf{H}}$ since it there is only one monic polynomial of third degree annihilating ${\mathsf{H}}$.
If one is not able to factor the resulting minimal polynomial, one needs to check whether the minimal polynomial and its derivative $m^{\prime}\mathsf{_{H}}(\lambda)$ have a common factor which can be achieved by applying the Euclidean algorithm to this pair (cf. [@abate97; @weigert05a]). In this present case, this amounts to writing $m\mathsf{_{H}}(\lambda)=(\lambda-\alpha)m_{\mathsf{H}}^{\prime}(\lambda)/3+R_{1}(\lambda)$, implying that $\alpha=0$ and $R_{1}(\lambda)=(2/3)(\xi^{2}-2)(\lambda+1/2)$. Two different cases arise: if $\xi\neq2$, one finds that the only common factor of the minimal polynomial and the derivative is equal to one - thus, the minimal polynomial is of the form (\[genminpolydiag\]) and the matrix $\mathsf{H}$ must have three different eigenvalues making it diagonalizable. If $\xi^{2}=2,$ the algorithm immediately stops and thus identifies $\lambda^{2}$ as the highest common factor of $m_{\mathsf{H}}(\lambda)$ and its derivative. Consequently, the minimal polynomial has a three-fold root $\lambda(=0)$, indicating that $\mathsf{H}$ is not diagonalizable.
Discussion and Outlook\[sec:Summary-and-Discussion\]
====================================================
The properties of $\mathsf{H}$ at the exceptional points defined by $\xi_{\pm}=\pm\sqrt{{2}}$ deserve a brief discussion. It is not difficult to see that the *geometric* multiplicity of the eigenvalue 0 is one at the exceptional points ($\mathsf{H}$ has only one non-zero eigenvector) while its *algebraic* multiplicity equals three (zero is a triple root of $m_{\mathsf{H}}$). Contrary to previously studied cases, *three* eigenvalues coalesce for $\xi\rightarrow\xi_{\pm}$. For all values of $\xi$, the vector $(1,-i\xi,-1)$ is an eigenstate of $\mathsf{H}(\xi)$ with eigenvalue 0, and it does not exhibit any particular behaviour at the critical values $\xi_{\pm}$. Therefore, it seems reasonable to say that the eigenstates associated with the two $\lambda$-dependent eigenvalues disappear at the exceptional point. It is not obvious from a physical point of view why this scenario is preferred over the familiar situation of just one disappearing eigenstate.
The natural question to ask now is whether one can expect algorithmic tests for diagonalizability to exist for a quantum system living in a Hilbert space accomodating a countable infinity of states. As one needs to potentially perform an infinite number of steps, the idea of a useful algorithm gets somewhat blurred. Nevertheless, it is likely that one can search for systematic properties of finite-dimensional approximations which, hopefully, behave smoothly in the limit of infinite dimension.
[00]{} T. Kato: [*Perturbation Theory for Linear Operators.*]{} (Springer, New York 1984).
M. Abate: [*Amer. Math. Monthly*]{} [**104**]{}, 824 (1997).
St. Weigert: [*An Algorithmic Test for the Diagonalizability of Finite-Dimensional PT-Invariant Systems.*]{} ([*quant-ph/0506042*]{})
M. Znojil: [*Phys. Lett.*]{} A [**285**]{}, 7 (2001); [*J. Math. Phys.*]{} [**45**]{}, 4418 (2004)
L. Schatzer and St. Weigert: [*Phys. Rev. A*]{} [**57**]{}, 68 (1998)
R. P. Feynman: *The Feynman Lectures on Physics.* (Addison Wesley, Reading, MA 1965) Vol. III, Ch. 13-1.
P. Lancaster and M. Tismenetsky: [*The Theory of Matrices.*]{} (Academic Press, Orlando $^2$1985)
R. A. Horn and C. R. Johnson: *Matrix Analysis.* (Cambridge University Press, New York 1985)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that compatible almost-complex structures on symplectic manifolds correspond to optimal quantizations.'
author:
- 'Louis Ioos$^1$, David Kazhdan and Leonid Polterovich'
title: 'Berezin-Toeplitz quantization and the least unsharpness principle'
---
Introduction
============
A Riemannian metric on a symplectic manifold $(M,\omega)$ is $\omega$-compatible if it can be written as $$\label{eq-metricg}
g_{\omega,J}(\cdot,\cdot):= \omega(\cdot,J\cdot)\;,$$ where $J$ is an almost complex structure on $M$. Vice versa, an almost complex structure $J$ is $\omega$-compatible if the bilinear form is a Riemannian metric. Compatible geometric structures were introduced as an effective auxiliary tool for detecting rigidity phenomena on symplectic manifolds [@Gromov]. In the present paper we show that these structures naturally arise from the perspective of mathematical physics. Loosely speaking, they correspond to “optimal" quantizations, the ones minimizing a natural physical quantity called *unsharpness*, which is one of the main characters of this paper (see Section \[sec-coc\] below).
Quantization is a mathematical recipe behind the quantum-classical correspondence, a fundamental physical principle stating that quantum mechanics contains classical mechanics in the limiting regime when the Planck constant $\hbar$ tends to zero [@D]. There exist two different, albeit related mathematical models of this principle. Assume that the classical phase space is represented by a closed (i.e., compact without boundary) symplectic manifold $(M,\omega)$. The first model, [*deformation quantization*]{}, is a formal associative deformation $$f*g= fg+\hbar c_1(f,g) + \hbar^2 c_2(f,g) + \cdots$$ of the multiplication on the space $C^\infty(M)$ of smooth real functions on $M$ such that $f*g-g*f= i\hbar\{f,g\}+{\mathcal{O}}(\hbar^2)$, where $\{\cdot,\cdot\}$ stands for the Poisson bracket [@BFFLS]. The operation $*$ is called [*the star-product*]{} and the Planck constant $\hbar$ plays the role of a formal deformation parameter.
The second model, [*geometric quantization*]{}, is described as a linear correspondence $f \mapsto T_\hbar(f)$ between classical observables, i.e., real functions $f$ on the phase space $M$, and quantum observables, i.e., Hermitian operators on a complex Hilbert space. This correspondence is assumed to respect, in the leading order as $\hbar\to 0$, a number of basic operations. In the present paper, we focus on [*Berezin-Toeplitz quantizations*]{} [@Berezin; @BMS; @Gu; @BU; @MM; @Sc; @C], whose distinctive feature is to send non-negative functions to non-negative operators (see Section \[sec-bt\]). The known models of Berezin-Toeplitz quantization on closed symplectic manifolds (see the discussion following Theorem \[BQ\]) determine a deformation quantization [@BMS; @Sch; @Gu], and are provided by certain auxiliary data involving in particular an almost complex structure $J$ compatible with the symplectic form on the phase space. While deformation quantizations of closed symplectic manifolds are completely classified up to a natural equivalence, the classification of Berezin-Toeplitz quantizations is not yet completely understood (see however [@KS01] for the relation between the two).
The main finding of the present paper is that conversely, any Berezin-Toeplitz quantization, defined through natural axioms presented in Section \[sec-bt\], gives rise in a canonical way to a Riemannian metric on the phase space. Specifically, we make the natural assumption that there exists of a complex-valued bi-differential operator $c: C^\infty(M) \times C^\infty(M) \to C^\infty(M,{{\mathbb{C}}})$ such that the ${{\mathbb{C}}}$-linear extension of $T_\hbar$ satisfies $$\label{eq-expan}
T_\hbar(f)T_\hbar(g) = T_\hbar(fg)+ \hbar\,T_\hbar(c(f,g))+ {\mathcal{O}}(\hbar^2)\;,$$ for all $f,\,g\in C^\infty(M)$ as $\hbar\to 0$, and we show that this induces a Riemannian metric $G$ on $M$ by the formula $$c(f,g) = -\frac{1}{2}G(\operatorname{sgrad}f, \operatorname{sgrad}g)\,,$$ where $\operatorname{sgrad}f$ stands for the Hamiltonian vector field of a function $f$ on $M$. (see Theorem \[thm-main-new-1\](I) below). Leaving precise definitions for Section \[sec-main\], let us discuss the above-mentioned results informally and present a motivation coming from physics. To this end recall that it is classically known, starting from the Groenewold-van Hove theorem, that a Berezin-Toeplitz correspondence cannot be a genuine morphism between the Lie algebras of functions and operators. We focus on yet another constraint on the precision of Berezin-Toeplitz quantizations, which we call [*unsharpness*]{}, and which is governed by the Riemannian metric $G$ defined above. The notion of unsharpness is closely related to the Heisenberg uncertainty principle. It comes from an analogy between quantization and measurement based on the formalism of positive operator valued measures (POVMs), which serves both subjects, and which we briefly recall in Section \[sec-bt\]. The unsharpness metric is a particular instance of the noise operator [@Busch] describing, loosely speaking, the increment of variances in the process of quantization (see the discussion p.13).
In this language, we propose [*the least unsharpness principle*]{}, a variational principle selecting quantizations whose unsharpness metric has minimal possible volume on phase space. It turns out that the least unsharpness volume equals the symplectic volume (see Theorem \[thm-main-new-1\](II) below), so that a quantization satisfying the least unsharpness principle determines a compatible almost complex structure $J$ on $(M,{\omega})$, in the sense that its unsharpness metric satisfies $G=g_{{\omega},J}$ as in . We refer to Section \[sec-coc\] for basic properties of unsharpness, while existence of the unsharpness metric and the least unsharpness principle are stated in Section \[sec-main\] and proved in Section \[sec-proofs\].
The unsharpness metric is a natural geometric invariant of a Berezin-Toeplitz quantization, and can be seen as a first step towards classification. As a case study, we show in Section \[sec-equiv\] that for $SU(2)$-equivariant quantizations of the two-dimensional sphere, the unsharpness metric completely determines the quantization up to conjugation and up to second order as $\hbar\to 0$. Further comments on classification can be found in Section \[subsec-class\].
Some historical remarks are in order. A canonical appearance of Riemannian geometry in quantization was discussed on a number of occasions in physical literature. Anandan and Aharonov [@AA] and Ashtekar and Schilling [@AS] developed a geometric approach to quantum mechanics based on the Fubini-Study metric on the projective space of pure quantum states. Klauder (see, e.g., [@Klauder]) studied a model of a path-integral quantization where the role of a metric was to define a Brownian motion on the phase space. The idea of selecting optimal quantizations as those possessing the least uncertainty goes back to Gerhenstaber [@G]. He deals with quantizations which do not necessarily preserve positivity, and his least uncertainty principle implies that unsharpness identically vanishes on some restricted class of observables (see Section \[subsec-ger\] for further discussion). Finally, while classification of equivariant quantizations is known in the context of deformation quantization [@AL; @BBG], our setting, including the notion of equivalence, is substantially different. The case of $SU(2)$-equivariant Berezin-Toeplitz quantizations of the sphere which we settle in Section \[sec-equiv\] is, to the best of our knowledge, the first one where a complete classification is currently available.
Berezin-Toeplitz quantization {#sec-bt}
=============================
Given a finite-dimensional complex Hilbert space $H$, we write ${{\mathcal{L}}}(H)$ for the space of all Hermitian operators [(representing quantum observables), and ${{\mathcal{S}}}(H) \subset {{\mathcal{L}}}(H)$ for the subset of all non-negative trace-one Hermitian operators (representing quantum states)]{}.
\[POVMdef\]
An ${{\mathcal{L}}}(H)$-valued *positive operator valued measure* (POVM) on a set $M$ equipped with a $\sigma$-algebra ${{\mathcal{C
}}}$ is a map ${\Gamma}: {{\mathcal{C
}}}\to {{\mathcal{L}}}(H)$ which satisfies the following conditions:
- ${\Gamma}(\emptyset) = 0~\text{and}~{\Gamma}(M)=\id$;
- ${\Gamma}(X) \geq 0$ for every $X \in {{\mathcal{C
}}}$;
- ($\sigma$-additivity) ${\Gamma}\left(\bigsqcup_{i\in{{\mathbb{N}}}} X_i\right) = \sum_{i\in{{\mathbb{N}}}} {\Gamma}(X_i)$ for any sequence of pair-wise disjoint subsets $\{X_i \in {{\mathcal{C
}}}\}_{i\in{{\mathbb{N}}}}$.
According to [@CDS],
for every ${{\mathcal{L}}}(H)$-valued POVM measure ${\Gamma}$, there exist a probability measure $\alpha$ on $(M, {{\mathcal{C
}}})$ and a measurable function $F: M \to {{\mathcal{S}}}(H)$ such that
$$\label{eq-POVM-density}
d{\Gamma}= n\,Fd\alpha\;,$$
where $n= \dim_{{{\mathbb{C}}}} H$.
\[cohstate\] [In the context of quantization, the state $F_{x} \in {{\mathcal{S}}}(H)$ is called the [*coherent state*]{} associated with $x\in M$, and describes the quantization of a classical particle sitting at $x\in M$ in phase space.]{}
For any real classical observable $f \in L^1(M,\alpha)$, we define the *quantization* $T(f)$ as the integral $$\label{quantmap}
T(f):= \int_M f\,d{\Gamma}\in {{\mathcal{L}}}(H)\,.$$ The dual map $T^*: {{\mathcal{L}}}(H) \to L^\infty(M)$ with respect to the scalar product $((A,B)):= {{\rm tr}}(AB), A,\,B\in{{\mathcal{L}}}(H)$ satisfies $T^*(A)(x) = n\,{{\rm tr}}(AF(x))$, for any $x\in M$ and $A\in{{\mathcal{L}}}(H)$.
[For a quantum observable $A$, the function $T^*(A)\in L^\infty(M)$ has a natural interpretation as the classical observable whose value at $x\in M$ is the expectation value of $A$ at the associated coherent state $F_x$. Thus, we call $T^*(A)$ the *dequantization* of the quantum observable $A \in{{\mathcal{L}}}(H)$.]{}
\[Bdef\] [The composition $$\label{eq-BT}
\begin{split}
{{\mathcal{B}}}:= \frac{1}{n}T^*T: L^1(M,\alpha) &\longrightarrow L^\infty(M)\,,\\
f(x)~& \longmapsto ~{{\rm tr}}\left(T(f)F_x\right)
\end{split}$$ is called the [*Berezin transform*]{} associated to the POVM $\Gamma$.]{}
[The Berezin transform can be interpreted as quantization followed by dequantization. It is a measure of the blurring induced by quantization.]{}
To study the quantum-classical correspondence, we need to introduce a parameter $\hbar$ in the above story, which can be thought as the *Planck constant*, and from which we recover the laws of classical mechanics as $\hbar\to 0$. This is given a precise meaning via the following definition.
\[BTdef\] [Let $(M,\omega)$ be a closed connected symplectic manifold of dimension $2d$ and ${{\mathcal{C
}}}$ be the $\sigma$-algebra of its Borel sets in $M$.]{}
A *Berezin-Toeplitz quantization* of $M$ is the following data:
- a subset $\Lambda \subset {{\mathbb{R}}}_{>0}$ having $0$ as limit point ;
- a finite-dimensional complex Hilbert space $H_\hbar$ for each $\hbar \in \Lambda$ ;
- an ${{\mathcal{L}}}(H_\hbar)$-valued positive operator valued measures ${\Gamma}_\hbar$ on $M$ for each $\hbar\in\Lambda$,
such that the *Toeplitz map* $T_\hbar: C^\infty(M,{{\mathbb{C}}}) \to {\textup{End}}(H_\hbar)$ induced for all $\hbar\in\Lambda$ by the quantization map is surjective and satisfies the following estimates, [*uniformly*]{} in the $C^N$-norms of $f,\,g\in C^\infty(M)$ for some $N\in{{\mathbb{N}}}$ :
- [**(norm correspondence)**]{} $$\|f\|- {\mathcal{O}}(\hbar) \leq \|T_\hbar(f)\|_{{{\textrm{op}}}} \leq \|f\|\;,$$ where $\|\cdot\|_{op}$ is the operator norm and $\|f\|:=
\max_{x\in M} |f(x)|$ ;
- [**(bracket correspondence)**]{} $$\left\| -\frac{i}{\hbar} [T_{\hbar}(f),T_{\hbar}(g)] - T_\hbar (\{f,g\})\right\|_{{{\textrm{op}}}} ={\mathcal{O}}(\hbar)\;,$$ where $[\cdot,\cdot]$ stands for the commutator and $\{\cdot,\cdot\}$ for the Poisson bracket[^1];
- [**(quasi-multiplicativity)**]{} There exists a bi-differential operator\
$c: C^\infty(M) \times C^\infty(M) \to C^\infty(M,{{\mathbb{C}}})$ such that $$\|T_\hbar(f)T_\hbar(g) - T_\hbar(fg)-\hbar T_\hbar(c(f,g))\|_{op}= {\mathcal{O}}(\hbar^2)\;.$$
- [**(trace correspondence)**]{} $${{\rm tr}}(T_\hbar(f)) = (2\pi\hbar)^{-d}\int_M f\,R_\hbar\, d\mu\;,$$ where $R_\hbar\in C^\infty(M)$ satisfies $R_\hbar = 1+ {\mathcal{O}}(\hbar)$, and $d\mu = \frac{\omega^d}{d!}$ is the symplectic volume on $M$;
- [**(reversibility)**]{} The maps ${{\mathcal{B}}}_\hbar:C^\infty(M)\to C^\infty(M)$ induced by the Berezin transform satisfy $${{\mathcal{B}}}_\hbar (f) = f + {\mathcal{O}}(\hbar)\,.$$
By uniformly in the $C^N$-norms of $f,\,g\in C^\infty(M)$, we mean that there exists a constant $C>0$ such that, in axioms (P2) and (P3) and for $k=1,\,2$ respectively, the remainders satisfy $$\left|{\mathcal{O}}(\hbar^k)\right|\leq C\hbar^k\,\|f\|_{C^N}\|g\|_{C^N}\,,$$ while in axioms (P1) and (P5), the remainders satisfy $\left|{\mathcal{O}}(\hbar)\right|\leq C\hbar\,\|f\|_{C^N}$. Writing the density associated to ${\Gamma}_\hbar$ in the form $$\label{eq-POVM-density-1}
d{\Gamma}_{\hbar}(x) = n_\hbar\,F_{\hbar,x}\,d\alpha_\hbar(x)\;,$$ the trace correspondence (P4) implies $$n_\hbar= \frac{\text{Vol}(M,\omega) + {\mathcal{O}}(\hbar)}{(2\pi \hbar)^d}\;,$$ and $$\label{eq-measure}
d\alpha_\hbar = \frac{1+{\mathcal{O}}(\hbar)}
{\text{Vol}(M,\omega)}d\mu\;.$$ where $\text{Vol}(M,{\omega})>0$ denotes the symplectic volume of $(M,\omega)$.
The existence of a Berezin-Toeplitz quantization is a highly non-trivial result. To discuss it, recall that an almost complex structure $J$ on $M$ is
$\omega$-[*compatible*]{} if the form $G_J:=\omega(\cdot,J\,\cdot)$ is a Riemannian metric on $M$. We refer to $(M,\omega, J, G_J)$ as an *almost-Kähler structure* on $M$.
\[BQ\] If the closed symplectic manifold $(M,\omega)$ is *quantizable*, i.e., if the cohomology class $[\omega]/(2\pi)$ is integral, then every $\omega$-compatible almost-complex structure $J$ defines a Berezin-Toeplitz quantization of $(M,{\omega})$, with $\Lambda=\{1/k\}_{k\in{{\mathbb{N}}}}$.
In the case of *Kähler manifolds*, i.e., if we assume additionally that the almost complex structure $J$ is integrable, there is a canonical construction of a Berezin-Toeplitz quantization, where the Hilbert spaces $H_\hbar$ consist of the global *holomorphic sections* of a holomorphic Hermitian line bundle with Chern curvature equal to $-2\pi i k{\omega}$, with $\hbar=1/k$, and the associated Toeplitz map $T_\hbar$ sends $f\in C^\infty(M)$ to the multiplication by $f$ followed by the orthogonal $L_2$-projection on holomorphic sections. In this context, Theorem \[BQ\] has been established by Bordemann, Meinrenken and Schlichenmaier in [@BMS], using the theory of Toeplitz structures developed by Boutet de Monvel and Guillemin in [@BdMG81]. The fact that this theory extends to the almost-Kähler case was proved in a series of papers by Guillemin [@Gu], Borthwick and Uribe [@BU], Schiffman and Zelditch [@SZ02], Ma and Marinescu [@MM], Charles [@C] and the first named author, Lu, Ma and Marinescu [@I2]. The dependence of the remainders in terms of the derivatives of the functions is discussed in [@CP].
Unsharpness cocycle {#sec-coc}
===================
In this section, we study general properties of the bi-differential operator $c: C^\infty(M) \times C^\infty(M) \to C^\infty(M)$ from the quasi-multiplicativity property (P3) of a Berezin-Toeplitz quantization. First note that norm correspondence (P1) implies that, if an asymptotic expansion such as the one appearing in (P3) holds, then it is unique, and in particular, the bi-differential operator $c$ is uniquely defined. Then the associativity of the composition of operators implies that $c$ is a *Hochschild cocycle*, meaning that for all $f_1,\,f_2,\,f_3 \in C^\infty(M)$, we have $$\label{Hochcocycle}
f_1\,c(f_2,f_3)-c(f_1f_2,f_3)+c(f_1,f_2f_3)-c(f_1,f_2)\,f_3=0\;.$$ Denote by $c_-$ and $c_+$ its anti-symmetric and symmetric parts, respectively: $$c_-(f,g) := \frac{c(f,g)-c(g,f)}{2}\quad\text{and}
\quad c_+(f,g) := \frac{c(f,g)+c(g,f)}{2}\;.$$ By the bracket correspondence (P2), we [have $T_\hbar(2c_-(f,g) - i\{f,g\}) = {\mathcal{O}}(\hbar)$]{}, and hence by the norm correspondence (P1), we get the formula $$\label{eq-cminus}
c_-(f,g)= \frac{i}{2}\{f,g\}\;.$$ Thus the anti-symmetric part $c_-$ (responsible for the non-commutativity of quantum observables) does not depend on a choice of a quantization. In contrast , the symmetric part $c_+$ does depend on a choice of a quantization. By the quasi-multiplicativity property (P3), the cocycle $c_+$ associated to a Berezin-Toeplitz quantization measures its failure of being a multiplicative morphism on Poisson-commutative subspaces of $C^\infty(M)$. From formula and basic properties of the Poisson bracket, we know that $c^-$, hence also $c^+$, satisfy formula for a Hochschild cocycle.
\[def-unsharp\] [We say that $c_+$ is the *unsharpness cocycle* of a quantization or simply its *unsharpness*.]{}
Note that by formula , the operator $T_\hbar(f)\in{\textup{End}}(H_\hbar)$ is Hermitian if and only if $f\in C^\infty(M,{{\mathbb{C}}})$ is real valued, and as the square of a Hermitian operator is Hermitian, the quasi-multiplicativity property (P3) then shows that $c^+:C^\infty(M)\times C^\infty(M)\to C^\infty(M)$ is a real-valued bi-differential operator. It is also a symmetric Hochschild cocycle, so that as explained e.g. in [@RG Prop.2.14], it is a differential Hochschild *coboundary*. This means that there exists a real-valued differential operator $a:C^\infty(M)\to C^\infty(M)$ such that $$\label{eq-a}
c_+(f,g) = a(fg)-f\,a(g)-g\,a(f)\;.$$ for all $f,\,g\in C^\infty(M)$. Since $T_\hbar(1)=\id$, we have that $c_+(1,1)=0$, and therefore $a(1)=0$. Note that $a$ is determined up to its first order part. The following result shows that the positivity preserving property imposes a strong condition on $c_+$.
\[thm-order\] The bi-differential operator $c_+$ is of order $(1,1)$.
The proof is given in Section \[sec-order-proof\] below. Theorem \[thm-order\] sheds light on the differential operator $a$ appearing in the coboundary formula . In fact, let us choose some Darboux coordinates $U\subset X$, and take $f,\,g\in C^\infty(M)$ with compact support in $U$. In these coordinates, we can write $$\label{c+coord}
c_+(f,g)=\sum_{j,\,k=1}^{2d} a_{jk}\,\partial_j f\,
\partial_k g\,,$$ with smooth $a_{jk}=a_{kj}\in C^\infty(U)$ for each $1\leq j,\,k\leq 2d$. Then one can choose the differential operator $$\label{acoord}
a:=\frac{1}{2}\sum_{j,\,k=1}^{2d}\partial_j \left(a_{jk}\partial_k\right)\,.$$ in the coboundary formula . Using integration by parts, we see that is symmetric with respect to the canonical $L^2$-scalar product on $C^\infty(M)$ associated to the symplectic volume, and as the differential operator $a$ is determined up to its first order part, it is the unique such choice.
\[exam-1\][Assume that $(M,{\omega})$ is quantizable and equipped with an almost-Kähler structure $(M,{\omega},J,G_J)$. Then the induced Berezin-Toeplitz quantization of Theorem \[BQ\] satisfies $$\label{eq-cpluskahler1}
c_+(f,g) = -\frac{1}{2} (\nabla f,\nabla g)\;,$$ where the gradient and the product are defined with respect to $G_J$. Using that $$\label{Deltacocycle}
\Delta(fg) + 2(\nabla f,\nabla g) = f\Delta g + g\Delta f\;,$$ where $\Delta$ is the (positive) Laplace-Beltrami operator associated with $G_J$, the differential operator in the coboundary formula can then be chosen to be $a=\Delta/4$, and by basic properties of $\Delta$, it is the unique $L^2$-symmetric choice with respect to the symplectic volume form, as it coincides with the Riemannian volume form of $G_J$. Formula can be found in [@Xu p.257] for the Kähler case and in [@I1; @I2] for the almost-Kähler case. Using the $J$-invariance of the metric and the relation $J\,\text{sgrad} f= - \nabla f$ between Hamiltonian vector field and gradient of a function $f\in C^\infty(M)$ for an $\omega$-compatible metric, formula translates into $$\label{eq-cpluskahler}
c_+(f,g) = -\frac{1}{2} G_J(\text{sgrad} f,\text{sgrad}g)\;.$$ ]{}
\[exam-2\] [We now give an example of a Berezin-Toeplitz quantization whose unsharpness cocycle $c_+$ is not of the form for some almost-Kähler structure on $(M,{\omega})$. This example serves as a paradigm for the construction presented in the proof of one of our main results, Theorem \[thm-main-new-1\](III) below. Assume $(M,{\omega})$ quantizable and equipped with an almost-Kähler structure $(M,{\omega},J,G_J)$, and consider the induced Berezin-Toeplitz quantization of Theorem \[BQ\]. Fix $t>0$, and using the notations of Example \[exam-2\], consider for any $\hbar\in\Lambda=\{1/k\}_{k\in{{\mathbb{N}}}}$ the map $T_\hbar^{(t)}: C^\infty(M) \to {{\mathcal{L}}}(H_\hbar)$ defined for any $f\in C^\infty(M)$ by $$T^{(t)}_\hbar(f) := T_\hbar(e^{-t\hbar\Delta} f)\;.$$ Observe that the heat flow preserves positivity, so that $T^{(t)}_\hbar$ is in fact the quantization map induced by a POVM construction. Then from the classical small time asymptotic expansion of the heat kernel (see e.g. [@BGV04 Th.2.29, (2.8)]), as $\hbar\to 0$, we have $$e^{-t\hbar\Delta}f = f -t\hbar\Delta\,f+{\mathcal{O}}(\hbar^2)\,
\|f\|_{C^4}
\,,$$ and this implies in particular that all the axioms of Definition \[BTdef\] hold. Let us now calculate the associated unsharpness cocycle, denoted by $c_+^{(t)}$. For any $\hbar\in\Lambda$ and $A,\,B\in{\textup{End}}(H_\hbar)$, put $A \bullet B := \frac{1}{2}(AB+BA)$, and recall formula for the Laplace-Beltrami operator. Then as $\hbar\to 0$, we have $$\label{computexam1}
\begin{split}
&T^{(t)}_\hbar(f)\bullet T^{(t)}_\hbar(g)=T_\hbar(f) \bullet T_\hbar(g) -
t \hbar T_\hbar(f\Delta g + g\Delta f)+ {\mathcal{O}}(\hbar^2)\\
&= T_\hbar(fg) - \hbar T_\hbar\left(\frac{1}{2}(\nabla f,\nabla g) +
t\left(\Delta(fg) + 2(\nabla f,\nabla g)\right)\right) + {\mathcal{O}}(\hbar^2)\\
&= T_\hbar(fg) + \hbar
T^{(t)}_\hbar\left(-\left(\frac{1}{2}+2t\right) (\nabla f,\nabla g)\right) + {\mathcal{O}}(\hbar^2)\;,
\end{split}$$ so that, recalling that the quasi-multiplicativity property (P3) determines the unsharpness cocycle uniquely via norm correspondence (P1), we get $$\begin{split}
c^{(t)}_+(f,g)&=-\left(\frac{1}{2}+2t\right) (\nabla f,\nabla g)\\
&=-\frac{1}{2}\left(1+4t\right)G_J(\text{sgrad} f,\text{sgrad} g)\;.
\end{split}$$ In particular, we see that $c^{(t)}_+$ is of the form for the Riemannian metric $G^{(t)}:=(1+4t)\,G_J$ on $M$, whose volume is strictly bigger than the volume of the almost-Kähler metric $G_J$. As the volume of an almost-Kähler metric is always equal to the symplectic volume of $(M,{\omega})$, we see from that $c^{(t)}_+$ is not the unsharpness cocycle of a Berezin-Toeplitz quantization coming from Theorem \[BQ\]. ]{}
The least unsharpness principle {#sec-main}
===============================
In this section, we state the main theorem on unsharpness of Berezin-Toeplitz quantizations, which we call the *least unsharpness principle*, and discuss its physical meaning.
Recall from Theorem \[thm-order\] that the unsharpness cocycle $c_+$ of a Berezin-Toeplitz quantization is a bi-differential operator of order $(1,1)$, so that there exists a bilinear symmetric form $G$ on $TM$ such that $$\label{eq-G}
c_+(f,g) =: -\frac{1}{2}G(\text{sgrad}f,\text{sgrad}g)\,,$$ where $\text{sgrad}f,\,\text{sgrad}g$ denote the Hamiltonian vector fields of $f,\,g\in C^\infty(M,{{\mathbb{R}}})$. Our main result provides a description of this bilinear form $G$.
\[thm-main-new-1\] Let $(M,\omega)$ be a closed symplectic manifold.
For every Berezin-Toeplitz quantization of $M$, the form $G$ is a Riemannian metric on $M$ which can be written as
the sum $$\label{metric}
G={\omega}(\cdot,J\cdot)+\rho(\cdot,\cdot)\,,$$ where $J\in{\textup{End}}(TM)$ is a compatible almost complex structure on $(M,{\omega})$ and $\rho$ is a non-negative symmetric bilinear form on $TM$.
We have $\textup{Vol}(M,G) \geq \textup{Vol}(M,\omega)$, with equality if and only if $\rho \equiv 0$.
If $(M,\omega)$ is quantizable, then every Riemannian metric of the form arises from some Berezin-Toeplitz quantization.
The proof is given in Section \[sec-proofs\]. Let us mention that the proof of item (III) of the theorem is modeled on Example \[exam-2\] above and is constructive. We produce the desired Berezin-Toeplitz quantization with the unsharpness metric given by as the composition of the almost-Kähler quantization associated to $(\omega,J)$ and an explicit, albeit non-canonical, Markov operator depending on all the data including $\rho$.
\[rem-rho\][For a given metric $G$ on $M$, the decomposition is in general not unique. However, as the proof of Theorem \[thm-main-new-1\](I) will show, [ there exists a unique $\omega$-compatible almost complex structure $J$ which additionally is $G$-orthogonal, i.e., $G(J\xi,J\eta)=G(\xi,\eta)$ for all $\xi,\eta\in TM$. Furthermore, for such a $G$, the symmetric bilinear form $\rho(\xi,\eta) = G(\xi,\eta) - \omega(\xi,J\eta)$ is non-negative, thus providing decomposition . ]{}]{}
Before going to the proof of Theorem \[thm-main-new-1\] in the next Section, let us first discuss the physical meaning of the unsharpness cocyle $c^+$ associated with a Berezin-Toeplitz operator, which shows from general principles that it is at least non-negative. [With every quantum state $\theta \in {{\mathcal{S}}}(H_\hbar)$ one associates a classical state (called [*the Husimi measure*]{}), which is the probability measure $\mu_{\theta}$ on $M$ such that $$\label{Husimi}
\int_M f\,d\mu_\theta = {{\rm tr}}(T_\hbar(f)\,\theta)\;, f \in C^\infty(M)\,.$$ ]{} This equality can be interpreted as follows: the expectation of any classical observable $f$ in the classical state $\mu_\theta$ coincides with the expectation of the corresponding quantum observable $T_\hbar(f)$ in the state $\theta$. What happens with variances? It turns out that the quantum variance is in general bigger than the classical one. More precisely, we have that $$\mathbb{V}ar (f,\mu_\theta) = \int_M f^2 d\mu_\theta - \left(\int fd\mu_\theta\right)^2\;,$$ $$\mathbb{V}ar (T_\hbar(f),\theta) = {{\rm tr}}(T_\hbar(f)^2\theta) - \left({{\rm tr}}(T_\hbar(f)\theta)\right)^2\;,$$ and hence $$\label{noiseincrement}
\mathbb{V}ar (f,\mu_\theta) = \mathbb{V}ar (T_\hbar(f),\theta) + {{\rm tr}}(\Delta_\hbar(f)\theta)\;,$$ where $$\label{noisedef}
\Delta_\hbar(f) := T_\hbar(f^2)-T_\hbar(f)^2\;.$$ The operator $\Delta_\hbar(f)$ is called the [*noise operator*]{} (see e.g. [@Busch]), whose main property is that it is a *non-negative operator*. It describes the increase of variances, which can be interpreted as the unsharpness of the quantization. Then by the quasi-multiplicativity property (P3), we have $$\label{eq-noise}
\Delta_\hbar(f) = - \hbar\,T_\hbar\left(c_+(f,f)\right)+ {\mathcal{O}}(\hbar^2)\,.$$ Look at the expectation of $\Delta_\hbar(f)$ at the coherent state $F_{\hbar,x}$ of Remark \[cohstate\] associated to $\Gamma_\hbar$, $$\label{noisefla}
\begin{split}
{{\rm tr}}&\left(\Delta_\hbar(f)F_{\hbar,x}\right)= -\hbar\,{{\rm tr}}\left(T_\hbar\,(c_+(f,f))F_{\hbar,x}\right)+{\mathcal{O}}(\hbar^2)\\
&=-\hbar\,{{\mathcal{B}}}_\hbar\left(c_+(f,f)\right)(x)+{\mathcal{O}}(\hbar^2)\\
&=-\hbar\,c_+(f,f)(x)+{\mathcal{O}}(\hbar^2)\;.
\end{split}$$ This explains the name of unsharpness cocycle for $c^+$. Since the noise operator is non-negative, we get the following fundamental property of the unsharpness cocycle, $$\label{eq-positivity}
- c_+(f,f)(x) \geq 0\quad\text{for all}\quad x \in M\;.$$ This shows that the symmetric bilinear form $G$ defined in equation is at least semi-positive. This property is the first step towards the proof of Theorems \[thm-main-new-1\], showing that $G$ is in fact a Riemannian metric, called the *unsharpness metric* of the quantization. Note that this property is also at the basis of the proof of Theorem \[thm-order\].
Define the [*total unsharpness*]{} of a Berezin-Toeplitz quantization as the volume of the phase space $M$ with respect to the unsharpness metric. With this language, statement (II) of Theorem \[thm-main-new-1\] can be interpreted as the [*least unsharpness principle*]{}: the minimal possible total unsharpness equals the symplectic volume, and the least unsharpness metrics are induced by compatible almost-complex structures on $M$.
\[rmkvar\]
Let us assume that the Berezin transform admits an asymptotic expansion up to the first order as $\hbar\to 0$ of the following form for all $f \in C^\infty(M)$, $$\label{BTexp}
{{\mathcal{B}}}_\hbar(f) = f + \hbar Df + {\mathcal{O}}(\hbar^2)\,,$$ where $D$ is a differential operator, stenghtening the reversibility property (P5). Then by Definition \[Bdef\] of the Berezin transform, formula for the quantization map and the expansion for $\alpha_\hbar$, for all $f,\,g \in C^\infty(M)$, we have $$\begin{split}
\frac{1}{n_\hbar} {{\rm tr}}(T_\hbar(f)T_\hbar(g))&= \int_M {{\mathcal{B}}}_\hbar(f)\,g
\,d\alpha_\hbar\\
&=
\int_M fg\,d\alpha_\hbar + \hbar \int_M \left(Df\right) g\,d\mu + {\mathcal{O}}(\hbar^2)\;.
\end{split}$$ On the other hand, by the quasi-mutliplicativity property (P3), using formula and basic properties of the Poisson bracket, we get $$\begin{split}
&\frac{1}{n_\hbar} {{\rm tr}}(T_\hbar(f)T_\hbar(g))= \int_M (fg+\hbar c(f,g)+ {\mathcal{O}}(\hbar^2))\,d\alpha_\hbar\\
& = \int_M fg\,d\alpha_\hbar + \hbar \int_M c_+(f,g)\, d\mu+{\mathcal{O}}(\hbar^2)\;.
\end{split}$$ Then taking $f,\,g\in C^\infty(M)$ with compact support in Darboux coordinates, using formulas and and intergration by parts, we then get $$\label{eq-bda}
D=-2a\,,$$ where $a:C^\infty(M)\to C^\infty(M)$ is the unique $L^2$-symmetric differential operator on $C^\infty(M)$ with respect to symplectic volume satisfying the coboundary formula . In light of Example \[exam-1\], this fact generalizes the Karabegov-Schlichenmaier expansion [@KS01; @IKPS] for the Berezin-Toeplitz quantizations of Theorem \[BQ\].
Another consequence of the improvement of the reversibility property (P5) is that “unsharpness equals variance on coherent states”. To see this, recall definition of the Husimi measure on the coherent state $F_{\hbar,x}\in{{\mathcal{S}}}(H_\hbar)$ of Remark \[cohstate\]. Then the discussion above implies $$\begin{split}
\mathbb{V}ar (f,\mu_{F_{x,\hbar}})
&={{\mathcal{B}}}_\hbar(f^2)-{{\mathcal{B}}}_\hbar(f)^2\\
&=-2\hbar\,c^+(f,f)+{\mathcal{O}}(\hbar^2)\,.
\end{split}$$ Thus by formula and , we get $$\label{var=c+}
\mathbb{V}ar (T_\hbar(f),F_{\hbar,x}) = - \hbar\,c_+(f,f)(x) + {\mathcal{O}}(\hbar^2)\;,$$ so that the variance of a quantized observable at coherent states is equal to its unsharpness.
In their geometric formulation of quantum mechanics, Ashtekar and Shilling [@AS §3.2.3,(26)] consider the projectivization $\mathbb{P}(H_\hbar)$ as a “quantum phase space": a line $\xi\in \mathbb{P}(H_\hbar)$ is identified with the pure state given by the rank-one projector to $\xi$. In this setting, they give a physical interpretation of the Fubini-Study metric $g^{FS}$ on $\mathbb{P}(H_\hbar)$ in terms of the variance of a quantum observable $A\in{{\mathcal{L}}}(H_\hbar)$ at a pure state $\xi \in\mathbb{P}(H_\hbar)$. Specifically, write $v_A$ for the vector field on $\mathbb{P}(H_\hbar)$ induced by the infinitesimal action of $iA\in\mathfrak{u}(H_\hbar)$, seen as an element of the Lie algebra of the group of unitary operators $U(H_\hbar)$ acting on $\mathbb{P}(H_\hbar)$. Then the variance of $A$ at $\xi$ is given by $$\mathbb{V}ar (A,\xi) = \frac{1}{2}\,
g^{FS}_{\xi}(v_A,v_A)\;.$$ Back to the quantization, assume further that the coherent states $F_{\hbar,x}\in{{\mathcal{S}}}(H_\hbar)$ are pure for all $x\in M$. Consider the induced map $$\label{Kodemb}
F_{\hbar}:M\longrightarrow\mathbb{P}(H_\hbar)\;.$$ Then equation says that the Fubini-Study length of the vector field $v_A$ induced by the quantum observable $A:= T_\hbar(f)\in{{\mathcal{L}}}(H_\hbar)$ at the coherent state $F_{\hbar,x}\in\mathbb{P}(H_\hbar)$ approaches, as $\hbar\to 0$, the length of the Hamiltonian vector field $\operatorname{sgrad}f$ at a point $x\in M$ with respect to our unsharpness metric. In the case of the Kähler quantizations of Theorem \[BQ\], the map coincides with the *Kodaira map*. Then the picture described above is closely related to a theorem of Tian [@Ti] showing that the pullback of the Fubini-Study metric by the Kodaira map approaches the Kähler metric as $k\to+\infty$.
Proof of the main Theorem {#sec-proofs}
=========================
In this Section, we prove Theorem \[thm-main-new-1\]. To this end, first recall from the previous section that the non-negativity of the noise operator leads to the semi-positivity property for the unsharpness. To establish the stronger statement (i) of Theorem \[thm-main-new-1\], we will use a stronger property of noise operators coming from the general theory of POVM-based quantum measurements, called the *noise inequality*. It appears in several sources [@Ozawa], [@Hyashi Theorem 7.5], [@PR Theorem 9.4.16], albeit none of them contains the version we need. For the sake of completeness, we present a proof which closely follows [@PR] and is based on an idea from [@Janssens].
Consider a set $M$ equipped with a $\sigma$-algebra ${{\mathcal{C
}}}$ together with a finite-dimensional Hilbert space $H$, and let $F$ be an ${{\mathcal{L}}}(H)$-valued POVM in the sense of Definition \[POVMdef\]. For a bounded function $u\in L^\infty(M)$, we define the *noise operator* $$\Delta_F(u) := \int_M u^2\,dF - \left(\int_M u\,dF\right)^2\;.$$ For a pair of bounded functions $u,\,v\in L^\infty(M)$, set $$U := \int_M u\,dF,\quad V:=\int_M v\,dF\,.$$
\[lem-unsharp\] For every state $\theta \in {{\mathcal{S}}}(H)$, we have the inequality $$\label{eq-unsharp}
{{\rm tr}}\left(\Delta_F(u)\theta\right)\,{{\rm tr}}\left(\Delta_F(v)\theta\right)\geq
\frac{1}{4}\,\left|{{\rm tr}}\left( [U,V]\theta\right) \right|^2\;.$$
By the *Naimark dilation theorem* (see e.g. [@PR Theorem 9.4.6]), there exists an isometric embedding of $H$ into a (possibly infinite-dimensional) Hilbert space $H'$ and an ${{\mathcal{L}}}(H')$-valued [*projector valued*]{} measure $P$ such that for every subset $X \subset {{\mathcal{C
}}}$, we have $$\label{Naimdil}
F(X)=\Pi P(X)\Pi^*\in{{\mathcal{L}}}(H)\,,$$ where $\Pi : H' \to H$ is the orthogonal projector, so that $\Pi^*:H\to H'$ is the inclusion. Here ${{\mathcal{L}}}(H')$ stands for the space of all bounded Hermitian operators on $H'$, and an ${{\mathcal{L}}}(H')$-projector valued measure is by definition a map $P:{{\mathcal{C
}}}\to{{\mathcal{L}}}(H')$ satisfying the axioms of Definition \[POVMdef\], and such that the operators $P(X)$, $X \in {{\mathcal{C
}}}$, are pair-wise commuting orthogonal projectors.
Define a pairing $$q: {{\mathcal{L}}}(H') \times {{\mathcal{L}}}(H') \to {\textup{End}}(H)\;,$$ $$q(S,T) := \Pi S(1-\Pi^*\Pi)T\Pi^*\;.$$ We claim that for every state $\theta \in {{\mathcal{S}}}(H)$ and all $S,\,T \in {{\mathcal{L}}}(H')$, we have $$\label{eq-ST-1}
{{\rm tr}}(q(S,S)\theta)\, {{\rm tr}}( q(T,T)\theta) \geq \left|{{\rm tr}}(q(S,T)\theta)\right|^2\;.$$ To see this, note that $1-\Pi^*\Pi:H'\to H'$ is the orthogonal projector on the orthogonal complement of $H$, so that $${{\rm tr}}(q(S,T)\theta)=
{{\rm tr}}((1-\Pi^*\Pi)T\,\Pi^*\theta\Pi\,S(1-\Pi^*\Pi))\,.$$ Then follows from Cauchy-Schwarz inequality applied to the semi-norm on the space ${\textup{End}}(H')$ of bounded operators of $H'$ defined by $(A,B)_\theta:={{\rm tr}}(A\,\Pi^*\theta\Pi\,B^*)$, for all $A,\,B\in{\textup{End}}(H')$.
Set now $S= \int u\,dP$ and $T = \int v\,dP$. Since $S$ and $T$ commute, we have $[\Pi S\Pi^*,\Pi T\Pi^*] = q(T,S)-q(S,T)$. On the other hand, by definition , we have $\Pi S\Pi^*= U$ and $\Pi T\Pi^*= V$, while using an approximation by simple functions, one computes that $q(S,S) = \Delta_F(u)$ and $q(T,T) = \Delta_F(v)$. The statement of the lemma then directly follows from .
[Proof of (I):]{} Applying Lemma \[lem-unsharp\] to the Berezin-Toeplitz POVM $\Gamma_\hbar$ of Definition \[BTdef\], for every state $\theta \in {{\mathcal{S}}}(H_\hbar)$ and observables $u,\,v \in C^\infty(M)$, we get $$\label{eq-surprising}
{{\rm tr}}\left( \Delta_\hbar(u)\theta \right)\,
{{\rm tr}}\left(\Delta_\hbar(v)\theta \right)\\ \geq \frac{1}{4}\,
\left|{{\rm tr}}([T_\hbar(u),T_\hbar(v)]\theta)\right|^2\;.$$ Now by Definition \[Bdef\] of the Berezin transform and the expression for the noise operator, we know that for all $u\in C^\infty(M)$, $$\label{eq-surp-vsp}
{{\rm tr}}\left ( \Delta_\hbar(u)F_{\hbar,x}\right) = -\hbar\,{{\mathcal{B}}}(c_+(u,u))(x) + {\mathcal{O}}(\hbar^2)\;,$$ and for all $u,\,v \in C^\infty(M)$, $$\begin{split}
-i\,{{\rm tr}}([T_\hbar(u),T_\hbar(v)]F_{\hbar,x})&= \hbar\,{{\rm tr}}(T_\hbar(\{u,v\})F_{\hbar,x}) +{\mathcal{O}}(\hbar^2)\\
&= \hbar\,B_\hbar(\{u,v\})(x)+{\mathcal{O}}(\hbar^2)\;.
\end{split}$$ Thus applying the noise inequality with $\theta$ being the coherent state $F_{\hbar,x}$, we get that $$\label{eq-Btr-bound}
{{\mathcal{B}}}_\hbar(c_+(u,u))(x)~{{\mathcal{B}}}_\hbar(c_+(v,v))(x) \geq \frac{1}{4}|{{\mathcal{B}}}_\hbar(\{u,v\})(x)|^2\;,$$ so that the reversibility property (P5) yields $$\label{eqBtrpre}
c_+(u,u)~c_+(v,v) \geq \frac{1}{4}|\{u,v\}|^2\;.$$ Thus for all $\xi,\,\eta\in T_xM$, picking functions $u,\,v \in C^\infty(M)$ with $\text{sgrad}\,u(x)=\xi$, $\text{sgrad}\,v(x)=\eta$ and by definition of the bilinear form $G$, we get $$\label{eq-Btr-bound-1}
G(\xi,\xi)~G(\eta,\eta) \geq |\omega(\xi,\eta)|^2 \;.$$ Now thanks to the non-negativity of the noise operator, which follows from Lemma \[lem-unsharp\], we already know that $G$ is a semi-positive symmetric bilinear form by formula . Inequality then shows that $G$ is in fact positive, so that it defines a Riemannian metric on $M$.
Let $K\in{\textup{End}}(TM)$ the $G$-antisymmetric operator defined by $$G(\cdot,\cdot)={\omega}(\cdot,K\cdot)\,.$$ Then there exists an orthonormal basis $\{e_j,f_j\}_{1\leq j\leq\dim M}$ of $TM$ such that $Ke_j=\alpha_j f_j$ and $Kf_j=-\alpha_je_j$, for $\alpha_j\geq 0$ for all $1\leq j\leq\dim M$. Define an almost complex structure $J\in{\textup{End}}(TM)$ by the formula $$Je_j=f_j~~\text{and}~~Jf_j=-e_j\,.$$ By definition, this almost complex structure is compatible with ${\omega}$, and $G$ is $J$-invariant. Set $$\label{g=G-om}
\rho(\cdot,\cdot):=G(\cdot,\cdot)-{\omega}(\cdot,J\cdot)\,.$$ We then need to show that for any $\xi\in TM$, we have $$\label{pos}
\rho(\xi,\xi)\geq 0\,.$$ But using , we know that $$G(\xi,\xi)=G(\xi,\xi)^{1/2}\,G(J\xi,J\xi)^{1/2}\geq{\omega}(\xi,J\xi)\,,$$ which readily implies by definition of $\rho$.
[Proof of (II):]{} Recall that the volume of an $\omega$-compatible metric is always equal to the symplectic volume ${\textup{Vol}}(M,{\omega})$. Then the statement (II) follows from the general form of an unsharpness metric $G$ given by formula .
[Proof of (III):]{} The construction below is a modification of the one in Example \[exam-2\]. Instead of dealing with the heat semigroup, which becomes elusive when the form $\rho$ is degenerate, we construct an explicit family of Markov kernels such that the desired quantization is the composition of the almost-Kähler quantization associated with $J$ from formula with the corresponding Markov operator. [^2] Let us pass to precise arguments.
All the estimates in the proof are meant uniformly in $x_0\in M$. Let $J\in{\textup{End}}(TM)$ be a compatible almost complex structure on $(M,{\omega})$ and let $\rho$ be a non-negative symmetric bilinear form on $TM$. Consider the Riemannian metric $g$ over $M$ defined by the formula $$g(\cdot,\cdot)={\omega}(\cdot,J\cdot)\,.$$ For any $t>0$, we define a smooth endomorphism of the tangent bundle $TM$ by the formula $$\label{A}
A_{t}:=t\left(-\pi J\rho_g J
+t\id\right)\in{\textup{End}}(TM)\,,$$ where $\rho_g\in{\textup{End}}(TM)$ is the non-negative symmetric endomorphism defined by $$\label{rhog}
g(\rho_g\cdot,\cdot)=\rho\,.$$ Then $A_{t}$ is positive symmetric with respect to $g$, for all $t>0$.
Let $\epsilon>0$ be smaller than the injectivity radius of $(X,g)$. For any $x_0\in X$, consider an isometric identification $(T_{x_0}X,g)\simeq({{\mathbb{R}}}^{2d},\langle\cdot,\cdot\rangle)$, where $\langle\cdot,\cdot\rangle$ is the standard Euclidean product of ${{\mathbb{R}}}^{2d}$, and let $Z=(Z_1,\cdots Z_{2d})\in{{\mathbb{R}}}^{2d}$ be the induced normal coordinates on the geodesic ball $B(x_0,\epsilon)\subset X$ of radius $\epsilon$ centered at $x_0$. We write $dZ$ for the Lebesgue measure on ${{\mathbb{R}}}^{2d}$. Let $\varphi:[0,+\infty)\rightarrow[0,1]$ be a smooth function identically equal to $1$ over $[0,\epsilon/2)$ and to $0$ over $[\epsilon,+\infty)$. We define an operator $K_t^\rho$ acting on $f\in C^\infty(X,{{\mathbb{R}}})$ by the following formula in normal coordinates around $x_0\in X$, $$\label{Krhodef}
K_t^\rho f\,(x_0):=\frac{1}{\alpha_{t}(x_0)}\int_{B(x_0,\epsilon)}
\varphi(|Z|)f(Z)\,e^{-\pi\left\langle A_{t}^{-1}Z,Z\right\rangle}
\,dZ\,,$$ where $\alpha_{t}(x_0):={\displaystyle{\int}}_{B(x_0,\epsilon)}\varphi(|Z|)\,
e^{-\pi\left\langle A_{t}^{-1}Z,Z\right\rangle}
\,dZ$ is chosen so that $K_t 1\equiv 1$ for all $t>0$. Note that $f\geq 0$ implies $K_tf\geq 0$ for all $t>0$.
Fix $x_0\in X$, and consider the isometric identification $(T_{x_0}X,g)\simeq({{\mathbb{R}}}^{2d},\langle\cdot,\cdot\rangle)$ in which $A_t$ is diagonal, so that using definition , we can write $$\label{Adiag}
A_{t,x_0}=\textup{diag}\big(t(\lambda_1+t),\cdots,t(\lambda_{2d}+t)
\big)\,,$$ where $\{\lambda_j\geq 0\}_{1\leq j\leq 2d}$ are the eigenvalues of $-\pi J\rho_gJ$ over $T_{x_0}X$. Using the multi-index notation $\alpha=(\alpha_1,\cdots,\alpha_{2d})\in{{\mathbb{N}}}^{2d}$, we will use the following Taylor expansion of $f$ up to order $4$ as $|Z|\rightarrow 0$, $$\label{Taylorfk}
\begin{split}
f&(Z) =\sum_{0\leq|\alpha|\leq 3}
\frac{\partial^{|\alpha|} f}{\partial Z^\alpha}(x_0)\frac{Z^\alpha}{\alpha!}
+O(|Z|^4)\|f\|_{C^{4}}\,.\\
\end{split}$$ On the other hand, using the change of variables $Z_j\mapsto
Z_j/t^{1/2}(\lambda_j+t)^{1/2}$ for each $1\leq j\leq 2d$ and the exponential decrease of the Gaussian function, we get a constant $\delta>0$ for any $\alpha\in{{\mathbb{N}}}^{2d}$ such that the following estimate holds as $t\rightarrow 0$, $$\label{exptrick}
\begin{split}
&\int_{B(x_0,\epsilon)}
\varphi(|Z|)Z^{\alpha}\,e^{-\pi\left\langle A_{t}^{-1}Z,Z\right\rangle}
\,dZ=\int_{{{\mathbb{R}}}^{2d}}Z^{\alpha}\,
e^{-\pi\sum_{j=1}^{2d}\left(t^{-1}(\lambda_j+t)^{-1}Z_j^2\right)}
dZ\\
&\quad\quad\quad\quad-\int_{{{\mathbb{R}}}^{2d}}(1-\varphi(|Z|))Z^{\alpha}\,
e^{-\pi\sum_{j=1}^{2d}\left(t^{-1}(\lambda_j+t)^{-1}Z_j^2\right)}
\,dZ\\
&=\prod_{j=1}^{2d}t^{1/2}(\lambda_j+t)^{1/2}
\left(t(\lambda_j+t)\right)^{\alpha_j/2}
\int_{{{\mathbb{R}}}^{2d}}Z^{\alpha}\,
e^{-\pi|Z|^2}dZ+O(e^{-\delta/t})\,.
\end{split}$$ Note that we can then explicitly evaluate the integral in the last line of using basic properties of the Gaussian function, and it vanishes as soon as there is an odd monomial inside $Z^{\alpha}$. Then considering the Taylor expansion inside the right hand side of equation and using the estimate , we get as $t\rightarrow 0$, $$\begin{gathered}
\label{comput1}
K_t^\rho f\,(x_0)=f(x_0)\\
+\frac{\prod_{k=1}^{2d}t^{1/2}(\lambda_k+t)^{1/2}}{\alpha_{t}(x_0)}\,
\left(\sum_{j=1}^{2d}\frac{t(\lambda_j+t)}{4\pi}
\frac{\partial^2 f}{\partial Z_j^2}(x_0)
+O(t^2)\,\|f\|_{C^{4}}\right)\,.\end{gathered}$$ On the other hand, it follows from the definition of $\alpha_t$ and the estimate that as $t\rightarrow 0$, we have $$\alpha_{t}(x_0)=\prod_{j=1}^{2d}t^{1/2}(\lambda_j+t)^{1/2}(1+
O(e^{-\delta/t}))\,.$$ Then we get from equation that as $t\rightarrow 0$, $$\label{comput2}
\begin{split}
K_t^\rho f\,(x_0)&=f(x_0)+t
\sum_{j=1}^{2d}\frac{\lambda_j+t}{4\pi}\,
\frac{\partial^2 f}{\partial Z_j^2}(x_0)
+O(t^2)\,\|f\|_{C^{4}}\\
&=f(x_0)+t
\sum_{j=1}^{2d}\frac{\lambda_j}{4\pi}\,
\frac{\partial^2 f}{\partial Z_j^2}(x_0)
+O(t^2)\,\|f\|_{C^{4}}\,.
\end{split}$$ Then writing $T_\hbar$ for the Berezin-Toeplitz quantization of $(M,{\omega},J)$, the quantization $T_\hbar^\rho$ defined for all $f\in C^\infty(X,{{\mathbb{R}}})$ by $$T_\hbar^\rho(f):=T_\hbar\left(K_\hbar^\rho
f\right)\,,$$ has unsharpness metric $G$ given by formula : in fact, for any $u,\,v\in C^\infty(X,{{\mathbb{R}}})$, writing $\nabla_g u,\,\nabla_g v$ for their gradient with respect to $g$ and in normal coordinates around $x_0\in X$ as above, we get from the last line of and following the computations of that the unsharpness cocycle $c_+^\rho$ associated with $T_\hbar^\rho$ satisfies $$\begin{split}
c_+^\rho(u,v)(x_0)&=-\frac{1}{2}\sum_{j=1}^{2d}\left(
\partial_j u(x_0)\,\partial_j v(x_0)
+\frac{\lambda_j}{\pi}\,\partial_j u(x_0)\,\partial_j v(x_0)\right)\\
&=-\frac{1}{2}\big(g_{x_0}(\nabla_g u,\nabla_g v)-
g_{x_0}(J\rho_gJ\nabla_g u,\nabla_g v)\big)\\
&=-\frac{1}{2}\big(g_{x_0}(\textup{sgrad}\,u,\textup{sgrad}\,v)+
\rho_{x_0}(\textup{sgrad}\,u,\textup{sgrad}\,v)\big)\,.
\end{split}$$ This shows that $G=g+\rho$, as required.
Case study: $SU(2)$ - equivariant quantizations {#sec-equiv}
===============================================
\[defin-equiv\][Two Berezin-Toeplitz quantizations $T_\hbar$ and $T_\hbar'$ with families of Hilbert spaces $\{H_\hbar\}$ and $\{H'_\hbar\}$, ${\hbar \in \Lambda}$, respectively, are called [*equivalent*]{} if there exists a sequence of unitary operators $U_\hbar: H_\hbar \to H'_{\hbar}$ such that for all $f\in C^\infty(M)$, $$\label{eq-equiv}
\|U_\hbar T_\hbar(f) U_\hbar^{-1} - T'_\hbar(f)\|_{op} = {\mathcal{O}}(\hbar^2)\;.$$ ]{}
Observe that if two quantizations are equivalent, their unsharpness metrics coincide. In this section we prove a converse statement in the context of $SU(2)$-equivariant quantizations of the two-dimensional sphere (see Section \[subsec-class\] below for further discussion). We consider the standard Kähler metric on the two-sphere $S^2$ normalized so that the total area equals $2\pi$. We denote by $L$ the line bundle dual to the tautological one, and by $H_k$ the $k+1$-dimensional space of holomorphic sections of its $k$-th tensor power $L^k$. One can identify $H_k$ with the space of homogeneous polynomials of two variables, so the group $SU(2)$ acts on $H_k$ via an irreducible unitary representation. Furthermore, $SU(2)$ acts on the space of Hermitian operators ${{\mathcal{L}}}(H_k)$ by conjugation. On the other hand the space $C^\infty(S^2)$ carries the natural action of $SU(2)$ by the change of variables. A quantization $Q_\hbar: C^\infty(S^2) \to {{\mathcal{L}}}(H_k)$, $\hbar\in\Lambda:=\{1/k\}_{k\in{{\mathbb{N}}}}$, is called [*$SU(2)$-equivariant*]{} if it intertwines the corresponding (real) representations. For instance, the standard Berezin-Toeplitz quantization $T_\hbar$ sending $f \in C^\infty(S^2)$ to the multiplication by $f$ followed by the orthogonal projection to the space of holomorphic sections is $SU(2)$-equivariant, and the same holds true for its images $T_\hbar^{(t)}$ under diffusion as defined in Example \[exam-2\]. Note that the quantizations $T_\hbar^{(t)}$ are pair-wise non-equivalent for different values of $t$ as the corresponding unsharpness metrics are different.
\[thm-equiv\] Every $SU(2)$-equivariant quantization of $S^2$ is equivalent to $T_\hbar^{(t)}$ for some $t \geq 0$.
[Step 1 (Applying Schur lemma):]{} Given any $SU(2)$-equivariant quantization $Q_\hbar$, pass to its complexification (denoted by the same letter) $$Q_\hbar: C^\infty(S^2,{{\mathbb{C}}}) \to {{\mathcal{L}}}(H_k) \otimes {{\mathbb{C}}}= H_k^* \otimes H_k\;.$$ On the one hand, $C^\infty(S^2,{{\mathbb{C}}})$ splits into the direct sum of irreducible summands $V_j$, $j=0,1,\dots$ corresponding to the eigenspaces of the Laplace-Beltrami operator associated to the Kähler metric with the eigenvalue $2j(j+1)$, with each $V_j$ isomorphic to $H_{2j}$ as an $SU(2)$-representation. On the other hand $$H_k^* \otimes H_k = H_{2k} \oplus H_{2k-2} \oplus \cdots \oplus H_0\;.$$ By the Schur Lemma, when $\hbar = 1/k$, we have that $Q_\hbar(V_j) \subset H_{2j}$ with respect to this decomposition, and furthermore there exists a constant $\alpha_{\hbar, j}\in{{\mathbb{C}}}$ such that, up to conjugation, we have $$\label{eq-QT}
Q_\hbar = (1+\alpha_{\hbar, j})T_\hbar \;\; \text{on}\;\; V_j\;.$$
[Step 2 (Legendre polynomials):]{} In what follows we introduce another parameter, $n \in {{\mathbb{N}}}$. We call a sequence $\{b_{\hbar,n}\}_{n\in{{\mathbb{N}}}}$ of the class ${\mathcal{O}}_N(\hbar^m)$ with $m,N \in {{\mathbb{N}}}$ if for some $c>0$ we have $|b_{\hbar,n}| \leq c\,\hbar^m (n+1)^N$ for all $n$. In the sequel, the dependence on $\hbar$ of such sequences will be made implicit. Denote by $P_n(z)$ the $n$-th Legendre polynomial considered as a function on the unit sphere $S^2 = \{x^2+ y^2 +z^2 =1\}$ lying in $V_n$. We write $\nabla$ for the gradient with respect to the standard metric on $S^2$ normalized so that the total area equals $2\pi$. The standard formulas for the Legendre polynomials (see e.g. formulas (43) and (44) in [@W]) readily yield, $$\label{Leg-1}
P_1P_n = q_n P_{n+1} + r_nP_{n-1},\;\; q_n=\frac{n+1}{2n+1},\;\; r_n= 1-q_n\;,$$ and $$\label{Leg-2}
(\nabla P_1, \nabla P_n) = s_n ( -P_{n+1} + P_{n-1}),\;\; s_n=\frac{2n(n+1)}{2n+1}\;.$$ We shall use that there exists $c>0$ such that $$\label{eq-Legendre-deriv}
\forall r \in {{\mathbb{N}}}\;\; \exists R\in {{\mathbb{N}}}\;: \|P_n\|_{C^r} \leq c(n+1)^R\;.$$ This (with $R = r$) follows immediately from the general result about the growth of $C^r$-norms of the Laplace-Beltrami eigenfunctions on Riemannian manifolds, see [@BinXu Corollary 1.1]. Using the fact that $\max_{x\in S^2} P_n=1$ by [@Ka Chapter 7, Theorem 17(i)], the norm correspondence property (P1), which holds uniformly in $C^N$-norm for some $N \in {{\mathbb{N}}}$, together with formula implies $$\|Q_\hbar(P_n)\|_{op} = 1-{\mathcal{O}}_N(\hbar),\;\; \|T_\hbar(P_n)\|_{op} = 1-{\mathcal{O}}_N(\hbar)\;.$$ Since $Q_\hbar(P_n) = (1+ \alpha_{\hbar,n})T_\hbar(P_n)$ by , it follows that $$\label{eq-alpha-O} \alpha_{\hbar,n} = {\mathcal{O}}_N(\hbar)\;.$$ In the course of the proof, we shall increase the value of $N$ according to our needs.
[Step 3 (Main calculation) :]{} Since $Q_\hbar$ is $SU(2)$-equivariant, the corresponding unsharpness metric equals $\mu$ times the standard one, for some constant $\mu \geq 1$. Thus the quasi-multiplicativity property (P3), which holds uniformly in $C^N$-norm for some $N\in{{\mathbb{N}}}$, together with formula yields $$\label{star-enhanced-3}
Q_\hbar(P_1)Q_\hbar(P_n) = Q_\hbar\left(P_1P_n- \frac{\mu}{2}\hbar(\nabla P_1,\nabla P_n) + {\mathcal{O}}_N(\hbar^2)\right)\;.$$ At the same time $$\label{star-enhanced-4}
T_\hbar(P_1)T_\hbar(P_n) = T_\hbar\left(P_1P_n- \frac{1}{2}\hbar(\nabla P_1,\nabla P_n) + {\mathcal{O}}_N(\hbar^2)\right)\;,$$ mind that here $\mu$ is replaced by $1$. By we have $$\label{eq-QT-1}
Q_\hbar(P_i) = (1+\alpha_{\hbar, i})T_\hbar(P_i)\;.$$ Identities and combined with , and enable us to express $T_\hbar(P_1)T_\hbar(P_n)$ as a linear combination of $T_\hbar(P_{n+1})$ and $T_\hbar(P_{n-1})$ in two different ways. The calculation is straightforward, and we obtain the result: $$\label{eq-vsp-PP-1}
A_nT_\hbar(P_{n+1})+ B_n T_\hbar(P_{n-1}) = A'_nT_\hbar(P_{n+1})+ B'_n T_\hbar(P_{n-1})+ {\mathcal{O}}_N(\hbar^2)\;,$$ for some $A_n,\,A_n'\in{{\mathbb{C}}}$ and $B_n,\,B_n'\in{{\mathbb{C}}}$, $n\in{{\mathbb{N}}}$, where $$B_n = (1+\alpha_{\hbar, 1})^{-1} (1+\alpha_{\hbar, n})^{-1}(1+\alpha_{\hbar,n-1})(r_n -\hbar\mu s_n/2)\;,$$ $$B'_n= r_n - \hbar s_n /2\;.$$ Projecting equation to the space $H_{2n-2}$ (which contains $T_\hbar(V_{n-1})$) and using that the operator norm of $T_\hbar(P_{n-1})$ is bounded away from zero (see Step 2), we get that $$B_n - B'_n = {\mathcal{O}}_N(\hbar^2)\;.$$ By using and explicit expressions for $q_n,r_n,s_n$ we get $$\label{eq-recur-2}
\alpha_{\hbar,n-1} - \alpha_{\hbar, n} - \alpha_{\hbar, 1} = (n+1)(\mu-1)\hbar + {\mathcal{O}}_N(\hbar^2)\;.$$ Substituting $n=1$ into we get that $$\alpha_{\hbar,1}= -(\mu-1)\hbar + {\mathcal{O}}_N(\hbar^2)\;.$$ Now we get a recursive formula $$\alpha_{\hbar,n} = \alpha_{\hbar,n-1} -n(\mu-1)\hbar + {\mathcal{O}}_N(\hbar^2)\;.$$ Noticing that $(n+1)\,{\mathcal{O}}_N(\hbar^2)= {\mathcal{O}}_{N+1}(\hbar^2)$ and redefining $N \mapsto N+1$ we conclude that $$\label{eq-alpha-complete}
\alpha_{\hbar, n} = - \frac{n(n+1)}{2}(\mu-1)\hbar + {\mathcal{O}}_N(\hbar^2)\;.$$
[Step 4 (Finale) :]{} Recall that $2n(n+1)$ is the eigenvalue of the Laplacian corresponding to the eigenspace $V_n$. Let $V = \oplus_{n=0}^\infty V_n$ be the space of all finite linear combinations of spherical harmonics. By norm correspondence (P1) and formula , for every $\phi_n \in V_n$ we have $$Q_\hbar(\phi_n) = \left(1- \frac{n(n+1)}{2}(\mu-1)\hbar \right) T_\hbar (\phi_n) + {\mathcal{O}}_N(\hbar^2)\,\|\phi_n\|_{C^N}$$ $$=T_\hbar(e^{-t\hbar\Delta} \phi_n) + {\mathcal{O}}_N(\hbar^2)\,\|\phi_n\|_{C^N} = T^{(t)}_\hbar(\phi_n)+ {\mathcal{O}}_N(\hbar^2)\,\|\phi_n\|_{C^N}$$ with $t = (\mu-1)/4$ in Example \[exam-2\].
Take now any $f \in C^\infty(S^2)$, and decompose it by spherical harmonics: $f = \sum_n \phi_n$. Since $f$ is smooth, the $C^N$-norms $\|\phi_n\|_{C^N}$ decay faster than any power of $n$ as $n\to+\infty$, so that $$\|Q_\hbar(f)- T_\hbar^{(t)}(f)\|_{op} \leq c\sum_{n\in{{\mathbb{N}}}} n^N \|\phi_n\|_{C^N}\hbar^2 \leq c'\hbar^2\;.$$ This shows that the quantizations $Q_\hbar$ and $T_\hbar^{(t)}$ are equivalent.
The unsharpness cocycle is of order $(1,1)$ {#sec-order-proof}
===========================================
In this Section we prove Theorem \[thm-order\]
For every $d\in{{\mathbb{N}}}$, we use the standard multi-index notation $\alpha=(\alpha_1,\cdots,\alpha_d)\in{{\mathbb{N}}}^d$, where for any sequence of symbols $x_1,\cdots, x_d$, we write $x_\alpha:=x_{\alpha_1}\cdots x_{\alpha_d}$, so that in particular $\alpha!:=\alpha_1!\cdots\alpha_d!\in{{\mathbb{N}}}$, and write $|\alpha|:=\alpha_1+\cdots+\alpha_d$.
Note that by , to show that $c_+$ is a bi-differential operator of bi-degree $(1,1)$, we need to show that the differential operator $a$ contains only terms of order $1$ and $2$. Note that $T_\hbar(1)=\id$ implies $c_+(1,1)=0$, so that $a$ cannot contain terms of order $0$. Let us show that $a$ cannot be of order $k>2$.
Assume by contradiction that $a$ is of order $k>2$. Let $x_0\in X$ be the center of local coordinates $(Z_1,\cdots,Z_{2n})\in U\subset{{\mathbb{R}}}^{2n}$ be such that for all $f\in C^\infty(X,{{\mathbb{R}}})$, $$\label{Ldef}
af(x_0)=\sum_{1\leq|\alpha|\leq k}a_\alpha\frac{\partial^{|\alpha|}f}
{\partial Z^\alpha}(x_0)\,,$$ where the sequence $\{a_\alpha\in{{\mathbb{R}}}\}_{1\leq|\alpha|\leq k}$ is such that $a_\beta\neq 0$ for some $\beta=(\beta_1,\cdots,\beta_{2n})\in{{\mathbb{N}}}^{2n}$ of length $|\beta|=k$. Fix $1\leq j\leq 2n$ such that $\beta_j\neq 0$, and writing $\hat\beta=(\beta_1,\cdots,\beta_j-1,\cdots,
\beta_{2n})\in{{\mathbb{N}}}^{2n-1}$, take $f\in C^{\infty}(X,{{\mathbb{R}}})$ satisfying $$\label{fcoord}
f(Z)=\frac{c}{\hat\beta!}Z^{\hat\beta}+Z_j\,,$$ for $Z\in U\subset{{\mathbb{R}}}^{2n}$ in the coordinates around $x_0\in X$ considered above and for some $c\in{{\mathbb{R}}}$ to be fixed later. Then this function $f$ and all its derivatives vanish at $x_0\in X$, except for $$\label{fcoord-1}
\frac{\partial^{|\hat\beta|}f}
{\partial Z^{\hat\beta}}(x_0)=c\quad\quad\text{and}
\quad\quad\frac{\partial f}{\partial Z_j}(x_0)=1\,.$$ Then by equations and , considering the multi-index $\gamma\in{{\mathbb{N}}}^{2n}$ of length $|\gamma|=2$ such that $\gamma_j=2$, one gets that for any $f\in C^\infty(X,{{\mathbb{R}}})$ satisfying , $$\begin{split}
c_+(f,f)(x_0)&=2a_\beta\,\frac{\partial^{|\hat\beta|}f}
{\partial Z^{\hat\beta}}(x_0)\,\frac{\partial f}{\partial Z_j}(x_0)
+2a_\gamma\frac{\partial f}{\partial Z_j}(x_0)\frac{\partial f}{\partial Z_j}(x_0)\\
&=2a_\beta c+2a_\gamma\,.
\end{split}$$ Thus if $f\in C^\infty(X,{{\mathbb{R}}})$ satisfies for $c\in{{\mathbb{R}}}$ such that $\text{sign}(a_\beta)c > -a_\gamma/|a_\beta|$, we get that $c_+(f,f)(x_0)> 0$. This contradicts the fact that $c_+(f,f)\leq 0$ for all $f\in C^\infty(X,{{\mathbb{R}}})$, which holds for every Berezin-Toeplitz quantization by the semi-positivity property .
Discussion and questions {#sec-disc}
========================
Historical remarks on unsharpness {#subsec-ger}
---------------------------------
The unsharpness cocycle appeared in earlier literature which, to the best of our knowledge, focussed on its elimination, of course, by the price of losing the positivity of a quantization. Let us elaborate on this point. Assume $(M,\omega)$ is a quantizable symplectic manifold equipped with a compatible almost-Kähler structure. Consider the induced Berezin-Toeplitz quantization of Theorem \[BQ\]. Using the notations of Example \[exam-2\], define for any $f\in C^\infty(M)$ and $\hbar\in\Lambda=\{1/k\}_{k\in{{\mathbb{N}}}}$, $$\label{metaKS}
Q_\hbar(f) := T_\hbar\left(f+\frac{\hbar}{4} \Delta f\right)\;.$$ This gives rise to a collection of maps $Q_\hbar:C^\infty(M)\to{{\mathcal{L}}}(H_\hbar)$ parametrized by $\hbar\in\Lambda$ and satisfying the axioms (P1)-(P4) of Definition \[BTdef\], but which does not preserve positivity, so that they do not come from a POVM construction via formula . Then following the computation in Example \[exam-1\], we see that the associated unsharpness cocycle $c_+^Q$, defined from the quasi-multiplicativity property (P3) as in the beginning of the section, satisfies $$\label{c+=0}
c_+^Q(f,g)=0\,,$$ for all $f,\,g\in C^\infty(M)$. As noted for instance by Charles in [@Cha §1.4] [^3], the quantization is, up to twisting with a line bundle, the *metaplectic Kostant-Souriau quantization*, which possesses remarkable sub-principal properties, a fact which is explained conceptually by the vanishing unsharpness property .
In the flat case $M={{\mathbb{C}}}$ with the standard symplectic form, Gerstenhaber considers in [@G] deformation quantizations parametrized by $\lambda\geq 0$ which, up to the second order in $\hbar$, correspond to the quantization maps parametrized by $\hbar>0$ defined for any smooth function $f: {{\mathbb{C}}}\to {{\mathbb{R}}}$ of polynomial growth by $$\label{Gerquant}
Q_\hbar^{(\lambda)}(f) := T_\hbar\left(f+ \frac{1-\lambda}{2}\hbar
\Delta f\right)\;.$$ Here $T_\hbar$ is the standard Toeplitz quantization of ${{\mathbb{C}}}$, sending $f$ to the multiplication by $f$ followed by the orthogonal $L_2$-projection on the space of holomorphic functions which are square integrable with respect to a Gaussian measure. Gerstenhaber formulates a *least uncertainty principle* for deformation quantization, which implies in particular that unsharpness vanishes on the classical harmonic oscillator. He then shows that the quantization satisfies this least uncertainty principle if $\lambda=1/2$, which corresponds to the flat version of the quantization .
Note that in the flat case $M={{\mathbb{C}}}$, the classical harmonic oscillator is a sum of squares of the coordinate functions. On the other hand, the quasi-multiplicativity property (P3) implies that for all $f\in C^\infty(M)$ as $\hbar\to 0$, $$T_\hbar(f)^2-T_\hbar(f^2)=\hbar T_\hbar(c_+(f,f))+{\mathcal{O}}(\hbar^2)\,.$$ We then see that unsharpness measures in particular the deviation of the quantum harmonic oscillator, defined as a sum of squares of the quantum coordinate operators, from the quantization of the classical harmonic oscillator. This explains in particular the standard justification of the metaplectic correction, as giving the “correct” quantum harmonic oscillator on flat space.
Least unsharpness surfaces and pseudo-holomorphic curves
--------------------------------------------------------
Let $G$ be the unsharpness metric associated to a Berezin-Toeplitz quantization of a closed symplectic manifold $(M,\omega)$ (see Section \[sec-main\]). [*A least unsharpness surface*]{} $\Sigma \subset M$ is a two-dimensional oriented submanifold with $\text{Area}_G (\Sigma) = \int_\Sigma \omega$. Repeating the the proof of Theorem \[thm-main-new-1\] we see that for such surfaces, the restriction of the Riemannian area form coincides with the restriction of the symplectic form. If $G$ has the minimal possible total unsharpness and hence by Theorem \[thm-main-new-1\](II) comes from some compatible almost-complex structure $J$ on $M$, the least unsharpness surfaces in $M$ are $J$-holomorphic curves (cf. [@S]). For instance, for the complex projective plane $M= {{\mathbb{C}}}P^2$, Gromov’s theory of pseudo-holomorphic curves predicts that for every compatible $J$, through every two distinct points $A,B \in M$ passes unique such curve $\Sigma$ in the homology class of $[{{\mathbb{C}}}P^1]$.
It is enticing to interpret $\Sigma$ as a worldsheet of the topological string theory describing a path joining constant loops $A$ and $B$. Note that the metric $G$ on our “space-time" $M$ is canonically associated to a Berezin-Toeplitz quantization of $M$, and the “total unsharpness " $\text{Area}_G (\Sigma)$ of a worldsheet $\Sigma$ is nothing else but the Nambu-Goto action up to a multiplicative constant. If the total unsharpness of $(M,G)$ is minimal possible, i.e., coincides with the symplectic volume of $M$, the least unsharpness surfaces are $J$-holomorphic curves for a compatible almost complex structure $J$ defining $G$, and hence represent “worldsheet instantons". Does there exist an interpretation of this picture in physical terms?
On classification of Berezin-Toeplitz quantizations {#subsec-class}
---------------------------------------------------
We conclude the paper with a discussion on classification of Berezin-Toeplitz quantizations up to equivalence in the sense of Definition \[defin-equiv\]. In Section \[sec-equiv\] we classified $SU(2)$-equivariant quantizations of the two-dimensional sphere. It would be interesting to extend this to equivariant quantizations for more general co-adjoint orbits equipped with the canonical symplectic structure. In the general (not necessarily equivariant) case, the problem is widely open.
In fact, establishing (non)-equivalence of quantizations is a non-trivial problem even for the Kähler quantizations of Theorem \[BQ\], where the holomorphic line bundles defining the quantization of $(M,{\omega})$ could be non-isomorphic. For instance, their Chern classes could differ by torsion even though the associated spaces of holomorphic sections have same dimension. Are the corresponding quantizations equivalent?
Another interesting example is as follows. According to Remark \[rem-rho\], there exist metrics $G$ on $M$ admitting different decompositions of the form . Each such decomposition determines a Berezin-Toeplitz quantization given by almost-Kähler quantization followed by diffusion, as explained in the proof of Theorem \[thm-main-new-1\](III). Are the quantizations corresponding to different decompositions of the same metric equivalent?
Let us address the question about invariants of quantizations with respect to equivalence. In addition to the unsharpness metric, there is another invariant coming from the trace correspondence, see item (P4) in Definition \[BTdef\]. Recall that the latter states that $${{\rm tr}}(T_\hbar(f)) = (2\pi\hbar)^{-d}\int_M f\,R_\hbar\, d\mu\;,$$ where $\dim M = 2d$, the function $R_\hbar\in C^\infty(M)$ satisfies $R_\hbar = 1+ {\mathcal{O}}(\hbar)$, and $d\mu = \frac{\omega^d}{d!}$ is the symplectic volume on $M$. Roughly speaking, since the trace is invariant under conjugation, the convergence rate of the sequence of differential forms $R_\hbar d\mu$ to the symplectic volume $d\mu$ as $\hbar \to 0$ does not change up to ${\mathcal{O}}(\hbar^2)$ under equivalence. For the sake of simplicity, let us, until the end of the paper, enhance axiom (P4) by assuming that there exists a function $r \in C^\infty(M)$ such that $$R_\hbar = 1 + \hbar\,r + {\mathcal{O}}(\hbar^2)\;.$$ We shall refer to the form $rd\mu$ as [*the Rawnsley form*]{}. Thus, [*equivalent quantizations possess the same Rawnsley form*]{}. Put $$\langle r \rangle := \text{Vol}(M)^{-1} \int_M r d\mu\;.$$
[Substituting $f=1$ into (P4), we get that $\langle r \rangle $ appears in the dimension formula $$\dim H_\hbar = \text{Vol}(M)(2\pi \hbar)^{-d}\left(1 +
\hbar\,\langle r \rangle + {\mathcal{O}}(\hbar^2)\right)\;.$$ Let us mention that for Kähler quantizations, an alternative asymptotic expression for the dimension of $H_\hbar$ is given by the Hirzebruch-Riemann-Roch theorem. Comparing coefficients at $\hbar$ one gets a simple topological interpretation of $\langle r \rangle$: $$\langle r \rangle = 2\pi\,\frac{\left\langle [\omega]^{d-1}\cup c_1(TM), [M]\right\rangle}{2(d-1)! \text{Vol}(M)} \;,$$ where $c_1(TM)$ stands for the first Chern class of $M$. ]{}
\[exam-rawn\] [Let $v$ be a vector field on the manifold $M$ generating a flow $\phi_t$. Given a Berezin-Toeplitz quantization $T_\hbar$ on $M$, define a new quantization by setting $T^{(v)}_{\hbar}(f) : = T_\hbar(f\circ \phi_{-\hbar})$. A direct calculation based on the expansion $T^{(v)}_{\hbar}(f) = T_\hbar\left(f -\hbar\,df(v)\right) + {\mathcal{O}}(\hbar^2)$ shows that this is a Berezin-Toeplitz quantization whose unsharpness metric coincides with the one of $T_\hbar$, and whose Rawnsley form is given by $(r + \text{div} (v))d\mu$, where $\text{div} (v)$ stands for the divergence of $v$ with respect to the symplectic volume. In particular, it follows that by choosing an appropriate vector field $v$, one can always achieve the Rawnsley form being equal to $\langle r \rangle$. ]{}
\[question-3\] [Consider a pair of quantizations with the Hilbert spaces of the same dimension. Suppose that their unsharpness metrics and the Rawnsley forms coincide. Are these quantizations equivalent?]{}
The answer in the general (not necessarily equivariant) case is at the moment unclear.
[**Acknowledgement.**]{} We thank Pavel Etingof for useful references. We are grateful to Jordan Payette for pointing out a number of inaccuracies and helpful suggestions on the presentation.
[a]{}
Anandan, J., and Aharonov, Y., [*Geometry of quantum evolution,*]{} Phys. Rev. Letters, [**65**]{} (1990), p.1697.
Alekseev, A., and Lachowska, A., [*Invariant $*$-products on coadjoint orbits and the Shapovalov pairing,*]{} Comment. Math. Helv. [**80**]{} (2005), 795–810.
Ashtekar, A., and Schilling, T.A., *Geometrical Formulation of Quantum Mechanics,*, On Einstein’s Path, 23–65, Harvey A., Springer, New York, NY, 1999.
Bayen, F., Flato, M., Fronsdal, C., Lichnerowicz, A., and Sternheimer, D., [*Quantum mechanics as a deformation of classical mechanics,*]{} Lett. Math. Phys. [**1**]{} (1977), no. 6, 521–530.
Berezin, F., [*General concept of quantization,*]{} Comm. Math. Phys. [**40**]{} (1975), 153–174.
Berline, N., Getzler, E., and Vergne, M., *Heat kernels and [D]{}irac operators*, Grundlehren Text Editions, Springer-Verlag, Berlin, 2004, Corrected reprint of the 1992 original.
Bertelson, M., Bieliavsky, P., and Gutt, S., [*Parametrizing equivalence classes of invariant star products,*]{} Lett. Math. Phys. [ **46**]{} (1998), 339–345.
Bin, X., [*Derivatives of the spectral function and Sobolev norms of eigenfunctions on a closed Riemannian manifold,*]{} Ann. Global Anal. Geom. [**26**]{} (2004), 231–252.
Bordemann, M., Meinrenken, E., and Schlichenmaier, M., [*Toeplitz quantization of Kähler manifolds and $gl(N), N \to \infty$ limits,*]{} Comm. Math. Phys. [**165**]{} (1994), 281–296.
Borthwick, D., and Uribe, A., [*Almost complex structures and geometric quantization,*]{} Math. Res. Lett. [**3**]{} (1996), 845–861.
Boutet de Monvel, L., and Guillemin, V., *The spectral theory of [T]{}oeplitz operators*, Annals of Mathematics Studies, vol. 99, Princeton University Press, Princeton, NJ; University of Tokyo Press, Tokyo, 1981.
Busch, P., Lahti, P., Pellonpää, J.P., and Ylinen, K., [*Quantum Measurement,*]{} Springer, 2016.
Charles, L., *Symbolic calculus for [T]{}oeplitz operators with half-form*, J. Symplectic Geom. **4** (2006), no. 2, 171–198.
Charles, L., *Quantization of compact symplectic manifolds*, J. Geom. Anal. **26** (2016), no. 4, 2664–2710.
Charles, L., and Polterovich, L., [*Sharp correspondence principle and quantum measurements,*]{} St. Petersburg Math. J., [**29**]{} (2018) 177 – 207.
Chiribella, G., D’Ariano, G.M., and Schlingemann, D., [*How continuous quantum measurements in finite dimensions are actually discrete,*]{} Phys. Rev. Lett. [**98**]{} (2007), no. 19, 190403.
Dirac, P.A.M., [*The Principles of Quantum Mechanics,*]{} Oxford University Press.
Gerstenhaber, M., [*Least uncertainty principle in deformation quantization,*]{} J. Math. Phys. [**48**]{} (2007), no. 2, 022103, 15 pp.
Gromov, M., [*Pseudo holomorphic curves in symplectic manifolds,*]{} Invent. Math. [**82**]{} (1985), 307–347.
Guillemin, V., [*Star products on compact pre-quantizable symplectic manifolds,*]{} Lett. Math. Phys. [**35**]{} (1995), 85–89.
Hayashi, M., [*Quantum Information Theory,*]{} Graduate Texts in Physics, Springer. 2017.
Ioos, L., *On the composition of [B]{}erezin-[T]{}oeplitz operators on symplectic manifolds*, Math. Z. **290** (2018), no. 1-2, 539–559.
Ioos, L., Lu, W., Ma, X., and Marinescu, G., *[Berezin-Toeplitz quantization for eigenstates of the Bochner-Laplacian on symplectic manifolds]{}*, J. Geom. Anal (2018), 32pp, DOI 10.1007/s12220-017-9977-y.
Ioos, L., Kaminker, V., Polterovich, L., and Shmoish, D.,[*Spectral aspects of the Berezin transform,*]{} arXiv:1811.03155, 2018, to appear in Ann. H. Lebesgue.
Janssens, B., [*Unifying decoherence and the Heisenberg principle,*]{} Lett. Math. Phys. [**107**]{} (2017), 1557–1579.
Kaplan, W., [*Advanced Calculus*]{}, 5th edition.
Karabegov, A.V., and Schlichenmaier, M., [*Identification of Berezin-Toeplitz deformation quantization,*]{} J. Reine Angew. Math. [**540**]{} (2001), 49–76.
Klauder, J.R., [*Quantization is geometry, after all,*]{} Ann. Physics [**188**]{} (1988), 120–141.
Ma, X., and Marinescu, G., [*Toeplitz operators on symplectic manifolds,*]{} J. Geom. Anal. [**18**]{} (2008), 565–611. Ozawa, M., [*Quantum limits of measurements and uncertainty principle,*]{} in [*“Quantum Aspects of Optical Communications",*]{} 1991, pp. 1–17. Springer, Berlin, Heidelberg.
Polterovich, L., and Rosen, D., [*Function theory on symplectic manifolds,*]{} CRM Monograph Series, 34. American Mathematical Society, Providence, RI, 2014.
Rawnsley, J., and Gutt, S. *Equivalence of star products on a symplectic manifold: An introduction to Deligne’s Cech cohomology classes*, J. Geom. Phys. 29, no. 4 (1999), 347–392.
Sanborn, B.A. [*The uncertainty principle and the energy identity for holomorphic maps in geometric quantum mechanics,*]{} arXiv:1710.09344 (2017).
Schlichenmaier, M., *Deformation quantization of compact [K]{}ähler manifolds by [B]{}erezin-[T]{}oeplitz quantization*, Conférence [M]{}oshé [F]{}lato 1999, [V]{}ol. [II]{} ([D]{}ijon), Math. Phys. Stud., vol. 22, Kluwer Acad. Publ., Dordrecht, 2000, pp. 289–306.
Schlichenmaier, M., [*Berezin-Toeplitz quantization for compact Kähler manifolds. A review of results,*]{} Adv. Math. Phys. (2010), 927280.
Shiffman, B., and Zelditch, S., *Asymptotics of almost holomorphic sections of ample line bundles on symplectic manifolds*, J. Reine Angew. Math. **544** (2002), 181–222.
Tian, G., *On a set of polarized Kähler metrics on algebraic manifolds,* J. Differential Geom., **32** (1990), 99–130.
Weisstein, E.W., [*Legendre Polynomial*]{}, From MathWorld–A Wolfram Web Resource.\
<http://mathworld.wolfram.com/LegendrePolynomial.html>
Xu, H., [*An explicit formula for the Berezin star product,*]{} Lett. Math. Phys. [**101**]{} (2012), 239 – 264.
<span style="font-variant:small-caps;">School of Mathematical Sciences, Tel Aviv University, Ramat Aviv, Tel Aviv 69978 Israel</span>
*E-mail address*: `[email protected]`
<span style="font-variant:small-caps;">Einstein Institute of Mathematics, Hebrew University, Givat Ram, Jerusalem, 91904, Israel</span>
*E-mail address*: `[email protected]`
<span style="font-variant:small-caps;">School of Mathematical Sciences, Tel Aviv University, Ramat Aviv, Tel Aviv 69978 Israel</span>
*E-mail address*: `[email protected]`
[^1]: Our convention for the Poisson bracket is $\{f,g\}:=-{\omega}(\text{sgrad} f,\text{sgrad} g)$ for all $f,\,g\in C^\infty(M)$, where $\text{sgrad} f$ is the Hamiltonian vector field of $f$ defined by $\iota_{\text{sgrad} f}{\omega}+df=0$.
[^2]: In the language of quantum measurement theory, the POVM of the quantization constructed below is a smearing of the Berezin-Toeplitz POVM of Theorem \[BQ\] by the explicitly constructed Markov operator.
[^3]: [@Cha] uses the holomorphic Laplacian, which is half the Laplace-Beltrami operator.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'High time resolution radio surveys over the last few years have discovered a population of millisecond-duration transient bursts called Fast Radio Bursts (FRBs), which remain of unknown origin. FRBs exhibit dispersion consistent with propagation through a cold plasma and dispersion measures indicative of an origin at cosmological distances. In this paper we perform Monte Carlo simulations of a cosmological population of FRBs, based on assumptions consistent with observations of their energy distribution, their spatial density as a function of redshift and the properties of the interstellar and intergalactic media. We examine whether the dispersion measures, fluences, derived redshifts, signal-to-noise ratios and effective widths of known FRBs are consistent with a cosmological population. Statistical analyses indicate that at least 50 events at Parkes are required to distinguish between a constant comoving FRB density, and a FRB density that evolves with redshift like the cosmological star formation rate density.'
author:
- |
M. Caleb$^{1,2,3}$[^1], C. Flynn$^{2,3}$, M. Bailes$^{2,3}$, E.D. Barr$^{2,3}$, R.W. Hunstead$^{4}$ E.F. Keane$^{5,2,3}$, V. Ravi$^{2}$, W. van Straten$^{2}$\
\
$^{1}$ Research School of Astronomy and Astrophysics, Australian National University, ACT, 2611, Australia\
$^{2}$ Centre for Astrophysics and Supercomputing, Swinburne University of Technology, P.O. Box 218, Hawthorn, VIC 3122, Australia\
$^{3}$ ARC Centre of Excellence for All-sky Astrophysics (CAASTRO)\
$^{4}$ Sydney Institute for Astronomy (SIfA), School of Physics, The University of Sydney, NSW 2006, Australia\
$^{5}$ SKA Organisation, Jodrell Bank Observatory, Cheshire, SK11 9DL, UK
bibliography:
- 'Arxiv.bib'
title: 'Are the distributions of Fast Radio Burst properties consistent with a cosmological population? '
---
\[firstpage\]
general – cosmology : intergalactic medium – pulsars : general
Introduction
============
Fast Radio Bursts (FRBs) are bright (few Jy), radio pulses occurring with time-scales of order milliseconds. Eighteen bursts have been discovered to date (@Lorimer [-@Lorimer]; @Thornton [-@Thornton]; @Spitler [-@Spitler]; @Spolaor-Bannister [-@Spolaor-Bannister]; @Petroff_FRB [-@Petroff_FRB]; @Ravi [-@Ravi]; @Champion [-@Champion]; @Masui [-@Masui]; Keane et al. in prep; Ravi et al. in prep) The integrated electron densities along the lines of sight to these bursts (called dispersion measures, or DMs) lie in the range of 375 to 1600 $\mathrm{pc\,cm^{-3}}$. This is greatly in excess of the expected contribution from the Galaxy via the Interstellar Medium (ISM) [@Cordes] along such lines-of-sight, which typically lie in the range of 20 to 50 $\mathrm{pc\,cm^{-3}}$.
For all FRBs discovered to date, the arrival time delay associated with the dispersion closely follows a $\mathrm{\nu^{-2}}$ frequency dependence, and the pulse width evolution follows a $\mathrm{\nu^{-4}}$ frequency dependence for those FRBs where the signal-to-noise ratio (S/N) has permitted frequency-dependent width measurements [@Thornton]. Both properties are consistent with propagation through a sparse, non-relativistic plasma.
A few years after the publication of the first FRB [@Lorimer] another population of sources (dubbed Perytons) was identified at the Parkes 64m radio telescope that were clearly not of extra-terrestrial origin. The Perytons [@Burke-spolaor] also show swept-frequency properties, although they tend to be broader, mimic interstellar scattering and typically occur during meal-times onsite. Unlike the FRBs, the Perytons appeared in all 13 beams of the Parkes multibeam receiver, indicative of a terrestrial origin. The Perytons were ultimately shown to be originating from improperly shielded microwave ovens [@Petroff_peryton]. Given their large DMs, [@Lorimer] and [@Thornton] have proposed that FRBs lie at cosmological distances, and that their DMs are dominated by propagation through the Intergalactic Medium (IGM), with minor contributions from the ISM in the Milky Way and the ISM in a putative host galaxy. Redshift estimates to the sources are made by subtracting the ISM component of the DM, and ascribing the rest to propagation through the IGM for which the electron density is available from cosmological models. For the 15 published FRBs, the redshift estimates are in the range $0.2 < z < 1.5$, firmly placing the sources at cosmological distances. As the IGM is thought to contain 90% of the Universe’s baryons, [e.g @Fukugita; @Savage], measuring the DMs of FRBs at high redshifts is potentially a novel way to probe this important cosmological component. Furthermore, if placed at such distances, the unbeamed (isotropic) energies of the observed FRBs lie in the range $\mathrm{10^{31}}$ to $\mathrm{10^{33}}$ J [@KeanePetroff]. The observed FRBs also have brightness temperatures well in excess of thermal emission ($\mathrm{T_{b}} > 10^{33}$ K), strongly suggesting coherent emission [@Katz; @Luan].\
Four events were found by [@Thornton] in the high-latitude component of the High Time Resolution Universe (HTRU) survey at Parkes [@Keith]. From these events, a rate of $\gtrsim
1.0^{+0.6}_{-0.5}\times10^4$ events $\mathrm{sky^{-1}}$ $\mathrm{day^{-1}}$ was estimated. If the redshifts ascribed to the bursts are valid, the volumetric rate to which this corresponds is $\sim 2 \times 10^4$ events $\mathrm{Gpc^{-3} yr^{-1}}$, which is similar to the volumetric rate for Soft Gamma-ray Repeaters (SGRs) ($<
2.5 \times 10^4$ events $\mathrm{Gpc^{-3} yr^{-1}}$), and within an order of magnitude of the volumetric rate of core collapse (Type II) supernovae ($\sim 2 \times 10^5$ events $\mathrm{Gpc^{-3} yr^{-1}}$) [@Kulkarni].
A cosmological origin for the excess DM of FRBs is not the only possibility, as FRBs could be Galactic objects in high electron density environments which electron density models for the Milky Way do not capture. This has been discussed by [@Loeb], who propose FRBs originate from low mass main sequence “flare stars”. No consensus has emerged regarding the progenitors of FRBs no matter whether Galactic or extra-Galactic, with possibilities including flare stars [@Loeb] (Galactic) and extra-Galactic sources such as annihilating blackholes [@Keaneburst], giant flares from soft gamma-ray repeaters [@Popov; @Thornton; @Lyubarsky], binary white dwarf mergers [@Kashiyama], neutron star mergers [@Totani], collapsing supramassive neutron stars [@Falcke], radio emission from pulsar companions [@Zarka], dark matter induced collapse of neutron stars [@Fuller] and the radio emission from pulsars [@Wasserman; @Connor]. In this paper we concentrate explicitly on an extra-Galactic origin for FRBs.
We present here simulations of a cosmological population of FRBs, under assumptions about their energy distribution, their spatial density as a function of redshift and the properties of the ISM and IGM (Section \[sec:MC\]), finding they are broadly consistent with origin at cosmological distances. The analysis of the models and the results are discussed in Section \[sec:AR\], in comparison with data from the HTRU survey. We present log$N$-log$\mathcal{F}$ curves and discuss the FRB rates at Parkes and UTMOST in Section \[sec:AS\] and finally our summary and conclusions in Section \[sec:CON\].
Monte Carlo Simulations {#sec:MC}
=======================
The High Time Resolution Universe (HTRU) survey at Parkes samples the transient radio sky with 64 $\upmu$s resolution at 1352 MHz and has a bandwidth of 340 MHz. The observing band is sub-divided into 390.625 kHz frequency channels. HTRU is composed of three sub-surveys at low, intermediate and high Galactic latitudes. The simulations in this paper are of the high latitude (Hilat) region of the survey — 34099, 270-sec pointings at declinations $\delta$ $< 10^\circ$ — where 9 of the 18 known FRBs have been discovered (@Thornton [-@Thornton]; @Champion [-@Champion]), and of the intermediate latitude (Medlat) region, which yielded no FRBs [@Petroff]. [@Petroff] and [@Hassall] have carried out studies similar to ours, to model the detectability of FRBs using simulations and analytic methods respectively. [@Petroff] simulated the effects of dispersion smearing which is the pulse broadening caused by the adopted frequency resolution, interstellar scattering and sky temperature on FRB sensitivity at Parkes, in the Medlat region. [@Hassall] used analytical methods to derive the detection rates at various telescopes operating over a wide range of frequencies. Our simulations are of FRB events at cosmological distances under assumptions about their co-moving density with redshift, and include the effects of ISM scattering, IGM scattering, dispersion smearing, sky temperature and telescope beam pattern. We produce estimates of the energy, fluence, signal-to-noise ratio (S/N), pulse width, DM and redshift distributions for FRBs with our models, and compare them to the 9 FRBs detected in Hilat. We perform two classes of simulations:
1. in section \[sec:MC\] we generate numerous events such that the Poisson noise of the simulations in Figures \[fig:parkes\_observed\] and \[fig:parkes\_inferred\] is negligible compared to the noise of the 9 hilat events,
2. in section \[sec:MH\] we generate thousands of short runs with an average of 9 events per simulation to estimate and compare the slopes of their log$N$-log$\mathcal{F}$ curves with the slope of the log$N$-log$\mathcal{F}$ curve of the 9 hilat FRBs.
For simplicity, FRB events in our simulations are assumed to be radiating isotropically at the source with a flat spectrum to be consistent with what is seen at 20 cm with Parkes. Their intrinsic energy distribution is assumed to be log-normal. We adopt a $\Lambda$CDM model with matter density $\Omega_{m}$ = 0.27, vacuum density $\Omega_\Lambda$ = 0.73 and Hubble constant $H_{0}$ = 71 km $\mathrm{s^{-1}} \mathrm{Mpc^{-1}}$ [@Wright]. The comoving number density distribution of FRBs in the simulations is assumed to be either a constant, or proportional to the cosmic star formation history (SFH). We adopt the SFH from the review paper of [@Hopkins] as typical of cosmic SFH measurements, which show a rise in the star formation rate of about an order of magnitude between the present ($z = 0$) and redshifts of $z \sim$ 2 (see their Figure 1). It has the parametric form $\dot{\rho_{*}} =
(a+bz)h/[1+(z/c)^{d}]$ where $h = 0.7$, $a = 0.0170$, $b = 0.13$, $c =
3.3$ and $d = 5.3$ (see their Section 4). We do not explicitly set the comoving number density of FRBs in the simulation : we compute the maximum in the product of SFH and comoving volume of each shell of width $dz$ as a function of $z$, and generate Monte Carlo events under this function. This allows the simulation to generate events at the maximum rate, which is important as our run times can be quite long (c.f. section \[sec:AR\]).
The total DM for any given FRB is assumed to arise from a component due to the IGM, a component due to the ISM in a putative host galaxy and a component due to the ISM of the Milky Way:
$$\label{eq:dmtot}
\mathrm{DM_{tot}} = \mathrm{DM_{IGM}} + \mathrm{DM_{ISM}} + \mathrm{DM_{host}}.$$
These different DM components are modeled as follows:
1. the DM due to the IGM is assumed to be related to the redshift of the source via the simple scaling relation $\mathrm{DM_{IGM}} =
1200 z \, \mathrm{pc} \, \mathrm{cm^{-3}}$ with a 1$\sigma$ scatter of order $\sim20\%$ over the redshift range and DM range of interest (DM$>$100, $0.5\lesssim z \lesssim 2$) [@Ioka; @Inoue].
2. The contribution due to the ISM of the Milky Way along the line of sight to each event is taken from the NE2001 model of [@Cordes] which includes the electron density distributions in the thin disk, thick disk, spiral arms and Galactic Center components. For the high Galactic latitude regions simulated, this is generally, $\lesssim $ 50 pc cm$^{-3}$.
3. The DM contribution of a putative host galaxy will depend on galaxy type, the FRB site within it and the viewing angle. [[@Xu] have modeled the DM distributions due to the ISM for FRBs arising in elliptical, dwarf and spiral galaxies. They scale the NE2001 model of the Milky Way ISM to the integrated $\mathrm{H}{\alpha}$ intensity maps for such hosts, to represent their electron density distributions. The ensemble average DM distribution for dwarf galaxies is 45 $\mathrm{pc} \,
\mathrm{cm^{-3}}$ and for elliptical galaxies is 37 $\mathrm{pc}
\, \mathrm{cm^{-3}}$]{}. For spirals, they derive the weighted average of the DM distribution over a range of inclination angles ($0^{\circ}, 30^{\circ}, 60^{\circ}, 75^{\circ}, 90^{\circ}$) to be 142 $\mathrm{pc} \, \mathrm{cm^{-3}}$. Noting that there may be more than one type of FRB progenitor [@Masui] conclude that their particular FRB could have occurred in a high density or star forming region of a host galaxy due to its high linear polarisation. Observationally, the galaxy stellar mass function distribution peaks near the Milky Way mass [@Robles] (their figure 9), and we assume the DM properties of the Milky Way are typical of a host FRB galaxy. Probing many random lines of sight through the NE2001 model, we derive a median DM of $\sim 70$ $\mathrm{pc} \, \mathrm{cm^{-3}}$ for the Milky Way. Given the wide range of DM estimates above, and the uncertainty even as to what typical host galaxies are and the sites of FRBs within them, we have decided to follow [@Thornton] and [@Xu], and assume a DM value of $\mathrm{DM_{host}} \sim100 \, \mathrm{pc} \,
\mathrm{cm^{-3}}$ as typical over a range of hosts and inclination angles. This assumption is somewhat *ad hoc*, but does have the advantage of facilitating comparison with previous work. The assumed DM of the host is a small fraction of the total DM to FRBs both in our observed samples and in the simulations, and we could vary this host galaxy DM over the full range discussed above ($40
\la {\mathrm DM} \la 140$) and not affect the conclusions of the paper.
{width="6"}
{width="6"}
In the simulation, events are generated out to a redshift $z = 3.0$ in shells of width $dz$ = 0.01, each populated in proportion to the co-moving volume of the shell and weighted by the star formation rate (SFR) at its redshift $z$ (in “SFH” type models). Events are distributed randomly over the sky surveyed by Hilat in proportion to the total time spent on sky (i.e. the product of the number of pointings and the integration time per pointing). No events are generated north of declination $\delta$ = +10$^ \circ$, the Northern limit of the survey performed at Parkes.
The fluence $\mathcal{F}$ (in Jy ms) at the telescope is derived from the energy at the source $E$, the luminosity distance in the $\Lambda$CDM cosmology and a factor of $(1+z)$ representing the redshifting of the observed frequency range, given by:
$$\mathcal{F} = \frac {10^{29} \, E } {4 \pi {D_{L}}^{2}(z) \, {B}\, (1+z)} \, \, \mathrm {Jy \,ms}$$
where $z$ is the redshift; $D_{L}(z)$ is the luminosity distance in pc; $E$ is isotropic emitted energy in J; $\textit{B}$ is the bandwidth of the receiver system in Hz. The S/N of each event is determined using the radiometer equation,
$$\mathrm{S/N} = \beta \, \frac {S \, G \, {\sqrt{B \, t \, N_\mathrm{p}}}} {T_\mathrm{rec} + T_\mathrm{sky}},$$
where ${S}$ is the flux of the signal in Jy, $\beta$ is the digitisation factor $\simeq$ 1.0, $B$ is the bandwidth in Hz, ${N_\mathrm{p}}$ is the number of polarisations, $t$ is the pulse width in seconds, ${T_\mathrm{rec}}$ and ${T_\mathrm{sky}}$ are the receiver and sky temperatures in K respectively, and $G$ is the system gain in K Jy$^{-1}$.
Additional simulations of the FRB rates in other surveys are made later in the paper, and the parameters adopted in those simulations are shown in Table \[tab:specs\].
The brightest FRB in [@Thornton], namely FRB110220 was detected with S/N of $\sim$ 50 and has an estimated energy $E$ = 10$^{32.5}$ J at source, a pulse width of $W$ = 5.6 ms, redshift of $z = 0.81$ and a luminosity distance of $D_{L}(z) = 5.1$ Gpc. [@Thornton] assumed the FRBs were radiating into 1 steradian (that is with a beaming fraction of 1/4$\pi$), whereas we assume isotropic radiation instead for simplicity. Accounting for this factor means that the isotropic energy of FRB110220 in the rest-frame is $E$ = 10$^{33.6}$ J and its fluence is 7.3 Jy ms.
Scattering
----------
Along the path from source to receiver, a radio pulse may be broadened in several ways. We assume the scatter-broadening time ($\tau$) of a pulsed signal passing through the ISM is related to the DM by the empirical function derived by [@Bhat]:
$$\label{eq:ISM}
\begin{split}
\textrm{log}(\tau_{\mathrm{ISM}}) = -6.5 + 0.15 \, \textrm{log($\mathrm{DM_{ISM}}$)} \\
+ 1.1 \,\textrm{log($\mathrm{DM_{ISM}}$)}^{2} - 3.9\,\textrm{log}{\nu}
\end{split}$$
where $\tau_\mathrm{ISM}$ is in ms and $\nu$ is in GHz. Rescaling the scatter-broadening time through the ISM for the IGM, [@Lorimer2013] arrived at an upper limit to the average amount of scattering as a function of DM, with the scattering due to the IGM being 3 orders of magnitude smaller than that due to the ISM, i.e.
$$\label{eq:IGM}
\textrm{log}(\tau_\mathrm{IGM}) = \textrm{log}(\tau_\mathrm{ISM}) - 3.0.$$
This rescaling on scattering in the IGM is still consistent with the observed widths of the majority of the FRBs discovered to date [@Lorimer2013].
Additionally, the pulse is broadened or smeared across frequency channels because of the adopted frequency resolution ${\tau_\mathrm{DM} = 8.3
\,\Delta\nu \, \mathrm{DM} \,\nu^{-3}\, (\mathrm{\mu\, s}})$ where DM is in $\mathrm{pc\, cm^{-3}}$, $\Delta\nu$ is the channel bandwidth in MHz and $\nu$ is in GHz. The observed width $W$ of the FRB taking into account the different contributing components is:
$$\label{eq:width}
W^2 = \mathrm{\tau^2_{IGM}} + \tau^2_{\mathrm{ISM}} +
\tau^2_{\mathrm{int}} + \tau^2_{\mathrm{DM}} + \tau^2_{\mathrm{\delta
DM}} + \tau^2_{\mathrm{samp}} + \tau^2_{\mathrm{\delta\nu}},$$
where the first two components are the scattering times due to the IGM and ISM, $\tau_{\mathrm{int}}$ is the (unknown) intrinsic width of the pulse, $\tau_{\mathrm{DM}}$ is due to the DM smearing, $\tau_{\mathrm{\delta DM}}$ is the second order correction to the DM smearing, $\tau_{\mathrm{samp}}$ is due to the adopted sampling time and $\tau_{\mathrm{\delta\nu}}$ is is the filter response of an individual frequency channel [@Cordes2003]. The $\tau_{\mathrm{\delta DM}}$ and $\tau_{\mathrm{\delta\nu}}$ terms are typically negligible in the context of our modelling. For the FRBs discovered at Parkes to date, $\tau_{IGM}$ ranges from $\sim$2 $\upmu$s to $\sim$40 ms and $\tau_{ISM}$ ranges from $\sim$40 ns to $\sim$10 ms. Previous studies dealing with FRB detectability have assumed either a “no scattering" scenario or a strong ISM-like scattering scenario for the IGM, as its properties are highly uncertain. [@Macquart] have suggested that if the latter scenario was true, the FRB pulses will be rendered undetectable at current telescopes, concluding that the IGM scattering was likely weak ($\leqslant 1\, \mathrm{ms}$). We may therefore be sampling a highly-selected population of FRBs, both in terms of luminosity and scattering.
The total width of a simulated event $W$ affects its S/N ratio, scaling it down by a factor proportional to $\sqrt{W}$. This essentially limits the horizon of the HTRU survey to $z \sim 2$ as dispersive effects beyond this redshift rapidly degrade the S/N of even the brightest events to well below the adopted threshold of 10. Consequently, we use $z = 3.0$ as the high redshift cut-off in the simulations. This is sufficiently far to sample the dispersion measure space of the known FRBs.
Measured signal-to-noise ratios
-------------------------------
The sky temperature additionally degrades the S/N particularly for sources close to the Galactic plane. We adopt a receiver temperature[^2] of ${T_\mathrm{rec}}$ = 21 K at Parkes and estimate the sky temperature (${T_\mathrm{sky}}$) at the Galactic longitude and latitude $(l,b)$ of the source from [@Haslam] who mapped the sky temperature at 408 MHz with a resolution of $0.85^{\circ} \times
0.85^{\circ}$. We scaled the survey frequency of 408 MHz to the HTRU frequency of 1.4 GHz by adopting a spectral index of $-$2.6 for the Galactic emission [@Reich], i.e.,
$$T_\mathrm{sky} = T_\mathrm{sky_{(l,b)}} \bigg(\frac{\nu}{408.0 \, \mathrm{MHz}}\bigg)^{-2.6}.$$
The S/N of each FRB event is then reduced by the additional factor ${T_\mathrm{rec}}/({T_\mathrm{rec}}+{T_\mathrm{sky}})$. For most sources this is a negligible correction, becoming important only near the Galactic centre and low in the Galactic plane. $T_\mathrm{sky}$ at high latitudes is typically $\sim 1$ K and lies between 3 and 30 K over the intermediate latitude regions.
The S/N is finally degraded depending on a randomly chosen position in the beam pattern. For Parkes, each beam of the multi-beam receiver is represented as an Airy disk with a 14.4 arc-minute full-width half-maximum. It should be noted that the effect of the beam pattern is quite significant on the distribution of both the event S/N and the apparent luminosity; this is discussed in detail in Section \[sec:lognlogf\].
Analysis and Results {#sec:AR}
====================
We have simulated FRBs in two models for their co-moving number density (either following the SFH, or simply constant density) and either including or excluding the effects of scattering. In each model, we adopt a log-normal source luminosity distribution, centered on a mean energy $E_{0}$ and spread $\sigma_\mathrm{logE}$. The average energy of the 4 [@Thornton] events in [@KeanePetroff] correcting for the beaming fraction is $10^{32.8}$ J. We initially adopt log$E_{0} = 32.8$ and a spread of $\sigma_\mathrm{logE} = 1.0$, as this is the order of magnitude scatter seen on the [@Thornton] events. Within a given model choice for the source density with redshift, $E_{0}$ and $\sigma_\mathrm{logE}$ are the 2 free parameters.
The simulations were run on 12 CPU cores with runtimes of a few days on the gSTAR national facility at the Swinburne University of Technology. Millions of FRBs are typically generated in the runs, the vast majority of which are too dim to see. We ran the simulations until we had $\sim$5000 FRBs that passed the selection criteria, to ensure good statistical sampling. The distributed properties of these FRBs are normalized and compared to the 9 observed Hilat events.
Slightly different selection criteria have been used by various authors to find FRBs. [@Thornton] used S/N$>$9 and DM$>$100 $\mathrm{pc\,cm^{-3}}$ and [@Champion] use the same selection criteria as [@Petroff], notably S/N$\geqslant$10, DM$\geqslant$ 0.9 $\times$ DM$_\mathrm{MW}$ and $W$$\leqslant$16.3 ms. We use the criteria S/N$\geqslant$10 and $W$$\leqslant$32.786 ms for the selection of the candidates in the simulations. We adopt an upper limit of 32.786 ms for the width motivated by the fact that the broadest FRB discovered in Hilat has a width of $\sim$19 ms, and in any case broader events still would have to be extremely bright to have S/N$>$10. We do not apply a DM threshold for the Hilat region as we are only sensitive to DM$>$100 $\mathrm{pc\,cm^{-3}}$ in keeping with [@Thornton], due to assuming the value of DM$_\mathrm{host}$ to be 100 $\mathrm{pc\,cm^{-3}}$. Tests showed that the differences in these selection criteria are minor and have negligible effect on our basic results. In particular every observed FRB fits each of these criteria.
Monte Carlo results for Parkes {#sec:SIM}
------------------------------
Figures \[fig:parkes\_observed\] and \[fig:parkes\_inferred\] display the results of the simulations of the cosmic SFH ($\rho_\mathrm{\text{\tiny{FRB}}}(z)$ = $\rho_\mathrm{\text{\tiny{SFH}}}(z)$) and constant-density ($\rho_\mathrm{\text{\tiny{FRB}}}(z)$ = constant) models with scattering included, as seen by Parkes, overlaid on histograms of the 9 observed Hilat FRBs (@Thornton [-@Thornton]; @Champion [-@Champion]). Figure \[fig:parkes\_observed\] shows observational parameters for each burst and Figure \[fig:parkes\_inferred\] displays the parameters that are derived. We quantify the goodness of fit of the model to the observations in Section \[sec:SA\].
The fluence distribution of our simulated events is displayed in panel A of Figure \[fig:parkes\_observed\] and their S/N distribution in panel B, each compared to the 9 Hilat events. All the observed Hilat FRBs have fluences between 0.7 and 7 Jyms. Both models peak at $\sim$ 0.5 Jyms and are reasonable matches to the observations. The S/N distributions of both models contain a large number of events just above the detection threshold of 10 and then gradually decline towards higher values; both appear to agree with the observations reasonably. Panel C shows the DM distribution of the models and the observations: this is similar to the panel showing the redshift distribution, since they are closely related. Both the cosmic SFH and constant density models are in agreement with the observed data.
The width of an FRB pulse affects its detection S/N. In the observer’s rest frame, the width results from the sum of contributions from scattering due to the ISM and IGM [Equations \[eq:ISM\] and \[eq:IGM\], see @Bhat; @Lorimer2013] DM smearing time and the intrinsic width. Panel D of Figure \[fig:parkes\_observed\] displays the distributions of the observed widths of the sources. We found neither model to agree with the data very well and may be a result of our simplistic model of intergalactic scattering discussed below.
The adopted model for the spatial density of the sources in Figure \[fig:parkes\_inferred\] panel E does not have much effect on their redshift distribution, with only a small excess of sources produced at $0 < z < 0.5$ for the constant density model compared to the cosmic SFH model. As expected, we see a tendency for more events at higher redshift for the SFH model compared to the constant density model. Panel F shows the energy distribution (at source and in-band) of the FRBs and the models. Both models are only sensitive to the bright tail of the adopted log-normal energy distribution function and have similar mean values to that of the 9 observed FRBs. Since the mean energy $E_{0}$ of the adopted luminosity function is a free parameter we adjust this to achieve good fits to the observed luminosities in panel F of Figure \[fig:parkes\_observed\]. Acceptable fits are obtained for both models by adopting $E_{0} = 10^{31.2}$ J, with a log normal-scatter of $\sigma_\mathrm{logE} = 1.0$. This adopted luminosity function is a parameterised luminosity function only and does not possess any physical significance.
Statistical analysis {#sec:SA}
--------------------
Kolmogorov-Smirnov tests (K-S) were performed on all the distributions in Figures \[fig:parkes\_observed\] and \[fig:parkes\_inferred\] and the resulting probability statistics $p$ are given in Table \[tab:p-values\]. A $p$-value of $<0.05$ is our criterion for deciding if the two distributions differ. Each model was compared against the data for the 9 FRBs.
The $p$-values show that both models are consistent with the observed distributions of redshift, energy, fluence S/N and DM but, as already noted above, we have difficulty modelling the effect of scattering on the FRBs. The $p$-values of 0.013 (density of FRBs proportional to the cosmic SFH with redshift) and 0.001 (density of FRBs constant with redshift) reject the hypotheses that both models and the FRB data are from the same population. The present sample of 9 events is thus insufficient to distinguish between these models per se (the poor match to the distribution of pulse widths in both models notwithstanding). For FRBs discovered at Parkes, our simulations indicate that of order 50 FRBs are required to distinguish between the two FRB number density models at the 95$\%$ confidence level. This certainly highlights the need to discover FRBs more efficiently, as the present discovery rate is only of order 1 per 12 days on sky at Parkes.
To better understand the effective widths, the 14 FRBs at Parkes as a function of scattering time is shown in Figure \[fig:scatter\]. The estimated widths of the events due to IGM scattering and a possible intrinsic width ($\mathrm{\tau^2_{IGM}} +
\tau^2_{\mathrm{int}}$ = $W^2 - \tau^2_{\mathrm{DM}} -
\tau^2_{\mathrm{ISM}}$ from Equation \[eq:width\]) are plotted against our estimate of the contribution to the total DM due to the IGM alone ($\mathrm{DM_{IGM}} = \mathrm{DM_{tot}} -
\mathrm{DM_{ISM}} - \mathrm{DM_{host}}$ from Equation \[eq:dmtot\]). We see that the scattering times are inconsistent with Equation \[eq:IGM\], and show considerable scatter around it.
![Adopted model of the scattering time due to the IGM (solid) at 1.4 GHz for Parkes versus estimated dispersion measures. Stars represent the 9 Hilat events and other markers represent FRBs discovered in various other surveys. The shaded region around the fitted line to the equation represents the order of magnitude scatter adopted in the simulation. Note that one Hilat FRB lies below the IGM scattering relation (see Equation \[eq:IGM\]) but is still within the adopted 1-sigma spread.[]{data-label="fig:scatter"}](Fig3.pdf){width="3.1"}
----------------- ------------ ------------------
Cosmic SFH Constant density
Redshift 0.543 0.048
Energy 0.884 0.186
Fluence 0.047 0.106
S/N 0.258 0.078
DM 0.730 0.053
Effective width 0.013 0.001
\[1ex\]
----------------- ------------ ------------------
: K-S test results for the model distributions against data in Figures \[fig:parkes\_observed\] and \[fig:parkes\_inferred\][]{data-label="tab:p-values"}
This result highlights the basic difficulty with the IGM model, apparent in the data (Figure \[fig:scatter\]), that the pulse widths of the observed FRBs scatter around the adopted functional form for the IGM (Equation \[eq:IGM\]). This behaviour is also seen for pulsars being scattered by the ISM, for which there is at least an order of magnitude scatter in the data around the observed pulse width trend [Equation \[eq:ISM\], see @Bhat]. If we assume that there is a similar scatter around $\tau_\mathrm{{IGM}}$ of an order of magnitude, we still do not acquire satisfactory fits to the data within 2$\sigma$ confidence. This suggests that the scattering is not due to a line-of-sight dependent inhomogeneous IGM. It may be due to interaction with the ISM of an intervening galaxy or an intracluster medium along the line-of-sight, although the probability of intersection at the redshifts modelled is quite low and only a small fraction of lines of sight may be affected [@Macquart]. We have not attempted to model such effects: our aim is to test a much simpler model before adding in difficult to test assumptions about the properties of the IGM.
If we assume that our simulated events have a mean intrinsic width of 3 ms (with a standard deviation of 3 ms, truncated at 0 ms), the resulting width distributions are found to be in good agreement with the observed 9 hilat FRBs. The intrinsic width assumption is motivated by FRB121002 and FRB130729 [@Champion] which have hints of double, rather than single peaked pulse profiles. This is a rather *ad hoc* assumption and further work on this is required once the population is expanded. The disagreement of the distribution of event widths with the observations is the weakest point in our modelling. Clearly, there is a need for more FRBs to resolve this problem.
The log$N$-log$\mathcal{F}$ of the Hilat events {#sec:lognlogf}
-----------------------------------------------
In a Euclidean Universe populated with events (or objects) of fixed luminosity (i.e. standard candles) and uniform number density, the number $N$ detected above some flux limit $S$ varies as $N
\propto S^{\alpha}$, where $\alpha=-3/2$. In our model, the FRBs have a very broad luminosity distribution and are sufficiently distant that non-Euclidean effects are important. Consequently we do not expect to see $\alpha = -3/2$.
The very wide range of luminosities of the observed events suggests they are not particularly good standard candles, and until we have a redshift of an FRB host galaxy, or some other independent distance indicator for an FRB, their luminosities are highly dependent on the assumption that DM is a proxy for redshift. The luminosities are dependent on each line of sight being equal to the average line of sight in a $\Lambda$CDM Universe. In fact it is the small deviations from this that we will use to do some cosmology, when we have a lot of FRBs with real redshifts. In any case, our FRB simulations are for a $\Lambda$CDM cosmology, which affects $\alpha$. In a $\Lambda$CDM cosmology, $\alpha$ varies smoothly from a slope of $-3/2$ for the nearby universe, gradually becoming flatter as further distances are probed. To illustrate, at a redshift of $z \sim
0.7$, typical of FRBs found to date, standard candles yield a relation with a slope of $\alpha \sim -1$. There are additional factors which affect $\alpha$. Firstly, the HTRU survey is “fluence incomplete” in the sense that events with the same fluence are easier to detect if they have narrower pulse widths. Secondly, propagation of FRB pulses through the IGM causes the pulses to broaden, reducing their S/N, so that a S/N selected sample effectively has a distance horizon beyond which pulses are too scattered to see. This will flatten the relation as we probe to dimmer events.
It is possible to select a “fluence complete” sample of the FRBs, and compare these to simulation events selected in the same way, but this would reduce our sample of 9 events to just 4 events. For a S/N of 10, the fluence completeness limit for Hilat is $\sim 2$ Jyms [@KeanePetroff]. This is an observational selection, and affects the slope $\alpha$, of the relation. It is straightforward to include this effect in the simulations, however, due to our already small sample of events we prefer to compare to the full set of 9 events selected by S/N, rather than a fluence complete set of 4 events.
The log$N$-log$\mathcal{F}$ plot of the 9 Hilat events is shown in Figure \[fig:9\_events\] — note that we use the fluence $\mathcal{F}$ in Jy ms (since FRB detections are width dependent) for what would normally be flux density $S$ in Jy. The cumulative log$N$-log$\mathcal{F}$ relation is reasonably linear for the 9 events, and has a slope of $\alpha = -0.9\pm0.3$. For the cumulative curve of only 9 events, sample variance is likely to be a significant factor. We use the simulations of Hilat (as described in section \[sec:MC\] with selection criteria described in Section \[sec:AR\]), which were set up to yield of order 9 events per run to estimate the error on $\alpha$. Those realisations which had exactly 9 events were used for comparison with the 9 observed Hilat events. We have fitted slopes ($\alpha$) to these simulated 9 event samples and show the distribution of $\alpha$ in Figure \[fig:slope\_dist\]. The typical error on $\alpha$ is $\pm$0.1 which is the adopted bin size in Figure \[fig:slope\_dist\]. The median slope obtained is $\alpha = -0.8$ for the SFH case and $\alpha = -0.7$ for the constant density case, but with significant scatters (the 1 $\sigma$ limits are shown as dashed lines) of order $\pm0.3$ for the SFH and $\pm0.2$ for the constant density around the mean. Our observed slope of $\alpha =
-0.9\pm0.3$ is consistent with both models.
We conclude that the slope of the log$N$-log$\mathcal{F}$ relation of the 9 observed events is consistent to within the uncertainties of both the simulated models, indicating that our measured log$N$-log$\mathcal{F}$ slope is consistent with FRBs being of cosmological origin. This is in agreement with the conclusion of [@Katz3] that the log$N$-log$S$ and N vs. DM distributions are consistent (except for the anomalously bright Lorimer burst) with cosmological distances inferred from their DM in a simple approximation to standard cosmology.
![The log$N$-log$\mathcal{F}$ curves for the 9 Hilat FRBs and the simulation samples. The left panel displays the cosmic SFH [@Hopkins] scenario and the right panel displays the constant density scenario. Stars represent the 9 Hilat FRBs and the solid line connects the medians of the number densities as a function of fluence for the simulation sample. The dashed and dotted lines represents the 1$\sigma$ and 2$\sigma$ limits around the median for each N. The inset in the left panel exhibits the 9 observed FRBs and a fitted slope of $\alpha = -$3/2 for comparison.[]{data-label="fig:9_events"}](Fig4.pdf){width="3.5"}
![The histograms display the slopes $\alpha$, of the simulation samples containing exactly 9 events each. The left panel represents the cosmic SFH scenario and the right panel represents the constant density scenario. The medians of the histograms are represented by the solid lines and the 1$\sigma$ scatter from the median is marked by the dashed lines. The slope of the 9 FRBs $\alpha = -0.9\pm0.3$ is found to be consistent with the simulations within the uncertainties.[]{data-label="fig:slope_dist"}](Fig5.pdf){width="8.6"}
Medlat vs Hilat {#sec:MH}
---------------
The intermediate-latitude component of the HTRU survey consists of 540-sec pointings in the range $-120^{\circ} < l < 30^{\circ}$ and $|b| < 15^\circ$. [@Petroff] found no FRBs in this region of the survey. Under the assumption that FRBs are isotropically distributed, scaling from Hilat, and accounting for a slight reduction in their detectable source density in the Medlat region due to the smearing effects of the ISM, they estimate the probability of this occurring by chance as only of order 0.5%. We simulate both the Medlat and Hilat regions (adopting 100$\%$ of Hilat and 100$\%$ of Medlat as the surveyed completeness for the regions for FRBs) to determine the likelihood of finding zero FRBs in Medlat for 9 discovered FRBs in Hilat. The simulation for Medlat is otherwise identical to the one described in Section \[sec:MC\] except for the survey parameters i.e. number of pointings, region of sky surveyed, $T_\mathrm{sky}$ corresponding to the region of sky surveyed and integration time per pointing. The same selection criteria as described in Section \[sec:AR\] are used for selection of candidates in both Medlat and Hilat.
We obtain an average of $\sim3\pm2$ events in our Medlat simulations for every 9 events in the Hilat simulations, finding no events just 5.1$\%$ of the time (Figure \[fig:poisson\]). The estimated probability of zero events being seen in Medlat 0.5$\%$ of the time made by [@Petroff] is based on the 4 events detected in the 24$\%$ of the Hilat survey which had been searched at the time. The higher probability we estimate of finding no events in Medlat in our simulations is due to our using the lower all sky rate, now that Hilat has been completely searched and it only yielded 9 FRBs.
![Number of FRBs expected in Medlat normalized to the 9 events in Hilat. Both surveys are assumed to be fully searched for FRBs. The histogram represents the number of FRBs expected in Medlat for a corresponding 9 FRBs detected in Hilat. A Poissonian curve is fitted to the data. The number of FRBs found in Medlat is zero 5.1% of the time.[]{data-label="fig:poisson"}](Fig6.pdf){width="2.63"}
------------------------- ---------------------- -------------------- ---------------- --------------------------------
Parameter Unit Parkes MB [@Keith] Parkes PAF[^3] UTMOST (Bailes et al. in prep)
\[0.5ex\] Field of View $\mathrm{deg^2}$ 0.55 2.2 $4.64\times2.14$
Central beam Gain K $\mathrm{Jy^{-1}}$ 0.7 0.9 3.5
Central beam $T_{sys}$ K 21 50 70
Bandwidth MHz 340 340 16
Frequency MHz 1352 1352 843
Channel width MHz 0.390625 $\sim$1 0.78125
No. of polarisations – 2 2 1
Polarisation feeds – Dual linear Dual linear Right circular
\[1ex\]
------------------------- ---------------------- -------------------- ---------------- --------------------------------
Telescope/Receiver $S_\mathrm{min}$ (Jy) Rate (events $\mathrm{day^{-1}}$)
-------------------- ----------------------- -----------------------------------
Parkes MB 0.4 0.08$\pm$0.03
Parkes PAF 0.6 0.10$\pm$0.04
UTMOST 1.6 0.16$\pm$0.06
\[1ex\]
: Minimum detectable flux density for a 10$\sigma$, 1 ms event and event rate assuming a Euclidean scaling for the Parkes multibeam, Parkes PAF and UTMOST []{data-label="tab:rates"}
The log$N$-log$\mathcal{F}$ of FRB events {#sec:AS}
=========================================
Our simulations have been used to generate FRB events at 2 facilities – Parkes and UTMOST (Bailes et al. in prep). UTMOST is the recently upgraded Molonglo Observatory Synthesis Telescope located about 300 km south-west of Sydney, near Canberra, and is a field station of the University of Sydney. We generate events for the specifications of UTMOST and Parkes for the soon to be installed phased array feed (PAF) receiver in comparison with the current multibeam receiver (MB) at Parkes. The FRB co-moving density models, and energy distributions are the same as those described in Section \[sec:MC\]. The effective pulse width of each event is computed using Equation \[eq:width\]. The S/N of the events were reduced by a factor of 4 for the events at UTMOST to account for the fact that it is less sensitive than the MB receiver at Parkes (Caleb et al. in prep). The Parkes PAF is estimated to have $\sim 50\%$ of the sensitivity of the multibeam[^4] which is accounted likewise. The S/Ns at both telescopes were further reduced by $\sqrt{W}$ before making the cut-off of S/N$\geqslant$10 and $W$$\leqslant$32.786 ms.
Figure \[fig:logNlogS\] shows the cumulative log$N$-log$\mathcal{F}$ curves at UTMOST and at Parkes for both the MB and phased array feed PAF. These curves do not include the effects of fluence completeness. All curves have been normalised to their respective FRB rates in Table \[tab:rates\], which have been calculated assuming a Euclidean Universe where the cumulative number density scales as $\propto \mathcal{F}^{\alpha}$ where $\alpha = -3/2$ (Caleb et al. in prep). This is a conservative option, as the slope of this relation is most likely flatter (as seen in the previous section), and underestimates the number of events expected.
{width="6"}
Discussion and Conclusions {#sec:CON}
==========================
We have simulated observational and derived properties of a cosmologically distributed population of FRBs, for comparison with the 9 FRBs seen in the HTRU/Hilat survey conducted at Parkes from 2008-2014. Two models for the spatial number density of the FRBs are examined: firstly, where the co-moving density is a constant, and secondly, where the number of FRBs is proportional to the cosmic SFH. The properties of the ISM in the Milky Way and a putative host galaxy for the FRB are taken into account, and conservative assumptions are made about the properties of the IGM, the spectral index of FRBs and their luminosity function.
The simulated distributions of redshift, energy, DM, S/N, fluence and effective widths for both the cosmic SFH and constant density models were compared to the 9 observed FRBs. We achieved reasonable matches to the data for all these properties except the event widths, by adjusting only the typical FRB event energy at source (and scatter around this energy) i.e. by adjusting only their luminosity function. It proved difficult to fit the distribution of FRB widths without making *ad hoc* assumptions about scattering in the IGM or the intrinsic widths of the pulses. The simulations are intended to look at FRB properties with as simple an assumption set as possible; adding in poorly constrained properties as these for the FRBs and the IGM for the sake of fitting the pulse widths was not pursued. As the pulse widths probe completely different properties of FRBs and the IGM, this may prove more fruitful to understanding their origin as more FRBs are found.
The most interesting property of the simulated events is their distribution of log$N$-log$\mathcal{F}$, where $N$ is the number of events detected above some fluence $\mathcal{F}$. If the sources have an even approximately typical luminosity (i.e. are standard candle-like) then the slope of this relation is a probe of their spatial distribution. For standard candles of flux $S$ distributed uniformly in empty, Euclidean space, the slope of the closely related log$N$-log$S$ relation is well known to be exactly $-3/2$. For FRBs, the slope of the relation is affected substantially for 3 main factors: firstly by cosmology (space is non-Euclidean); secondly by propagation through the IGM (i.e. space is not empty) and thirdly by selection effects at the telescope (narrower events are detected more readily than broader ones). A major aim of the simulation is to quantify these effects.
The observed slope $\alpha$ of the log$N$-log$\mathcal{F}$ of the 9 FRBs analysed is $\alpha = -0.9 \pm 0.3$. Our simulations are able, in both scenarios for the number density of the sources with redshift, to match this slope well, yielding $\alpha = -0.8 \pm 0.3$ for the cosmic SFH and $\alpha = -0.7 \pm 0.2$ for the constant density case. We conclude that the properties of the observed FRBs are consistent with arising from sources at cosmological distances, with the important caveat that the pulse width distribution does not match our simulation results particularly well.
The luminosity function of the FRBs is the main free parameter in the simulations. We adopt a log-normal luminosity function (LF) and adjust the mean energy $E_{0}$ and spread in energy $\sigma_\mathrm{logE}$. It is clear from the 9 observed events that a narrow, standard-candle like LF is an unacceptable fit, since their inferred intrinsic lumninosities has a spread of about an order of magnitude. We measure a mean energy $E_0$ of $\sim 10^{31.2}$ J with a spread of a factor of 10 in energy. As the observed FRBs very much sample only the high luminosity tail of this distribution, other choices for the LF, such as a truncated power law would also adequately match the data. Our studies show that the beam pattern of the telescope has a strong effect only when the number of FRBs is large ($\ga {\mathrm few} \times 100$), which is then sensitive to the high luminosity tail of events. The LF choice affects the distributions strongly even for small samples : an LF with a significant spread in luminosity is required to model the 9 events. Finally, our simulations show that the adopted comoving density models for the FRBs has weak effects, and large sample sizes ($\ga 100$) are required to probe this.
Future work could implement other LF choices and investigate the extent to which the LF and SFH and beam pattern affect the observed distributions analysed in this paper: the small number of FRBs detected to date do not warrant such work here.
Our simulations show that at least 50 FRB events are required to distinguish, at the 95$\%$ confidence level, between our two tested models for their cosmological spatial distributions for the specifications of the Parkes telescope. This argues strongly for projects to increase the detection rate of FRBs by using wide field of view instruments, such as UTMOST and CHIME [@Bandura]. Even more important in the immediate future is to localise events on the sky (to find putative host galaxies for FRBs) and a number of experiments are ongoing to do this (eg: SUPERB project at Parkes).
We have applied our simulations to the Medlat survey at Parkes (which is part of the HTRU survey), which surveyed a lower Galactic latitude region of the sky with longer integrations. Our simulations of this survey supports the conclusion of [@Petroff] that the sky rate of FRBs in [@Thornton] is overestimated by about 50$\%$, or that FRBs are not distributed isotropically on the sky.
We simulate FRB rates at two other facilities : at UTMOST (first survey results of which are in a companion paper Caleb et al. in prep) and at Parkes with the planned Phased Array Feed (PAF), under conservative assumptions about the spectral index of FRBs, and the sensitivity of the instruments. UTMOST has the capability, at full design sensitivity to dominate the FRB detection rate. Uncertainty in the final PAF design sensitivity makes prediction difficult, but its wide sky coverage has the potential to increase the FRB discovery rate of FRBs close to the fluence limit. The fully sensitive UTMOST will dominate the event detection rate at all fluences.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors gratefully acknowledge valuable discussions with Emily Petroff, J-P Macquart and Alan Duffy. This work used the gSTAR national facility which is funded by Swinburne and the Australian Governments Education Investment Fund. Parts of this research were conducted by the Australian Research Council Centre for All-Sky Astrophysics (CAASTRO), through project number CE110001020.
[^1]: Email: [email protected]
[^2]: www.parkes.atnf.csiro.au/observing/documentation/user$\_$guide
[^3]: http://www.atnf.csiro.au/management/atuc/2013dec/science$\_$meeting/ATUC$\_$PKS$\_$receivers.pdf
[^4]: http://www.atnf.csiro.au/management/atuc/2013dec/ science$\_$meeting/ATUC$\_$PKS$\_$receivers.pdf
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Jun Xu
- Mouyuan Sun
- Yongquan Xue
- Junyao Li
- Zhicheng He
title: 'X-ray absorption and 9.7 silicate feature as a probe of AGN torus structure'
---
Introduction {#sect:introduction}
============
The dusty torus, which is widely believed to be responsible for obscuring the broad emission-line region and the central engine, is a vital component of the unification models [e.g., @Antonucci1993; @Urry1995; @Netzer2015] of active galactic nuclei (AGNs). However, the structure of the dusty torus remains largely undetermined partly because it cannot be readily resolved, despite of a few successful attempts [e.g., @Imanishi2018; @Garcia-Burillo2019; @Gravity2020]. Some phenomenological torus models in which the morphology and distribution of dusty clouds are predefined are proposed. Then the efforts are focused on solving the sophisticated radiative transfer and obtaining the corresponding spectral energy distributions (SEDs) at near- to mid-infrared bands. These models can be roughly divided into three categories, i.e., a smooth torus [e.g., @Fritz2006], a clumpy torus [e.g., @Nenkova2008p1; @Nenkova2008p2], or a mixture of these two [e.g., @Siebenmorgen2015].
Observationally speaking, the dusty torus manifests itself by various multi wavelength spectroscopic signatures. For instance, gas in the dusty torus can induce heavy X-ray obscuration (with X-ray column density $N_\mathrm{H}^\mathrm{X}$ to be more than $10^{24}\ \mathrm{cm}^{-2}$) if our line of sight is nearly edge-on. The dusty torus can absorb a significant fraction of AGN UV-to-optical continuum emission and re-emit mainly at mid-infrared (MIR) bands, thereby making the AGN intrinsic UV-to-optical SEDs much redder. In addition to the strong MIR continuum emission, the inner torus region can emit prominent silicate emission lines at $9.7\ \mathrm{\mu m}$ and $18\ \mathrm{\mu m}$ due to Si-O stretching and bending modes. Such features have indeed been detected by the *Spizter*/IRS spectroscopic observations [e.g., @Siebenmorgen2005; @Hao2005; @Shi2006].
According to the simplest AGN unification model, if our line of sight is roughly face-on, we can directly detect emission from the central engine, the broad emission-line region, and the inner torus region. Therefore, we expect such AGNs to show unambiguous broad emission lines in their UV/optical spectra (classified as optical type-1), unobscured X-ray power-law emission and the $9.7\ \mathrm{\mu m}$ and $18\ \mathrm{\mu m}$ silicate emission features. In contrast, if the viewing angle is nearly edge-on, our line of sight is obscured by the dusty torus. Hence, such AGNs show X-ray spectra with heavy obscuration and lack broad emission lines (classified as optical type-2). Their silicate features are also expected to be observed as absorption [e.g., @Siebenmorgen2004; @Shi2006].
In reality, the relations among these dusty torus signatures are much more complex than expected. For example, optical type 2 AGNs with silicate emission features have been observed [@Sturm2006; @Nikutta2009]. [@Shi2006] systematically investigated the X-ray absorption and the silicate feature of $97$ AGNs with various types. They found that there is a connection between $N_\mathrm{H}^\mathrm{X}$ (inferred mostly from hardness ratios, which might be biased; see, e.g., @Li2019) and the silicate feature (see Eq. \[eq:Ssi\] for the definition) which is consistent with the expectations of the simplest AGN unification model. However, the scatters of the connection are quite large. Therefore, they proposed that the torus structure should be complex and clumpy. In addition, [@Goulding2012] studied $20$ nearby Compton-thick AGNs and pointed out that, at least for Compton-thick AGNs, the observed silicate absorption feature might be caused by galaxy-scale dust rather than a compact dusty torus near the central engine.
The study of dusty torus multi wavelength signatures can benefit from more complete AGN multi wavelength surveys. One of such surveys is the *Swift*/BAT all-sky survey [@Baumgartner2013]. The resulting AGN sample is complete with respect to X-ray absorption since the *Swift*/BAT ultra-hard ($14$–$195$ keV) band X-ray can penetrate through obscuring gas clouds with $N_\mathrm{H}^\mathrm{X}>10^{24}\ \mathrm{cm^{-2}}$. Thanks to the wide X-ray spectroscopic coverage ($0.3$–$195$ keV), the X-ray properties (especially $N_\mathrm{H}^\mathrm{X}$) of each AGN were robustly measured [@Ricci2015; @Ricci2017b]. Meanwhile, the spectroscopic follow-up surveys were performed for a large fraction of this sample [hereafter BASS;[^1] @Koss2017]. Furthermore, a significant fraction of the BASS AGNs also have *Spitzer*/IRS spectroscopic coverage which enables us to examine their silicate features. Therefore, this sample is ideal for us to explore the nature and structure of the dusty torus.
This paper is laid out as follows. In Section \[sect:data\], we describe our sample construction and data analyses. In Section \[sect:Results\], we present our results. In Section \[sect:dis\], we discuss the implications of our results. Our conclusions are summarized in Section \[sect:sum\].
Sample construction and data reduction {#sect:data}
======================================
![The distribution of our sample in the 14-195 keV X-ray luminosity vs. redshift diagram.[]{data-label="fig:Lx_z"}](fig1.pdf){width="75.00000%"}
Our parent sample consists of 836 AGNs from the first $70$-month observations of the unprecedented deep ultra-hard X-ray (14–195 keV) survey of the Burst Alert Telescope on the *Swift* Space satellite. Thanks to the wide X-ray spectroscopic coverage ($0.3$–$195$ keV), the X-ray properties (e.g., $N_\mathrm{H}^\mathrm{X}$ and the intrinsic X-ray luminosity, $L_{14-195\ \mathrm{keV}}$) of all 836 AGNs were well determined and the resulting catalog is publicly available [@Ricci2015; @Ricci2017b]. Therefore, compared with [@Shi2006], our AGNs have more reliable $N_\mathrm{H}^\mathrm{X}$ measurements.
For these 836 AGNs, we use their published counterparts [see @Baumgartner2013] to cross-match (by name) with the Combined Atlas of Sources with Spitzer IRS Spectra (CASSIS[^2]) database [@Lebouteiller2011] to construct a new sample, which consists of 208 AGNs. Five sources are rejected since their *Spitzer*/IRS spectra do not have spectroscopic coverage around the 9.7 silicate feature. The remaining 203 AGNs are then cross-matched (by name) with the BASS catalog [@Koss2017] to obtain their optical spectroscopic measurements. Most of the AGNs ($185/203$) in our sample have optical spectroscopic coverage. However, for ten out of the 185 AGNs, their optical measurements and types are absent. Therefore, we reject these ten sources. Our final sample, which consists of 175 AGNs, enables us to study the dusty torus in three spectroscopic respects, i.e., X-ray absorption, optical type, and the silicate feature. Their luminosity and redshift ranges are presented in Figure \[fig:Lx\_z\].
We perform the $9.7\ \mathrm{\mu m}$ silicate feature measurements by using *DeblendIRS*[^3] [@Hernan2015] to fit *Spitzer*/IRS spectra. *DeblendIRS* is an *IDL* package that fits the MIR spectra with a linear combination of three spectral templates, i.e., a“pure" AGN template, a “pure" stellar template, and a “pure" Polycyclic Aromatic Hydrocarbon (PAH, which accounts for the interstellar emission) template. The templates are constructed from real *Spitzer*/IRS spectra which are dominated by a single physical component (i.e., AGN, stellar, or interstellar). For each AGN template, the silicate strength and the slope ($\alpha$) of a power-law continuum between $8.1$and $12.5$are pre-measured. The silicate strength is defined as $$S_{\mathrm{Sil}}=\mathrm{ln}\ \frac{F(\lambda_p)}{F_C(\lambda_p)} \\,$$ where $F(\lambda_p)$ and $F_C(\lambda_p)$ stand for the maximum flux density of the silicate line profile near 9.7 and the correspondingflux density of the underlying continuum profile, respectively. Note that for sources with negative values of $S_{\mathrm{Sil}}$, we expect the optical depth of the silicate absorption $\tau_{9.7}=-S_{\mathrm{Sil}}$. Therefore, *DeblendIRS* can provide the best-fitting results and uncertainties for the contribution of AGN emission at rest-frame $6$, 12, and $5$–$15$(hereafter $L_6$ AGN fraction, $L_{12}$ AGN fraction and rAGN, respectively), the stellar contribution at rest-frame $12$and $5$–$15$(hereafter $L_{12}$ SB fraction and rSTR, respectively), the interstellar contribution at rest-frame $5$–$15$(hereafter rPAH), $S_{\mathrm{Sil}}$ and $\alpha$ [for more technical details, refer to Section 2 of @Hernan2015]. An example of our *DeblendIRS* fitting results is shown in the upper panel of Figure \[fig:deblendIRS\]. It is evident that the best-fitting model explains the data well.
![A typical `deblendIRS` decomposed AGN IRS spectrum (top). The lower panels show the probability distributions of the fitting parameters (for their definitions, refer to Section \[sect:data\]), where the red solid and blue dashed lines represent the best-fitting results and the expectations of the distributions. The shaded regions indicate the 1$\sigma$ uncertainties.[]{data-label="fig:deblendIRS"}](fig2.pdf){width="75.00000%"}
Unlike our work, [@Ichikawa2019] adopted another *IDL* routine *DecompIR* [@Mullaney2011] to decompose the multi-band IR (from $\sim 3$to $\sim 200$) photometric data and neglected the *Spitzer*/IRS spectra. There are $160$ sources in both our final sample and the catalog of [@Ichikawa2019]. To further justify our fitting results, we compare our best-fitting 12 $\mu$m monochromic luminosities (hereafter $\lambda L_\lambda(12\mathrm{\mu m})$) with those of [@Ichikawa2019] (see Figure \[fig:12um\_comparison\]). Our results are well consistent with those of [@Ichikawa2019]; indeed, the median ratio between our and their $\lambda L_\lambda(12\mathrm{\mu m})$ is $1.06$. Therefore, we argue that our AGN measurements are reliable.
![Comparison between our measurements of $\lambda L_\lambda(12\mathrm{\mu m})$ with those of [@Ichikawa2019]. The red dashed line indicates the one-to-one relation. Our results are in good aggrement with those of [@Ichikawa2019].[]{data-label="fig:12um_comparison"}](fig3.pdf){width="75.00000%"}
Results {#sect:Results}
=======
Silicate strength and X-ray absorption {#sect:nhx}
--------------------------------------
Following [@Shi2006], we first explore the relationship between the silicate feature and $N_\mathrm{H}^\mathrm{X}$. [@Shi2006] defined the following quantity, $$R_{9.7}=\frac{F(\lambda_p)-F_C(\lambda_p)}{F_C({\lambda_p})}
\label{eq:Ssi}$$ where negative/positive $R_{9.7}$ suggests silicate absorption/emission. It is straightforward to show that $R_{9.7}=\exp{(S_{\mathrm{Sil}})}
- 1$ and approaches $S_{\mathrm{Sil}}$ if $S_{\mathrm{Sil}}$ is close to zero. Figure \[fig:Shi2006fig\] plots $R_{9.7}$ as a function of $N_\mathrm{H}^\mathrm{X}$ for our final sample. We confirm that, consistent with the result of [@Shi2006], there is a weak anti-correlation between $R_{9.7}$ and $N_\mathrm{H}^\mathrm{X}$ (the Spearman’s $\rho=-0.63$ and the corresponding $p$-value is $4.5\times 10^{-21}$). That is, heavily X-ray obscured AGNs tend to show silicate absorption and absence of broad emission lines, and vice versa. We fit the data with the linear relation $R_{9.7}=A + B\log N_\mathrm{H}^\mathrm{X}$ via the MCMC algorithm.[^4] The fitting results are $A=4.1^{+0.6}_{-0.5}$ and $B=-0.20^{+0.03}_{-0.03}$, which is in agreement with the best-fitting results of [@Shi2006]. However, the scatters of the anti-correlation are quite large. The large scatters might be caused by several different factors. For example, the gas-to-dust ratio may vary among different AGNs. As pointed out by [@Shi2006], the scatter of this ratio should be more than two orders of magnitude, which is unlikely to be the case here. Another possibility is that the heavy X-ray absorption is caused by the gas lying closer to the central engine than the dusty torus, e.g., the broad emission-line gas. However, this scenario cannot explain the fact that many of these AGNs, which show heavy X-ray obscuration but different silicate strengths, are actually optical type-1.9 or type-2 sources, i.e., sources with strong dust extinction. Other possibilities are that the dusty torus is not a smooth “donut” but a highly clumpy one, the line-of-sight absorption is time-dependent [e.g., @Yang2016; @Jaffarian2020], and/or the radiative transfer may also play a role.
![$R_{9.7}$ as a function of $N_{\rm H}^{\rm X}$ for our selected sources. The blue open circles, yellow open diamonds, green open squares, and cyan open diamonds represent type 1-1.2, type 1.5, type 1.9, and type 2 AGNs, respectively.[]{data-label="fig:Shi2006fig"}](fig4.pdf){width="75.00000%"}
Silicate strength and broad-H$\alpha$-line-inferred extinction {#sect:bha}
--------------------------------------------------------------
For a subsample of AGNs with broad H$\alpha$ emission lines, [@Shimizu2018] adopted the empirical relation between X-ray and broad H$\alpha$ luminosities and the absorption-corrected X-ray luminosities to obtain the intrinsic broad H$\alpha$ luminosities. By comparing the intrinsic broad H$\alpha$ luminosities with the observed ones, [@Shimizu2018] estimated the optical extinction of the broad-line region (hereafter $\mathrm{A_V}[\mathrm{bH}\alpha]$). Then, they explored the relation between $\mathrm{A_V}[\mathrm{bH}\alpha]$ and $N_\mathrm{H}^\mathrm{X}$ and found that a significant fraction of AGNs have orders of magnitude higher $N_\mathrm{H}^\mathrm{X}$ than the $\mathrm{A_V}[\mathrm{bH}\alpha]$-inferred values by assuming a Galactic ratio of $N_\mathrm{H}$ to $\mathrm{A_V}$ [@Draine2011]. The population of optical type 1 AGNs with heavy X-ray absorption has also been explored by [@Merloni2014] who used the $1310$ XMM-COSMOS AGNs as well as several previous works [e.g., @Burtscher2016; @Schnorr2016].
![The optical extinction of the broad $H\alpha$ line ($A_{V}[\mathrm{bH}\alpha]$) versus the silicate strength-inferred $V$-band extinction ($A_V[\mathrm{Si}]$). The blue and red symbols indicate X-ray unobscured (i.e., $N_\mathrm{H}^\mathrm{X}
<10^{21.5}\ \mathrm{cm^{-2}}$) and obscured sources (i.e., $N_\mathrm{H}^\mathrm{X}\geq 10^{21.5}\ \
\mathrm{cm^{-2}}$), respectively. Note that the uncertainties of $A_{V}[\mathrm{bH}\alpha]$ are caused by both the measurement errors of the broad $H\alpha$ fluxes and a systematic uncertainty of $1.2$ mag [@Shimizu2018]. The dashed line indicates the one-to-one relation. []{data-label="fig:avsil"}](fig5.pdf){width="75.00000%"}
We first investigate the relation between $A_{V}[\mathrm{bH}\alpha]$ and the silicate strength-inferred $V$-band extinction (hereafter $A_V[\mathrm{Si}]$) where $A_V[\mathrm{Si}]$ is estimated from $S_{\mathrm{Sil}}$ following the methodology presented in Section 3 of [@Shi2006]. Note that, for sources with silicate emission lines, we set $A_V[\mathrm{Si}]=0$. For the 93 AGNs in our final sample that have broad $H\alpha$ measurements, we follow [@Merloni2014] and classify them into two categories according to $N_\mathrm{H}^\mathrm{X}$, i.e., the X-ray unobscured (i.e., $N_\mathrm{H}^\mathrm{X} <
10^{21.5}\ \mathrm{cm^{-2}}$) and obscured sources (i.e., $N_\mathrm{H}^\mathrm{X}\geq 10^{21.5}\ \mathrm{cm^{-2}}$). The results are shown in Figure \[fig:avsil\]. For X-ray unobscured AGNs, their $A_{V}[\mathrm{bH}\alpha]$ and $A_V[\mathrm{Si}]$ values are both small. However, for X-ray obscured AGNs, $A_{V}[\mathrm{bH}\alpha]$ and $A_V[\mathrm{Si}]$ are large and the former is on average much smaller than the latter (see Section \[sect:add\] for the discussions of the possible physical reasons).
![Relative fractions of AGN types as a function of the AGN $14$-$195$ keV X-ray luminosity. The blue shaded region represents type-11 sources (optical type 1 and X-ray unobscured). The purple shaded region represents type-12 sources (optical type 1 and X-ray obscured). The yellow shaded region represents type-21 sources (optical type 2 and X-ray unobscured). The red shaded region represents type-22 sources (optical type 2 and X-ray obscured). The median silicate strength values for type-11, type-12 and type-22 are annotated, where positive values indicate silicate emission and vice versa. There is only one type-21 AGN and its silicate strength is not shown. []{data-label="fig:M12"}](fig6.pdf){width="75.00000%"}
Following [@Merloni2014], we further classify our final sample into four categories, i.e., type-11 (optical type 1 and $N_\mathrm{H}^\mathrm{X}
< 10^{21.5}\ \mathrm{cm^{-2}}$), type-12 (optical type 1 and $N_\mathrm{H}^\mathrm{X}\geq 10^{21.5}\ \mathrm{cm^{-2}}$), type-21 (optical type 2 and $N_\mathrm{H}^\mathrm{X}< 10^{21.5}\ \mathrm{cm^{-2}}$), and type-22 (optical type 2 and $N_\mathrm{H}^\mathrm{X}\geq 10^{21.5}\
\mathrm{cm^{-2}}$). Their relative fraction as a function of X-ray luminosities are presented in Figure \[fig:M12\]. Unlike [@Merloni2014], we also can calculate the median value of $S_{\mathrm{Sil}}$ (i.e., $-\tau_{9.7}$) for each type. Consistent with our expectations, type-11/type-22 sources on average have silicate emission/absorption features. However, type-12 AGNs tend to show prominent silicate absorption, which is consistent with the result of Figure \[fig:Shi2006fig\]. Among them, the silicate absorption in optical type 1.5 is weak or nearly absent; the silicate absorption in optical type 1.9 is rather strong. Our results indicate that, at least for the optical type 1.9 AGNs in the type-12 population, the excess of X-ray absorption might not be caused by dust-free broad-line gas.
In conclusion, the results of Figures \[fig:Shi2006fig\]-\[fig:M12\] along with previous works [e.g., @Shi2006; @Merloni2014] suggest that the distributions and structures of obscuring gas and extinction dust are very complex.
Discussion {#sect:dis}
==========
As demonstrated in Section \[sect:Results\], $S_{\mathrm{Sil}}$, $N_\mathrm{H}^\mathrm{X}$, and $A_{V}[\mathrm{bH}\alpha]$ often show discrepant results, which might be caused by various factors as discussed below.
Smooth torus vs. clumpy torus
-----------------------------
One possible explanation is that the dusty torus is not a smooth “donut” but a highly clumpy one. To test this scenario, we compare our results in Figures \[fig:Shi2006fig\]-\[fig:M12\] with a smooth torus model of [@Fritz2006] and a clumpy torus model of [@Nenkova2008p1; @Nenkova2008p2], respectively.
### Testing the smooth torus model {#sect:smooth}
One popular smooth torus model is introduced by [@Fritz2006]. In this model, the dust mass density is a function of both radius (with respect to the central black hole) and inclination angle ($i$), i.e., $$\rho(r, i)=\rho_{0} r^{-q} \mathrm{e}^{-\gamma|\cos (i)|} \\,$$ where $\rho_0$ is determined by the equatorial-plane-dust optical depth at $9.7$($\tau^0_{9.7}$), $q$ is the radial power-law index, and $\gamma$ is the polar exponential index, respectively. The torus inner radius is determined by the dust sublimation radius; the ratio of the outer radius to the inner radius ($Y$) is allowed to vary as a free parameter. Another parameter is the angular region occupied by the dust ($\Theta$). The smooth torus is then illuminated by a central isotropic point AGN emission with a fixed SED of [@Schartmann2005]. The radiation emitted by the smooth torus is calculated by solving the radiative transfer equations [for more details, see Section 2 of @Fritz2006].
The explored parameter ranges, which are introduced by [@Feltre2012], are listed in Table \[tbl:smooth\]; the covered physical space is wider than the original work of [@Fritz2006]. For each of the $24000$ smooth-torus SED templates, we first estimate its $S_{\rm Sil}$ by following the methodology in Section 5.2 of [@Hernan2015]. Second, we calculate the corresponding line-of-sight $N_{\rm H}$ and the $V$-band extinction as follows. The line-of-sight optical depth at $9.7$is $$\tau_{9.7}= \tau_{9.7}^0 \times\mathrm{e}^{-\gamma|\cos (\theta)|} \\,$$ and the corresponding extinction is $$\begin{aligned} \frac{A_{9.7}}{\operatorname{mag}} & \equiv 2.5 \log _{10}\left[F_C(\lambda_p) / F(\lambda_p)\right] \\ &=2.5 \log _{10}
\left[e^{\tau_{9.7}} \right] =1.086 \tau_{9.7} \end{aligned} \\.
\label{eq:A_T}$$ The $V$-band extinction is assumed to be $A_{\mathrm{V}}=19\times A_{9.7}$ [@Roche1985]. Then, the corresponding line-of-sight $N_{\rm H}$ is estimated by considering the dust-to-gas ratio of $A_{\mathrm{V}}/N_{\mathrm{H}}=0.62\times10^{-21}\ \mathrm{mag\;cm^{2}}$ [@Savage1979].
$\Theta$ (degree) $60^\circ$, $100^\circ$, $140^\circ$
------------------- ------------------------------------------------------------------------------------------------------------------------
$q$ 0.00, 0.25, 0.50, 0.75, 1.0
$i$ (degree) $0^\circ$, $10^\circ$, $20^\circ$ , $30^\circ$, $40^\circ$, $50^\circ$, $60^\circ$, $70^\circ$, $80^\circ$, $90^\circ$
$\gamma$ 0.0, 2.0, 4.0, 6.0
$\tau^0_{9.7}$ 0.1, 0.3, 0.6, 1.0, 2.0, 3.0, 6.0, 10.0
$Y$ 10, 30, 60, 100, 150
: The parameter space of the smooth torus model of [@Fritz2006].\[tbl:smooth\]
The relation between $N_{\mathrm{H}}$ and $S_{\rm Sil}$ for the smooth torus model is presented in Figure \[fig:smooth-nh\]. It is evident that SEDs with silicate absorption (i.e., negative $S_{\rm Sil}$) always correspond to significant line-of-sight $N_{\mathrm{H}}$ (i.e., $>10^{22.5} \mathrm{cm^{-2}}$). Therefore, the smooth torus model cannot explain AGNs with small $N_{\mathrm{H}}^{X}$ but evident silicate absorption. The relation between $A_{\mathrm{V}}$ and $S_{\rm Sil}$ is presented in Figure \[fig:smooth-av\]. Again, the smooth torus model cannot explain our observations.
![$S_{\rm Sil}$ as a function of $N_{\mathrm{H}}$. The blue-open circles, yellow-open-thin diamonds, green-open squares, and cyan-open diamonds represent the observations (i.e., $S_{\rm Sil}$ and $N_{\mathrm{H}}^X$) of optical Type 1-1.2, Type 1.5, Type 1.9, and Type 2 AGNs, respectively. The grey histogram represents the two-dimensional distribution of $S_{\rm Sil}$ and $N_{\mathrm{H}}$ for the smooth torus model of [@Fritz2006]. The number of SEDs in each bin is labeled. The smooth torus model cannot account for AGNs with small $N_{\mathrm{H}}^{X}$ but evident silicate absorption. []{data-label="fig:smooth-nh"}](fig7.pdf){width="75.00000%"}
![$S_{\rm Sil}$ as a function of $A_{\mathrm{V}}$. The blue dots represent the observations (i.e., $S_{\rm Sil}$ and $A_{\mathrm{V}}[\rm bH_{\alpha}]$) of our final sample. The grey histogram represents the two-dimensional distribution of $S_{\rm Sil}$ and $A_{\mathrm{V}}$ for the smooth torus model of [@Fritz2006]. The number of SEDs in each bin is indicated by the color bar. The smooth torus model cannot account for AGNs with small $V$-band extinction but evident silicate absorption.[]{data-label="fig:smooth-av"}](fig8.pdf){width="75.00000%"}
### Testing the clumpy torus model {#sect:clumpy}
A clumpy torus model is presented by [@Nenkova2008p1; @Nenkova2008p2]. According to this model, in the radial direction, dusty clouds with the same $V$-band extinction ($\tau_V$) are distributed as $r^{-q}$; in the azimuthal direction, the dusty clouds follow a Gaussian distribution, i.e., $e^{-((90-i)/\sigma)^2}$, where $\sigma$ is the Gaussian width. Other parameters are the ratio of the outer radius to the inner one ($Y$) and the cloud number along the equatorial plane ($N_0$). The clumpy torus is then illuminated by a central AGN source with an SED shape of [@Rowan-Robinson1995].
![Same as Figure \[fig:smooth-nh\] but for the clumpy torus model of [@Nenkova2008p1].[]{data-label="fig:clumpy-nh"}](fig9.pdf){width="75.00000%"}
The explored parameter space of the clumpy torus model is listed in Table \[tbl:clumpy\]. Then, for each of the $1247400$ resulting SEDs, we estimate the corresponding $S_{\mathrm{Sil}}$. The line-of-sight $N_{\rm H}$ and $V$-band extinction are calculated as follows. First, the line-of-sight cloud number is $$n_{\mathrm{los}} = N_0 \times \mathrm{e}^{-(\frac{90-i}{\sigma})^2} \\.$$ Considering again the dust-to-gas ratio of [@Savage1979], the observed hydrogen column density $N_\mathrm{H}$ is $$N_{\mathrm{H}} = \frac{n_{\mathrm{los}}\times 1.086\tau_V}{0.62\times 10^{-21}}\ \mathrm{cm^{-2}} \\.$$
[c|c]{} $\sigma$ (degree) &
--------------------------------------------------------------------------------------------------------
$15^\circ$, $20^\circ$, $25^\circ$, $30^\circ$,$35^\circ$,$40^\circ$,$45^\circ$,$50^\circ$,$55^\circ$,
$60^\circ$,$65^\circ$,$70^\circ$
--------------------------------------------------------------------------------------------------------
: The parameter space of the [@Nenkova2008p1] clumpy torus template library.\[tbl:clumpy\]
\
$q$ & 0.0, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0\
$i$ (degree) & $0^\circ$, $10^\circ$, $20^\circ$ , $30^\circ$, $40^\circ$, $50^\circ$, $60^\circ$, $70^\circ$, $80^\circ$, $90^\circ$\
$\tau_\mathrm{v}$ & 10, 20, 40, 60, 80, 120, 160, 200, 300\
$N_0$ & 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15\
$Y$ & 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100\
The relation between $N_{\mathrm{H}}$ and $S_{\rm Sil}$ ($A_{\mathrm{V}}$ and $S_{\rm Sil}$) for the clumpy torus model is presented in Figure \[fig:clumpy-nh\] (Figure \[fig:clumpy-av\]). Compared with the smooth torus model, the clumpy torus model can explain the observations for most of our sources, especially AGNs with evident silicate absorption and small line-of-sight $N_{\mathrm{H}}$ or $V$-band extinction.
![Same as Figure \[fig:smooth-av\] but for the clumpy torus model of [@Nenkova2008p1].[]{data-label="fig:clumpy-av"}](fig10.pdf){width="75.00000%"}
Additional gas and dust obscuration {#sect:add}
-----------------------------------
The discrepancy between $S_{\mathrm{Sil}}$ and $N_\mathrm{H}^\mathrm{X}$ or $A_{V}[\mathrm{bH}\alpha]$ might also be caused by other effects. For instance, as pointed out by [@Goulding2012], the silicate absorption might be contributed by dust located in the host galaxy rather than the AGN torus. If so, we would expect that sources with larger inclination angles should have stronger silicate absorption. Following [@Goulding2012], we use the ratio of the major isophotal diameter to the minor one (hereafter $R25$) as introduced in the Third Reference Catalog of Bright Galaxies [@Corwin1994] to probe the galaxy inclination angle (sources with smaller $R25$ values might tend to be more face-on). For the $153$ sources in the final sample, we can obtain their $R25$. We then divide these sources into two groups according to their positions relative to the best-fitting relation between $R_{9.7}$ and $N_\mathrm{H}^\mathrm{X}$ (see Section \[sect:nhx\]). That is, we calculate the difference ($\Delta R_{9.7}$) between the observed $R_{9.7}$ and the predicted one from the best-fitting relation. Group 1 (2) sources have positive (negative) $\Delta R_{9.7}$. The distributions of $\log R25$ for the two groups are presented in Figure \[fig:nhx-r25\]. We also perform the Anderson-Darling test to check the differences between the two distributions. We find that the null hypothesis that the two distributions are drawn from the same population cannot be rejected (i.e., the $p$-value of the null hypothesis is much larger than $0.05$). Therefore, we conclude that the discrepancy between $S_{\mathrm{Sil}}$ and $N_\mathrm{H}^\mathrm{X}$ are unlikely to be caused by galaxy-scale dust absorption.
![Distributions of $\log R25$ for two groups of sources with different $\Delta R_{9.7}$. The distributions are normalized to ensure that the total area within each histogram is equal to one. The distribution of $\log R25$ for sources with positive $\Delta R_{9.7}$ (i.e., less silicate absorption) is similar to the distribution of $\log R25$ for sources with negative $\Delta R_{9.7}$. []{data-label="fig:nhx-r25"}](fig11.pdf){width="75.00000%"}
We then check the discrepancy between $S_{\mathrm{Sil}}$ and $A_{V}[\mathrm{bH}\alpha]$ as a function of $R25$. We define $\delta A_{\mathrm{V}}=(A_V[\mathrm{Si}]-A_{V}[\mathrm{bH}\alpha])/A_{V}[\mathrm{bH}\alpha]$ and divide the sources with non-zero $A_V[\mathrm{Si}]$ and $A_{V}[\mathrm{bH}\alpha]$ into two groups, i.e., group 1 with $\delta A_{\mathrm{V}}>3$ and group 2 with $\delta A_{\mathrm{V}}<3$. Again, we find that the two groups share the same distribution of $\log R25$. Therefore, the discrepancy between $S_{\mathrm{Sil}}$ and $A_{V}[\mathrm{bH}\alpha]$ is also unlikely to be mainly driven by galaxy-scale dust absorption.
![Same as Fig. \[fig:nhx-r25\], but for two groups of sources with different $\delta A_{\mathrm{V}}$. The two distributions are statistically consistent.[]{data-label="fig:ha-r25"}](fig12.pdf){width="75.00000%"}
The discrepancy between $S_{\mathrm{Sil}}$ and $A_{V}[\mathrm{bH}\alpha]$ for X-ray obscured sources might be explained as follows. Firstly, as mentioned in Section \[sect:bha\], the estimates of $A_{V}[\mathrm{bH}\alpha]$ rely on the empirical relation between X-ray and broad H$\alpha$ luminosities, which is obtained for type 1.0 or 1.2 Seyferts. However, type 1.0 or 1.2 Seyferts might also suffer dust extinction to some degree; therefore, $A_{V}[\mathrm{bH}\alpha]$ is almost always an under-estimation of true dust extinction. Secondly, for X-ray obscured sources, their hidden broad emission-line fluxes might be scattered into our line of sight (indeed, spectropolarimetry observations revealed high-polarization broad emission lines in at least some of our X-ray obscured sources, e.g., Mrk 3, Mrk 348, and Mrk 1210; see, e.g., @Miller1990 [@Tran1992]). That is, the observed broad $H\alpha$ flux is larger than the direct broad H$\alpha$ flux (which is heavily absorbed by the dusty torus), i.e., $F_{\mathrm{obs}}(\mathrm{bH}\alpha)=
F_{\mathrm{int}}(\mathrm{bH}\alpha)\exp(-\tau(H\alpha)) + F_{\mathrm{sct}}(\mathrm{bH}\alpha)$, where $F_{\mathrm{int}}(\mathrm{bH}\alpha)$, $\tau(H\alpha)$, and $F_{\mathrm{sct}}(\mathrm{bH}\alpha)$ are the intrinsic broad H$\alpha$ flux, the optical depth of H$\alpha$, and the scattered broad H$\alpha$ flux, respectively. Note that the scatted light is assumed to be not absorbed by the dusty torus. If $\tau$ is larger than $9$ (i.e., $A_V\geq 11.95$ for the extinction law of @Cardelli1989), the direct broad-line flux is extinguished and $F_{\mathrm{obs}}(\mathrm{bH}\alpha)\cong
f_{\mathrm{sc}} F_{\mathrm{int}}(\mathrm{bH}\alpha)$, where $f_{\mathrm{sc}}$ is the ratio of $F_{\mathrm{sct}}(\mathrm{bH}\alpha)$ to $F_{\mathrm{int}}(\mathrm{bH}\alpha)$. If so, for such sources, the inferred $A_{V}[\mathrm{bH}\alpha]$ is about $4.02 + 3.32\log (0.05/f_{\mathrm{sc}})$, i.e., $A_{V}[\mathrm{bH}\alpha]$ might thus be systematically underestimated compared to the actual one given that $f_{\mathrm{sc}}$ is about a few percent [e.g., @Reynolds1997]. Thirdly, it is also possible that the broad H$\alpha$ in some type 1.9 AGNs might be produced by strong outflows (i.e., they are actually type 2 AGNs). One such example is 2MASX J07595347+2323241, whose broad H$\alpha$ is surprisingly narrow. In fact, its H$\alpha$-inferred virial black hole mass is two orders of magnitude smaller than the expectation of the $M_{\mathrm{BH}}-\sigma$ relation or the infrared broad Pa$\beta$-inferred one [@Ricci2017a]. Therefore, possibly due to the combined effect of the aforementioned factors, $A_{V}[\mathrm{bH}\alpha]$ is expected to be systematically smaller than the silicate strength-inferred $A_{V}$, at least for some of our X-ray obscured AGNs.
Dust-free gas (e.g., gas in the broad-line regions) can provide additional X-ray obscuration. However, they cannot contribute to silicate absorption or $H\alpha$ extinction. As mentioned in Section \[sect:Results\] (also see Figures \[fig:Shi2006fig\] and \[fig:M12\]), dust-free gas cannot fully account for the discrepancy between $S_{\mathrm{Sil}}$ and $N_\mathrm{H}^\mathrm{X}$ since many sources are type-2 sources.
Summary of conclusions {#sect:sum}
======================
We decompose the *Spitzer*/IRS spectra of $175$ BASS AGNs to measure their silicate strengths and compare them with absorption signatures in X-rays and broad $\rm H\alpha$ emission. Our results are summarized as follows.
1. Consistent with previous work [@Shi2006], we confirm a weak relation between the silicate strength and $N_{\mathrm{H}}^X$ using more accurate $N_{\rm H}^{\rm X}$ measurements; however, the scatter of the relation is quiet large (see Figure \[fig:Shi2006fig\] and Section \[sect:nhx\]).
2. For X-ray unobscured AGNs, the silicate strength and the $\rm H\alpha$-inferred $V$-band extinction are both small; while for X-ray obscured ones, the silicate strength is much stronger than the expectation of the $\rm H\alpha$-inferred $V$-band extinction (see Figure \[fig:avsil\] and Section \[sect:bha\]). This result and the previous one suggest that the distributions and structures of obscuration gas and extinction dust are very complex.
3. We test our data against two popular torus models, i.e., the smooth torus model (see Section \[sect:smooth\]) of [@Fritz2006] and the clumpy torus model (see Section \[sect:clumpy\]) of [@Nenkova2008p1]. We find that the clumpy torus model is more consistent with our observations than the smooth one.
ACKNOWLEDGEMENTS
================
We thank the referee for his/her helpful comments that improved the manuscript. J.X., M.Y.S., Y.Q.X., and J.Y.L. acknowledge the support from NSFC-11973002, NSFC-11890693, NSFC-11421303, the China Postdoctoral Science Foundation (2016M600485), and the CAS Frontier Science Key Research Program (QYZDJ-SSW-SLH006), and the K.C. Wong Education Foundation.
The Combined Atlas of Sources with Spitzer IRS Spectra (CASSIS) is a product of the IRS instrument team, supported by NASA and JPL. CASSIS is supported by the “Programme National de Physique Stellaire” (PNPS) of CNRS/INSU co-funded by CEA and CNES and through the “Programme National Physique et Chimie du Milieu Interstellaire” (PCMI) of CNRS/INSU with INC/INP co-funded by CEA and CNES. \[sect:acknowledgement\]
[99]{} Antonucci, R. 1993, , 31, 473
Baumgartner, W. H., Tueller, J., Markwardt, C. B., et al. 2013, , 207, 19
Burtscher, L., Davies, R. I., Graci[á]{}-Carpio, J., et al. 2016, , 586, A28
Cardelli, J. A., Clayton, G. C., & Mathis, J. S. 1989, , 345, 245
Corwin, H. G., Buta, R. J., & de Vaucouleurs, G. 1994, , 108, 2128
Draine, B. T. 2011, Physics of the Interstellar and Intergalactic Medium by Bruce T. Draine. Princeton University Press
Fritz, J., Franceschini, A., & Hatziminaoglou, E. 2006, , 366, 767
Feltre, A., Hatziminaoglou, E., Fritz, J., et al. 2012, , 426, 120
Garc[í]{}a-Burillo, S., Combes, F., Ramos Almeida, C., et al. 2019, , 632, A61
Goulding, A. D., Alexander, D. M., Bauer, F. E., et al. 2012, , 755, 5
Gravity Collaboration, Pfuhl, O., Davies, R., et al. 2020, , 634, A1
Hao, L., Spoon, H. W. W., Sloan, G. C., et al. 2005, , 625, L75
Hern[á]{}n-Caballero, A., Alonso-Herrero, A., Hatziminaoglou, E., et al. 2015, , 803, 109
Ichikawa, K., Ricci, C., Ueda, Y., et al. 2019, , 870, 31
Imanishi, M., Nakanishi, K., Izumi, T., et al. 2018, , 853, L25
Jaffarian, G. W., & Gaskell, C. M. 2020, , 493, 930
Koss, M., Trakhtenbrot, B., Ricci, C., et al. 2017, , 850, 74
Lebouteiller, V., Barry, D. J., Spoon, H. W. W., et al. 2011, , 196, 8
Li, J., Xue, Y., Sun, M., et al. 2019, , 877, 5
Merloni, A., Bongiorno, A., Brusa, M., et al. 2014, , 437, 3550
Miller, J. S., & Goodrich, R. W. 1990, , 355, 456
Mullaney, J. R., Alexander, D. M., Goulding, A. D., & Hickox, R. C. 2011, , 414, 1082
Nenkova, M., Sirocky, M. M., Ivezi[ć]{}, [Ž]{}., & Elitzur, M. 2008a, , 685, 147
Nenkova, M., Sirocky, M. M., Nikutta, R., Ivezi[ć]{}, [Ž]{}., & Elitzur, M. 2008b, , 685, 160
Netzer, H. 2015, , 53, 365
Nikutta, R., Elitzur, M., & Lacy, M. 2009, , 707, 1550
Reynolds, C. S., Ward, M. J., Fabian, A. C., et al. 1997, , 291, 403
Ricci, C., Ueda, Y., Koss, M. J., et al. 2015, , 815, L13
Ricci, F., La Franca, F., Marconi, A., et al. 2017a, , 471, L41
Ricci, C., Trakhtenbrot, B., Koss, M. J., et al. 2017b, , 233, 17
Roche, P. F., & Aitken, D. K. 1985, , 215, 425
Rowan-Robinson, M. 1995, , 272, 737
Savage, B. D., & Mathis, J. S. 1979, , 17, 73
Schartmann, M., Meisenheimer, K., Camenzind, M., et al. 2005, , 437, 861
Schnorr-M[ü]{}ller, A., Davies, R. I., Korista, K. T., et al. 2016, , 462, 3570
Shi, Y., Rieke, G. H., Hines, D. C., et al. 2006, , 653, 127
Shimizu, T. T., Davies, R. I., Koss, M., et al. 2018, , 856, 154.
Siebenmorgen, R., Kr[ü]{}gel, E., & Spoon, H. W. W. 2004, , 414, 123
Siebenmorgen, R., Haas, M., Kr[ü]{}gel, E., et al. 2005, , 436, L5
Siebenmorgen, R., Heymann, F., & Efstathiou, A. 2015, , 583, A120
Sturm, E., Hasinger, G., Lehmann, I., et al. 2006, , 642, 81
Tran, H. D., Miller, J. S., & Kay, L. E. 1992, , 397, 452
Urry, C. M., & Padovani, P. 1995, , 107, 803
Yang, G., Brandt, W. N., Luo, B., et al. 2016, , 831, 145
\[lastpage\]
[^1]: For more details, refer to <https://www.bass-survey.com>
[^2]: For more details, refer to <https://cassis.sirtf.com/>.
[^3]: For more details, refer to <http://www.denebola.org/ahc/deblendIRS/>.
[^4]: We use *lnr.py* to perform the fit. This *Python* code is available at <https://www.astro.princeton.edu/~sifon/pycorner/lnr/>.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The complex pattern of presence and absence of many genes across different species provides tantalising clues as to how genes evolved through the processes of gene genesis, gene loss and lateral gene transfer (LGT). The extent of LGT, particularly in prokaryotes, and its implications for creating a ‘network of life’ rather than a ‘tree of life’ is controversial. In this paper, we formally model the problem of quantifying LGT, and provide exact mathematical bounds, and new computational results. In particular, we investigate the computational complexity of quantifying the extent of LGT under the simple models of gene genesis, loss and transfer on which a recent heuristic analysis of biological data relied. Our approach takes advantage of a relationship between LGT optimization and graph-theoretical concepts such as tree width and network flow.\
author:
- 'Leo van Iersel, Charles Semple and Mike Steel [^1]\'
title: 'Quantifying the Extent of Lateral Gene Transfer Required to Avert a ‘Genome of Eden’'
---
[**Keywords:**]{} tree, phylogenetic network, lateral gene transfer, tree-width
[**Email:**]{} [email protected], [email protected], [email protected]
INTRODUCTION
============
Modern sequencing technology is providing an increasingly detailed picture of the distribution of genes across a wide array of taxa. Some molecular biologists have used these data to argue that unless ancestral genomes were considerably larger than present-day ones, extensive lateral gene transfer (LGT) must be invoked to explain the current distribution of genes [@dag07], [@dag08], [@mir03]. LGT is a process by which a gene (or genes) from one species is transferred into the genotype of another species by various genetic mechanisms. The extent of LGT is controversial, but it has been argued to be widespread in prokaryotes (e.g. bacteria) and during the earlier epochs of evolution, suggesting in turn that a network, rather than a tree, best describes the evolution of life [@doo07].
Although the pattern of presence and absence of different genes across a set of species can suggest that LGT events occurred in the evolution of these species, another explanation is that certain genes are simply lost in different lineages. As a result, various attempts to quantify the extent of LGT based on gene content have been developed, typically based either on most-parsimonious scenarios or on stochastic models of gene genesis, loss and transfer (see, for example, [@dag07], [@jin07], [@spe06]). Attempts to reconstruct evolutionary histories under the assumption that no LGT events have occurred (and that genes arise just once) imply that some common ancestors of the considered species must have had far more genes than their current-day descendants. Doolittle [*et al.*]{} [@doo03] refer to such an unlikely all-encompassing ancestral genome as the ‘genome of Eden’ hypothesis. Allowing LGT events reduces the need for genes to be present at earlier species, as illustrated for a single gene in Fig. \[intro\].
In this paper, we exploit the combinatorial structure that underlies a key biological insight on which a recent heuristics analysis of data was based by [@dag07] (see also [@dag08], [@mir03]). This insight is that simple models of gene evolution, in which a gene typically arises just once (gene genesis) but can be lost multiple times, imply lower bounds on the extent of LGT simply to prevent hypothetical ancestral genomes from becoming unfeasibly large. For such a model, we aim to bound the number of gene transfer events that have occurred in the evolution of a set of taxa, based on the presence/absence patterns of genes in each of these taxa, assuming that ancestral genomes are bounded by a given size.
Notice that we wish to count transfer events (rather than the total number of genes that are transferred), since in each transfer event, several genes may be transferred from one species into another. Thus our count of LGTs is conservative, and recognizes that genes are not independently transferred and that a transfer event may insert a section of the genome (with several genes) into an individual organism of a different species.
The structure of this paper is as follows. In the next section, we define the model of gene genesis, loss and transfer precisely, and summarize our main results. We then provide proofs of these results in subsequent sections, and end with some concluding comments and a conjecture.
MATHEMATICAL MODEL AND SUMMARY OF MAIN RESULTS
==============================================
Definitions and model specification
-----------------------------------
We begin by recalling some notation concerning digraphs, and phylogenetic trees and networks.
Let $v$ be a vertex of a digraph $D$. The [*indegree*]{} of $v$ is the number of arcs directed into $v$, while the [*outdegree*]{} of $v$ is the number arcs directed out of $v$. The indegree of $v$ is denoted by $d^-(v)$ and the outdegree of $v$ is denoted by $d^+(v)$. The *degree* of $v$ is $d^-(v)+d^+(v)$. Furthermore, $u$ is an [*in-neighbour*]{} of $v$ if $(u,v)$ is an arc in $D$, while $w$ is an [*out-neighbour*]{} of $v$ if $(v,w)$ is an arc in $D$. A digraph $D$ is [*rooted*]{} if there exists a vertex, $\rho$ say, of indegree zero such that, for each vertex $v$ in $D$, there exists a directed path from $\rho$ to $v$.
Throughout the paper, ${{\mathcal X}}$ will denote a finite set of taxa and ${{\mathcal G}}$ will denote a finite set of genes. A [*phylogenetic tree (on ${{\mathcal X}}$)*]{} is a rooted tree whose root has degree at least two and all other internal vertices have degree at least three, and whose leaf set is ${{\mathcal X}}$. More generally, a [*phylogenetic network $N$ (on ${{\mathcal X}}$)*]{} is a rooted acyclic digraph with the following properties:
- the root has outdegree at least two and, for all vertices $v$ with $d^+(v)=1$, we have $d^-(v)\ge 2$; and
- the set of vertices of outdegree zero is ${{\mathcal X}}$.
The elements of ${{\mathcal X}}$ are the [*leaves*]{} of $N$. For a subset $U$ of the vertex set of $N$, the sub-digraph of $N=(V,A)$ [*induced by $U$*]{} is the digraph whose vertex set is $U$, and whose arc set is the subset $\{(u,v): \mbox{$u,v\in U$ and $(u,v)\in A$}\}$ of $A$.
We now describe the model of gene genesis, loss, and transfer. For each taxon $x\in{{\mathcal X}}$, assume that the subset $G(x)$ of ${{\mathcal G}}$ consisting of the genes in ${{\mathcal G}}$ that have been observed in taxon $x$ is known. We refer to the associated map $G: {{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ as a [*genome assignment*]{}. Let $N=(V,A)$ be a phylogenetic network on ${{\mathcal X}}$. For a fixed positive integer $k$, and a genome assignment $G: {{\mathcal X}}\rightarrow 2^{{\mathcal G}}$, a [*$(G,k)$-gene labelling of $N$*]{} is a mapping $F:V\rightarrow 2^{{{\mathcal G}}}$ such that the following hold:
- $F(x)=G(x)$ for each $x\in{{\mathcal X}}$;
- $|F(v)|\le k$ for all $v\in V$;
- For each gene $g\in{{\mathcal G}}$, the sub-digraph of $N$ induced by $\{v\in V : g\in F(v)\}$ is rooted (and therefore connected).
Note that if $x\in {{\mathcal X}}$ and $|G(x)|>k$, then $N$ has no $(G,k)$-labelling. If $N$ has a $(G,k)$-labelling, we say that $N$ [*exhibits*]{} such a labelling. A gene labelling describes a possible evolution of the genes observed in the taxa under consideration. Property (I) says that each leaf of the network is labelled by the set of genes observed in the corresponding taxon. Property (II) demands that each vertex is labelled by a set of at most $k$ genes; the parameter $k$ thus bounds the sizes of the ancestral genomes. Lastly, (III), means that each gene in ${{\mathcal G}}$ is created once at most. There is no restriction on the number of times a gene is lost.
Any function $F$ which satisfies properties $(I)$ and $(III)$ we will call a [*$G$-gene labelling*]{}. With these definitions in hand we can now state the main results of this paper.
Bounding the number of gene transfers required
----------------------------------------------
Our first result establishes lower and upper bounds on the number of LGT events required to explain a given data set. Suppose our input is given by a rooted phylogenetic tree $T$ on ${{\mathcal X}}$ (“species tree”), a genome assignment $G: {{\mathcal X}}\rightarrow 2^{{\mathcal G}}$, and a positive integer $k$. Given a phylogenetic network $N$, we say that $N$ can be obtained from $T$ by *adding* $h$ *arcs*, if there is a subgraph $T'$ of $N$ that is a subdivision of $T$ (i.e. $T'$ can be obtained from $T$ by replacing arcs by directed paths) and $h$ arcs of $N$ are not arcs of $T'$. Here, one views these added arcs as LGT events.
We are interested in the minimum number of LGT events that must be added to $T$ in order for the resulting network to exhibit a $(G,k)$-gene labelling. We denote this minimum number by $\ell(T,G,k)$. Given the above input, Theorem \[prop:LB\] provides lower and upper bounds for $\ell(T,G,k)$. For a vertex $v$ of $T$, let $n(v)$ denote the number of genes $g\in{{\mathcal G}}$ for which there exist two leaves $x_1,x_2\in{{\mathcal X}}$ such that $g\in G(x_1)$, $g\in G(x_2)$ and the most recent common ancestor of $x_1$ and $x_2$ in $T$ is $v$.
\[prop:LB\] Let $T=(V,E)$ be a rooted phylogenetic tree on $\mathcal{X}$, let ${{\mathcal G}}$ be a set of genes, let $G: {{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment, and let $k$ be a positive integer. Then:
- $\ell(T,G,k) \geq \sqrt{\frac{2}{3}|\{v\in V : n(v)>k\}|}$.
- $\ell(T,G,k) \leq \Big\lceil\frac{|\mathcal{G}|-k}{k}\Big\rceil \cdot (|\mathcal{X}|+1)$.
The proof of Theorem \[prop:LB\] is given in Section \[boundproof\].
Hardness results
----------------
The next two results show that two fundamental decision questions concerning the existence of $(G, k$)-labellings are NP-complete. First, consider the following problem:
[lp[0.85]{}]{}\
*Given:* & A phylogenetic network $N$ on ${{\mathcal X}}$, a finite set $\mathcal G$ of genes, a genome assignment $G:{{\mathcal X}}\rightarrow 2^{{\mathcal G}}$, and a positive integer $k$.\
*Question:* & Does $N$ exhibit a $(G,k)$-labelling?\
\[prop:homeomorphism\] The decision problem [Gene Labelling]{} is NP-complete even if $k=1$.
A related problem, but concerning rooted phylogenetic trees, is the following:
[lp[0.85]{}]{}\
*Given:* & A finite set $\mathcal X$ of taxa, a finite set $\mathcal G$ of genes, a genome assignment $G:{{\mathcal X}}\rightarrow 2^{{\mathcal G}}$, and a positive integer $k$.\
*Question:* & Does there exist a rooted phylogenetic tree $N$ on ${{\mathcal X}}$ that exhibits a $(G,k)$-labelling?\
\[Gktree\] The decision problem [$(G,k)$-Tree]{} is NP-complete.
The proofs of these two theorems are established Section \[unravel\].
Algorithms
----------
Despite the apparent intractability of the two problems described above, there are instances for which there exist polynomial-time algorithms. Several such instances are described in Section \[easysec\]. One in particular is given next.
Let $N$ be a phylogenetic network on ${{\mathcal X}}$. A sequence of vertices and arcs is an [*underlying cycle*]{} of $N$ if it is a cycle of the underlying graph (i.e the undirected graph obtained by ignoring the directions of the arcs). A phylogenetic network $N$ on ${{\mathcal X}}$ is a [*galled tree*]{} if, for each pair $C$ and $D$ of underlying cycles, the vertex sets of $C$ and $D$ are disjoint. Each such cycle is called a [*gall*]{}. Theorem \[prop:level1\] shows that restricting the phylogenetic networks in [Gene Labelling]{} to galled trees, the decision problem becomes polynomial-time solvable.
\[prop:level1\] Let $N$ be a galled tree on ${{\mathcal X}}$, let ${{\mathcal G}}$ be a set of genes, let $G: {{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment, and let $k$ be a positive integer. Then there is a polynomial-time algorithm for deciding whether or not $N$ exhibits a $(G,k)$-gene labelling.
Theorem \[prop:level1\], together with the following corollary, is established in Section \[easysec\].
\[gallcor\] Let $T$ be a rooted phylogenetic tree on ${{\mathcal X}}$, let ${{\mathcal G}}$ be a set of genes, let $G: {{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment, and let $k$ be a positive integer. If $h$ is a fixed positive integer, then there is a polynomial-time algorithm for deciding whether or not there is a galled tree $N$ on ${{\mathcal X}}$ that can be obtained from $T$ by adding at most $h$ arcs and which exhibits a $(G,k)$-gene labelling.
HOW MANY GENE TRANSFERS ARE NEEDED? {#boundproof}
===================================
In this section, we prove Theorem \[prop:LB\].
[*Proof of Theorem \[prop:LB\].*]{} For the proof of (i), suppose that a network $N$ admitting a $(G,k)$-gene labelling can be obtained by adding $\ell(T,G,k)$ arcs to $T$. It follows that there exists a tree $T'$ that is a subdivision of $T$ and a subgraph of $N$. In other words, $T'$ is an embedding of $T$ in $N$. An arc of $N$ is said to be an *lgt-arc* if it is not an arc of $T'$. Consider two leaves $x_1,x_2$ and their lowest common ancestor $v$ in $T'$. Suppose that for a gene $g\in\mathcal{G}$ we have $g\in G(x_1)$ and $g\in G(x_2)$. Since network $N$ admits a $(G,k)$-gene labeling $F$, there has to be an undirected path from $x_1$ to $x_2$ in $N$ containing only vertices $u$ with $g\in F(u)$. Furthermore, at least one such undirected path has to consist of two directed paths, one ending in $x_1$ and one ending in $x_2$, since the subgraph of $N$ induced by $\{v\in V | g\in F(v)\}$ is rooted and hence contains a rooted tree. There are four possibilities. Firstly, it is possible that this undirected $x_1-x_2$-path passes through $v$, implying that $g\in F(v)$. The remaining three cases are illustrated in Fig. \[fig:bound\]. The first case is that the undirected $x_1-x_2$-path uses an lgt-arc $(a,d)$ between two vertices $a,d$ that have $v$ as their lowest common ancestor in $T'$. A second possibility is that the path uses two lgt-arcs $(a,b)$ and $(c,d)$ such that $v$ is the lowest common ancestor of $a$ and $d$ in $T'$. Finally, it is also possible that the path uses two lgt-arcs $(b,a)$ and $(c,d)$ such that $v$ is the lowest common ancestor of $a$ and $d$ in $T'$. Thus, for any vertex $v$ with $n(v)>k$, there has to be either an lgt-arc $(a,d)$ or two lgt-arcs $(a,b),(c,d)$ or two lgt-arcs $(b,a),(c,d)$, with $a$ and $d$ two vertices that have $v$ as their lowest common ancestor in $T'$.
Given a vertex $v$ of $T$, we say that an lgt-arc $(s,t)$ *satisfies* $v$ if $v$ is the lowest common ancestor of $s$ and $t$ in $T'$. Since in a tree there is a unique lowest common ancestor, each single lgt-arc satisfies at most one vertex. Furthermore, we say that a pair of lgt-arcs $\{(s,t),(s',t')\}$ *satisfies* $v$ if $v$ is the lowest common ancestor of either $s$ and $t'$, or of $s'$ and $t$ or of $t$ and $t'$ in $T'$. It follows directly that each pair of lgt-arcs satisfies at most three vertices. Since there are $\ell(T,G,k)$ lgt-arcs, in total at most $3{\ell(T,G,k)\choose 2} + \ell(T,G,k)$ vertices $v$ with $n(v)>k$ can be satisfied. From the previous paragraph we know that each vertex $v$ with $n(v)>k$ needs to be satisfied, either by a single lgt-arc or by a pair of lgt-arcs. It follows that there can be at most $3{\ell(T,G,k)\choose 2} + \ell(T,G,k)$ vertices $v$ with $n(v)>k$. Part (i) follows by generously bounding $3{\ell(T,G,k)\choose 2} + \ell(T,G,k)$ by $\frac{3}{2}\ell(T,G,k)^2$.
![Illustration for the proof of Theorem \[prop:LB\]. The three cases apply, without loss of generality, whenever $g\in G(x_1)$, $g\in G(x_2)$, but $g\not\in F(v)$, where $v$ is the lowest common ancestor of $x_1$ and $x_2$ in $T'$. Straight lines denote arcs, while curves denote paths. Solid curves are in $T'$, while dotted lines/curves can be either in $T'$ or only in $N$.[]{data-label="fig:bound"}](bound.pdf){width="\textwidth"}
For (ii), we can construct a network $N$ admitting a $(G,k)$-gene labelling as follows. We select a set $G^0$ of $k$ arbitrary genes in $\mathcal{G}$ and set $F(v)=G^0$ for each internal vertex $v$ of $T$. The third property of a $(G,k)$-gene labelling is now satisfied for the genes in $G^0$. For the remaining $|\mathcal{G}|-k$ genes we do the following. We introduce $f = \lceil\frac{|\mathcal{G}|-k}{k}\rceil$ additional isolated vertices $v_1,\ldots ,v_f$ and label these vertices by disjoint sets $F(v_1),\ldots ,F(v_f)$ that partition $\mathcal{G}\setminus G^0$ and contain at most $k$ genes each. Finally, we add arcs from the root to each $v_i$ and from each $v_i$ to each leaf $x$ with $G(x)\cap F(v_i)\neq\emptyset$. This leads to the claimed upper bound.
To improve upon this simple upper bound turns out to be challenging. This can perhaps be explained by the results in the next section, in which we show that, even if the network $N$ is given and $k=1$, it is NP-complete to decide if a $(G,k)$-gene labelling of $N$ exists.
UNRAVELLING LATERAL GENE TRANSFER IS HARD {#unravel}
=========================================
We begin this section by first showing that [Gene Labelling]{} is NP-complete. First consider the following decision problem:
[lp[0.85]{}]{}\
*Given:* & Directed acyclic graphs $D=(V_D,E_D)$ and $P=(V_P,E_P)$ with $V_P\subseteq V_D$.\
*Question:* & Is $P$ homeomorphic to a subgraph of $D$?\
A graph $P$ is *homeomorphic* to a graph $H$ if $H$ can be obtained from $P$ by replacing arcs $(u,v)$ by internally vertex-disjoint directed $u-v$ paths. Hence, <span style="font-variant:small-caps;">DASH</span> can be seen as a disjoint-paths problem. The graph $P$ is called the “pattern graph”. It was observed by Fortune et al. [@FortuneEtAl1980] that NP-hardness of <span style="font-variant:small-caps;">DASH</span> follows from a result of Even, Itai, and Shamir [@EvenEtAl1976] on multi-commodity flows.
[**Theorem \[prop:homeomorphism\].**]{} [*The decision problem [Gene Labelling]{} is NP-complete even if $k=1$.*]{}
[*Proof.*]{} The reduction is from [DASH]{}. Let $(D,P)$ be an instance of <span style="font-variant:small-caps;">DASH</span>. We begin by showing that we may assume, for each vertex $u$ in $P$, we have $d_P^-(u)+d_P^+(u)=1$. To see this, let $D'$ and $P'$ be the digraphs obtained from $D$ and $P$, respectively, by iteratively doing the following for each vertex $v$ in $P$:
- Let $\{s_1,s_2,\ldots,s_i\}$ be the set of in-neighbours of $v$ in $P$ and let $\{t_1,t_2,\ldots,t_j\}$ be the set of out-neighbours of $v$ in $P$.
- In $P$, replace $v$ and the arcs $(s_1,v),\ldots,(s_i,v)$ and $(v,t_1),\ldots,(v,t_j)$ with the new vertices $v_1,v_2,\ldots,v_{i+j}$ and the new arcs $(s_1,v_1),\ldots,(s_i,v_i)$ and $(v_{i+1},t_1),\ldots,(v_{i+j},t_j)$.
- Let $\{x_1,x_2,\ldots,x_r\}$ be the set of in-neighbours of $v$ in $D$ and let $\{y_1,y_2,\ldots,y_s\}$ be the set of out-neighbours of $v$ in $D$.
- In $D$, replace $v$ and the arcs $(x_1,v),\ldots,(x_r,v)$ and $(v,y_1),\ldots,(v,y_s)$ with the new vertices $v_1,v_2,\ldots,v_{i+j}$ and the new arcs $$\begin{aligned}
(x_1,v_1),(x_2,v_1),\ldots, (x_r,v_1), (x_1,v_2), & (x_2,v_2),\ldots, (x_r,v_2), \\
& \ldots,(x_1,v_i),(x_2,v_i),\ldots,(x_r,v_i)\end{aligned}$$ and $$\begin{aligned}
(v_{i+1},y_1),(v_{i+1},y_2),\ldots, (v_{i+1},y_s), & (v_{i+2},y_1), (v_{i+2},y_2),\ldots, (v_{i+2},y_s),\\
& \ldots,(v_{i+j},y_1),(v_{i+j},y_2),
\ldots,(v_{i+j},y_s).\end{aligned}$$
At the end of this iterative construction, for each vertex $u$ in $P'$, we have $d_{P'}^-(u)+d_{P'}^+(u)=1$. Moreover, it is straightforward to check that $P'$ is homeomorphic to a subgraph of $D'$ if and only if $P$ is homeomorphic to a subgraph of $D$. It now follows that we may assume that our given instance $(D,P)$ of [DASH]{} is of the form at the completion of this construction.
We next describe a polynomial-time transformation of our instance $(D,P)$ of DASH into an instance of [Gene Labelling]{} with $k=1$. Set $k=1$. We define $N$, $\mathcal X$, $\mathcal G$, and the function $G:{\mathcal X}\rightarrow 2^{\mathcal G}$ iteratively as follows. Initially, set $\mathcal X$ and $\mathcal G$ to be both empty. Let $N$ be the phylogenetic network obtained from $D=(V,A)$ by applying the following sequence of operations:
- For each arc $a=(u,v)$ of $P$, add a new gene $g_a$ to $\mathcal G$, add new leaf vertices $\ell_u,\ell_v$ to $V$ and to $\mathcal X$, add new arcs $(u,\ell_u)$ and $(v,\ell_v)$ to $A$, and set $G(\ell_u)=G(\ell_v)=\{g_a\}$. Furthermore, delete all incoming arcs of $u$ from $A$. At the end of (I), the constructions of the sets $\mathcal X$ and $\mathcal G$, and the function $G:{\mathcal X}\rightarrow 2^{\mathcal G}$ are completed.
- Repeatedly remove all leaves of the resulting network not in $\mathcal X$ and repeatedly remove all vertices of indegree zero that do not have an element of $\mathcal X$ as a child.
- Finally, root the resulting network by choosing a vertex of indegree zero as a root and then adding an arc from this root to each other vertex of indegree zero. Setting $N$ to be the resulting phylogenetic network on $\mathcal X$, we have now constructed the desired instance of [Gene Labelling]{}.
An example of this construction is shown in Fig. \[fig:homeomorphism\]. Note that, while $D$ may not be connected, $N$ is connected because of (O-III). We complete the proof by showing that $N$ admits a $(G,1)$-gene labelling if and only if $P$ is homeomorphic to a subgraph of $D$.
![An example of the reduction in the proof of Theorem \[prop:homeomorphism\]. From an instance $(P,D)$ of <span style="font-variant:small-caps;">DASH</span>, a phylogenetic network $N$ is constructed with leaf-labelling $G(\ell_u)=G(\ell_v)=\{g_a\}$ and $G(\ell_{u'})=G(\ell_{v'})=\{g_{a'}\}$. Disjoint paths $u\rightarrow w_2\rightarrow v$ and $u'\rightarrow w_4\rightarrow v'$ in $D$ correspond to a labelling $F(\ell_u)=F(u)=F(w_2)=F(v)=F(\ell_v)=\{g_a\},F(\ell_{u'})=F(u')=F(w_4)=F(v')=F(\ell_{v'})=\{g_{a'}\}$.[]{data-label="fig:homeomorphism"}](homeomorphism_reduction.pdf){width=".8\textwidth"}
Suppose that $P$ is homeomorphic to a subgraph of $D$. Then, for each arc $a=(u,v)$ of $P$, there exists a directed $u-v$ path in $D$ such that all these directed paths are pairwise vertex disjoint. We first claim that for each such $u-v$ path in $D$, there exists a corresponding $u-v$ path in $N$. To see this, observe that, in the construction of $N$ from $D$, the only arcs deleted are those arcs directed into a vertex, $u$ say, for which $u$ is a vertex in $P$, and arcs incident with a vertex, $w$ say, for which either there is no directed path from $w$ to a vertex in $\mathcal X$ or there is no directed path from a parent of a vertex in $\mathcal X$ to $w$.
None of these deletions deletes an arc on any $u-v$ path in $D$ and so the claim holds. Now, for each arc $a=(u,v)$ of $P$ and for each vertex $w$ on the associated $u-v$ path in $N$, set $F(w)=\{g_a\}$. Since the children $\ell_u$ of $u$ and $\ell_v$ of $v$ are the only other vertices with a label containing $g_a$, the subgraph of $N=(V,A)$ induced by $\{w\in V | g_a\in F(w)\}$ is rooted and connected. Labelling all remaining vertices $w$ by $F(w)=\emptyset$ thus leads to a $(G,1)$-gene labelling of $N$.
Now suppose that $F$ is a $(G,1)$-gene labelling of $N$. It remains to show that $P$ is homeomorphic to a subgraph of $D$. Consider a gene $g_a\in\mathcal{G}$, and let $a=(u,v)$ be the associated arc of $P$. Since $F$ is a $(G,1)$-gene labelling, the subgraph of $N$ induced by $\{w\in V(N): g_a\in F(w)\}$ is connected. Furthermore, each of the arcs added in (O-III) in the construction of $N$ joins two vertices that are assigned distinct genes in $\mathcal G$ by $F$ as $F$ is a $(G,1)$-labelling of $N$. Thus none of these arcs are contained in the subgraph of $N$ induced by $\{w\in V(N): g_a\in F(w)\}$. Since $u$ has no other incoming arcs, $u$ has indegree zero in this subgraph. Since the child $\ell_v$ of $v$ is also labelled $F(\ell_v)=\{g_a\}$, it follows that $N$ contains a directed path from $u$ to $v$ whose vertices are assigned $\{g_a\}$ under $F$. This path is also a directed path in $D$. Moreover, for two distinct genes $g_a,g_b\in \mathcal G$, these paths are pairwise disjoints and so they are pairwise disjoint in $D$. The union of these paths in $D$ forms a subgraph $H$ of $D$ such that $P$ is homeomorphic to $H$. This completes the proof of the theorem.
We turn now to the proof of Theorem \[Gktree\], which is based on the concepts of tree-width and tree-decomposition from graph theory – we define these notions now; for further background the interested reader may wish to consult [@die].
A *tree decomposition* of a graph $H=(V_H,E_H)$ is a pair $(T,\{X_i : i\in I\})$ where $T=(I,E_T)$ is a tree and, for all $i\in I$, the set $X_i$ is a subset of $V_H$ such that:
- $\bigcup_{i\in I} X_i = V_H$;
- for each $(u,v)\in E_H$, there exists an $i\in I$ with $u,v\in X_i$;
- for each $v\in V_H$, the subgraph of $T$ induced by $\{i\in I : v\in X_i\}$ is connected.
The *width* of the tree decomposition is defined as $\max_{i\in I} |X_i|-1$.
We use the following NP-complete problem for the reduction in the proof of the theorem.
[lp[0.85]{}]{}\
*Given:* & An undirected graph $H=(V_H,E_H)$ and a natural number $k'$.\
*Question:* & Does there exist a tree decomposition of $H$ with width at most $k'$?\
[**Theorem \[Gktree\].**]{} [*The decision problem [$(G,k)$-Tree]{} is NP-complete.*]{}
[*Proof.*]{} The reduction is from [Treewidth]{}. Let $(H,k')$ be an instance of <span style="font-variant:small-caps;">Treewidth</span>, and set $\mathcal{X}=E_H$, $\mathcal{G}=V_H$, $G(x)=\{u,v\}$ for each edge $x=\{u,v\}\in E_H$, and $k=k'+1$. We complete the proof by showing that there exists a tree decomposition of $H$ with width at most $k'$ if and only if there exists a phylogenetic tree $N$ on $\mathcal{X}$ that admits a $(G,k)$-gene labelling.
Firstly, let $(T,\{X_i : i\in I\})$ be a tree decomposition of $H$ with width $k'$. For each $\{u,v\}\in E_H$, there exists an $i\in I$ with $u,v\in X_i$. Hence, for each taxon ${x\in\mathcal{X}}$, there exists a vertex $i$ of $T$ with $G(x)\subseteq X_i$. We construct $N$ from $T$ by choosing an arbitrary vertex as a root, directing all edges away from the root and, for each $x\in\mathcal{X}$, adding a leaf $x$ and an arc $(i,x)$ where $i$ is an arbitrary vertex of $T$ with $G(x)\subseteq X_i$. Repeatedly deleting leaves not in $\mathcal X$, set $N$ to be the resulting rooted phylogenetic tree on $\mathcal X$. We can now obtain a $(G,k)$-gene labelling $F$ of $N$ by setting $F(x)=G(x)$ for each leaf $x\in\mathcal{X}$ and ${F(i)=X_i}$ for each other vertex. For each gene $g\in\mathcal{G}$, the subgraph of ${N=(V,A)}$ induced by $\{v\in V : g\in F(v)\}$ is connected by property (iii) of a tree decomposition, and is rooted as $N$ is a rooted phylogenetic tree.
Now suppose that there exists a phylogenetic tree $N$ on $\mathcal{X}$ and a $(G,k)$-gene labelling $F$ of $N=(V,A)$. Then we can obtain a tree decomposition ${(T,\{X_i : i\in I\})}$ of $H$ by setting $I=V$ and $X_i=F(i)$ for all $i\in I$, and defining $T$ to be the tree obtained from $N$ by ignoring the rooting and thus orientation of each of the arcs. All properties of a tree decomposition are clearly satisfied, and the width is at most $k'=k-1$ because $|F(i)|\leq k$ by the definition of a $(G,k)$-gene labelling.
$\ldots$ BUT SOMETIMES IT IS EASY {#easysec}
=================================
Let $N$ be a galled tree on ${{\mathcal X}}$, let ${{\mathcal G}}$ be a set of genes, let $G:{{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment, and let $k$ be a positive integer. The main result of this section shows that there is a polynomial-time algorithm for deciding whether $N$ exhibits a $(G,k)$-labelling. If $N$ is a phylogenetic tree, then this problem is equivalent to deciding if $\ell(N,G,k)=0$.
Let $T$ be a phylogenetic tree on ${{\mathcal X}}$, let ${{\mathcal G}}$ be a set of genes, let $G:{{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment, and let $k$ be a positive integer. Then there is a polynomial-time algorithm for deciding whether $\ell(T,G,k)=0$. \[zero\]
Deciding whether $\ell(T,G,k)=0$ is equivalent to deciding if $T$ has a $(G,k)$-gene labelling. With this in mind, it is easily seen that the following $G$-gene labelling function $F$ of $T$ minimizes $k$. For all $v\in V$, the gene $g\in {{\mathcal G}}$ is in $F(v)$ precisely if $v$ is a vertex of the minimal subtree of $T$ that connects those leaves $x$ for which $g\in G(x)$. If $|F(v)|\le k$ for $v$, then $F$ is a $(G,k)$-gene labelling; otherwise there is no such gene labelling of $T$.
Proposition \[one\] (below) establishes the main result when $N$ has exactly one gall. We will use this proposition as the base case for an inductive proof of the main result. The proof of this proposition relies on the following construction. Let $N$ be a galled tree on ${{\mathcal X}}$ with exactly one gall. Thus the undirected graph underlying $N$ has exactly one cycle. Label (in order) the vertices of this cycle $w_1,w_2,\ldots, w_p$, where $w_p$ is the unique vertex in $N$ with two arcs directed into it.
Let $F^*$ be the following map from the vertex set $V$ of $N$ to $2^{{\mathcal G}}$. For each $v\in V$, the gene $g\in {{\mathcal G}}$ is in $F^*(v)$ precisely if, ignoring the direction of the arcs, either:
- there is a pair of leaves $x_1$ and $x_2$ with $g\in G(x_1)$ and $g\in G(x_2)$, and $v$ is on a path between $x_1$ and $x_2$ that avoids $w_p$, or
- there is a pair of leaves $x_1$ and $x_2$ with $g\in G(x_1)$ and $g\in G(x_2)$, and $v$ is on [*all*]{} paths between $x_1$ and $x_2$.
The following two observations are important for what follows. First, if $F$ is a $G$-gene labelling of $N$, then it is easily seen that $F^*(v)\subseteq F(v)$ for all $v\in V$. Second, $F^*$ is not necessarily a $G$-gene labelling of $N$. The exact reason for this is that there can be a gene $g\in {{\mathcal G}}$ such that the sub-digraph of $N$ induced by $\{v\in V : g\in F^*(v)\}$ consists of two rooted connected components; one lying below $w_p$ (more precisely, in the subgraph of $N$ induced by the vertices that are reachable from $w_p$ by a directed path) and at one lying above $w_p$ (more precisely, in the subgraph of $N$ induced by the vertices that are not reachable from $w_p$ by a directed path).
Now let ${{\mathcal G}}'$ be the subset of genes $g\in {{\mathcal G}}$ for which the sub-digraph of $N$ induced by $\{v\in V : g\in F^*(v)\}$ is disconnected. We extend $F^*$ to a $G$-gene labelling $F$ of $N$ by reformulating the problem as an undirected network flow problem and then using its solution to identify the extension. Here one can view each edge $\{a,b\}$ as the two arcs $(a,b)$ and $(b,a)$. We construct an undirected graph $U$ from $N$ by starting with the sub-digraph of $N$ induced by $\{w_1,w_2,\ldots,w_p\}$ and ignoring the direction of the arcs, adding a source vertex $s$, and, for each gene $g\in{{\mathcal G}}'$, adding a new vertex $s_g$ and the three edges $\{s,s_g\}$, $\{s_g,w_{i_1-1}\}$, and $\{s_g,w_{i_2+1}\}$, where $i_1$ and $i_2$ are the smallest and largest index $i\neq p$ for which $g\in F^*(w_i)$. Now assign each $s_g$ capacity $1$ and, for each $i\in \{1,2,\ldots,p-1\}$, assign $w_i$ capacity $k-|F^*(w_i)|$.
To illustrate the above construction, consider the galled tree $N$ shown in Fig. \[fig:exampleinput\]. Each leaf $x$ of $N$ is labelled by the set $G(x)$ of input genes observed in the corresponding taxon. The map $F^*$ is shown in Fig. \[fig:fstar\]. The undirected graph $U$ with $k=3$ is shown in Fig. \[fig:auxgraph\].
[ ![(a) A galled tree $N$ with one gall. Each leaf $x$ of $N$ is labelled by $G(x)$. (b) The initial labelling $F^*$ in which, for example, the sub-digraph of $N$ induced by $\{v\in V : 4\in F^*(v)\}$ (displayed by the dashed arcs and their end vertices) consists of two connected components. (c) Auxiliary graph $U$ with capacities in parentheses. (d) A $(G,3)$-gene labelling of $N$. This gene labelling corresponds to a maximum flow in $U$ which sends one unit of flow through $s_1$ and $w_1$ and one unit of flow through $s_4$ and $w_4$. []{data-label="fig:example"}](inputdigraph.pdf "fig:") \[fig:exampleinput\] ]{}
[ ![(a) A galled tree $N$ with one gall. Each leaf $x$ of $N$ is labelled by $G(x)$. (b) The initial labelling $F^*$ in which, for example, the sub-digraph of $N$ induced by $\{v\in V : 4\in F^*(v)\}$ (displayed by the dashed arcs and their end vertices) consists of two connected components. (c) Auxiliary graph $U$ with capacities in parentheses. (d) A $(G,3)$-gene labelling of $N$. This gene labelling corresponds to a maximum flow in $U$ which sends one unit of flow through $s_1$ and $w_1$ and one unit of flow through $s_4$ and $w_4$. []{data-label="fig:example"}](fstar.pdf "fig:") \[fig:fstar\] ]{}
[ ![(a) A galled tree $N$ with one gall. Each leaf $x$ of $N$ is labelled by $G(x)$. (b) The initial labelling $F^*$ in which, for example, the sub-digraph of $N$ induced by $\{v\in V : 4\in F^*(v)\}$ (displayed by the dashed arcs and their end vertices) consists of two connected components. (c) Auxiliary graph $U$ with capacities in parentheses. (d) A $(G,3)$-gene labelling of $N$. This gene labelling corresponds to a maximum flow in $U$ which sends one unit of flow through $s_1$ and $w_1$ and one unit of flow through $s_4$ and $w_4$. []{data-label="fig:example"}](Dprime.pdf "fig:") \[fig:auxgraph\] ]{}
[ ![(a) A galled tree $N$ with one gall. Each leaf $x$ of $N$ is labelled by $G(x)$. (b) The initial labelling $F^*$ in which, for example, the sub-digraph of $N$ induced by $\{v\in V : 4\in F^*(v)\}$ (displayed by the dashed arcs and their end vertices) consists of two connected components. (c) Auxiliary graph $U$ with capacities in parentheses. (d) A $(G,3)$-gene labelling of $N$. This gene labelling corresponds to a maximum flow in $U$ which sends one unit of flow through $s_1$ and $w_1$ and one unit of flow through $s_4$ and $w_4$. []{data-label="fig:example"}](fopt.pdf "fig:") \[fig:fopt\] ]{}
There exists an integer flow in $U$ from $s$ to $w_p$ with value $|{{\mathcal G}}'|$ if and only if there exists a $(G,k)$-gene labelling of $N$. Moreover, if there is such an integer flow, then it leads to a $(G,k)$-gene labelling of $N$. \[flow\]
First suppose that there exists such a flow $f$ with value $|{{\mathcal G}}'|$. Based on $f$, we show that there exists a $(G,k)$-labelling $F$ of $N$. For this existence proof, we assume that we know the path that each unit of flow takes. We will conclude the proof by showing how an actual $(G,k)$-labelling can be constructed.
Initially set $F=F^*$. Since $f$ has value $|{{\mathcal G}}'|$ and each $s_g$ has capacity $1$, there is exactly one unit of flow passing through $s_g$ from $s$ to $w_p$. Furthermore, as $f$ is integer, it uses exactly one of the two edges $\{s_g,w_{i_1-1}\}$ and $\{s_g,w_{i_2+1}\}$. If $f$ uses $\{s_g,w_{i_1-1}\}$, then the corresponding unit of flow either uses the vertices on the path from $w_{i_1-1}$ to $w_p$ through $\{w_1,w_p\}$ or the vertices on the path from $w_{i_1-1}$ to $w_p$ through $\{w_{p-1},w_p\}$. Depending on which of these paths this unit of flow takes, add $g$ to $F(w_i)$ for each of the vertices on this path. Similarly, if $f$ uses $\{s_g,w_{i_2+1}\}$, then the corresponding unit of flow either uses the vertices on the path from $w_{i_2+1}$ to $w_p$ through $\{w_1,w_p\}$ or the vertices on the path from $w_{i_2+1}$ to $w_p$ through $\{w_{p-1},w_p\}$. Depending on which of these paths this unit of flow takes, add $g$ to $F(w_i)$ for each of the vertices on this path. Doing this for each $g\in {{\mathcal G}}'$, we claim that the resulting map $F:V\rightarrow 2^{{{\mathcal G}}}$ is a $(G,k)$-labelling of $N$. Clearly, $F$ satisfies (III). Furthermore, as each vertex $w_i$ has capacity $k-|F^*(w_i)|$, the cardinality of $F(w_i)$ is at most $k$. Thus $F$ satisfies (II). It now follows that $F$ is a $(G,k)$-labelling of $N$.
Now suppose that there exists a $(G,k)$-gene labelling $F$ of $N$. By one of the two observations earlier, $F^*(v)\subseteq F(v)$ for all $v\in V$. Consider a gene $g\in {{\mathcal G}}'$. The sub-digraph of $N$ induced by $\{v\in V : g\in F^*(v)\}$ consists of two rooted connected components. However, by (III), the sub-digraph of $N$ induced by $\{v\in V : g\in F(v)\}$ is rooted and connected. Therefore, there is a path on the cycle consisting of vertices $w_i$ with $g\in F(w_i)-F^*(w_i)$ that connects the two components. Sending one unit of flow from $s$ to the first vertex on this path via $s_g$, and then along this path to $w_p$ for each $g\in {{\mathcal G}}'$ gives a desired integer flow.
We have now shown that if there is an integer flow $f$ from $s$ to $w_p$ with value $|{{\mathcal G}}'|$, then there is a $(G,k)$-labelling of $N$. This does not directly give such a labelling as we can make no distinction on the flow units. In particular, it is not directly clear which of the two paths a flow unit takes once it reaches a vertex $w_i$ in the cycle. This can be rectified as follows. Let $f$ be such a flow and let $g\in {{\mathcal G}}'$. Ignoring the vertices $w_{i_1},\ldots,w_{i_2}$, either the flow unit through $s_g$ takes the path from $w_{i_1-1}$ to $w_p$ via $w_1$ or the path from $w_{i_2+1}$ to $w_p$ via $w_{p-1}$. To make this decision, consider the following modification of the integer flow problem. Extend $F^*$ to $F^*_g$ by adding $g$ to each of $F^*(w_{i_1-1}),\ldots,F^*(w_1)$ and, for each of these vertices, subtract one from their capacities. If there is an integer flow from $s$ to $w_p$ in $U{\backslash}s_g$ of $|{{\mathcal G}}'|-1$ units, then we may assume that the unit of flow through $s_g$ in $U$ follows the path from $w_{i_1-1}$ to $w_p$ via $w_{p-1}$. In this case, replace $F^*$ with $F^*_g$ and $U$ with $U{\backslash}s_g$, and repeat for another element in ${{\mathcal G}}'-g$. If there is no such integer flow in $U{\backslash}s_g$, then the unit of flow through $s_g$ in $U$ follows the path from $w_{i_2+1}$ to $w_p$ via $w_{p-1}$. In this second case, replace $F^*$ with that obtained by adding $g$ to each of $F^*(w_{i_1-1}),\ldots,F^*(w_1)$ and, for each of these vertices, subtract one from their capacities, and replace $U$ with $U{\backslash}s_g$. Continuing in this way, we eventually obtain a $(G,k)$-labelling of $N$.
To illustrate Lemma \[flow\] and its proof, consider the example prior to the lemma, illustrated in Fig. \[fig:example\]. In $U$, a maximum flow could send either two units of flow through $w_1$ or one unit of flow through vertex $w_1$ and one unit of flow through vertex $w_4$. From the latter option, one can for example obtain the $(G,3)$-gene labelling shown in Fig. \[fig:fopt\].
Let $N$ be a phylogenetic network on ${{\mathcal X}}$, let ${{\mathcal G}}$ be a set of genes, let $G:{{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment, and let $k$ be a positive integer.
- If $N$ is a galled tree with exactly one gall, then there is a polynomial-time algorithm for deciding whether $N$ exhibits a $(G,k)$-labelling, in which case, such a labelling can also be found in polynomial time.
- If $T$ is a phylogenetic tree, then there is a polynomial-time algorithm for deciding whether $\ell(T,G,k)=1$.
\[one\]
First note that a maximum-valued integer flow can be found in $O(n^{1.5}\log(n\cdot k))$ time [@GoldbergRao1998]. Thus, (i) follows from Lemma \[flow\]. For (ii), if $|{{\mathcal X}}|=n$, then there is $O(n^2)$ possible ways of adding a single arc to $T$. Applying Lemma \[flow\] to each such way gives the desired algorithm. This completes the proof of the proposition.
We now extend Proposition \[one\](i) to all galled trees using induction on the number of galls. Let $N$ be a galled tree on ${{\mathcal X}}$, let ${{\mathcal G}}$ be a set of genes, let $G:{{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment and let $k$ be a positive integer. If $N$ has either no galls or exactly one gall, then we have such an algorithm by Propositions \[zero\] and \[one\], so we may assume that $N$ has at least two galls. In this case there exists a vertex $u_1$ of $N$ with the property that, for some gall, each of the vertices in the vertex set of this gall are descendants of $u_1$ and no vertex that is a proper descendant $u_1$ has this property. Let $N_1$ be the phylogenetic network obtained from $N$ by replacing $u_1$ and all of its descendants with a single vertex $q_1$. Let $Q_1$ be the phylogenetic network obtained from $N$ by deleting all of the vertices of $N$ that are not descendants of $u_1$ and adjoining a parent vertex $r_1$ to $u_1$ with one further child other than $u_1$. Call the additional child vertex $v_1$. Let $L_{Q_1}$ denote the leaf set of $Q_1$. Effectively, we have partitioned $N$ into two phylogenetic networks $N_1$ and $Q_1$. See Figure \[fig:galledtree\] for an example. Let $$G(q_1)= G(v_1)=\big(\bigcup_{x\in L_{Q_1}-\{v_1\}} G(x)\big)\cap
\big(\bigcup_{x\in {{\mathcal X}}-L_{Q_1}}G(x)\big).$$ The proof of the following lemma is straightforward, and so the details are omitted.
![A galled tree $N$ and the decomposition of $N$ into $Q_1$ and $N_1$ described in the text.[]{data-label="fig:galledtree"}](galled_tree.pdf){width="11cm"}
The galled tree $N$ has a $(G,k)$-labelling if and only if each of $N_1$ and $Q_1$ has a $(G,k)$-labelling. \[repeat\]
By Proposition \[one\](i), there is a polynomial-time algorithm for deciding whether or not $Q_1$ has a $(G,k)$-labelling. If there is no such labelling, then, by Lemma \[repeat\], $N$ has no $(G,k)$-labelling. On the other hand, if $Q_1$ has a $(G,k)$-labelling, then one needs to check if $N_1$ has a $(G,k)$-labelling. Now repeat the above construction with $N$ replaced by $N_1$. Continuing in this way, we either find a galled tree with a single gall that does not exhibit a $(G,k)$-labelling, and thereby show that $N$ has no such labelling, or we find no such galled tree and conclude that $N$ has a $(G,k)$-labelling. Note that the number of galls in $N$ is polynomial in the size of the vertex set of $N$. In particular, we have established the following results.
[**Theorem \[prop:level1\]**]{} [*Let $N$ be a galled tree on ${{\mathcal X}}$, let ${{\mathcal G}}$ be a set of genes, let $G:{{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment, and let $k$ be a positive integer. Then there is a polynomial-time algorithm for deciding whether $N$ exhibits a $(G,k)$-gene labelling.*]{}
\[gallcor\] Let $T$ be a rooted phylogenetic tree on ${{\mathcal X}}$, let ${{\mathcal G}}$ be a set of genes, let $G:{{\mathcal X}}\rightarrow 2^{{\mathcal G}}$ be a genome assignment, and let $k$ be a positive integer. If $h$ is a fixed non-negative integer, then there is a polynomial-time algorithm for deciding whether or not there is a galled tree $N$ on ${{\mathcal X}}$ that can be obtained from $T$ by adding at most $h$ arcs and which exhibits a $(G,k)$-gene labelling.
Suppose that $N$ is a galled tree on ${{\mathcal X}}$ that can be obtained from $T$ by adding at most $h$ arcs. Then there is an embedding $T'$ of $T$ in $N$. Notice that since $N$ is a galled tree, it follows that all vertices of $N$ are contained in $T$ and thus that $N$ can be obtained from $T$ by subdividing at most $2h$ arcs and adding at most $h$ arcs.
Hence, given $T$, we can try each possible way of subdividing at most $2h$ arcs and adding at most $h$ arcs. For each such possibility, we check if the resulting network is a galled tree. In each such case we can check if a $(G,k)$-gene labelling of this network exists, by Theorem \[prop:level1\]. The time needed is polynomial in the size of the input, for each fixed $h$.
CONCLUDING COMMENTS
===================
The analysis of this paper rests on a number of assumptions concerning gene evolution. Perhaps the most restrictive is the requirement that gene genesis is a unique event. This requirement reflects the fact that a gene is typically a long and fairly precise sequence of nucleotides, and the probability that a similar sequence could evolve independently in a different part of the tree is small. This seems reasonable if DNA sequence evolution is described by a neutral model [@kim], but, in some cases, natural selection will, no doubt, direct the evolution of DNA sequences towards certain genes that confer higher fitness. Thus, simple arguments based on neutrality need to be treated with caution. It would be interesting to extend the analysis of this paper to allow for a small frequency of independent gene genesis events.
A related question is what degree of sequence similarity is required in order to classify two sequences as coding for the same gene. Insisting on exact sequence identity is too severe, since it is well known that different species typically encode a gene with slightly different sequences that result from random site substitutions (indeed these differences have been the main signal used for phylogenetic tree reconstruction [@fels]). This question of gene identity is also relevant to the probability of independent gene genesis: a region of DNA that codes for a gene could, in principle, accumulate sufficient site mutations to put it just outside the range of being identified with that gene, but could then mutate back within range, giving the appearance of a second gene genesis event.
Other aspects of the model that may be criticized are the assumptions that the species tree is known with certainty (or, indeed, that it is meaningful to talk of a ‘species tree’ [@doo07]), and that the model does not penalize gene losses at all.
Our computational complexity results highlight that many problems are surprisingly difficult, even for a tree, and some questions still remain to be explored further. One that seems particularly interesting is described as follows, along with our conjecture as to its possible resolution.
Given a rooted phylogenetic tree $T$, a set of genes $G(x)$ for each leaf $x$ of $T$, and natural numbers $k$ and $h$, consider the problem of deciding whether it is possible to add at most $h$ arcs to $T$ to obtain a phylogenetic network $N$ that admits a $(G,k)$-gene labelling.
\[con:networkfromtree\] This problem is NP-hard in general, but for each fixed $h$, it admits a polynomial-time algorithm.
[50]{}
T. Dagan and W. Martin, (2007), Ancestral genome sizes specify the minimum rate of lateral gene transfer during prokaryote evolution, Proc. Natl. Acad. Sci. USA. 104, 870–875.
T. Dagan, Y. Artzy-Randrup, and W. Martin, (2008), Modular networks and cumulative impact of lateral transfer in prokaryote genome evolution, Proc. Natl. Acad. Sci. USA. 105, 10039–10044.
R. Diestel (2006) Graph Theory (3rd ed.), Springer.
W.F. Doolittle and E. Bapteste, (2007), Pattern pluralism and the Tree of Life hypothesis, Proc. Natl. Acad. Sci. USA. 104, 2043–2049.
W.F. Doolittle, Y. Boucher, C. . Nesbø, C.J. Douady, J. O. Andersson, and A. J. Roger, (2003), How big is the iceberg of which organellar genes in nuclear genomes are but the tip? Philos. Trans. Roy. Soc. B. 358, 39–58.
A.V. Goldberg and S. Rao, (1998), Beyond the flow decomposition barrier, J. ACM. 45(5), pp. 783–797.
S. Even, A. Itai, and A. Shamir (1972). On the complexity of timetable and multi-commodity flow problems, SIAM J. Comput. 1(2), 188–202.
J. Felsenstein (2004) Inferring phylogenies. Sinauer Press, Sunderland, MA.
S. Fortune, J. Hopcroft, and J. Wyllie, (1980), The directed subgraph homeomorphism problem, Theor. Comput. Sci. 10, 111–121.
G. Jin, L. Nakhleh, S. Snir, and T. Tamir, (2007), Inferring phylogenetic networks by the maximum parsimony criterion: A case study, Mol. Biol. Evol. 24, 324–337.
M. Kimura (1983) The neutral theory of evolution. Cambridge University Press.
B.G. Mirkin, T. I. Fenner, M.Y. Galperin, and E.V. Koonin, (2003), Algorithms for computing parsimonious evolutionary scenarios for genome evolution, the last universal common ancestor and dominance of lateral gene transfer in the evolution of prokaryotes, BMC Evol. Biol. 3, 2.
M. Spencer, E. Susko, and A.J. Roger, (2006), Modelling prokaryote gene content, Evol. Bioinf. Online. 2, 157–178.
[^1]: We thank the Allan Wilson Centre for Molecular Ecology and Evolution, and the New Zealand Marsden Fund for helping fund this work.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We consider the phenomenon of forced symmetry breaking in a symmetric Hamiltonian system on a symplectic manifold. In particular we study the persistence of an initial relative equilibrium subjected to this forced symmetry breaking. We see that, under certain nondegeneracy conditions, an estimate can be made on the number of bifurcating relative equilibria . [*To cite this article: F. Grabsi et al., C. R. Acad. Sci. Paris, Ser. I 338 (2004).*]{}
0.5
[**Résumé**]{} 0.5[**Bifurcation et brisure forcée de symétrie dans les systèmes hamiltoniens.** ]{} Nous considérons le phénomène de brisure forcée de symétrie dans un système hamiltonien symétrique défini sur une variété symplectique. Plus précisement, nous étudions la persistance d’un équilibre relatif soumis à une brisure de symétrie. Nous verrons que, sous certaines hypothèses de non-dégénéréscence, on peut donner une estimation du nombre d’équilibres relatifs persistants après la brisure. [*Pour citer cet article : F. Grabsi et al., C. R. Acad. Sci. Paris, Ser. I 338 (2004).*]{}
address:
- 'Institut Non Linéaire de Nice, UMR 129 CNRS-UNSA, 1361 route des , 06560 Valbonne, France'
- 'Department of Mathematics. UMIST. PO Box 88, Manchester M60 1QD, United Kingdom'
- 'Centre National de la Recherche Scientifique, Département de Mathématiques de Besançon, Université de Franche-Comté. 16, route de Gray, 25030 Besançon cedex, France'
author:
- Féthi Grabsi
- James Montaldi
- 'Juan-Pablo Ortega'
title: Bifurcation and forced symmetry breaking in Hamiltonian systems
---
Differential Geometry/Dynamical Systems
Version française abrégée {#version-française-abrégée .unnumbered}
=========================
Nous nous intéressons au phénomène de la brisure forcée de symétrie dans les systèmes hamiltoniens symétriques. On considère une action libre du tore ${\mathbb T}^n$ de dimension $n$ sur une variété symplectique $({\mathcal M},
\omega)$ qui admet une application moment (nécessairement invariante) ${\mathbf
J}_{{\mathbb T}^n} : {\mathcal M} \longrightarrow (\mathfrak{t} ^n)^\ast \simeq
\mathbb{R} ^n$. Soit $H _0 $ une fonction hamiltonienne invariante par rapport à l’action de ${\mathbb T}^n$ et dont le champ de vecteurs $X_{H _0}$ associé présente, par hypothèse, un équilibre relatif (<span style="font-variant:small-caps;">er</span>). Nous rappelons qu’un <span style="font-variant:small-caps;">er</span> $m$ de $X_{H_0}$ est défini par la condition $
X_{H_0}(m)=\xi_{{\mathcal M}}(m)$, pour un élément $\xi$ dans l’algèbre de Lie de $\mathbb{T} ^n $ dénommé la vitesse de l’<span style="font-variant:small-caps;">er</span>.
Soit $ {{\mathbb T}}^r\subset {{\mathbb T}}^n$ un sous-tore de ${{\mathbb T}}^n$ et $H_{\varepsilon}$ une famille de perturbations ${{\mathbb T}}^r$-invariantes de l’hamiltonien $H_0$, paramétrée d’une manière lisse par $\varepsilon\in
\mathbb{R}$. En principe, les <span style="font-variant:small-caps;">er</span> du système $({\mathcal M}, \omega, H_0,
{{\mathbb T}}^n, {\bf J}_{{{\mathbb T}}^n} )$ ne seront plus en général des <span style="font-variant:small-caps;">er</span> pour le système $({\mathcal M}, \omega, H_{\varepsilon}, {{\mathbb
T}}^r, {\bf J}_{{{\mathbb T}}^r})$. Notre intérêt est de déterminer sous quelles conditions l’<span style="font-variant:small-caps;">er</span> de départ $m$ continue à être un <span style="font-variant:small-caps;">er</span> du système $({\mathcal M}, \omega, H_{\varepsilon}, {{\mathbb T}}^r, {\bf
J}_{{{\mathbb T}}^r})$. Plus précisement, nous montrons que les <span style="font-variant:small-caps;">er</span> qui sont non-dégénérés dans un certain sens et dont la vitesse appartient à l’algèbre de Lie du sous-tore $ \mathbb{T}^r$ persistent. De plus, le théorème suivant nous donne une estimation de leur nombre :
Soit $H_\varepsilon$ une famille de fonctions hamiltoniennes paramétrisée d’une manière lisse par $\varepsilon \in \mathbb{R}$ et définie sur la variété symplectique $({\mathcal M},\omega)$. Supposons que $H_0$ est invariant par rapport à une action libre et canonique du tore $\mathbb{T}^n$ et que le champ de vecteurs associé $X_{H_0} $ a un équilibre relatif $m$ avec vitesse $\xi \in
\mathfrak{t}^n $ et moment $\mu:=\mathbf{J}_{\mathbb{T}^n}(m)\in (\mathfrak{t}^n)
^\ast \simeq \mathbb{R}^n$. $\mathbf{J}_{\mathbb{T}^n}:\mathcal{M}\rightarrow
(\mathfrak{t} ^n) ^\ast $ est une application moment associée à l’action de $\mathbb{T}^n$.
Soit $\mathbb{T}^r\subset \mathbb{T}^n$ un sous-tore dont l’action restreinte associée sur $\mathcal{M}$ laisse invariantes les foncions $H_{\varepsilon}$, pour tout $\varepsilon$. Notons par $i:\mathfrak{t}^r\hookrightarrow \mathfrak{t}^n$ l’inclusion de l’algèbre de Lie de $\mathbb{T}^r$ dans celle de $\mathbb{T}^n$ et par $i^*:(\mathfrak{t}^n) ^\ast \rightarrow (\mathfrak{t}^r) ^\ast $ son dual. Si $m$ est $i^*\mu$-non-dégénéré dans le sens de la Définition \[nondegeneracy definition\] et sa vitesse appartient à l’algèbre de Lie de $\mathbb{T}^r$ alors pour chaque valeur du paramètre $\varepsilon$ suffisamment proche de $0$, les équilibres relatifs de $X_{H_{\varepsilon}}$ sont en correspondance bijective avec les points critiques d’une fonction lisse $\overline{h} _\varepsilon : \mathbb{T}^{n-r}\rightarrow \mathbb{R}$. Par conséquent, il existe au moins $n-r+1$ équilibres relatifs de $X_{H_{\varepsilon}}$ avec moment $i^*\mu$ et vitesse proche de $\xi$. De plus, si les points critiques de $\overline{h} _\varepsilon$ sont tous non-dégénérés le nombre d’équilibres relatifs bifurqués est au moins $2^{n-r}$.
Pour établir ce théorème, nous utilisons la technique de réduction symplectique par rapport au tore ${\mathbb T} ^r $ (espace de Marsden-Weinstein ${\mathcal M}_\alpha$) dans les coordonnées de Marle-Guillemin-Sternberg correspondant à l’action de $ \mathbb{T}^n $. Nous y caractérisons les équilibres relatifs bifurquant de $m$ après brisure forcée de symétrie comme points critiques de l’hamiltonien réduit $h _\varepsilon$ sur ${\mathcal
M}_\alpha$. Moyennant une condition de non-dégénérescence sur l’équilibre relatif initial $m$, nous obtenons de $h _\varepsilon$ une fonction $\overline{h}
_\varepsilon$ sur le tore ${\mathbb T}^{n-r}$ dont les points critiques sont en correspondance bijective avec les <span style="font-variant:small-caps;">er</span> recherchés. Il s’en suit par la théorie des points critiques d’une fonction à variable réelle sur une variété compacte les estimations sur leur nombre (catégorie de Ljusternik-Schnirelmann, théorie de Morse).
Introduction
============
[r]{}[100pt]{} {height="66pt" width="80pt"}
Forced symmetry breaking in dynamical systems is a phenomenon that takes place when we add to a symmetric system a perturbation with less symmetry. In this note we study this phenomenon in the context of globally Hamiltonian dynamical systems, that is, symmetric Hamiltonian systems to which a momentum map can be associated. The particular problem of study is the “survival” or persistence of relative equilibria of the fully symmetric system after the symmetry breaking perturbation is added to it. Our motivation relies strongly upon the fact that this phenomenon is naturally present in many systems. For instance, consider a spherical pendulum whose bob of mass $m$ has been charged with a positive charge $q$ (see the figure). Suppose now that right below the point of suspension of the pendulum we place a charge identical to that of the pendulum (position (a) in the picture). If the repulsive electrostatic force is strong enough, the stable downright equilibrium of the spherical pendulum becomes unstable and a ring of equilibria appears. Suppose now that the circular symmetry of the system is broken by slightly sliding the charge to a side (position (b) in the picture). It can be seen that only two of the equilibria in the ring survive. Our main goal in this paper is the formulation of a general theorem capable of predicting such behaviour. The kind of systems we are interested in can be mathematically described by considering a finite dimensional symplectic manifold $({\mathcal M}, \omega )$ acted freely and canonically upon by the $n$-torus ${{\mathbb T}}^n$. We assume that this action has a momentum map $
{\bf J}_{{{\mathbb T}}^n} : {\mathcal M} \longrightarrow ( \mathfrak{t}^n)
^\ast\simeq \mathbb{R} ^n $. Let $\mathbb{T} ^r\subset \mathbb{T} ^n $ be a subtorus, $H_\varepsilon$ a family of Hamiltonian functions on ${\mathcal M}$ parametrized by $\varepsilon \in \mathbb{R} $, and assume that $H _0 $ is $\mathbb{T}^n $-invariant whereas $H _\varepsilon $ is only $\mathbb{T}^r
$-invariant, for all $\varepsilon \in \mathbb{R}$. The problem that we discuss in this note is under what conditions a given relative equilibrium $m \in {\mathcal
M}$ of $H _0 $ with respect to its $\mathbb{T}^n $-symmetry [*persists*]{} to relative equilibria of the Hamiltonian vector fields associated to $H _\varepsilon
$, for $\varepsilon$ sufficiently small, with respect to their $\mathbb{T}^r
$-symmetry. We recall that a relative equilibrium (<span style="font-variant:small-caps;">re</span>) of a $\mathbb{T}^n
$-equivariant dynamical system $X $ on ${\mathcal M} $ is a point $m$ for which there exists an element $\xi $ in the Lie algebra $\mathfrak{t}^n $ of $\mathbb{T}
^n$ (called the [*velocity*]{} of the <span style="font-variant:small-caps;">re</span>) such that $X (m)= \xi_{{\mathcal
M}}(m) $. The symbol $\xi_{{\mathcal M}}(m):= \left.\frac{d}{dt}\right|_{t=0} \exp t
\xi \cdot m $ denotes the infinitesimal generator of the $\mathbb{T}^n $-action associated to the element $\xi$.
Preliminaries
=============
Throughout we assume that $(\mathcal{M},\omega)$ is a finite-dimensional symplectic manifold with a free Hamiltonian action of the torus ${\mathbb T}^n$ with momentum map ${\mathbf J}_{{\mathbb T}^n}:\mathcal{M}\to( \mathfrak{t}^n) ^\ast$. Such a momentum map is necessarily *invariant*: ${\mathbf J}(g\cdot m) = {\mathbf
J}(m)$ (with $g\in {\mathbb T}^n$ and $m\in\mathcal{M}$). We fix a torus subgroup ${\mathbb T}^r\subset {\mathbb T}^n$, and let $i:\mathfrak{t}^r\hookrightarrow\mathfrak{t} ^n$ be the inclusion of Lie algebras. The momentum map for the restricted action by ${\mathbb T}^r$ is ${\mathbf
J}_{{\mathbb T}^r} = i^*\circ{\mathbf J}_{{\mathbb T}^n}$, where $i^*:(\mathfrak{t}^n)^*\rightarrow(\mathfrak{t}^r)^*$ is the dual map to $i$.
We also assume that $m\in \mathcal{M}$ is a relative equilibrium for the Hamiltonian system $({\mathcal M}, \omega, H_0, {{\mathbb T}}^n, {\mathbf J}_{{{\mathbb
T}}^n})$, with momentum $\mu:={\mathbf J}_{{{\mathbb T}}^n}(m)\in (\mathfrak{t} ^n)
^\ast $ and velocity $\xi\in \mathfrak{t}^n $. We recall that this amounts to the point $m$ being a critical point of the *augmented Hamiltonian* $H _0-
\mathbf{J}_{\mathbb{T} ^n} ^\xi $, that is, $D(H_0-{\mathbf J}_{{{\mathbb
T}}^n}^{\xi})(m)=0$.
In order to formulate the main hypothesis of the theorem we need to recall the [**Witt-Artin decomposition**]{} of the tangent space $T_m{\mathcal M}$; define $V_m$ as a complement in $\ker T_m{\mathbf J}_{{{\mathbb T}}^n}$ to the tangent space ${\mathfrak t}^n\cdot m$ at $m$ of the $\mathbb{T} ^n $-group orbit, that is, $ \ker
T_m{\mathbf J}_{{{\mathbb T}}^n}= V_m\oplus {\mathfrak t}^n\cdot m $. The space $V_m$ is called the [**symplectic normal space**]{} at $m$. Notice that ${\mathfrak t}^n\cdot m
\subset (V _m)^{\omega(m)} $. Let $W$ be a Lagrangian complement to ${\mathfrak
t}^n\cdot m$ in $(V _m)^{\omega(m)}$. The decomposition $$\label{witt artin decomposition} T_m{\mathcal M}=V_m\oplus {\mathfrak
t}^n\cdot m\oplus W$$ is called a Witt-Artin decomposion of the tangent space $T_m{\mathcal M}$. We will refer to $W$ as the [**orbital complement**]{} at $m$ of the Witt-Artin decomposition (\[witt artin decomposition\]). To finish these preliminaries, we give a definition which we will use in our result.
\[nondegeneracy definition\] With the notation as above, a [**nondegeneracy space**]{} ${\mathcal N}_\alpha $ at $m$ associated to the momentum $\alpha\in
(\mathfrak{t} ^r) ^\ast $ is defined as $${\mathcal N}_{\alpha}=A_{\alpha}\oplus
V_m$$ where $A_{\alpha}:=\left\{ w \in W\ \vert \ i^*(\mu +T_m{\mathbf J}_{{{\mathbb
T}}^n}(w))=\alpha \right\}$. Let $H \in C^{\infty}({\mathcal M})^{\mathbb{T}^n} $ be a smooth $\mathbb{T} ^n $-invariant function on ${\mathcal M} $ that exhibits a critical point at $m$, that is, $DH (m)=0 $. We say that $m$ is an $\alpha$-[**nondegenerate**]{} critical point of $H$ when the symmetric bilinear form $$D^2H (m)|_{{\mathcal N}_\alpha\times {\mathcal N}_\alpha}$$ is nondegenerate.
The $\alpha $-nondegeneracy of a critical point given in the previous definition depends only on the value $\alpha\in (\mathfrak{t} ^r) ^\ast $ and not on the specific Witt-Artin decomposition used to verify this condition.
**Proof. **It suffices to show that our nondegeneracy condition is independent of the choice of $V _m$ and $W$ in the Witt-Artin decomposition. Assume that $H$ is $\alpha$-nondegenerate at $m$ for a fixed choice of $V _m$ and $W $. Let $V _m' $ be another choice of symplectic normal space at $m$, $W' $ a complement to ${\mathfrak t}^n\cdot m $ in $(V_m')^{\omega(m)}$, and ${\mathcal N}_\alpha' $ the associated nondegeneracy space. Let $v _1+ w _1, v _2 + w _2 \in {\mathcal N}_
\alpha $ be arbitrary with $v _1, v _2 \in A _\alpha $ and $w _1, w _2 \in V _m $. The Witt-Artin decomposition of $T _m {\mathcal M} $ implies the existence of unique elements $\xi, \eta \in \mathfrak{t} ^n $, $v _1', v _2' \in A _\alpha' $, $w _1',
w _2' \in V _m ' $ such that $$v _1+ w _1= \xi _{{\mathcal M}} (m)+v _1'+ w _1' \quad\textrm{and}\quad
v _2+ w _2= \eta _{{\mathcal M}} (m)+v _2'+ w _2'.$$ The $\mathbb{T}^n $-invariance of $H$ implies that $$D^2H (m)(v _1+ w _1,v _2+ w _2)=D^2H (m)(v _1'+ w _1',v _2'+ w
_2').$$ Given that the map $v + w \in {\mathcal N}_\alpha \longmapsto v' + w' \in {\mathcal
N}_\alpha' $ is an isomorphism, the result follows. $\Box$
Theorem on forced symmetry breaking
===================================
The goal of this section is to prove the following theorem:
\[main theorem on bifurcation symmetry breaking\] Let $({\mathcal M},
\omega,{\mathbb T}^n,{\mathbf J}_{{\mathbb T}^n}, {\mathbb T}^r, H_0,m,\xi,\mu )$ be as above. Let $H_\varepsilon$ be a family of ${\mathbb T}^r$-invariant Hamiltonian functions on ${\mathcal M}$ smoothly parametrized by $\varepsilon \in \mathbb{R} $ with $H _0 $ $\mathbb{T}^n $-invariant. Suppose that the relative equilibrium $m \in {\mathcal M} $ has velocity $\xi$. If
(i)
: $\xi \in i(\mathfrak{t}^r) $ and
(ii)
: $m$ is a $i ^\ast \mu
$-nondegenerate critical point of $H _0- \mathbf{J} ^\xi _{\mathbb{T}^n} $, where $\mu:= \mathbf{J}_{\mathbb{T} ^n} (m) $,
then for any value of the parameter $\varepsilon$ close enough to zero, the relative equilibria of the Hamiltonian vector field $X _{H _\varepsilon} $ are in bijective correspondence with the critical points of a smooth function $\overline{h}_\varepsilon: \mathbb{T}^{n-r} \rightarrow \mathbb{R} $. Consequently, under these hypotheses there exist at least $n-r+1$ relative equilibria of $X_{H_{\varepsilon}}$ with momentum $i^*\mu$ and velocity close to $\xi$. Additionally, if the critical points of $\overline{h} _\varepsilon\in
C^{\infty}(\mathbb{T}^{n-r}) $ are all nondegenerate the number of bifurcated relative equilibria is at least $2^{n-r}$.
**Proof. ** The local character of the result that we want to prove permits us to use the local model around the $\mathbb{T}^n $-orbit of $m$ given by the [*Marle-Guillemin-Sternberg*]{} (MGS) normal form [@normal1; @normal2]. We recall that this result provides a $\mathbb{T}^n $-equivariant symplectomorphism between a $\mathbb{T}^n $-invariant neighborhood of the orbit $\mathbb{T}^n \cdot m $ and the product $Y:=\mathbb{T}^n\times
(\mathfrak{t}^n) ^\ast
\times V _m $, considered as a $\mathbb{T}^n $-symplectic space with the $\mathbb{T}^n $-action given by $g \cdot (h, \eta, v):=(gh, \eta, v) $, $g,h
\in \mathbb{T} ^n $, $ \eta \in (\mathfrak{t}^n) ^\ast$, $v \in V _m $ and with a symplectic form with respect to which the momentum map associated to this $\mathbb{T}^n $-action has the form ${\mathbf J}_{\mathbb{T}^n} (g, \eta , v)=
\mu+ \eta $. In this model, the point $m\in {\mathcal M}$ is represented by $(e,0,0)\in Y$ and the space $V _m $ is one of the symplectic normal spaces at $m$ that we have previously defined. We will carry out the proof of our theorem in these coordinates by looking for the critical points of the reduced Hamiltonians $h _{\varepsilon, \alpha} $ on the $ \mathbb{T} ^r $-Marsden-Weinstein reduced space ${\mathcal M}_\alpha:= {\mathbf J} ^{-1}_{\mathbb{T} ^r}(\alpha)/
\mathbb{T}^r $ defined by $h _{\varepsilon, \alpha}\circ \pi_\alpha=H _\varepsilon \circ i
_\alpha $, where $\alpha:= i ^\ast \mu $, $i _\alpha: {\mathbf J}
^{-1}_{\mathbb{T} ^r}(\alpha) \hookrightarrow {\mathcal M} $ is the injection, and $\pi _\alpha: {\mathbf J} ^{-1}_{\mathbb{T} ^r}(\alpha) \rightarrow {\mathcal
M}_\alpha $ is the projection. A straighforward computation in MGS coordinates shows that $$\label{eq4} {\mathbf J}^{-1}_{{{\mathbb T}}^r}(\alpha)={{\mathbb T}}^n\times
A_{\alpha}\times V_m$$ where $A_{\alpha}$ is the vector subspace of $(\mathfrak{t}^n) ^\ast$ given by $A_{\alpha}:=\left\{ \eta \in (\mathfrak{t}^n) ^\ast\ \vert \ i^*(\mu +\eta)=\alpha
\right\}=\ker i ^\ast $ and that ${\mathcal N}_\alpha:=A _\alpha \times V _m $ is a $\alpha$-nondegeneracy space at $m$. From expression (\[eq4\]) it is clear that $${\mathcal M}_{\alpha}={{\mathbb T}}^n\times A_{\alpha}\times
V_m/{{\mathbb T}}^r\simeq{{\mathbb T}}^{n-r}\times {\mathcal N}_{\alpha}.$$ The problem of finding the relative equilibria in the statement of the theorem is now equivalent to the search of the critical points of the real-valued functions $h _{\varepsilon, \alpha} $ defined on the Marsden-Weinstein reduced space ${\mathcal M}_{\alpha}={{\mathbb T}}^{n-r}\times {\mathcal N}_{\alpha}$.
The hypothesis on the $\alpha$-nondegeneracy of $m$ as a critical point of $H _0-
\mathbf{J} ^\xi_{\mathbb{T} ^n} $ implies that the quadratic form $D^2{h_{0,
\alpha}}(e,0,0)|_{{\mathcal N} _\alpha \times {\mathcal N}_\alpha}$ is nondegenerate. In order to lighten the notation we will omit the symbol $\alpha $ in the function $h _{\varepsilon, \alpha} $ in all that follows. With this notation, we need to find the triples $(k, \tilde{v}, \varepsilon)\in {{\mathbb
T}}^{n-r}\times{\mathcal N}_{\alpha}\times{\mathbb R}$ such that $$\label{eq7}
D h_{\varepsilon}(k,\tilde{v})=0$$ We proceed by using the Implicit Function Theorem to eliminate the parameter $\tilde{v}\in {\mathcal N}_{\alpha}$ from the equation (\[eq7\]) by writing it in terms of the $ {{\mathbb T}}^{n-r}$ and ${\mathbb R}$ variables. Indeed, consider the following map $$\begin{array}{ccc} {\mathcal F} : {{\mathbb T}}^{n-r}\times{\mathcal N}_{\alpha}\times{\mathbb R}
&\longrightarrow & \left({\mathcal N}_{\alpha}\right)^*\simeq {\mathcal N}_{\alpha}\\
(k, \tilde{v},\varepsilon) &\longmapsto &D_{{\mathcal
N}_{\alpha}}h_{\varepsilon}(k,\tilde{v})\\
\end{array}$$
Since $m\equiv (e,0,0)$ is a $\mathbb{T} ^n $-relative equilibrium for $H_0$ we have ${\mathcal F} (g,0,0)=0$, for all $g\in {{\mathbb T}}^{n-r}$. Moreover, since the partial derivative $D_{{\mathcal N}_{\alpha}}{\mathcal F}(g,0,0) : {\mathcal
N}_{\alpha}\longrightarrow \left({\mathcal N}_{\alpha}\right)^*\simeq {\mathcal N}_{\alpha}$ of ${\mathcal F}$ with respect to the ${\mathcal N}_{\alpha}$-factor, evaluated at $(g,0,0)$ is given by $D_{{\mathcal N}_{\alpha}}{\mathcal F}(g,0,0)=D^2_{{\mathcal
N}_{\alpha}}h_0(g,0)$ then the hypothesis on the $\alpha$-nondegeneracy of $m$ as a critical point of $H _0- \mathbf{J} ^\xi_{\mathbb{T} ^n} $ implies that $D_{{\mathcal N}_{\alpha}}{\mathcal F}(g,0,0) : {\mathcal N}_{\alpha}\longrightarrow {\mathcal
N}_{\alpha}$ is injective. Consequently, $D_{{\mathcal N}_{\alpha}}{\mathcal
F}(g,0,0)$ is an isomorphism and we can then define via the Implicit Function Theorem a smooth map $\tilde{v}_g: {\mathcal U}_{g}\times {\mathcal W}_g \longrightarrow
\left({\mathcal N}_{\alpha}\right)_g\subset {\mathcal N}_{\alpha}$ defined in an open neighborhood of $(g,0)\in {{\mathbb T}}^{n-r}\times{\mathbb R}$ such that, for any $(k,\varepsilon)\in
{{\mathbb T}}^{n-r}\times{\mathbb R}$ in that neighborhood, we have that: $$\label{eqv}
D_{{\mathcal N}_{\alpha}}h_{\varepsilon}(k,\tilde{v}_g(k,\varepsilon))=D_{{\mathcal
N}_{\alpha}}h_{\varepsilon}(k,{\tilde{v}}_{g,\varepsilon}(k))=0$$ Given that this argument can be repeated for any $g \in \mathbb{T}^{n-r}$ we can invoke the compactness of $\mathbb{T}^{n-r} $ to build a finite family of functions ${\tilde{v}}_{g_i}:{\mathcal U}_{g_i}\times {\mathcal W}_{g_i} \longrightarrow
\left({\mathcal N}_{\alpha}\right)_{g_i}$, $i \in \{1, \ldots, \ell\} $, satisfying (\[eqv\]) and such that $\bigcup _{i=1}^\ell{\mathcal U}_{g_i}=
\mathbb{T}^{n-r}$. Let us define
$$\begin{array}{ccc} {\tilde{v}} : {{\mathbb T}}^{n-r}\times
\bigcap_{i=1}^\ell{\mathcal W}_{g_i} &\longrightarrow & \bigcup_{i=1}^\ell
\left({\mathcal N}_{\alpha}\right)_{g_i}\\ (g,\varepsilon) &\longmapsto
&{\tilde{v}}_{g _i}(g,\varepsilon) \
\ \mbox{if } g\in {\mathcal U}_{g_i}.\\
\end{array}$$ This map is well defined by the uniqueness of the maps ${\tilde{v}}_{g _i}$ obtained from the Implicit Function Theorem. Taking into account this new map, our bifurcation equation (\[eq7\]) is now equivalent to: $$\label{eq9}
Dh_{\varepsilon}(k,{\tilde{v}}(k, \varepsilon))=0.$$ The solutions of this equation coincide with the critical points of the function ${\overline h}_{\varepsilon}(k):=h_{\varepsilon}(k,{\tilde{v}}(k, \varepsilon))$ defined, for each value of the parameter $\varepsilon$, on the compact manifold ${{\mathbb T}}^{n-r}$. Indeed, using (\[eqv\]), we have $$\begin{aligned}
D{\overline
h}_{\varepsilon}(t)&=&D_{{{\mathbb T}}^{n-r}}
h_{\varepsilon}(t, \tilde{v}(t,\varepsilon)) +D_{{\mathcal N}_{\alpha}}h_{\varepsilon}(t,
\tilde{v}(t,\varepsilon))\cdot D_{{{\mathbb T}}^{n-r}}\tilde{v}(t,\varepsilon))=D_{{{\mathbb T}}^{n-r}}
h_{\varepsilon}(t, \tilde{v}(t,\varepsilon)) \nonumber \\
&=&D_{{{\mathbb T}}^{n-r}}
h_{\varepsilon}(t, \tilde{v}(t,\varepsilon))+D_{{\mathcal N}_{\alpha}}h_{\varepsilon}(t,
\tilde{v}(t,\varepsilon))=Dh_{\varepsilon}(t,{\tilde{v}}(t, \varepsilon)).
\nonumber\end{aligned}$$ Consequently, the pair $(t,{\tilde{v}}(t, \varepsilon))$ is a solution of (\[eq9\]) if and only if $t\in {{\mathbb T}}^{n-r}$ is a critical point of ${\overline h}_{\varepsilon} $. A lower bound for the number of these critical points is provided by the Ljusternik-Schnirelmann category ${\rm Cat}({{\mathbb
T}}^{n-r})=n-r+1$ of the torus ${{\mathbb T}}^{n-r} $ (see for instance [@Ljusternik; @1966]), which proves the statement of the theorem. Additionally, if we know in advance that the critical points of ${\overline h}_{\varepsilon} \in
C^{\infty}(\mathbb{T}^{n-r})$ are all nondegenerate, the Morse inequalities guarantee that this function has at least $b^0(\mathbb{T}^{n-r})+b^1(\mathbb{T}^{n-r})+ \cdots+b^{n-r}(\mathbb{T}^{n-r})$ critical points, where $b^i (\mathbb{T}^{n-r}) $, $i \in \{0, \ldots, n-r \} $, is the $i$-th Betti number of the torus $\mathbb{T}^{n-r} $. Since $b^i
(\mathbb{T}^{n-r}) = \pmatrix{n-r\cr i} $, $i \in \{0, \ldots, n-r \} $, we have $$b^0(\mathbb{T}^{n-r})+b^1(\mathbb{T}^{n-r})+ \cdots+b^{n-r}(\mathbb{T}^{n-r})=\sum
_{i=0}^{n-r} \pmatrix{n-r\cr i} =2^{n-r}$$ and hence the second estimate in the statement follows. $\Box$
Symmetry breaking using Poisson reduction
=========================================
In the previous theorem we confined our search for bifurcated relative equilibria to the momentum level set $\mathbf{J}_{\mathbb{T}^r}^{-1}(i ^\ast\mu)$. This fact appears in the proof of that result when we use the symplectic reduced space ${\mathcal M}_{i ^\ast \mu} $. If instead of using ${\mathcal M}_{i ^\ast \mu} $ we consider the Poisson reduced space $\widetilde{{\mathcal M}}:={\mathcal M}/
\mathbb{T}^r $ we can obtain another bifurcation result where the predicted relative equilibria could, in principle, have a momentum different from that of the given <span style="font-variant:small-caps;">re</span>. This is obtained at the expense of imposing a more demanding nondegeneracy condition.
\[poisson\] Let $({\mathcal M}, \omega,{\mathbb T}^n,{\mathbf J}_{{\mathbb T}^n}, {\mathbb
T}^r, H_0,m,\xi,\mu )$ be as in Section 2. Let $H_\varepsilon$ be a family of Hamiltonian functions on ${\mathcal M}$ parametrized by $\varepsilon \in \mathbb{R}
$ and assume that $H _0 $ is $\mathbb{T}^n $-invariant whereas $H _\varepsilon $ is only $\mathbb{T}^r $-invariant, for all $\varepsilon \in \mathbb{R}$. Suppose that the point $m \in {\mathcal M} $ is a $ \mathbb{T}^n $-relative equilibrium of the Hamiltonian vector field $X_{H _0}$ with velocity $\xi\in i(\mathfrak{t}^r)$. Suppose moreover that $$D^2\left.(H _0- \mathbf{J} ^\xi _{\mathbb{T}^n})(m) \right\vert_{{\mathcal N}\times
{\mathcal N}}$$ is a nondegenerate quadratic form, where ${\mathcal N}:=W\times V_m $, for the symplectic normal space $V _m $ and orbital complement $W $ corresponding to some Witt-Artin decomposition of $T _m {\mathcal M} $. Then for any value of the parameter $\varepsilon$ close enough to zero, the relative equilibria of the Hamiltonian vector field $X _{H _\varepsilon} $ are in bijective correspondence with the critical points of a smooth function $[h_{\varepsilon}] : {{\mathbb
T}}^{n-r}\longrightarrow {\mathbb R}$. Consequently, under these hypotheses there exist at least $n-r+1$ $\mathbb{T}^r $-relative equilibria near $m$ with momentum close to $i^*\mu$ and velocity close to $\xi$. Additionally, if the critical points of $[h_{\varepsilon}]\in C^{\infty}(\mathbb{T}^{n-r}) $ are all nondegenerate the number of these bifurcated relative equilibria is at least $2^{n-r}$.
**Proof. ** This mimics the proof of Theorem \[main theorem on bifurcation symmetry breaking\] where the reduced space ${\mathcal M} _\alpha $ has been replaced by ${\widetilde{\mathcal M}} $. Note that in the MGS normal form coordinates we can write, locally, $${\widetilde{\mathcal M}}=\left({{\mathbb T}}^n\times (\mathfrak{t} ^n) ^\ast \times
V_m\right)/{{\mathbb T}}^r\simeq {{\mathbb T}}^{n-r}\times (\mathfrak{t} ^n)
^\ast\times V_m.$$ -6mm
$\Box$
Acknowledgements {#acknowledgements .unnumbered}
================
This research was partially supported by the European Commission through funding for the Research Training Network *Mechanics and Symmetry in Europe* (MASIE).
[00]{}
Guillemin, V., and Sternberg, S. \[1984\] A normal form for the moment map. In *Differential Geometric Methods in Mathematical Physics*. S. Sternberg (editor). Mathematical Physics Studies, [**6**]{}. D. Reidel Publishing Company.
Ljusternik, L. A. \[1966\] [*The Topology of the Calculus of Variations in the Large.*]{} Translations of Mathematical Monographs, Vol. [**16**]{}. American Mathematical Society.
Marle, C.-M. \[1984\] Le voisinage d’une orbite d’une action hamiltonienne d’un groupe de Lie. *Séminaire Sud–Rhodanien de Géométrie II* (P. Dazord, N. Desolneux–Moulis eds.) pages 19–35.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Among those nearly incompressible vector fields $\v:\R^n\to\R^n$ with $|x|\log|x|$ growth at infinity, we give a pointwise characterization of the ones for which $\curl\v= D\v-D^t\v$ belongs to $L^\infty$. When $n=2$ we can go further and describe, still in pointwise terms, the vector fields $\v:\R^2\to\R^2$ for which $|\div\v|+|\curl\v|\in L^\infty$.'
author:
- 'Albert Clop, Banhirup Sengupta'
title: |
Pointwise descriptions\
of nearly incompressible vector fields\
with bounded curl
---
Introduction
============
Following [@Rei], we will say that a continuous vector field $\v:\R^n\to\R^n$ is of *Reimann’s type*, and write $\v\in Q$, if there is a constant $C_0\geq 0$ such that for each $x,h, k\in\R^n$ with $|h|=|k|\neq 0$ one has $$\left|\frac{\langle \v(x+h)-\v(x),h\rangle}{|h|^2}-\frac{\langle \v(x+k)-\v(x),k\rangle}{|k|^2}\right|\leq C_0.$$ The best possible value of $C_0$ is denoted $\|\v\|_Q$. This class of vector fields was introduced by H.M. Reimann in [@Rei]. Even though every Lipschitz vector field belongs to the $Q$ class, there exist many vector fields of Reimann type which are not Lipschitz. Indeed, every element of $Q$ belongs to the Zygmund class. Thus, by the classical ODE theory, the autonomous initial value problem $$\begin{cases}\frac{d}{dt}X(t,x)=\v(X(t,x)),\\X(0,x)=x.\end{cases}$$ has a well defined, unique flow of time-dependent solutions $X(t,x)$. Moreover, in the space variable $x$, this solution is a Hölder continuous homeomorphism. If $\v=\v(t,x)$ is not autonomous and also depends on time, then the same conclusion holds if one assumes $\sup_t\|\v(t,\cdot)\|_Q<\infty$.\
\
The relevance of Reimann’s vector fields in Geometric Function Theory was first proven in [@Rei] with the quasisymmetry of the flow maps $x\mapsto X(t,x)$. At the same time, it is quite remarkable the fact that these maps enjoy a significant degree of Sobolev regularity in the space variable, as a consequence of the quasisymmetry. This fact puts Reimann’s $Q$ class into a very narrow and unstable borderline: the one between the classical ODE theory and a much more recent result by Jabin [@J] (see also [@ACM]). Roughly, in the first theory Lipschitz vector fields are proven to produce bilipschitzian flows. The second theory refers to vector fields in the Sobolev space $W^{1,p}$ ($p<\infty$), and asserts that no Sobolev smoothness (even fractional) can be expected for their flow.\
\
Among the tools for proving the Sobolev regularity of the flow of a given $\v\in Q$, there is the following differential characterization from [@Rei Theorem 3], $$\label{Reimannequi}
\v\in Q\hspace{1cm}\Longleftrightarrow\hspace{1cm}S\v\in L^\infty( \R^n)\text{ and }\frac{|\v(x)|}{|x|\,\log(e+|x|)}\leq C,$$ as well as its quantitative formulation $\|\v\|_Q\simeq \|S\v\|_{L^\infty}$. Here $S\v$ denotes the traceless symmetric differential of $\v$, $$S\v = \frac{D\v + D^t\v}2-\frac{\div\v}{n}\,\Id.$$ When $n=2$, $S\v$ reduces to $\overline\partial\v$, the classical Cauchy-Riemann derivative from complex analysis, $$\overline\partial\v = \frac{(\partial_x+i\,\partial_y)(v^1+i\,v^2)}2\equiv\frac12\left(\begin{array}{c}\partial_xv^1-\partial_yv^2\\\partial_xv^2+\partial_yv^1\end{array}\right).$$ From , one deduces that if $\v\in Q$ then the flow map $x\mapsto X(t,x)$ is quasiconformal at every time. The authors address the interested reader to the monographs [@AIM] or [@IM2] for a self-contained background in quasiconformality. Roughly, quasiconformal maps are a relatively compact class of Sobolev homeomorphisms, and their trascendence goes beyond Geometric Function Theory to many areas in mathematics. In particular, when $n=2$ their optimal degree of Sobolev regularity can be obtained from Astala’s Area Distortion Theorem [@A].\
\
It turns out a similar situation occurs in several active scalar models, an apparently disconnected area. For instance, the planar Euler system for incompresible, inviscid fluids, in vorticity form $$\label{euler}
\begin{cases}
\omega_t + (\v\cdot\nabla)\omega=o\\
\v(t,\cdot)=\frac{i}{2\pi z}\ast \omega(t,\cdot)\\
\omega(0,\cdot)=\omega_0
\end{cases}$$ was proven to be well posed by Yudovich [@Y] in the class of vector fields with bounded curl. More precisely, given a compactly supported $\omega_0:\R^2\to\R$ with $\omega_0\in L^\infty$, Yudovich [@Y] proved existence and uniqueness of a solution $\omega=\omega(t,z)$ of belonging to $L^\infty((0,\infty)\times \R^2)$. This, together with the incompressibility, provides us with a vector field $\v=\v(t,z)$ such that $\partial\v\in L^\infty((0,\infty)\times \R^2)$. Here $\partial\v$ denotes the complex derivative of the velocity field $\v$, $$\partial\v= \frac{(\partial_x-i\,\partial_y)(v^1+i\,v^2)}2\equiv\frac12\left(\begin{array}{c}\partial_xv^1+\partial_yv^2\\\partial_xv^2-\partial_yv^1\end{array}\right)=\frac{\div\v+i\,\curl\v}2.$$ A similar situation is given in the aggregation model (in which the convolution kernel from is replaced by $\frac{1}{2\pi z}$). In analogy with Reimann, it was recently shown in [@CJ] that, at least for small times, vector fields $\v$ satisfying $\partial\v\in L^\infty$ admit a well defined flow which is Sobolev regular in the space variable, with a Sobolev exponent that may vary with time. In [@BN], this result was improved and obtained a degree of Sobolev regularity for the flow for every time.\
\
Although conditions $\overline\partial\v\in L^\infty$ and $\partial\v\in L^\infty$ may look analytically similar, they have a significant difference. In the first case, for a general non-autonomous $\v$, the flow map $X(t,\cdot)$ belongs to the Sobolev space $W^{1,p}_{loc}$ whenever $$p<\frac{2}{1-\exp\left(-2\int_0^t\|\overline\partial\v(s,\cdot)\|_{L^\infty}\,ds\right)},$$ as a consequence of both Reimann’s [@Rei] and Astala’s [@A] Theorems. In contrast, this remains being an open problem in the second case. In accordance, it was conjectured in [@CJ] that if $\partial\v\in L^\infty$ then for each $t>0$ the flow map $X(t,\cdot)$ belongs to the Sobolev space $W^{1,p}_{loc}$ whenever $$p<\frac{2}{1-\exp\left(-2\int_0^t\|\partial\v(s,\cdot)\|_{L^\infty}\,ds\right)}.$$ The asymptotic behavior of this conjecture as $t\to 0$ was proven to be the right one in [@CJ]. Moreover, when $\v$ arises from , this conjecture says that $p<\frac{2}{1-e^{-t\,\|\omega_0\|_{L^\infty}}}$. By the Sobolev embedding, this gives a Hölder exponent strictly below $e^{-t\,\|\omega_0\|_{L^\infty}}$, as shown by Bahouri and Chemin [@Chex].\
\
Geometric Function Theory has proven to be very useful in obtaining the optimal Sobolev regularity in Reimann’s case, and therefore it is natural to try to face Euler’s case with a similar scheme, as it was done in the works [@CJMO3; @CJ]. In this paper, we continue this line of research by focusing our attention in the pointwise characterization of [@Rei Theorem 3]. We investigate the existence of similar pointwise characterizations of the condition $\partial\v\in L^\infty$, both in the plane and in higher dimensions.\
\
In the plane, we introduce the class $\bar{Q}$ of functions $\v: \R^2\to\R^2$ for which there is a constant $C_0\geq 0$ such that for each $x\in\R^2$ and every $h,k\neq0$ with $|h|=|k|$ one has $$\left|\frac{\langle \v(x+h)-\v(x),\bar{h}\rangle}{|h|^2}-\frac{\langle \v(x+k)-\v(x),\bar{k}\rangle}{|k|^2}\right|\leq C_0.$$ Here $\bar{h}$ and $\bar{k}$ mean complex conjugates. By $\|\v\|_{\bar{Q}}$ we denote the best possible value of $C_0$. Similarly, we denote by $R$ the set of vector fields $\v:\R^2\to\R^2$ for which there is a constant $C_0\geq 0$ such that for each $x\in\R^2$, every $h,k\neq0$ with $|h|=|k|$, and every $\theta\in[0,2\pi]$, one has $$\left|\frac{\langle \v(x+h)-\v(x),e^{i\theta}k\rangle}{|h||k|}-\frac{\langle \v(x+k)-\v(x),e^{i\theta}h\rangle}{|h||k|}\right|\leq C_0.$$ Again, $\|\v\|_{R}$ denotes the best possible constant $C_0$. Our first result is the following one.
Let $\v:\R^2\to\R^2$ be a continuous vector field. The following are equivalent:
- $\v \in \bar{Q}$.
- $\v\in R$.
- $\v$ is differentiable a.e., $\partial \v\in L^\infty$, and $\frac{|\v(x)|}{|x|\,\log(e+|x|)}\leq C$.
If one of them holds true, then $\|\v\|_{\bar{Q}}\simeq \|\v\|_R\simeq\|\partial\v\|_{L^\infty}$.
The presence of complex conjugation in the definition of $\bar{Q}$ prevents us from extending it to higher dimensions, at least trivially. Extending the definition of $R$ to $\R^n$, $n\geq 2$, seems not an easy task either, because the set of rotations to be included is not obvious (see Lemma \[Rnfailure\]). It turns out that one may still get some $L^\infty$ estimates by removing all rotations, even in higher dimensions. Namely, let us introduce $R_0$ as the class of vector fields $\v:\R^n\to\R^n$ for which there is $C_0$ such that for each $x\in \R^n$ and each $h,k$ with $|h|=|k|\neq 0$ one has $$\left|\frac{\langle \v(x+h)-\v(x),k\rangle}{|h||k|}-\frac{\langle \v(x+k)-\v(x),h\rangle}{|h||k|}\right|\leq C_0.$$ As usually, $\|\v\|_{R_0}$ denotes the best possible constant $C_0$.
Let $\v\in R_0$. Then the distribution $D\v-D^t\v$ belongs to $L^\infty$, and $$\|D\v-D^t\v\|_{L^\infty}\leq C\,\|\v\|_{R_0}.$$ for some constant $C>0$.
As it was the case for $Q$, $\bar{Q}$ or $R$, the elements of $R_0$ belong as well to the Zygmund class. However, when $n=2$ the class $R_0$ is much larger than $R$, and one cannot guarantee its elements to be differentiable a.e.. This makes it more difficult to find higher dimensional counterparts to Theorem A. In the present paper we solve this by asking $\v$ to be nearly incompressible, that is, $\div\v\in L^\infty$. This allows to state the above mentioned counterpart, which is based in the differential operator $$A\v=\frac{D\v-D^t\v}2+\frac{\div\v}n\,\Id.$$ Note that for $n=2$ one has $A\v\equiv\partial\v$.
Let $\v:\R^n\to\R^n$ be a continuous vector field.
- If $\v\in \R_0$ and $\v$ is nearly incompressible, then $\v$ is differentiable a.e. and the estimate $$\|A\v\|_{L^\infty}\leq C\,(\|\div\v\|_{L^\infty}+\|\v\|_{R_0})$$ holds.
- If $A\v\in L^\infty$ and $\frac{|v(x)|}{|x|\,\log(e+|x|)}\leq C$ then $\v\in R_0$ and $$\|\v\|_{R_0}\leq C\,\|A\v\|_{L^\infty}.$$
As in Reimann’s setting, one of the main tools here is the fact that if $\v$ is a compactly supported vector fields with $A\v\in L^\infty$ then $\v$ has $BMO$ derivatives and, in particular, it is differentiable a.e. (see Lemma \[curldiv\]). For this reason, here one can relax the assumption $\div\v\in L^\infty$ to $\div\v\in L^p$ for some $p>n$. On the other hand, as a possible application, the above result can be used to describe in a pointwise way, among all the solutions to the Euler system of equations, the ones with bounded curl.\
\
The paper is structured as follows. In Section \[poisson\] we recall some basic facts about Poisson integrals that will be used in the rest of the paper. In Section \[barQsection\] we prove $(a)\Leftrightarrow(c)$ from Theorem A. In Section \[Rsection\] we prove $(b)\Leftrightarrow(c)$ from Theorem A. In Section \[R\^nsection\] we prove Theorems B and C.
**Notation**. Bold letters like $\b,\u,\v,\w,\g$ denote vector valued functions. After identifying planar vectors with complex numbers, the inner product in $\R^2$ can be represented as $\langle z,w\rangle=\Re(z\bar{w})$, where $\Re$ denotes real part and $\bar{w}$ stands for the complex conjugate of $w$, that is, if $w=(w_1, w_2)$ then $\bar{w}=(w_1, -w_2)$. If $A\simeq B$ then there is a constant $C\geq 0$ such that $\frac{B}{C}\leq A\leq CB$.
**Acknowledgements**. The authors warmly thank Artur Nicolau for letting us know about Lemma \[higherorder\]. Both authors are partially supported by projects MTM2016-81703-ERC, MTM2016-75390 (spanish Government) and 2017SGR395 (catalan Government).
Preliminaries {#poisson}
=============
In this section we recall some fundamental facts concerning harmonic functions on the upper half space. We refer the interested reader to [@St] for a more detailed review on this. We will be working with functions defined on $\R^{n+1}_+$, where points are represented as $(x,y)$ with $x\in\R^n$ and $y>0$. Let us recall that a function $u:\R^{n+1}_+\to\R$ is said to be harmonic if $$\Delta u (x,y)=0$$ where $\Delta = \Delta_x+\partial_{yy}^2=\sum_{i=1}^n\partial^2_{x_i,x_i}+ \partial_{yy}^2$. A typical way of constructing harmonic functions on the upper half space is through the Poisson integral of a function $g:\R^n\to\R$, $$u(x,y)= P_y\ast g(x)=\int_{\R^n}P_y(x-z)\,g(z)\,dz$$ where $$P(z,y)=P_y(z)=\frac{c_n y}{(|z|^2+y^2)^\frac{n+1}2}$$ is the Poisson kernel. Above, the constant $c_n$ is chosen so that $\|P_n\|_{L^1(\R^n)}=1$. For a vector valued $\g:\R^n\to\R^m$, then one interprets $\u=P_y\ast\g:\R^{n+1}_+\to\R^m$ componentwise. In either case, one often says that $u$ is the *Poisson integral of $g$*, and that $g$ represents $u$’s *boundary values*. The following result explains the latter terminology.
If $g\in C_c(\R^n)$ then $u=P_y\ast g$ is the only solution to the Dirichlet problem $$\begin{cases}
\Delta u=0&\R^{n+1}_+\\u(\cdot, 0)=g&\R^n.
\end{cases}$$
From$$\partial^2_{yy}P(z,y)=(n+1)P_y(z)\,\frac{ -3|z|^2+ny^2}{(|z|^2+y^2)^2}\hspace{1cm}\Delta_zP_y(z)=(n+1)\,P_y(z)\,\frac{3|z|^2-ny^2}{(|z|^2+y^2)^2}$$ it is immediate that $\Delta_z P_y(z)+\partial_{yy}^2P_y(z)=0$ and so $P_y(z)$ is harmonic on $\R^{n+1}_+$. As a consequence, $u$ is harmonic on $\R^{n+1}_+$. About the boundary condition, it suffices to observe that $P_y$ is an approximation of unity in $\R^n$, so one has $P_y\ast g\to g$ uniformly as $y\to 0$. In particular, $u$ is continuous on $\overline{\R^{n+1}_+}$ and $u(x,0)=g(x)$ for every $x\in\R^n$. Uniqueness follows from the maximum principle for harmonic functions.
One may ask if there are other harmonic functions in $\R^{n+1}_+$ that are not representable as $P_y\ast g$ for some $g$. The theory of Hardy spaces helps in this direction. Note that one may also define $P_y\ast g$ even when $g$ is a measure.
\[hardy\] Let $u:\R^{n+1}_+\to\R$ be harmonic.
- Given $1<p\leq\infty$, there is $g\in L^p(\R^n)$ such that $u=P_y\ast g$ if and only if $\sup_y \|u(\cdot, y)\|_{L^p}<\infty$, and moreover in this case one has $\|g\|_{L^p}=\sup_y \|u(\cdot, y)\|_{L^p}$.
- There is a finite Borel measure $\mu$ on $\R^n$ with $u=P_y\ast\mu$ if and only if $\sup_y \|u(\cdot, y)\|_1<\infty$, and moreover in this case one has $\|\mu\|=\sup_y \|u(\cdot, y)\|_1$. Furthermore, if $u>0$ then $\mu$ is non-negative.
Poisson integrals of $BMO$ functions can also be characterized, but its description involves a completely different quantity, as stated in the following Theorem by Carleson. Let us remind that $g:\R^n\to\R$ belongs to the $BMO$ class if $$\|g\|_\ast=\sup\left\{\frac{1}{|B|}\int_B\left|g-\frac{1}{|B|}\int_Bg\right|; B\subset\R^n\text{ is a ball}\right\}<\infty.$$
\[carlesonthm\] Let $u:\R^{n+1}_+\to\R$ be harmonic. Then $u=P_y\ast g$ for some $g\in BMO(\R^n)$ if and only if $$\|u\|_{\ast\ast}=\sup_{x_0\in\R^n,\delta>0}\frac{1}{|B(x_0,\delta)|}\int_0^\delta\int_{B(x_0,\delta)}(|D_xu(x,y)|^2+|\partial_yu(x,y)|^2)\,dx\,y\,dy <\infty.$$ Moreover, in case this happens, then $\|u\|_{\ast\ast}\simeq \|g\|_\ast$ with universal constants.
For a non continuous function $g$, calling it to be the *boundary values* of $P_y\ast g$ requires some explanation. Let us remind that the limit $\lim_{y\to 0}u(x,y)= g(x)$ is said to be taken *nontangentially at the point $x$* if and only if it happens when $(x,y)$ move within a cone with vertex $x$.
\[nontangential\] Let $g\in L^p(\R^n)$.
- If $1\leq p\leq\infty$, then $P_y\ast g\to g$ nontangentially at almost every point.
- If $1<p<\infty$, then $\|P_y\ast g-g\|_{L^p}\to 0$ as $y\to 0$.
- If $p=1$ or $p=\infty$ then there exists $g\in L^p(\R^n)$ such that $\|P_y\ast g-g\|_{L^p}\nrightarrow 0$ as $y\to 0$.
In the case of Borel measures, the situation is significantly different. To see this, if $
\delta_0$ is the Dirac Delta then $u(\cdot,y)=P_y\ast\delta_0=P_y$ so that $u(x,0)\equiv 0$. It turns out the following is true.
If $g$ is a finite Borel measure, singular w.r.t. $dx$, then $P_y\ast g$ has nontangential limit $0$ almost everywhere.
Combining propositions \[hardy\] and \[nontangential\], one sees that every bounded harmonic function $u:\R^{n+1}_+\to\R$ is precisely of the form $u=P_y\ast g$ for some $g\in L^\infty(\R^n)$, and moreover $u(\cdot,y)$ converges nontangentially to $g$ at almost every point. It is interesting to note that there is some control as well on the first order derivatives of $u$.
\[easyLinfty\] If $g\in L^\infty(\R^n)$ then $$\aligned
\|P_y\ast g\|_{L^\infty}&\leq \|g\|_{L^\infty} \\
\|\partial_y(P_y\ast g)\|_{L^\infty}&\leq n\,\frac{\|g\|_{L^\infty} }{y}\\
\|D_x(P_y\ast g)\|_{L^\infty}&\leq \frac{n+1}2\,\frac{\|g\|_{L^\infty} }{y}
%\|D_Z\partial_y(P_y\ast g)\|_{L^\infty}&\leq C(n)\,\frac{\|g\|_{L^\infty} }{y^2}\\
%\|\partial^2_{yy}(P_y\ast g)\|_{L^\infty}&\leq C(n)\,\frac{\|g\|_{L^\infty} }{y^2}\\
\endaligned$$
First, one easily sees that $|P_y\ast g(x)|\leq \|P_y\|_1\cdot\|g\|_{L^\infty}=\|g\|_{L^\infty}$ since $\|P_y\|_{L^1(\R^n)}=1$. Secondly, direct calculation shows that $$\partial_yP(z,y)= \frac{P_y(z)}{y}\,\frac{|z|^2-ny^2}{|z|^2+y^2}\hspace{2cm}D_zP_y(z)=\frac{P_y(z)}y\,\frac{-(n+1)\,yz}{|z|^2+y^2}$$ Thus, $|\partial_yP_y(z)|\leq\frac{n\,P_y(z)}{y} $ and hence $\left|(\partial_yP_y)\ast g(x)\right|\leq n\, \frac{\|g\|_{L^\infty}}{y} $. The bound for the spatial derivative follows in the same way, after observing that $|D_zP_y(z)|\leq \frac{n+1}2\,\frac{P_y(z)}{y}$.
Lemma \[easyLinfty\] motivates the introduction of the class $B$ of *harmonic Bloch functions*, which consists of functions $u:\R^{n+1}_+\to\R$ that are harmonic in $\R^{n+1}_+$ and whose gradient blows up as $y\to0$ like $\frac1y$, that is, $$u\in B\hspace{1cm}\Longleftrightarrow\hspace{1cm}u\text{ is harmonic and }\|u\|_B=\sup_{\R^{n+1}_+}\,y(|D_xu(x,y)|+|\partial_yu(x,y)|)<\infty.$$ Vector valued harmonic Bloch functions are defined componentwise. Examples of harmonic Bloch functions are, for instance, Poisson integrals of $L^\infty$ functions, as shown in Lemma \[easyLinfty\]. It turns out Poisson integrals of $BMO$ functions also belong to the Bloch class.
\[poissonBMO\] If $g\in BMO$ then $P_y\ast g\in B$, and moreover $$\aligned
\|\partial_y (P_y\ast g)\|_{L^\infty}&\leq \frac{C(n)\,\|g\|_\ast}y\\
\|D_x(P_y\ast g)\|_{L^\infty}&\leq \frac{C(n)\,\|g\|_\ast}{y}\\
%\|D\partial_y (P_y\ast g)\|_{L^\infty}&\leq \frac{C(n)\,\|g\|_\ast}{y^2}
\endaligned$$
One can find a proof in [@RR p. 86-87] or also in [@FJN Lemma 1.1]. We sketch the latter here for the reader’s convenience. Denote $u(x,y)=P_y\ast g(x)$. Then $u$ is harmonic in the upper half space, and therefore all its partial derivatives are harmonic as well. By the mean value property, if $r=\frac{y_0}4$ then $$\partial u(x_0,y_0)=\fint_{|x-x_0|^2+|y-y_0|^2<r^2} \partial u(x,y)\,dx\,dy$$ at any point $(x_0, y_0)\in\R^{n+1}_+$. Here $\partial$ denotes any element of the set $\{\partial_{x_1},\dots,\partial_{x_n},\partial_y\}$. We now observe that $$\aligned
|\partial u(x_0,y_0)|^2
&=\left|\fint_{|x-x_0|^2+|y-y_0|^2<r^2} \partial u(x,y)\,dx\,dy\right|^2\\
&\leq \fint_{|x-x_0|^2+|y-y_0|^2<r^2} |\partial u(x,y)|^2\,dx\,dy\\
&\leq \fint_{|x-x_0|^2<r^2}\,\fint_{[3y_0/4, 5y_0/4]}|\partial u(x,y)|^2\,dy\,dx\\
&\leq \fint_{|x-x_0|^2<r^2}\,\frac{8}{3y_0^2}\int_{[3y_0/4, 5y_0/4]}|\partial u(x,y)|^2\,y\,dy\,dx\\
&\leq \frac{c}{r^2} \fint_{|x-x_0|^2<(5r)^2}\,\int_{[0, 5r]}|\partial u(x,y)|^2\,y\,dy\,dx\leq \frac{c}{r^2} \,\|u\|_{\ast\ast}^2\leq \frac{c}{r^2}\,\|f\|_\ast^2=\frac{c}{y_0^2}\,\|f\|_\ast^2\endaligned$$ as claimed.
It is very rellevant for this paper the blow-up at the boundary of higher order derivatives of Poisson integrals. In this direction, we have the following fact from [@St Appendix].
\[higherorder\] If $u:\R^{n+1}_+\to\R$ is harmonic, then $$\sup_{(x,y)\in \R^{n+1}_+}\,\left(\sup_{1\leq i_1\leq\dots\leq i_k\leq n+1} y^k\,|\partial^k_{x_{i_1}\dots x_{i_k}} u(x,y)|\right)\leq C(n,k)\,\sup_{(x,y)\in \R^{n+1}_+}\,\sup_{1\leq i\leq n+1} y \,|\partial_{x_i} u(x,y)|.$$
In other words, the blow-up of the first order derivatives roughly determines that of the higher order ones. In particular, if $u$ is a harmonic Bloch function and $Hu(x,y)$ denotes its $(n+1)$-dimensional Hessian, $$Hu(x,y)=\left(\begin{array}{cc}D^2_xu(x,y)&D_x\partial_yu(x,y)\\D_x\partial_yu(x,y)&\partial^2_{yy}u(x,y)\end{array}\right)$$ then one has $$\label{blochy2}
y^2\,|Hu(x,y)|\leq C(n)\, \|u\|_B.$$ It turns out that the bound may be significantly improved if $u$ is the harmonic extension of a function in the Lipschitz class. Recall that $g:\R^n\to\R$ is Lipszhitz if $$\|g\|_{Lip}=\inf\left\{C\geq 0: |g(x)-g(y)|\leq C|x-y|\text{ for every }x,y\in\R^n\right\}<\infty.$$ Lipschitz functions are also characterized by having bounded derivatives. Thus, if $g\in Lip$ and $u=P_y\ast g$ then $D_xu=P_y\ast Dg$ and therefore combining Lemmas \[easyLinfty\] and \[higherorder\] one gets $$\label{blochy1}
y\,|Hu(x,y)|\leq C(n)\, \|Dg\|_{L^\infty}.$$ which certainly improves . Let us recall that $g:\R^n\to\R$ is an element of $Z$ if and only if $$\|g\|_Z=\inf\left\{C\geq 0: |g(x+h)+g(x-h)-2g(x)|\leq C|h|\text{ for every }x,h\in\R^n\right\}<\infty.$$ For instance, if $g$ has distributional derivatives $Dg\in BMO$ then $g\in Z$. The Zygmund class is a little larger than the Lipschitz class $Lip$. Indeed, one may think that $Z$ is to $Lip$ what $BMO$ is to $L^\infty$. Thus, the following result has an easy proof for functions in $Lip$, and a more complicate one for functions in $Z$.
\[poissonZ\] Let $g\in L^\infty(\R^n)$, and $u=P_y\ast g$. Let $\nabla u=(D_xu,\partial_yu)$ denote the $(n+1)$-dimensional gradient of $u$. Then $g\in Z$ if and only if $\nabla u \in B$, and moreover $$\frac1C\,\|g\|_Z\leq \|\nabla u\|_B\leq C\,\|g\|_Z$$ for some constant $C>0$.
A proof of this fact can be found in [@St p. 146]. As a consequence, if $g\in Z$ and $u=P_y\ast g$ then Lemma \[poissonZ\] tells that $$\label{hessian1/y}
y\,|Hu(x,y)|\leq C\,\|g\|_Z ,$$ which is better than . Moreover, one can combine this with Lemma \[higherorder\] and obtain that $y^k |\nabla^{k+1}u(x,y)|$ is bounded by a multiple of $\|g\|_Z$, for every $k=1,2,\dots$. Inequality can be proven, for instance, if $g$ is a function with $Dg\in BMO$. That proof requires the help of the classical $BMO-H^1$ duality (see also [@Rei p. 263, top Corollary]). However, functions with $BMO$ derivatives belong to the Zygmund class. For this reason, we prefered to state Lemma \[poissonZ\] and use the notion of harmonic Bloch gradients, which characterizes the class of Zygmund functions and at the same time allows for a more precise constant in the inequality.\
\
Finally, we include in this section the following result, which will be repeatedly used in the rest of the paper, and whose proof is implicit in [@IM]. Let us recall that if $\b:\R^n\to\R^n$ is a vector field, then one defines the divergence and the curl of $\b$, respectively, as $$\div\b =\tr(D\b)\hspace{1cm}\curl\b = D\b-D^t\b$$
\[curldiv\] Let $1<p<\infty$. If $\b:\R^n\to\R^n$ is continuous and compactly supported, and $\curl\b,\div\b\in L^p(\R^n)$, then also $D\b\in L^p(\R^n)$, with $$\|Db\|_{L^p}\leq C\,(\|\div\b\|_{L^p}+\|\curl\b\|_{L^p}).$$ If the assumptions hold with $p=\infty$, then one has $D\b\in BMO$, and $$\|Db\|_{\ast}\leq C\,(\|\div\b\|_{L^p}+\|\curl\b\|_{L^p}).$$
We write the proof for the reader’s convenience. When $n=2$, the assumptions say that $\b$ has complex derivative $\partial\b=\frac{\div\b+i\curl\b}2$ in $L^p$. Since $\b$ is continuous and compactly supported, we can write $\b = \frac{1}{\pi\bar{z}}\ast (\partial\b)$, whence $\overline\partial\b = p.v.\frac{-1}{\pi\bar{z}^2}\ast(\partial\b)$. But the convolution with $p.v.\frac{-1}{\pi\bar{z}^2}$ defines a Calderón-Zygmund operator, and thus $\overline\partial\b\in L^p$ (or $BMO$, if $p=\infty$) with $\|\overline\partial\b\|_{L^p}\leq C\,\|\partial\b\|_{L^p}$ (resp. $\|\overline\partial\b\|_{\ast}\leq C\,\|\partial\b\|_{L^\infty}$) as claimed.\
\
When $n>2$ the proof is a little bit delicate. We start by reminding that the second derivatives of a function $v$ vanishing at infinity can be recovered from its laplacian $\Delta v$ through the second order Riesz transforms, $$\frac{\partial^2v}{\partial x_j\partial x_k} = - R_jR_k\Delta v , \hspace{1cm} j,k=1,\dots, n.$$ where $\widehat{R_j v}(\xi)=-i\frac{\xi_j}{|\xi|}\,\widehat{v}(\xi)$ at the Fourier side. As Calderón-Zygmund operators, one has again that $R_j: L^p\to L^p$ is bounded if $1<p<\infty$, and that $R_j: L^\infty\to BMO$ is bounded. We now proceed first with the proof for $p\in(1,\infty)$. Since $\b$ is continuous and compactly supported, the Poisson equation $$\Delta \u =\b$$ has a unique solution $\u:\R^n\to\R^n$ vanishing at infinity. In particular, the distributional Hessian matrix $H\u$ of the solution $\u$ has all its entries in $L^s$ and $\|H\u\|_s\leq C\,\|\b\|_s$, for every $s\in (1,\infty)$. We now decompose $\b$ as follows, $$\label{decomp}
\b=\nabla \div\u + \div \curl\u$$ where we recall that $\curl \u=D\u - D^t\u$ is a matrix valued field. This is, indeed, the Hodge decomposition of $\b$ as the sum of a curl free vector field (i.e. $\nabla\div\u$) and a divergence free field (i.e. $\div\curl\u$). We now observe that $\curl\u$ solves the following Poisson equation, $$\label{eqcurl}\Delta(\curl \u)=\curl\b$$ because $\Delta(\curl \u)=\curl (\Delta\u)$. In particular, if $\curl\b\in L^p$ then the same holds for the hessian $H(\curl\u)$, and moreover $\|H(\curl\u)\|_{L^p}\leq C\,\|\curl\b\|_{L^p}$. Similarly, $\div\u$ solves the Poisson equation $$\label{eqdiv}\Delta(\div\u)=\div\b$$ because $\Delta(\div\u)=\div(\Delta\u)$. This shows that if $\div\b$ belongs to $L^p$ then also the hessian $H(\div\u)$ does, and we have the bound $\|H(\div\u)\|_{L^p}\leq C\,\|\div\b\|_{L^p}$. Summarizing, if both $\curl\b,\div\b\in L^p$, then both hessians $H(\curl\u)$ and $H(\div\u)$ have $L^p$ entries, whence both terms in the right hand side of belong to the homogeneous Sobolev space $\dot{W}^{1,p}$, and $$\aligned\|\b\|_{\dot{W}^{1,p}}&\leq \|\nabla \div\u \|_{\dot{W}^{1,p}}+ \|\div \curl\u\|_{\dot{W}^{1,p}}\\
&\leq\|H(\div\u)\|_{L^p}+\|H(\curl\u)\|_{L^p}\\
&\leq C\|\div\b\|_{L^p}+C\|\curl\b\|_{L^p}\endaligned$$ so the claim follows if $1<p<\infty$. In case that $\curl\b,\div\b\in L^\infty$, then the proof follows similarly, with the only difference that now $\curl\u$ and $\div\u$ have distributional hessian in $BMO$ instead, and therefore both terms in have first order derivatives in $BMO$, so $\b$ also does.\
\
It just remains to prove , which we do by direct calculation, $$\aligned
\nabla\div\u &+ \div \curl\u =\\
&= \left(\begin{array}{c}\partial_{x_1}\div\u\\\vdots\\\partial_{x_n}\div\u\end{array}\right)+\div\left(\begin{array}{cccc}
0&\partial_{x_2}u^1-\partial_{x_1}u^2&\dots&\partial_{x_n}u^1-\partial_{x_1}u^n\\
\partial_{x_1}u^2-\partial_{x_2}u^1&0&\dots&\partial_{x_n}u^2-\partial_{x_2}u^n\\
\vdots&\vdots&\ddots&\hdots\\
\partial_{x_1}u^n-\partial_{x_n}u^1&\partial_{x_2}u^n-\partial_{x_n}u^2&\dots&0\\
\end{array}\right)\\
&=\left(\begin{array}{c}\sum_j\partial^2_{x_1x_j}u^j\\\sum_j\partial^2_{x_2x_j}u^j\\\vdots\\\sum_j\partial^2_{x_nx_j}u^j\end{array}\right)+
\left(
\begin{array}{c}
\sum_{j\neq 1}\partial^2_{x_jx_j}u^1-\partial_{x_1}\sum_{j\neq 1}\partial_{x_j}u^j\\
\sum_{j\neq 2}\partial^2_{x_jx_j}u^2-\partial_{x_2}\sum_{j\neq 2}\partial_{x_j}u^j\\
\vdots\\
\sum_{j\neq n}\partial^2_{x_jx_j}u^2-\partial_{x_2}\sum_{j\neq n}\partial_{x_j}u^j
\end{array}
\right)=\Delta\u.
\endaligned$$ This is legitimate for $\u$ because it has locally integrable second order derivatives.
The planar setting: the class $\bar{Q}$ {#barQsection}
=======================================
With the spirit of finding a counterpart to Reimann’s $Q$ class, we introduce a class $\bar{Q}$ consisting of functions $\b:\R^2\to\R^2$ such that there is $C>0$ with $$\|\b\|_{\bar{Q}}=\sup_{z\in\R^2}\sup_{|h|=|k|\neq 0}\left|\frac{\langle \b(z+h)-\b(z),\bar{h}\rangle}{|h|^2}-\frac{\langle \b(z+k)-\b(z),\bar{k}\rangle}{|k|^2}\right|<\infty$$ It is not hard to see that Lipszchitz functions are elements of $\bar{Q}$. Also, arguing as in [@Rei], one can show that the elements of $\bar{Q}$ are, at every time $t$, elements of the Zygmund $Z$ class.
\[LipsbarQsZ\] If $\b:\R^2\to\R^2$, then one has $$\|\b \|_Z\leq C\,\|\b\|_{\bar{Q}}\leq C\, \|\b \|_{Lip}.$$ In particular, Lipschitz vector fields belong to $\bar{Q}$, and elements of $\bar{Q}$ are Zygmund vector fields. Also, if $\b\in\bar{Q}$ then it holds that $$\left|\frac{\langle \b(z+h)-\b(z),\bar{h}\rangle}{|h|^2}-\frac{\langle \b(z+k)-\b(z),\bar{k}\rangle}{|k|^2}\right|\leq C\,\left(1+\left|\log\frac{|h|}{|k|}\right|\right)$$ for all pairs $h,k\neq 0$, and with $C\leq c\,\|\b\|_{\bar{Q}}$, where $c$ is a constant independent of $\b$.
The proof of the above result follows the lines of [@Rei], and therefore we omit it. The interested reader is adressed to Propositon \[zygmund\] below, whose proof is very similar. In the following lemma we give a rather descriptive necessary condition for smooth elements of $\bar{Q}$.
\[smoothbarQ\] Let $\b:\R^2\to\R^2$ be smooth. If $\b\in \bar{Q}$, then $\partial \b\in L^\infty$ and $$\|\partial \b\|_{L^\infty}\leq \frac12\,\|\b\|_{\bar{Q}}.$$
In complex coordinates, the Taylor expansion of $\b$ at a differentiability point $z\in\R^2$ looks as follows, $$\b(z+h)-\b(z)=\partial \b(z)h+\overline\partial \b(z)\,\bar{h}+o(|h|).$$ Hence, if we now take inner product with $\bar{h}$, we obtain $$\aligned
\langle \b(z+h)-\b(z),\bar{h}\rangle
&= \langle \partial \b(z)h+\overline\partial \b(z)\,\bar{h},\bar{h}\rangle+\langle o(|h|), \bar{h}\rangle\\
&= \Re ((\partial \b(z)h+\overline\partial \b(z)\,\bar{h}),h)+\langle o(|h|), \bar{h}\rangle\\
&= \Re ((\partial \b(z)h^2))+\Re((\overline\partial \b(z))\,|h|^2)+\langle o(|h|), \bar{h}\rangle\\
&= \Re ((\partial \b(z)h^2))+\Re((\overline\partial \b(z))\,|h|^2)+\langle o(|h|), \bar{h}\rangle\\\endaligned$$ whence $$\aligned
\left|\frac{\langle \b(z+h)-\b(z),\bar{h}\rangle}{|h|^2}-\frac{\langle \b(z+k)-\b(z),\bar{k}\rangle}{|k|^2}\right|&=\Re\left(\partial \b(z)\left(\frac{h^2}{|h|^2}-\frac{k^2}{|k|^2}\right)\right)\\&+\frac{\langle o(|h|),h\rangle}{|h|^2}+\frac{\langle o(|k|),k\rangle}{|k|^2}
\endaligned$$ We now choose $h,k$ so that $k=ih$ and $h^2=\epsilon\,\overline{\partial\b(z)}$, and then let $\epsilon\to 0$. We get $$\label{motivation}
\limsup_{|h|=|k|\to 0}\left|\frac{\langle \b(z+h)-\b(z),\bar{h}\rangle}{|h|^2}-\frac{\langle \b(z+k)-\b(z),\bar{k}\rangle}{|k|^2}\right|\geq2|\partial \b(z)|,$$ and therefore $|\partial \b (z)|\leq \frac12\,\|\b\|_{\bar{Q}}$. If $\b$ is differentiable at every point $x$ the claim follows.
It is a well known fact that Zygmund functions admit a modulus of continuity of the form $\delta\,\log\frac1\delta$, but may fail to differentiable almost everywhere. Thus, removing the differentiability assumption in Lemma \[smoothbarQ\] does not seem automatic. Our next goal consists of proving this is actually the case.
\[Qimpliesboundedpartial\] Let $\b:\R^2\to\R^2$ belong to the class $\bar{Q}$. Then, $\b$ is differentiable almost everywhere, has $BMO$ distributional derivatives, and $\partial \b \in L^\infty$. Moreover, $$\|\partial \b \|_{L^\infty}\leq C\,\|\b\|_{\bar{Q}}$$ for some constant $C>0$.
We first prove that it is not restrictive to assume that $\b$ has compact support. To do this, let us assume that the theorem is proved under the extra assumption that $\b$ has compact support. Now, let us be given $\b\in\bar{Q}$ non compactly supported. and set $\b_t=g_t\,\b$, where $$\label{gt}
g_t(x)=\begin{cases}
1 & |x|\leq t\\
1-\frac1t\,\log\frac{\log |x|}{\log t}& t\leq |x|\leq t^{e^t}\\
0 & t^{e^t}\leq |x|.
\end{cases}$$ Clearly, $\b_t$ has compact support. We now prove that $\b_t\in\bar{Q}$. For proving this, we denote $\tau_h\g(x)=\g(x+h)$ and $\Delta_h\g(x)=\tau_h\g(x)-\g(x)$. Then we observe that $$\aligned
\frac{\langle\Delta_h\b_t,\bar{h}\rangle}{|h|^2}&-\frac{\langle\Delta_k\b_t,\bar{k}\rangle}{|k|^2}
= \tau_hg_t\,\frac{\langle\Delta_h\b,\bar{h}\rangle}{|h|^2}-\tau_kg_t\,\frac{\langle\Delta_k\b,\bar{k}\rangle}{|k|^2}+\Delta_hg_t\,\frac{\langle \b,\bar{h}\rangle}{|h|^2}-\Delta_kg_t\,\frac{\langle \b,\bar{k}\rangle}{|k|^2}\\
&=(\tau_hg_t-\tau_kg_t)\,\frac{\langle \Delta_h\b,\bar{h}\rangle}{|h|^2}+\tau_kg_t\left(\frac{\langle\Delta_h\b,\bar{h}\rangle}{|h|^2}-\frac{\langle\Delta_k\b,\bar{k}\rangle}{|k|^2}\right)+\langle \b,\frac{\bar{h}\Delta_hg_t}{|h|^2}-\frac{\bar{h}\Delta_kg_t}{|k|^2}\rangle\\
&=(\Delta_hg_t-\Delta_kg_t)\,\frac{\langle \Delta_h\b,\bar{h}\rangle}{|h|^2}+\tau_kg_t\left(\frac{\langle\Delta_h\b,\bar{h}\rangle}{|h|^2}-\frac{\langle\Delta_k\b,\bar{k}\rangle}{|k|^2}\right)+\langle \b,\frac{\bar{h}\Delta_hg_t}{|h|^2}-\frac{\bar{h}\Delta_kg_t}{|k|^2}\rangle
\endaligned$$ Now we use the Mean Value Theorem to deduce that $$|\Delta_h g_t(x)|\leq \frac{C|h|}{t\,|x|\,\log|x|}\hspace{1cm}\text{and}\hspace{1cm}|\Delta_k g_t(x)|\leq \frac{C|k|}{t\,|x|\,\log|x|}$$ We now recall that $\b\in\bar{Q}$ implies that $\b\in Z$, and therefore $\b$ has $|x|\log|x|$ growth at infinity. Having in mind that $|g_t|\leq 1$, we have for $|h|=|k|$ that $$\left|
\frac{\langle\Delta_h\b_t,\bar{h}\rangle}{|h|^2}-\frac{\langle\Delta_k\b_t,\bar{k}\rangle}{|k|^2}\right|\leq
\left|
\frac{\langle\Delta_h\b,\bar{h}\rangle}{|h|^2}-\frac{\langle\Delta_k\b,\bar{k}\rangle}{|k|^2}\right|+\frac{C}{t}$$ whence $g_t\b\in \bar{Q}$ and $\|g_t\b\|_{\bar{Q}}\leq \|\b\|_{\bar{Q}}+\frac{C}{t}$. We are now in situation to apply the theorem to $g_t\b$ and so $g_t\b$ is differentiable a.e. and moreover we have the bound $$\|\overline\partial (g_t\b)\|_{L^\infty}\leq C\,\|g_t\b\|_{\bar{Q}}\leq C\|\b\|_{\bar{Q}}+\frac{C}{t}$$ The proof now finishes easily, as for any fixed $x$ one can always find $t>0$ large enough so that $$\overline\partial \b(x)=g_t(x)\,\overline\partial \b(x)=\overline\partial (g_t\b)(x)-\b(x)\,\overline\partial g_t(x)$$ whence, after enlarging $t$ if needed, $$|\overline\partial \b(x)|\leq \|\overline\partial (g_t\b)\|_{L^\infty}+|\b(x)\,\overline\partial g_t(x)|\leq C\,\|g_t\b\|_{\bar{Q}}+\frac{C}{t}\leq C\|\b\|_{\bar{Q}}$$ as desired. Therefore, we can assume without loss of generality that $\b$ has compact support in $\R^2$.\
\
Through a dilation if needed, we will suppose that $\supp\b\subset\D$, where $\D$ denotes the unit disk on $\R^2$. Then, since $\b$ is continuous, the convolution $\u(z,y)=P_y\ast \b(z)$ is harmonic on $\R^2\times (0,+\infty)$ and continuous in $\R^2\times[0,+\infty)$. Also the complex derivative $\partial \u$ is harmonic in $\R^2\times (0,\infty)$, and as distributions one has $$\partial \u=\partial (P_y\ast \b)=\partial P_y\ast \b=P_y\ast \partial \b.$$ In particular, the last convolution is well defined, and from $\supp(\partial\b)\subset \D$ we have $$|\partial \u(z)| =|\partial P_y\ast\b(z)|\leq C\,\int_\D\frac{|\b(w)|}{|z-w|^3}\,dA(w)\leq \frac{C}{|z|^3}\hspace{1cm}\text{for each }z\notin2\D$$ uniformly for each $y>0$. In particular, $\partial \u\in L^p(\C\setminus 2\D)$ for each $\frac23<p<\infty$. From $\b\in \bar{Q}$ and $\|P_y\|_1=1$ we have that also $\u\in \bar{Q}$ and $\|\u\|_{\bar{Q}}\leq \|\b\|_{\bar{Q}}$, uniformly in $y>0$. So by Lemma \[smoothbarQ\], one has $2\|\partial \u\|_{L^\infty}\leq \| \u\|_{\bar{Q}}=\|\b\|_{\bar{Q}}$, and this uniformly in $y$. It then follows that $\partial \u\in L^p(\C)$ uniformly in $y$, for each $1<p<\infty$. As an element of the harmonic Hardy space $h^p(\R^2\times (0,+\infty))$, $p>1$, we know that $\partial \u$ has well defined boundary values $\g\in L^p(\R^2)$, and moreover one necessarily has $\partial \u=P_y\ast \g$. Since also $P_y\ast \partial\b=P_y\ast \g$, and $p>1$, it then follows that $\partial\b=\g$ and so $\partial\b$ is actually an $L^p(\R^2)$ vector field. By Lemma \[curldiv\] we obtain $D\b\in L^p$. This already gives that $\b$ is differentiable a.e., because one can take any $p>2$ (see for instance [@Kinn Theorem 2.21]). Once we know that $\partial\b\in L^p$ and $P_y\ast \partial\b\in L^\infty$ we immediately infer that $\partial\b\in L^\infty$ with $\|\partial\b\|_{L^\infty}\leq \|P_y\ast \partial\b\|_{L^\infty}= \|\partial\u\|_{L^\infty}\leq \frac12\|\b \|_{\bar{Q}}$, and this with no dependence on $\supp\b$. Using again Lemma \[curldiv\] we get $D\b\in BMO$. In particular, $\b$ is differentiable almost everywhere.
In the converse direction, an extra assumption on the growth of $\b$ is needed.
\[partialfboundedimpliesQ\] Let $\b\in W^{1,1}_{loc}(\R^2;\R^2)$ be a continuous vector field such that $$\label{xlogxgrowth}
\limsup_{|x|\to\infty}\frac{|\b(x)|}{|x|\,\log|x|}<\infty$$ and that $\partial \b\in L^\infty$. Then $\b\in\bar{Q}$ and $\|\b\|_{\bar{Q}}\leq C\,\|\partial \b\|_{L^\infty}$.
This proof follows the scheme of [@Rei Proposition 12]. So we first assume that $\b$ has compact support. Fix two unit vectors $\alpha,\beta\in \R^2$, and set $a=\alpha h, b=\beta h$ for some $h>0$. For each vector field $\g:\R^2\to\R^2$, we define $$\Delta \g(x)=\Delta_{a,b}\g(x)=\langle \g(x+a)-\g(x),\bar{\alpha}\rangle -\langle \g(x+b)-\g(x),\bar{\beta}\rangle.$$ Clearly, $\Delta=\Delta_{a,b}$ is a linear operator in $\g$, and $$\label{elinftybound}
|\Delta \g(x)|\leq 4\,\|\g\|_{L^\infty}$$ Moreover, $\g \in \bar{Q} $ if and only if $|\Delta \g|\leq C\,h$ for some constant $C$ that does not depend on $a$, $b$. We can represent $\Delta \g$ in terms of $\partial \g$ and $\overline\partial \g$ as follows, $$\aligned
\Delta \g(x)
&=\int_0^h \frac{d}{ds}\bigg(\langle \g(x+\alpha s), \bar\alpha\rangle-\langle \g(x+\beta s),\bar\beta\rangle\bigg)\,ds\\
&=\int_0^h \langle D\g(x+\alpha s)\,\alpha, \bar\alpha\rangle-\langle D\g(x+\beta s)\,\beta,\bar\beta\rangle \,ds\\
&=\Re \int_0^h (\partial \g(x+\alpha s)\alpha^2- \partial \g(x+\beta s)\beta^2)\, ds
+\Re \int_0^h(\overline\partial \g(x+\alpha s) -\overline\partial \g(x+\beta s) ) \,ds\\&=\Delta_\partial \g(x)+\Delta_{\overline\partial}\g(x)
\endaligned$$ where we set $$\aligned
\Delta_{\overline\partial}\g(x)&=\Re \int_0^h (\overline\partial \g(x+\alpha s)- \overline\partial \g(x+\beta s))\,ds\\
\Delta_\partial \g(x)&=\Re \int_0^h(\partial \g(x+\alpha s)\,\alpha^2 -\partial \g(x+\beta s)\,\beta^2 ) \,ds
\endaligned$$ We now proceed with the proof. We denote $\u(x,y)=P_y\ast \b(x)$, $x\in\C$, $y\geq 0$. We know that $\u$ is harmonic in $\R^{3}_+$ and continuous up to the boundary, since $\b\in C_c(\C)$. For each $y>0$, $$\aligned
\b(x)=\u(x,0)
&=\int_0^yt\,\partial^2_{yy}\u(x,t)\,dt-y\,\partial_y\u(x,y)+\u(x,y)\\
&\equiv\int_0^y t\,\w_t(x)\,dt-y\,\v_y(x)+\u_y(x)
\endaligned$$ where we wrote $\u_y(x)=\u(x,y)$, $\v_y(x)= \partial_y\u(x,y)$ and $\w_r(x)= \partial_{yy}^2\u(x,r)$. By the linearity of $\Delta$, which acts only on the $x$ variable, one has $$\label{3terms}
\Delta \b(x)=\int_0^y t\, \Delta \w_t(x)\,dt-y\,\Delta \v_y(x)+\Delta \u_y(x).$$ We now bound the three terms in the right hand side. For the first one, we use Lemma \[curldiv\] to see that $\partial \b\in L^\infty$ implies $D\b\in BMO$, which in turn guarantees that $\b\in Z$. Now, from Lemma \[poissonZ\] as well as equation , we deduce that $\|H \u\|_{L^\infty}\leq C\,\frac{ \|\b\|_Z}{y}$ which in turn gives us that $$\|\w_r\|_{L^\infty}\leq C\, \frac{\|\b\|_Z}{r}.$$ This fact, together with , implies for the first term in the bound $$\left|\int_0^y t\, \Delta \w_t(x)\,dt\right|\leq \int_0^y t\,4\|\w_t\|_{L^\infty}\,dt= C \,y\,\|\b\|_Z.$$ For the second and third terms in , we use that $\Delta =\Delta_{\overline\partial}+\Delta_\partial$, $$\aligned
y\,\Delta \v_y(x)&=y\,\Delta_{\overline\partial} \v_y(x)+y\,\Delta_\partial \v_y(x)\\
\Delta \u_y(x)&= \Delta_{\overline\partial} \u_y(x)+ \Delta_\partial \u_y(x)\\
\endaligned$$ and proceed first with the $\Delta_\partial$ terms. For each fixed $y$, Lemma \[easyLinfty\] gives us that $$\aligned
\partial_{x_i}\,\u_y= \partial_{x_i} \,(P_y\ast \b)=P_y\ast (\partial_{x_i}\b)
&\Longrightarrow \partial \u_y=P_y\ast \partial\b\\
&\Longrightarrow \|\partial\u_y\|_{L^\infty}=\|P_y\ast \partial\b\|_{L^\infty}\leq \|\partial\b\|_{L^\infty}\endaligned$$ On the other hand, since $u$ is smooth, we can argue similarly to get that $$\aligned
\partial_{x_i}\v_y = \partial^2_{y, x_i}\u = \partial_y\,\left(P_y\ast \partial_{x_i}\b\right)
&\Longrightarrow \partial\v_y = \partial_y(P_y\ast \partial\b)\\
&\Longrightarrow \|\partial\v_y\|_{L^\infty} =\|\partial_y(P_y\ast \partial\b)\|_{L^\infty} \leq C \,\frac{\|\partial\b\|_{L^\infty}}{y}.
\endaligned$$ Thus, from $|\Delta_\partial \g(x)|\leq 2h\,\|\partial \g\|_{L^\infty}$ one gets that $$\aligned
|\Delta_\partial\u_y(x)|&\leq 2h\,\|\partial\u_y\|_{L^\infty} \leq C \,h\,\|\partial\b\|_{L^\infty},\\
|y\,\Delta_\partial\v_y(x)|&\leq 2h y \|\partial\v_y\|_{L^\infty}\leq C \,h\,\|\partial\b\|_{L^\infty}.
\endaligned$$ Now we proceed with the $\Delta_{\overline\partial}$ terms. Calling $\gamma=\frac{\alpha-\beta}{|\alpha-\beta|}$, we see that $$\aligned
|\Delta_{\overline\partial}\g(x)|
&=\left|\Re \int_0^h \int_0^{s|\alpha-\beta|}\frac{d}{d\sigma}(\overline\partial \g(x+\beta s+\gamma\sigma) )\,d\sigma\,ds\right|\\
&=\left|\Re\int_0^h \int_0^{s|\alpha-\beta|} D(\overline\partial \g(x+\beta s+\gamma\sigma)\cdot\gamma )\,d\sigma\,ds\right|
\leq \frac{h^2\,|\alpha-\beta|}{2}\, \|D(\overline\partial \g)\|_{L^\infty}
\endaligned$$ After applying this to $\g=\u_y$ and to $\g=\v_y$, and putting al together in , one obtains $$\label{finalstep1}
|\Delta\b(x)|\leq C \,y\,\|\b\|_Z+C \,h\,\|\partial\b\|_{L^\infty}+\frac{h^2\,|\alpha-\beta| }2\left(\| D(\overline\partial\u_y)\|_{L^\infty}+y\| D(\overline\partial\v_y) \|_{L^\infty}\right)$$ Lemma \[curldiv\] tells that from $\partial\b\in L^\infty$ we get $\overline\partial\b\in BMO$ and so $P_y\ast(\overline\partial\b)$ is harmonic Bloch. This, together with Lemma \[poissonBMO\], implies that $$\aligned
\u_y=P_y\ast\b \hspace{.5cm}
&\Longrightarrow\hspace{.5cm}\overline\partial\u_y= P_y\ast \overline\partial\b\\
&\Longrightarrow\hspace{.5cm}D(\overline\partial\u_y)= D(P_y\ast \overline\partial\b)\\
&\Longrightarrow \hspace{.5cm}\|D(\overline\partial\u_y)\|_{L^\infty}=\|D(P_y\ast \overline\partial\b)\|_{L^\infty}\leq C \,\frac{\|\overline\partial\b\|_\ast}{y}\leq C \,\frac{\|\partial\b\|_{L^\infty}}{y}.
\endaligned$$ Similarly, $$\aligned
\v_y=\partial_y\u_y=\partial_yP_y\ast\b
&\hspace{.5cm}\Longrightarrow\hspace{.5cm}\overline\partial\v_y= \partial_yP_y\ast \overline\partial\b\\
&\hspace{.5cm}\Longrightarrow\hspace{.5cm}D(\overline\partial\v_y)= D(\partial_yP_y\ast \overline\partial\b)\\
&\hspace{.5cm}\Longrightarrow\hspace{.5cm} \|D(\overline\partial\v_y)\|_{L^\infty}=\|D(\partial_yP_y\ast \overline\partial\b)\|_{L^\infty}\leq C \,\frac{\|\overline\partial\b\|_\ast}{y^2}\leq C\, \frac{\|\partial\b\|_{L^\infty}}{y^2}
\endaligned$$ Plugging the above bounds into , we get $$|\Delta\b(x)|
\leq C \,y\,\|\b\|_Z+C \,h\,\|\partial\b\|_{L^\infty}+
C \,\frac{h^2\,|\alpha-\beta| }{2y}$$ and choose $y=h$ to get $|\Delta\b(x)|\leq C \,h\,\|\partial\b\|_{L^\infty}$. So $\b\in \bar{Q}$ and $\|\b\|_{\bar{Q}}\leq C\,\|\partial\b\|_{L^\infty}$. The claim follows in the case $\b\in C_c(\C)$.\
\
In order to remove the assumption on the compact support, we use again Reimann’s ideas. So we use the $g_t$ functions introduced at , and assume that $\partial \b\in L^\infty$ and $|\b(x)|\leq C\,|x|\,\log|x|$ as $|x|\to\infty$. For every fixed $t>0$, we have that $\partial (g_t\b)=\b\,\partial g_t + g_t\,\partial \b$ and so $\partial(g_t\b)\in L^\infty$. Moreover, $g_t\b$ has compact support. It then follows that $g_t\b\in \bar{Q}$ and $\|g_t\b\|_{\bar{Q}}\leq C\,\|\partial(g_t\b )\|_{L^\infty}$. However, from we see that $$\aligned
|\partial (g_t\b)(x)|
&\leq |\partial \b(x)| + |\b(x)||\partial g_t(x)|\\
&\leq |\partial \b(x)|+ C\,|x|\log|x|\frac{1}{t|x|\,\log|x|}\\
&\leq |\partial \b(x)| + \frac{C}t
\endaligned$$ Thus, we can always pick $t>0$ large enough so that $\|g_t\b\|_{\bar{Q}}\leq C\|\partial \b\|_{L^\infty}$. We now fix $x\in\R^2$. For every pair $|h|=|k|$ there is always $t>0$ large enough and such that $|x|,|x+h|,|x+k|< t^{e^t}$ so that $\b=g_t\b$ at $x$, $x+h$ and $x+k$. Thus, when evaluating the $\bar{Q}$ norm of $\b$ at $x, x+h$ and $x+k$ one reduces the differences of $\b$ to the differences of $g_t\b$, which are controlled by $\|g_t\b\|_{\bar{Q}}$, which is independent of $t$, $|h|$ and $|k|$. It follows that $\b\in \bar{Q}$ and $\|\b\|_{\bar{Q}}\leq C\,\|\partial \b\|_{L^\infty}$.
In the above proof, among all terms in the right hand side of , most of them admit the desired key bound precisely because $\b\in Z$, except the two $\Delta_\partial$ terms, which are the only ones requiring specifically that $\partial\b\in L^\infty$.\
\
On the other hand, one can deduce from the previous Theorem that $\bar{Q}$ contains many non-trivial, non-Lipschitz vector fields. At least, as many as non-Lipschitz solutions of the planar Euler system with bounded vorticity.
\[Qnonlipschitz\] $\bar{Q}$ contains many non-Lipschitzian vector fields.
Let us assume that $\omega_0:\C\to \R$ is a real valued, compactly supported function, such that $\omega_0\in L^\infty$. It follows from Yudovich Theorem [@Y] that the associate Euler system, in its vorticity form $$\begin{cases}
\omega_t+(\v\cdot\nabla)\omega=0\\
\v(t,\cdot)=\frac{1}{2\pi}\,\frac{(y,-x)}{x^2+y^2}\ast\omega(t,\cdot)\\
\omega(0,\cdot)=\omega_0\end{cases}$$ admits a unique solution $\omega$ global in time, belonging to $L^\infty((0,\infty); L^\infty(\C))$, and whose associate velocity field $v$ is such that $\curl\v=\omega$, that is, $2 \partial\v = i\omega$. In particular, $\partial\v(t,\cdot)\in L^\infty$ for every $t$. Therefore, $\v(t,\cdot)$ is an element of $\bar{Q}$ at every time. However, it is well known that not all bounded vorticities produce Lipschitz vector fields, see for instance the example by Bahouri and Chemin in [@Chex Theorem 1.3].
An alternative to $\bar{Q}$: the class $R$ {#Rsection}
==========================================
The class $\bar{Q}$ is an appropiate counterpart to Reimann’s $Q$ class when $n=2$, but seems not so convenient if $n>2$ due to the absence of complex conjugation. The following observation shows that there is another way to recover $|\partial \b(x)|$ from the Taylor development of $\b$ at $x$ that may be more convenient with higher dimensional counterparts.
\[boundedcurldiv\] Let $\b$ be a vector field in $\R^2$. Assume that $x$ is a differentiability point of $\b$. Then $$\limsup_{|h|,|k|\to 0} \sup_{0\leq\theta\leq 2\pi}\frac{|\langle \b(x+h)-\b(x), e^{i\theta} k\rangle-\langle \b(x+k)-\b(x), e^{i\theta} h\rangle |}{|h|\,|k|} = 2\left| \partial \b(x) \right|.$$
We first note that $$\aligned
\langle D\b(x)h, e^{i\theta} k\rangle-\langle D\b(x)k, e^{i\theta} h\rangle
&=\Re\bigg( (\partial \b(x) h+\overline\partial \b(x)\,\bar{h})e^{-i\theta}\bar{k}-(\partial \b(x) k+\overline\partial \b(x)\,\bar{k})e^{-i\theta}\bar{h}\bigg)\\
&=\Re\bigg( \partial\b(x)e^{-i\theta}( h \bar{k}- k \bar{h})\bigg)\\
&=-2\Im\bigg( \partial\b(x)e^{-i\theta}\bigg)\,\Im( h \bar{k} )\\
&= \bigg(-2 \Im(\partial\b(x))\,\cos\theta +2\,\Re(\partial\b(x))\,\sin\theta\bigg)\,\Im( h \bar{k} )
\endaligned$$ But since $\b$ is differentiable at $x$ we know that $$\limsup_{|h|\to 0}\frac{|\langle\b(x+h)-\b(x)-D\b(x)h, e^{i\theta}k\rangle|}{|h|\,|k|}=\limsup_{|k|\to 0}\frac{|\langle\b(x+k)-\b(x)-D\b(x)k, e^{i\theta}h\rangle|}{|h|\,|k|}=0$$ Thus $$\aligned
2\Im\bigg( \partial\b(x)e^{-i\theta}\bigg)\,\frac{\Im( h \bar{k} )}{|h|\,|k|}
&=-\frac{\langle D\b(x)h, e^{i\theta} k\rangle-\langle D\b(x)k, e^{i\theta} h\rangle}{|h|\,|k|}\\
&=-\frac{\langle\b(x+h)-\b(x), e^{i\theta} k\rangle-\langle\b(x+k)-\b(x), e^{i\theta} h\rangle}{|h|\,|k|}\\
&+\frac{\langle o(h), k\rangle}{|h|\,|k|}+\frac{\langle o(k), h\rangle}{|h|\,|k|}
\endaligned$$ so it is obvious that if we take first supremum in $\theta$ and then $\limsup$ in $h,k$ one gets $$\limsup_{h,k\to 0}\sup_\theta\left|\frac{\langle\b(x+h)-\b(x), e^{i\theta} k\rangle-\langle\b(x+k)-\b(x), e^{i\theta} h\rangle}{|h|\,|k|}\right|\leq \left|2 \partial\b(x) \right|.$$ For the converse inequality, just choose $k=ih$, then take supremum in $\theta$ and let $h\to 0$ then $$\left|2 \partial\b(x) \right| \leq \limsup_{h,k\to 0}\left|\frac{\langle\b(x+h)-\b(x), e^{i\theta} k\rangle-\langle\b(x+k)-\b(x), e^{i\theta} h\rangle}{|h|\,|k|}\right|.$$ The claim follows.
Lemma \[boundedcurldiv\] encourages us to introduce the following definition.
\[definitionRn=2\] We say that a continuous function $\b:\R^2\to\R^2$ is an element of the class $R$ if $$\sup_{x\in\R^2}\sup_{|h|=|k|\neq0}\sup_{0\leq\theta\leq 2\pi}\frac{|\langle\b(x+h)-\b(x), e^{i\theta} k\rangle-\langle\b(x+k)-\b(x), e^{i\theta} h\rangle |}{|h|\,|k|} \leq C.$$ The best constant $C$ will be denoted by $\|\b\|_{R}$.
It is not hard to see that we have the inequalities $$\|\b\|_{Z}\leq c\,\|\b\|_R\leq c\,\|\b\|_{Lip}.$$ As it was for $\bar{Q}$, these inequalities are actually a direct consequence of Propositon \[zygmund\], which will be proven in the next sections. Also, it is not hard to deduce from Lemma \[boundedcurldiv\] that if $\b\in R$ happens to be smooth then one has the bound $$\label{Rsmooth}
\|\partial \b\|_{L^\infty}\leq \frac12\,\|\b\|_R,$$ arguing as we did in Lemma \[smoothbarQ\]. As in the previous section, the difficulty is in proving that also holds true in absence of smoothness.
\[Rimpliesboundeddb\] Let $\b:\R^2\to\R^2$ belong to the class $R$. Then $\b$ is differentiable almost everywhere, it has $BMO$ distributional derivatives, and $\partial \b\in L^\infty$ with $\|\partial \b\|_{L^\infty}\leq\frac12\|\b\|_{R}$.
The proof of the above result follows the lines of the proof we have given in Theorem \[Qimpliesboundedpartial\], so we omit it.
The above sufficient condition for belonging to $R$ is also necessary, again with the growth condition.
\[partialfboundedimpliesR\] Let $\b\in W^{1,1}_{loc}(\R^2;\R^2)$ be a vector field such that $$\limsup_{|x|\to\infty}\frac{|\b(x)|}{|x|\,\log|x|}<\infty$$ and that $\partial \b\in L^\infty$. Then $\b\in R$ and $\|\b\|_R\leq C\,\|\partial \b\|_{L^\infty}$.
Even though he proof is similar to the proof of Theorem \[partialfboundedimpliesQ\], some modifications need to be done. As before, we only do it assuming that $\b$ has compact support (removing this assumption can be done as in Theorem \[partialfboundedimpliesQ\]), and start by fixing two unit vectors $\alpha,\beta\in \R^2$, and set $a=\alpha h, b=\beta h$ for some $h>0$. Given $\g:\R^2\to\R^2$, this time one sets $$\Delta \g(x)=\Delta_{a,b,\theta}\g(x)=\langle \g(x+a)-\g(x),e^{i\theta}\beta\rangle -\langle \g(x+b)-\g(x),e^{i\theta}\alpha\rangle.$$ Here, $\theta\in\{0,\pi/2\}$. Clearly, $\Delta=\Delta_{a,b,\theta}$ is a linear operator in $\g$, and $$\label{Linftybound}
|\Delta \g(x)|\leq 4\,\|\g\|_{L^\infty}$$ Moreover, $\g \in R$ if and only if $|\Delta \g|\leq C\,h$ for some constant $C$ that does not depend on $a$, $b$ or $\theta$. The representation of $\Delta \g$ in terms of $\partial \g$ and $\overline\partial \g$ changes a bit with respect to that in Theorem \[partialfboundedimpliesQ\], $$\aligned
\Delta \g(x)
&=\int_0^h \frac{d}{ds}\bigg(\langle \g(x+\alpha s),e^{i\theta}\beta\rangle-\langle \g(x+\beta s),e^{i\theta}\alpha\rangle\bigg)\,ds\\
&=\int_0^h \langle D\g(x+\alpha s)\,\alpha, e^{i\theta}\beta\rangle-\langle D\g(x+\beta s)\,\beta,e^{i\theta}\alpha\rangle \,ds\\
&=\Delta_{\overline\partial}\g(x)+\Delta_{ \partial}\g(x),
\endaligned$$ where we have set $$\aligned
\Delta_{\overline\partial}\g(x)&=\Re \left(e^{-i\theta} \int_0^h (\overline\partial \g(x+\alpha s)- \overline\partial \g(x+\beta s))\,\bar{\beta}\bar{ \alpha}\,ds\right),\\
\Delta_\partial \g(x)&=\Re\left( e^{-i\theta}\int_0^h(\partial \g(x+\alpha s)\,\alpha\bar{\beta}-\partial \g(x+\beta s)\,\beta\bar{\alpha} ) \,ds\right).\endaligned$$ The proof now follows as the one of Theorem \[partialfboundedimpliesQ\]. So for $\u(x,y)=P_y\ast \b(x)$ one knows that $\u$ is harmonic in $\R^{3}_+$ and continuous up to the boundary, since $\b\in C_c(\C)$. For each $t>0$, $$\aligned
\b(x)=\u(x,0)
&=\int_0^yt\,\partial^2_{yy}\u(x,t)\,dt-y\,\partial_y\u(x,y)+\u(x,y)\\
&\equiv\int_0^y t\,\w_t(x)\,dt-y\,\v_y(x)+\u_y(x)
\endaligned$$ where we wrote $\u_y(x)=\u(x,y)$, $\v_y(x)= \partial_y\u(x,y)$ and $\w_r(x)= \partial_{yy}^2\u(x,r)$. By the linearity of $\Delta$, which acts only on the $x$ variable, one has $$\Delta \b(x)=\int_0^y t\, \Delta \w_t(x)\,dt-y\,\Delta\v_y(x)+\Delta \u_y(x).$$ and now one proceeds term by term. For the $\w$ term, one can use Lemma \[poissonZ\] to see that $$\aligned
\partial \b\in L^\infty\hspace{1cm}
&\Longrightarrow\hspace{1cm}D\b\in BMO\\
&\Longrightarrow\hspace{1cm} \b\in Z\\
&\Longleftrightarrow\hspace{1cm} \|H \u\|_{L^\infty}\leq C\,\frac{ \|\b\|_Z}{y}\hspace{1cm}\Rightarrow \|\w_r\|_{L^\infty}\leq C\, \frac{\|\b\|_Z}{r}\endaligned$$ Hence $$\left|\int_0^y t\, \Delta \w_t(x)\,dt\right|\leq \int_0^y t\,4\|\w_t\|_{L^\infty}\,dt= C \,y\,\|\b\|_Z$$ As desired. For the other two terms, we use that $\Delta =\Delta_{\overline\partial}+\Delta_\partial$, $$\aligned
y\,\Delta \v_y(x)&=y\,\Delta_{\overline\partial} \v_y(x)+y\,\Delta_\partial \v_y(x)\\
\Delta \u_y(x)&= \Delta_{\overline\partial} \u_y(x)+ \Delta_\partial \u_y(x)\\
\endaligned$$ and proceed first with the $\Delta_\partial$ terms. For each fixed $y$, Lemma \[easyLinfty\] gives us that $$\aligned
\partial_{x_i}\,\u_y= \partial_{x_i} \,(P_y\ast \b)=P_y\ast (\partial_{x_i}\b)
&\Longrightarrow \partial \u_y=P_y\ast \partial \b\\
&\Longrightarrow \|\partial \u_y\|_{L^\infty}=\|P_y\ast \partial \b\|_{L^\infty}\leq C \,\|\partial \b\|_{L^\infty}\endaligned$$ On the other hand, since $\u$ is smooth, we can argue similarly to get that $$\aligned
\partial_{x_i} \v_y = \partial^2_{y, x_i} \u = \partial_y\,\left(P_y\ast \partial_{x_i} \b\right)
&\Longrightarrow \partial \v_y = \partial_y(P_y\ast \partial \b)\\
&\Longrightarrow \|\partial \v_y\|_{L^\infty} =\|\partial_y(P_y\ast \partial \b)\|_{L^\infty} \leq C \,\frac{\|\partial \b\|_{L^\infty}}{y}.
\endaligned$$ Thus $$\aligned
|\Delta_\partial \u_y(x)|&\leq 2h\,\|\partial \u_y\|_{L^\infty} \leq C \,h\,\|\partial \b\|_{L^\infty}\\
|y\,\Delta_\partial \v_y(x)|&\leq 2h y \|\partial \v_y\|_{L^\infty}\leq C \,h\,\|\partial \b\|_{L^\infty}
\endaligned$$ where $C$ is a constant. Concerning the $\Delta_{\overline\partial}$ terms, we call $\gamma=\frac{\alpha-\beta}{|\alpha-\beta|}$, and observe that $$\aligned
|\Delta_{\overline\partial}\g(x)|
&=\left|\Re \left(e^{-i\theta} \int_0^h \int_0^{s|\alpha-\beta|}\frac{d}{d\sigma}(\overline\partial \g(x+\beta s+\gamma\sigma) )\,\bar{\beta}\bar{ \alpha}\,d\sigma\,ds\right)\right|\\
&=\left|\Re \left(e^{-i\theta} \int_0^h \int_0^{s|\alpha-\beta|} D(\overline\partial \g(x+\beta s+\gamma\sigma)\cdot\gamma )\,\bar{\beta}\bar{ \alpha}\,d\sigma\,ds\right)\right|
\leq \frac{h^2\,|\alpha-\beta|}{2}\, \|D(\overline\partial \g)\|_{L^\infty}
\endaligned$$ After applying this to $\g=\u_y$ and to $\g=\v_y$, one obtains $$\label{finalstep}
|\Delta \b(x)|\leq C \,y\,\|\b\|_Z+C \,h\,\|\partial \b\|_{L^\infty}+\frac{h^2\,|\alpha-\beta| }2\left(\| D(\overline\partial \u_y)\|_{L^\infty}+y\| D(\overline\partial \v_y) \|_{L^\infty}\right)$$ We now use the first inequality in Lemma \[poissonBMO\] with $\g= \overline\partial \u_y$. Indeed, by the linearity of all the involved operators $$\aligned
\u_y=P_y\ast \b \hspace{1cm}
&\Rightarrow\hspace{1cm}\overline\partial \u_y= P_y\ast \overline\partial \b
\endaligned$$ Now, since $\partial \b\in L^\infty$ we have $\overline\partial \b\in BMO$ and therefore $\b\in Z$, so Lemma \[poissonZ\] applies, $$\|D(\overline\partial \u_y)\|_{L^\infty}=\|D(P_y\ast \overline\partial \b)\|_{L^\infty}\leq C \,\frac{\| \b\|_Z}{y}.$$ For $\g=\overline\partial \v_y$, we proceed similarly, and observe that $$\aligned
\v_y=\partial_y\u_y=\partial_yP_y\ast \b\hspace{1cm}
&\Rightarrow\hspace{1cm}\overline\partial \v_y= \partial_yP_y\ast \overline\partial \b.
\endaligned$$ Hence, one may combine Lemmas \[higherorder\] and \[poissonZ\] to get $$\|D(\overline\partial \v_y)\|_{L^\infty}=\|D(\partial_yP_y\ast \overline\partial \b)\|_{L^\infty}\leq C \,\frac{\|\b\|_Z}{y^2}.$$ We now plug the above bounds into , $$|\Delta \b(x)|
\leq C \,y\,\|\b\|_Z+C \,h\,\|\partial \b\|_{L^\infty}+
C \,\frac{h^2\,|\alpha-\beta| }{2y}$$ and choose $y=h$ to get $|\Delta \b(x)|\leq C \,h\,\|\partial \b\|_{L^\infty}$. So $\b\in R$ and $\|\b\|_{R}\leq C\,\|\partial \b\|_{L^\infty}$. The claim follows in the case $\b\in C_c(\C)$.
The following corollary, Theorem A in the introduction, is a way of putting together Theorems \[Qimpliesboundedpartial\], \[partialfboundedimpliesQ\], \[Rimpliesboundeddb\] and \[partialfboundedimpliesR\].
Let $\b:\R^2\to\R^2$ be a continuous vector field. The following conditions are equivalent:
- $\b\in R$
- $\b\in \bar{Q}$
- $\b$ is differentiable a.e., $\partial \b\in L^\infty$ and $|\b(x)|\leq C|x|\,\log|x|$ as $|x|\to\infty$.
Moreover, in case this happens, then $\|\b\|_{\bar{Q}}\simeq\|\b\|_R\simeq \|\partial \b\|_{L^\infty}\simeq \|\div\b\|_{L^\infty}+\|\curl\b\|_{L^\infty}$.
As explained at the beginning of this section, the absence of complex conjugation in $\R^n$ when $n>2$ seems to make the $R$ class more suitable for higher dimensional counterparts. In order to build them, one may replace the rotation factor $e^{i\theta}$ in Definition \[definitionRn=2\] by rotations not only in the $Ox_1,x_2$ plane, but on any of the coordinate planes $Ox_i,x_j$. For this, let us introduce the set ${\mathcal J}_n=\{J_{i,j}\}_{1\leq i<j\leq n}$ of matrices $J_{i,j}\in\R^{n\times n}$ defined by $$\aligned J_{i,j}e_i&=-e_j\\ J_{i,j}e_j &= e_i\\J_{i,j}e_k&=e_k\,,\,\,k\neq i,j\endaligned$$ where $e_1,\dots,e_n$ is the canonical basis in $\R^n$. When $n=2$, ${\mathcal J}_n$ contains only the matrix $$\left(\begin{array}{cc}0&1\\-1&0\end{array}\right)$$ which is nothing but the rotation $e^{-i\frac{\pi}2}$. More in general, ${\mathcal J}_n$ contains $\frac{n(n-1)}{2}$ elements.
\[Rnfailure\] Suppose that $n\geq 3$. Let $\b:\R^n\to\R^n$ be a vector field, and assume that $x$ is a differentiability point. If $$\label{Rnattempt}
\limsup_{|h|=|k|\to 0}\sup_{J\in {\mathcal J}_n\cup\{\Id\}} \left|\frac{\langle \b(x+h)-\b(x),Jk\rangle}{|h|\,|k|}-\frac{\langle \b(x+k)-\b(x), Jh\rangle}{|h|\,|k|}\right|\leq C_0$$ then also $|D\b(x)|\leq C\,C_0$ for some dimensional constant $C$.
Since $x$ is a differentiability point, $$\aligned
\limsup_{|h|=|k|\to 0}&\left|\frac{\langle \b(x+h)-\b(x),Jk\rangle}{|h|\,|k|}-\frac{\langle \b(x+k)-\b(x), Jh\rangle}{|h|\,|k|}\right|\\
&=\limsup_{|h|=|k|\to 0} \left|\frac{\langle D\b(x)h, Jk\rangle -\langle D\b(x)k, Jh\rangle}{|h|\,|k|}\right|\\
&=\sup_{|h|=|k|=1} |\langle D\b(x)h, Jk\rangle -\langle D\b(x)k, Jh\rangle|=\sup_{|h|=|k|=1} |\langle h, (D^t\b(x)J -J^t D\b(x))k \rangle|\\
\endaligned$$ When taking $J=\Id$ one recovers the curl matrix $D\b(x)-D^t\b(x)$, $$\langle D\b(x)h, Jk\rangle -\langle D\b(x)k, Jh\rangle = \langle h, (D^t\b(x) J-J^tD\b(x))k\rangle=\langle h, (D^t\b(x)-D\b(x))k\rangle$$ Let us now take $J=J_{i,j}$ for a given pair $1\leq i<j\leq n$. We get $$\aligned
\langle D\b(x)e_i, Je_j\rangle -\langle D\b(x)e_j, Je_i\rangle &= \langle \partial_i \b, e_i\rangle + \langle \partial_j \b, e_j\rangle= \partial_i\b_i+\partial_j\b_j\\
\langle D\b(x)e_i, Je_k\rangle -\langle D\b(x)e_k, Je_i\rangle &=\langle \partial_i \b, e_k\rangle+\langle \partial_k \b, e_j\rangle=\partial_i\b_k+\partial_k\b_j,\hspace{1cm}k\neq i,j\\
\langle D\b(x)e_j, Je_k\rangle -\langle D\b(x)e_k, Je_j\rangle &=\langle \partial_j \b, e_k\rangle-\langle \partial_k \b, e_i\rangle=\partial_j\b_k-\partial_k\b_i,\hspace{1cm}k\neq i,j
\endaligned$$ Suming up the second quantity with $-\partial_i\b_k+\partial_k\b_i$, and the third with $-\partial_j\b_k+\partial_k\b_j$ (both of which come from $D\b-D^t\b$), we get that both $\partial_k\b_j+\partial_k\b_i$ and $-\partial_k\b_i +\partial_k\b_j$ are bounded by multiples of $C_0$, which means that $\partial_k\b_i,\partial_k\b_j$ are bounded by multiples of $C_0$ whenever $k\neq i,j$. Moving now $i, j$ we obtain the same sort of boundedness for all non-diagonal elements of $Db$. Also, note that the boundedness of all pairs $\partial_i\b_i+\partial_j\b_j$ implies that of all diagonal elements. This finishes the proof.
The above result shows that the class of vector fields $\b:\R^n\to\R^n$ satisfying reduces, when $n>2$, to Lipschitz vector fields. In contrast, when $n=2$, this class is much larger: this can be deduced from Lemma \[Qnonlipschitz\], together with the fact that in the plane one has $\bar{Q}=R$. This suggests it is not a good idea to build higher dimensional counterparts to $R$ in this way, because the class of vector fields one obtains is included into the Lipschitz ones, which are well understood.
Extending to higher dimensions: the class $R_0$ {#R^nsection}
===============================================
Lemma \[boundedcurldiv\] gives light to another fact: one may separate $\curl\b$ from $\div\b$ by simply choosing different values of $\theta$. This is the starting point to our following observation. Let us fix an integer $n\geq 2$.
\[pointwisecurl\] Let $\b:\R^n\to\R^n$ be a vector field, and assume that $x$ is a differentiability point of $\b$. Then $$\limsup_{|h|,|k|\to 0} \frac{|\langle \b(x+h)-\b(x), k\rangle-\langle \b(x+k)-\b(x), h\rangle |}{|h|\,|k|}= |D\b(x)-D^t\b(x)|.$$
We first observe that if $\b$ is differentiable at $x$, then $$\aligned
\langle \b(x+h)-\b(x), k\rangle
&= \langle D\b(x)h, k\rangle+ \langle o(|h|),k\rangle \\
%&= \langle \frac{Db(x)+D^tb(x)}2 \, h, k\rangle + \langle \frac{Db(x)-D^tb(x)}2 \, h, k\rangle+\langle o(|h|),k\rangle \\
%&= \langle \left(\frac{Db(x)+D^tb(x)}2 -\frac{\div b}{n}\,\Id\,\right)\, h, k\rangle + \langle \frac{Db(x)-D^tb(x)}2 \, h, k\rangle +\frac{\div b(x)}{n}\,\langle h, k\rangle+\langle o(|h|),k\rangle
\endaligned$$ Now, after exchanging the roles of $h$ and $k$, we also have $$\aligned
\langle \b(x+k)-\b(x), h\rangle
&= \langle D\b(x) k, h\rangle+\langle o(|k|), h\rangle \\
%&= \langle \frac{Db(x)+D^tb(x)}2 \,k, h\rangle + \langle \frac{Db(x)-D^tb(x)}2 \,k, h\rangle+\langle o(|k|), h\rangle \\
%&= \langle \left(\frac{Db(x)+D^tb(x)}2 -\frac{\div b}{n}\,\Id\,\right)\, k, h\rangle + \langle \frac{Db(x)-D^tb(x)}2 \, k, h\rangle +\frac{\div b(x)}{n}\,\langle k,h\rangle+\langle o(|k|), h\rangle
\endaligned$$ Thus $$\frac{\langle \b(x+h)-\b(x), k\rangle-\langle \b(x+k)-\b(x), h\rangle }{|h|\,|k|}=
\frac{ \langle (D\b(x)-D^t\b(x))\, h, k\rangle }{|h|\,|k|}+\frac{\langle o(|h|),k\rangle}{|h|\,|k|}-\frac{\langle o(|k|),h\rangle}{|k|\,|h|}.$$ and therefore one immediately gets $$\limsup_{|h|,|k|\to 0}\frac{|\langle \b(x+h)-\b(x), k\rangle-\langle \b(x+k)-\b(x), h\rangle| }{|h|\,|k|}\leq |D\b(x)-D^t\b(x)|.$$ For the converse inequality, we recall that $$|D\b(x)-D^t\b(x)|=\sup_{h,k\neq 0}\frac{ \langle (D\b(x)-D^t\b(x))\, h, k\rangle }{|h|\,|k|}$$ so we can pick two sequences $h_m, k_m\to 0$ such that $$\aligned
|D\b(x)-D^t\b(x)|&=\lim_{m\to\infty}\frac{ \langle (D\b(x)-D^t\b(x))\, h_m, k_m\rangle }{|h_m|\,|k_m|}\\
&=\lim_{m\to\infty}\frac{|\langle \b(x+h_m)-\b(x), k_m\rangle-\langle \b(x+k_m)-\b(x), h_m\rangle |}{|h_m|\,|k_m|}\\
&\leq\limsup_{|h|,|k|\to 0}\frac{|\langle \b(x+h)-\b(x), k\rangle-\langle \b(x+k)-\b(x), h\rangle |}{|h|\,|k|}
\endaligned$$ and the claim follows.
The above result motivates the following definition.
We say that a continuous function $\b:\R^n\to\R^n$ belongs to the class $R_0$ if $$\frac{\left|\langle \b(x+h)-\b(x), k\rangle-\langle \b(x+k)-\b(x), h\rangle\right|}{|h|\,|k|}\leq C$$ for each pair $h,k$ such that $|h|=|k|\neq 0$. The best constant $C$ will be denoted as $\|\b\|_{R_0}$.
When $n=1$, the only options are $h=k$ (which gives nothing) or $h=-k$, which gives us that in fact $R_0$ is exactly the class of Zygmund functions. When $n=2$, though, the above definition suggests that $R_0$ is larger than $R$. For a general $n>1$, taking $k=-h$ in the definition we obtain that $$\frac{\left|\langle \b(x+h)+ \b(x-h)-2\b(x), h\rangle\right|}{|h|^2}\leq C$$ which suggests that there may be some connection between $R_0$ and $Z$.
\[zygmund\] One has $R_0\subset Z$, and moreover $\|\b\|_Z\leq 4\,\|\b\|_{R_0}$.
Let us assume for a while that $a,b\in\R^n$ are such that $\langle a, b\rangle=0$. Then by Pitagoras $|a+b|=|a-b|$ and thus we can use condition $R_0$ for $x=z+a$, $h=b-a$ and $k=-b-a$. We get $$\aligned
|\langle \b(z+b)-\b(z+a), -b-a \rangle &- \langle \b(z-b) -\b(z+a), b-a\rangle |\\
&= |\langle \b(z-b)-\b(z+b), a\rangle - \langle \b(z+b)+\b(z-b), b\rangle + 2\langle \b(z+a), b\rangle|\\
&\leq \|\b\|_{R_0}\,|-b-a|\,|b-a| =\|\b\|_{R_0}\,(|a|^2+|b|^2)\endaligned$$ Similarly, for $x=z-a$, $h=b+a$ and $k=-b+a$, $$\aligned
|\langle \b(z+b)-\b(z-a), -b+a \rangle &- \langle \b(z-b) -\b(z-a), b+a\rangle| \\
&= |-\langle \b(z-b)-\b(z+b), a\rangle - \langle \b(z+b)+\b(z-b), b\rangle + 2\langle \b(z-a), b\rangle|\\
&\leq \|\b\|_{R_0} \,|-b+a|\,|b+a|= \|\b\|_{R_0}\,(|a|^2+|b|^2)
\endaligned$$ Summing up the above inequalities, $$|\langle \b(z+a)+\b(z-a), b\rangle-\langle \b(z+b)+\b(z-b), b\rangle|\leq \|\b\|_{R_0}\,(|a|^2+|b|^2)$$ and as a consequence $$|\langle \b(z+a)+\b(z-a)-2\b(z), b\rangle-\langle \b(z+b)+\b(z-b)-2\b(z), b\rangle|\leq \|\b\|_{R_0}\,(|a|^2+|b|^2)$$ whence $$\aligned
|\langle \b(z+a)+\b(z-a)-2\b(z), b\rangle|
&\leq |\langle \b(z+b)+\b(z-b)-2\b(z), b\rangle|+ \|\b\|_{R_0}\,(|a|^2+|b|^2)\\
&=|\langle \b(z-b)-\b(z), b\rangle -\langle \b(z+b)-\b(z), -b\rangle|+ \|\b\|_{R_0}\,(|a|^2+|b|^2)\\
&\leq \|\b\|_{R_0}\,(|a|^2+2|b|^2)\\
\endaligned$$ Let us now take a vector $v\in\R^n$, and decompose it as $v=v_1+v_2$ with $v_1=\langle v, a\rangle \frac{a}{|a|^2}$. Then $\langle v_2, a\rangle=0$ so that taking $b=\frac{v_2}{|v_2|}\,|a|$ we certainly have $\langle a,b\rangle=0$ and $|a|=|b|$, and so we can apply what we proved before. Namely, $$\aligned
|\langle \b(z+a)&+\b(z-a)-2\b(z), v\rangle|\\
&\leq |\langle \b(z+a)+\b(z-a)-2\b(z), v_1\rangle|+ |\langle \b(z+a)+\b(z-a)-2\b(z), v_2\rangle|\\
&= |\langle \b(z+a)+\b(z-a)-2\b(z), a\rangle|\,\frac{|\langle v,a\rangle|}{|a|^2}+|\langle \b(z+a)+\b(z-a)-2\b(z), b\rangle|\,\frac{|v_2|}{|a|}\\
&\leq \|\b\|_{R_0}\,|a|^2\,\frac{|\langle v,a\rangle|}{|a|^2}+\|\b\|_{R_0}\,(|a|^2+2|b|^2)\,\frac{|v_2|}{|a|}\\
&\leq \|\b\|_{R_0}\, |\langle v,a\rangle| +\|\b\|_{R_0}\,3|a|\,|v_2|\leq 4\,\|\b\|_{R_0}\,|a|\,|v| \\
\endaligned$$ and the claim follows.
In the above proof, condition $R_0$ has only been used for precise pairs $h$ and $k$ for which either $h=k$ or $\langle h, k\rangle=0$ with $|h|=|k|$. It will be clear that the class of vector fields one obtains with this restriction is exactly the same $R_0$. This can be seen as a consequence of Theorems \[R0givesboundedcurl\] and \[AboundedimpliesR0\] below.
Among the consequences, we deduce that each element of $R_0$ has growth at most $|x|\,\log|x|$, as $|x|\to\infty$, and also that each element of $R_0$ has $t\,\log\frac1t$ local modulus of continuity. Arguing as in Reimann’s Proposition 5 for $n=1$, functions in the $R_0$ class can be shown to satisfy the following extended version of condition $R_0$, $$\label{log}
\frac{|(\b(x+h)-\b(x))k-(\b(x+k)-\b(x))h|}{|hk|}\leq \|\b\|_{R_0}\,\left(\frac32+\frac1{2\log 2}\,\left|\log\frac{|h|}{|k|}\right|\right)$$ provided that $h\cdot k>0$ (replace $3/2$ by $5/2$ in case you want to allow $h\cdot k<0$). The extension of this fact to functions in the higher dimensional $R_0$ class works as follows.
There exists $C=C(n)\geq 1$ such that if $\b\in R_0$ then $$\frac{|\langle \b(x+h)-\b(x),k\rangle -\langle \b(x+k)-\b(x),h\rangle |}{|h|\,|k|}\leq C\,\|\b\|_{R_0}\,\left(1+ \left|\log\frac{|h|}{|k|}\right|\right)$$ whenever $h, k\in\R^n$ are non-zero.
Let us fix two non-zero vectors $a,b\in\R^n$, choose $y=\frac{a}{|a|}$, and observe that $$\left|\frac{\langle \b(x+|a|y)-\b(x), |a|\frac{b}{|b|}\rangle}{|a|^2}-\frac{\langle \b(x+|a|\frac{b}{|b|})-\b(x),|a|y\rangle}{|a|^2}\right|\leq \|\b\|_{R_0}$$ while $$\label{diff}
\left|\frac{\langle \b(x+b)-\b(x), a\rangle }{|a||b|}-\frac{\langle \b(x+|a|\frac{b}{|b|})-\b(x),|a|y\rangle}{|a|^2}\right|
=\left|\langle\frac{\b(x+b)-\b(x)}{|b|}-\frac{\b(x+|a|\frac{b}{|b|})-\b(x)}{|a|}, \frac{a}{|a|}\rangle\right|$$ In order to control this quantity, we use the auxiliary function $g:\R\to\R$ defined as $g(t)=\langle \frac{\b(x+tb)}{|b|}, \frac{a}{|a|}\rangle$. Since $\b\in R_0$ implies $\b\in Z$, we deduce for each fixed $t$ that $$\aligned
\left|\frac{g(t+s)+g(t-s)-2g(t)}{s}\right|
&=\left| \frac{\b(x+(t+s)b)+\b(x+(t-s)b)-2\b(x+tb),a\rangle}{s|a||b|}\right|\\
&\leq \|\b\|_Z\endaligned$$ so that $g\in Z$ with $\|g\|_Z\leq \|\b\|_Z\leq 4\|\b\|_{R_0}$. As a consequence, and arguing as in Reimann’s proof of Proposition 5 (part C) we get for all $r\in\R$ that if $t,s>0$ then $$\left|\frac{g(r+t)-g(r)}{t}-\frac{g(r+s)-g(r)}{s}\right|\leq \|g\|_Z\,\left(\frac32+\frac1{2\log 2}\,\left|\log\frac{t}{s}\right|\right)$$ In particular, if $r=0$ and $s<0<t$, $$\aligned
\left|\frac{g(t)-g(0)}{t}-\frac{g(s)-g(0)}{s}\right|
&\leq
\left|\frac{g(t)-g(0)}{t}-\frac{g(-s)-g(0)}{-s}\right|+\left|\frac{g(-s)-g(0)}{-s}-\frac{g(s)-g(0)}{s}\right|\\
&\leq \|g\|_Z\,\left(\frac32+\frac1{2\log 2}\,\left|\log\frac{t}{-s}\right|\right)+\|g\|_Z\\
&= \|g\|_Z\,\left(\frac52+\frac1{2\log 2}\,\left|\log\frac{t}{-s}\right|\right).\endaligned$$ We now go back to , and apply the above estimate with $s=1$, $t=\frac{|a|}{|b|}$. We obtain $$\aligned
\left|\langle \frac{\b(x+b)-\b(x)}{|b|}, \frac{a}{|a|}\rangle-\langle \frac{\b(x+\frac{|a|}{|b|}b)-\b(x)}{|a|}, \frac{a}{|a|}\rangle\right|
&=\left|\frac{g(t)-g(0)}{t}-\frac{g(s)-g(0)}{s}\right|\\
&\leq 4\|\b\|_{R_0}\,\left(\frac52+\frac1{2\log 2}\,\left|\log\frac{|a|}{|b|}\right|\right)
\endaligned$$ and the claim follows.
As it was done in the previous sections for the classes $\bar{Q}$ and $R$, we are interested in a differential characterization of the class $R_0$. It is clear from Lemma \[pointwisecurl\] that if $\b$ is a smooth element of $R_0$ then $$\|D\b-D^t\b\|_{L^\infty}\leq \|\b\|_{R_0}$$ However, this time the situation for a non necessarily smooth $\b\in R_0$ is more delicate than in the previous sections, because differentiability points may not even exist. Indeed, with $D\b-D^t\b$ there is not enough information to control all of $D\b$. Observe also that if $n=2$ then $R_0$ is strictly larger than $R$.
\[R0givesboundedcurl\] Let $\b\in R_0$. Then the distribution $D\b-D^t\b$ is an element of $L^\infty(\R^n)$, and $$\|D\b-D^t\b\|_{L^\infty} \leq C(n)\,\|\b\|_{R_0}$$ for some constant $C(n)$ that depends only on $n$.
The proof of this result is very similar to that of Theorem \[Qimpliesboundedpartial\]. However, some special attention is needed to stop the argument at an earlier point.
We will first assume that $\b$ has compact support. Let us call $\u=P_y\ast \b$. We will write $\u=(u^1,\dots, u^n)$ and similarly $\b=(b^1,\dots,b^n)$. One immediately sees that $\partial_{x_i}b^j$ is a well defined distribution, because $\b$ has compact support. Moreover, this distribution can be easily extended to act against testing functions with polynomial decay, as for instance Poisson extensions of smooth compactly supported functions. So the action $\langle\partial_{x_i} b^j, P_y\ast\varphi \rangle$ is well defined whenever $\varphi\in C^\infty_c$. One has $$\aligned
\langle\partial_{x_i}u^j-\partial_{x_j}u^i,\varphi\rangle
&=-\langle u^j, \partial_{x_i}\varphi\rangle +\langle u^i, \partial_{x_j}\varphi \rangle\\
&=-\langle P_y\ast b^j, \partial_{x_i}\varphi\rangle +\langle P_y\ast b^i, \partial_{x_j}\varphi \rangle\\
&=-\langle b^j, P_y\ast\partial_{x_i}\varphi\rangle +\langle b^i, P_y\ast\partial_{x_j}\varphi \rangle\\
&=-\langle b^j, \partial_{x_i}(P_y\ast\varphi)\rangle+\langle b^i,\partial_{x_j}(P_y\ast\varphi)\rangle\\
&=\langle \partial_{x_i} b^j, P_y\ast\varphi \rangle-\langle \partial_{x_j} b^i, P_y\ast\varphi \rangle\\
&=\langle \partial_{x_i}b^j- \partial_{x_j} b^i, P_y\ast\varphi \rangle
\endaligned$$ In particular, we have the following equality of distributions, $$\label{eqdistr}
\partial_{x_i}u^j-\partial_{x_j}u^i = \partial_{x_i}(P_y\ast b^j)-\partial_{x_j}(P_y\ast b^i)=P_y\ast (\partial_{x_i}b^j-\partial_{x_j}b^i)$$ which is equivalent to say that $D\u-D^t\u=P_y\ast(D\b-D^t\b)$. The convolution operator $P_y\ast$ commutes with translations. Therefore it is not hard to see that $$\b\in R_0\hspace{1cm}\Rightarrow\hspace{1cm}\u(\cdot, y)\in R_0,\,\,\text{and}\,\,\|\u(\cdot, y)\|_{R_0}\leq \|P_y\|_{L^1(\R^n)}\,\|\b\|_{R_0}=\|\b\|_{R_0}$$ where we used that $\|P_y\|_{L^1(\R^n)}=1$. However, $\u$ is smooth. Thus, every point $x$ is a differentiability point of $\u(\cdot, y)$, and therefore by Lemma \[pointwisecurl\] $$\label{bd}
|D\u(x,y)-D^t\u(x,y)|\leq \|\u(\cdot, y)\|_{R_0}\leq \|\b\|_{R_0}.$$ In particular, this shows that each slice of $\partial_{x_i}u^j-\partial_{x_j}u^i$ belongs to $L^\infty(\R^n)$, and this happens uniformly in $y>0$. We now show that one also has $\partial_iu^j-\partial_ju^i\in L^p(\R^n)$ for some $p\in (1,\infty)$. Indeed, from $\u=P_y\ast \b$ we see that $\partial_{x_i} \u=(\partial_{x_i}P_y)\ast \b$ and therefore $$\aligned
|D \u(x,y)|
&\leq C\,\int_{\supp \b} |DP_y(x-z)|\,|\b(z)|\,dz \\
&= C\,\int_{\supp \b} \frac{|(n+1)c_n\,y(x-z)|}{(y^2+|x-z|^2)^\frac{n+3}{2}}\,|\b(z)|\,dz \\
&=C\,\int_{\supp \b} \frac{|(n+1)c_n\,y(x-z)|}{y^2+|x-z|^2}\,\frac{|\b(z)|}{(y^2+|x-z|^2)^\frac{n+1}{2}}\,dz\leq c_n\,\int \frac{|\b(z)|}{|x-z|^{n+1}}\,dz \\
\endaligned$$ As a consequence, if $\supp\b\subset B(0,R)$ and $|x|>2R$ then $$|\partial_{x_i}u^j(x,y)-\partial_{x_j}u^i(x,y)|\leq \frac{c_n\,\|\b\|_{L^\infty}}{|x|^{n+1}}$$ From this, if $p>\frac{n}{n+1}$ then $\|\partial_{x_i}u^j(\cdot,y)-\partial_{x_j}u^i(\cdot,y)\|_{L^p(\R^n\setminus B(0,2R))}$ is bounded uniformly in $y$. Combining this fact with , one gets that $$\sup_{y>0}\|\partial_{x_i}u^j(\cdot,y)-\partial_{x_j}u^i(\cdot,y)\|_{L^p(\R^n)}\leq C(n,R).$$ As a consequence, $\partial_{x_i}u^j -\partial_{x_j}u^i$ belongs to the Hardy space of harmonic functions $h^p(\R^{n+1}_+)$. As such, we can infer that there is $v_{i,j}\in L^p(\R^n)$ such that $\partial_{x_i}u^j -\partial_{x_j}u^i =P_y\ast v_{i,j}$ and moreover $$\lim_{y\to 0}\| (\partial_{x_i}u^j -\partial_{x_j}u^i )-(v_{i,j})\|_{L^p(\R^n)}=0.$$ In particular, there is a subsequence of heights $y_n\to 0$ for which the converge is pointwise, $$\label{pointwlimit}
\lim_{n\to \infty} \partial_{x_i}u^j -\partial_{x_j}u^i =v_{i,j}\hspace{1cm}a.e.$$ which combined with gives us that $v_{i,j}\in L^\infty(\R^n)$. Finally, since holds for all testing functions $\varphi\in L^{p'}(\R^n)$, we see that $$\lim_{y\to 0}\|P_y\ast (\partial_{x_i}b^j -\partial_{x_j}b^i )-v_{i,j}\|_{L^p(\R^n)}=0$$ This forces $\|P_y\ast(\partial_{x_i}b^j -\partial_{x_j}b^i )\|_{L^p}$ to remain bounded as $y\to 0$, which in turn forces the distribution $\partial_{x_i}b^j -\partial_{x_j}b^i$ to belong to $L^p(\R^n)$, and therefore by Fatou’s Theorem $v_{i,j}=\partial_{x_i}b^j -\partial_{x_j}b^i $ almost everywhere. Moreover, since $v_{i,j}\in L^\infty(\R^n)$ we also have $\partial_{x_i}b^j -\partial_{x_j}b^i \in L^\infty(\R^n)$, and $$\|\partial_{x_i}b^j -\partial_{x_j}b^i \|_{L^\infty}=\|v_{i,j}\|_{L^\infty}\leq \sup_{y>0}\|\partial_{x_i}u^j -\partial_{x_j}u^i \|_{L^\infty}\leq \|P_y\|_1\,\|\b\|_{R_0}$$ so the claim follows.\
\
In order to remove the assumption on $\supp\b$, we proceed as in Theorem \[Qimpliesboundedpartial\]. So we start by recalling that $\b\in R_0$ implies $\b\in Z$, whence $$L=\limsup_{|x|\to\infty}\frac{|\b(x)|}{|x|\,\log|x|}<\infty.$$ Setting $\Delta_h\varphi(x)=\varphi(x+h)-\varphi(x)$ and $\tau_h\varphi(x)=\varphi(x+h)$ and taking $g=g_t$ as in one has $$\aligned
\langle \Delta_h(g\b), k \rangle&-\langle \Delta_k(g\b),h\rangle\\
&= \tau_hg\,\langle \Delta_h\b, k\rangle -\tau_kg\,\langle \Delta_k\b,h\rangle
+ \langle \b,k\rangle\,\Delta_hg-\langle \b, h\rangle\,\Delta_kg\\
&= \tau_hg\,(\langle \Delta_h\b, k\rangle -\langle \Delta_k\b,h\rangle)+(\tau_hg-\tau_kg)\,\langle \Delta_k\b,h\rangle
+ \langle \b,k\rangle\,\Delta_hg-\langle \b, h\rangle\,\Delta_kg
\endaligned$$ If $|x|$ is large, from the mean value theorem we see that $$|\Delta_kg(x)|\leq \frac{C|k|}{t |x|\,\log|x|}$$ $$|\Delta_hg(x)|\leq \frac{C|h|}{t |x|\,\log|x|}$$ $$|\tau_hg(x)-\tau_kg(x)|\leq \frac{C|h-k|}{t|x|\,\log|x|}$$ This, together with the growth of $\b$ at infinity, gives $$\left|\frac{\Delta_h(g\b),k\rangle-\langle\Delta_k(g\b),h\rangle}{|h|\,|k|}\right|\leq \|\b\|_{R_0}+\frac{C}t$$ and so $g\b \in R_0$ and has compact support. From the first part of the proof, we deduce that $D(g\b)-D^t(g\b)\in L^\infty$, with norm les than $\|g\b\|_{R_0}$. But from $$D(g\b )-D^t(g\b)=\b\otimes \nabla g-\nabla g\otimes \b+ g\,(D\b-D^t\b)$$ one gets at points $|x|\leq t$ that $$D(g\b )-D^t(g\b)=D\b-D^t\b$$ whence for $|x|\leq t$ one has $$\aligned
|D\b(x)-D^t\b(x)|
&\leq \|D(g\b)-D^t(g\b)\|_{L^\infty} \\
&\leq \|g\b\|_{R_0} \\
&\leq \|\b\|_{R_0}+\frac{C}t\leq C(n) \|D\b-D^t\b\|_{L^\infty} + \frac{C}t.
\endaligned$$ The proof finishes by letting $t\to\infty$.
Theorem \[R0givesboundedcurl\] provides a sufficient condition for $L^\infty$ bounds for the distributional curl $D\b-D^t\b$. It says nothing about the differentiability of $\b$, nor the total pointwise differential $D\b$. For instance, if $u\in W^{1,1}(\R^n)$ and $\b=\nabla u$, then $D\b-D^t\b=0$ in the sense of distributions, but $\b$ may not be differentiable almost everywhere. That is, zero curl does not imply pointwise differentiability a.e.. Hence, at this point it is not clear why should any $\b\in R_0$ be differentiable almost everywhere. This absence of regularity makes it harder to state Theorem \[R0givesboundedcurl\] in the same terms we stated Theorems \[Qimpliesboundedpartial\] and \[Rimpliesboundeddb\] above. We solve this obstruction in the following result, which is a slight modification of Theorem \[R0givesboundedcurl\]. It refers to a slightly smaller subclass of $R_0$, given in terms of the divergence $\div\b$.
\[boundedcurl\] Let $\b:\R^n\to\R^n$ belong to the class $R_0$. Assume that $\div\b\in L^p_{loc}(\R^n)$.
- If $1<p<\infty$, then $\b$ has $L^p_{loc}(\R^n)$ distributional derivatives, and $\curl\b\in L^\infty(\R^n)$.
- If $p>n$, then one further has that $\b$ is differentiable almost everywhere.
Let us first assume that $\b$ has compact support. If $\b\in R_0$ then we know from Theorem \[R0givesboundedcurl\] that $D\b-D^t\b\in L^\infty$. Then, from the compact support we deduce that $D\b-D^t\b\in L^p$, and from $\div\b\in L^p$ and Lemma \[curldiv\] we get that $D\b\in L^p$. The rest is standard real analysis. If $\b$ has not compact support, then using the functions $g_t$ from we see that $g_t\b$ is an element of $R_0$ with $L^p$ divergence, so using again Lemma \[curldiv\] we get that $g_t\b$ has $L^p(\R^n)$ derivatives, which in turn ensures $D\b\in L^p_{loc}$. The differentiability a.e. is a well known result of classical real analysis, see for instance [@Ev].
We also obtain the following counterpart to Theorem \[boundedcurl\] in $\R^n$ for the case $p=\infty$. It states that vector fields in $R_0$ with bounded divergence must necessarily have $BMO$ derivatives and bounded curl. Let us recall that $$A\b = \frac{D\b-D^t\b}{2}+\frac{\div\b}{n}\,\Id.$$ It can also be seen as a counterpart to [@Rei Proposition 15], as well as to Theorems \[Qimpliesboundedpartial\] and \[Rimpliesboundeddb\] above.
Let $\b:\R^n\to\R^n$ belong to the class $R_0$. Assume that $\div\b\in L^\infty(\R^n)$. Then $\b$ is differentiable almost everywhere, has $BMO(\R^n)$ distributional derivatives, and $A\b\in L^\infty(\R^n)$.
We first assume that that $\b$ has compact support. Having also that $\b\in R_0$, we proved in Theorem \[R0givesboundedcurl\] that also $\curl\b\in L^\infty$. It then follows from Lemma \[curldiv\] that $D\b\in BMO$, and so the differentiability a.e. is automatic. The boundedness of $A\b$ is immediate. The proof for non compactly supported $\b$ goes similarly, since $g_t\b$ is compactly supported and also $g_t\b\in R_0$.
In the converse direction, we have the following result, which establishes a much netter counterpart to [@Rei Proposition 12] or Theorems \[partialfboundedimpliesQ\] or \[partialfboundedimpliesR\].
\[AboundedimpliesR0\] Let $\b \in W^{1,1}_{loc}(\R^n; \R^n)$ be a vector field. Assume that $\b $ is continuous, and that $|\b (x)|\leq O( |x|\log|x|)$ as $|x|\to\infty$. If there exists a constant $C>0$ such that $$\| A\b \|_{L^\infty} \leq C$$ then $\b $ belongs to the $R_0$ class, and $\|\b \|_{R_0}\leq C'$ for some constant $C'$ depending only on $C$.
Let us remind that $A\b\in L^\infty$ gives us bounds for $\div\b$ and $\curl\b$ in the $L^\infty$ norm. Again, we follow the steps in the proof of Theorem \[partialfboundedimpliesQ\], so we will first assume that $\b$ has compact support, and later on will remove this assumption. Given $\alpha,\beta\in \R^n$, $|\alpha|=|\beta|=1$, set $a=\alpha h, b=\beta h$ for some $h>0$. For each $\g :\R^n\to\R^n$, define $$\Delta \g (x)=\Delta_{a,b}\g (x)=\langle \g (x+a)-\g (x),\beta\rangle -\langle \g (x+b)-\g (x),\alpha\rangle.$$ Clearly, $\Delta=\Delta_{a,b}$ is a linear operator in $\g $, and $$\label{elinftybound}
|\Delta \g (x)|\leq 4\,\|\g \|_{L^\infty}$$ Moreover, $\g $ belongs to the class $R_0$ if and only if $|\Delta \g |\leq C\,h$ for some constant $C$ that does not depend on $a$ or $b$. Using that $D\g=S\g+A\g$, we can represent $\Delta \g $ as follows, $$\aligned
\Delta \g (x)
&=\int_0^h \frac{d}{ds}\bigg(\langle \g (x+\alpha s), \beta\rangle-\langle \g (x+\beta s),\alpha\rangle\bigg)\,ds\\
&=\int_0^h \langle D\g (x+\alpha s)\,\alpha, \beta\rangle-\langle D\g (x+\beta s)\,\beta,\alpha\rangle \,ds=\Delta_S\g(x)+\Delta_A\g(x)\endaligned$$ with $$\aligned
\Delta_S\g (x)&=\int_0^h \langle S\g (x+\alpha s)\,\alpha, \beta\rangle-\langle S\g (x+\beta s)\,\beta, \alpha\rangle\,ds\\
\Delta_A\g (x)&=\int_0^h\langle A\g (x+\alpha s)\,\alpha,\beta\rangle-\langle A\g (x+\beta s)\,\beta,\alpha\rangle \,ds
\endaligned$$ By construction, $\u(x,y)=P_y\ast \b (x)$ is harmonic in $\R^{n+1}_+$ and continuous up to the boundary, since $\b \in C_c(\R^n)$. For each $t>0$, $$\aligned
\b (x)=\u (x,0)
&=\int_0^yt\,\partial^2_{yy}\u (x,t)\,dt-y\,\partial_y\u (x,y)+\u (x,y)\\
&\equiv\int_0^y t\,\w _t(x)\,dt-y\,\v _y(x)+\u _y(x)
\endaligned$$ where we wrote $\u _y(x)=\u (x,y)$, $\v _y(x)= \partial_y\u (x,y)$ and $\w _r(x)= \partial_{yy}^2\u (x,r)$. By the linearity of $\Delta$, which acts only on the $x$ variable, one has $$\Delta \b (x)=\int_0^y t\, \Delta \w _t(x)\,dt-y\,\Delta \v _y(x)+\Delta \u _y(x).$$ We now proceed term by term. First, from Lemma \[curldiv\] we know that $A\b \in L^\infty$ implies $D\b \in BMO$, which in turn gives $\b \in Z$. Hence, from Lemma \[poissonZ\], $$\left|\int_0^y t\, \Delta \w _t(x)\,dt\right|\leq \int_0^y t\,4\|\w _t\|_{L^\infty}\,dt\leq \int_0^y t\,4\,\frac{C(n)\,\|\b \|_Z}{t}\,dt= C(n)\,y\,\|\b \|_Z.$$ For the second term, we use that $\Delta =\Delta_S+\Delta_A$, $$\aligned
y\,\Delta \v _y(x)&=y\,\Delta_S \v _y(x)+y\,\Delta_A\v _y(x)\\
\Delta \u _y(x)&= \Delta_S \u _y(x)+ \Delta_A\u _y(x)\\
\endaligned$$ and proceed first with the $\Delta_A$ terms. For each fixed $y$, Lemma \[easyLinfty\] gives us that $$\aligned
\partial_{x_i}\,\u _y= \partial_{x_i} \,(P_y\ast \b )=P_y\ast (\partial_{x_i}\b )
&\Longrightarrow A\u _y=P_y\ast A\b \\
&\Longrightarrow \|A\u _y\|_{L^\infty}=\|P_y\ast A\b \|_{L^\infty}\leq \|P_y\|_1\,\|A\b \|_{L^\infty}=\|A\b \|_{L^\infty}\endaligned$$ On the other hand, since $\u $ is smooth, we can argue similarly to get that $$\aligned
\partial_{x_i} \v _y = \partial^2_{y, x_i} \u = \partial_y\,\left(P_y\ast \partial_{x_i} \b \right)
&\Longrightarrow A\v _y = \partial_y(P_y\ast A\b )\\
&\Longrightarrow \|A\v _y\|_{L^\infty} \leq C(n)\,\frac{\|A\b \|_{L^\infty}}{y}.
\endaligned$$ Thus $$\aligned
|\Delta_A \u _y(x)|&\leq 2h\,\|A\u _y\|_{L^\infty} \leq 2h\,\|A\b \|_{L^\infty}\\
|y\,\Delta_A\v _y(x)|&\leq 2h y \|A\v _y\|_{L^\infty}\leq C(n)\,h\,\|A\b \|_{L^\infty}
\endaligned$$ for some dimensional constant $C(n)$. Now is time to proceed with the $\Delta_S$ terms. For any function $\g$, set $$(S\g)_{\alpha,\beta}(x)=\langle S\g(x)\cdot\alpha, \beta\rangle.$$ Using that $S\g$ is a symmetric matrix, and calling $\gamma=\frac{\alpha-\beta}{|\alpha-\beta|}$, $$\label{Scommut}
\aligned
\langle S\g(x+\alpha s)\,\alpha, \beta\rangle-\langle S\g(x+\beta s)\,\beta, \alpha\rangle
&=\langle\alpha, (S\g(x+\alpha s)-S\g(x+\beta s))\,\beta\rangle\\
&=\langle\alpha,\left(\int_0^{s|\alpha-\beta|}\frac{d}{d\sigma} (S\g(x+\beta s+\sigma\gamma))\,d\sigma\right)\beta\rangle\\
&= \int_0^{s|\alpha-\beta|}\frac{d}{d\sigma} \bigg(\langle\alpha,S\g(x+\beta s+\sigma\gamma)\,\beta\rangle\bigg) d\sigma\\
\endaligned$$ Therefore $$\aligned
|\Delta_S \g(x)|
&\leq \int_0^h \int_0^{s|\alpha-\beta|}\left|\frac{d}{d\sigma} (S\g)_{\alpha,\beta}(x+\beta s+\sigma\gamma) \right|\,d\sigma\, ds\\
&\leq \int_0^h \int_0^{s|\alpha-\beta|} | D((S\g)_{\alpha,\beta})(x+\beta s+\sigma\gamma) |\,d\sigma\, ds\\
&\leq \| D((S\g)_{\alpha,\beta})\|_{L^\infty}\,\int_0^h \int_0^{s|\alpha-\beta|} d\sigma\, ds= \| D((S\g)_{\alpha,\beta})\|_{L^\infty}\,\frac{h^2\,|\alpha-\beta| }2
\endaligned$$ After applying this to $\g=\u _y$ and to $\g=\v _y$, one obtains $$\label{finalstep}
|\Delta \b (x)|\leq C(n)\,y\,\|\b \|_Z+C(n)\,h\,\|A\b \|_{L^\infty}+\frac{h^2\,|\alpha-\beta| }2\left(\| D((S\u _y)_{\alpha,\beta})\|_{L^\infty}+y\| D((S\v _y)_{\alpha,\beta}) \|_{L^\infty}\right)$$ Next, we see that $$\aligned
\u _y=P_y\ast \b \hspace{1cm}
&\Rightarrow\hspace{1cm}D\u _y= P_y\ast D\b \\
&\Rightarrow\hspace{1cm} S\u _y=P_y\ast S\b \\
&\Rightarrow\hspace{1cm}(S\u _y)_{\alpha,\beta}=P_y\ast (S\b )_{\alpha, \beta}.\endaligned$$ Now, since $D\b \in BMO$ we have in particular that $(S\b )_{\alpha,\beta}\in BMO$, in particular $(S\u _y)_{\alpha,\beta}$ is harmonic Bloch. Lemma \[poissonBMO\] with $\g= (S\u _y)_{\alpha,\beta}$ gives us that $$\label{sb1}
\|D((S\u _y)_{\alpha,\beta})\|_{L^\infty}=\|D(P_y\ast (S\b )_{\alpha,\beta})\|_{L^\infty}\leq \frac{C(n)\,\|(S\b )_{\alpha,\beta}\|_\ast}{y}\leq \frac{C(n)\,\|S\b \|_\ast}{y}.$$ For $\g=(S\v _y)_{\alpha,\beta}$, we proceed similarly and note that $$\aligned
\v _y=\partial_y\u _y=\partial_yP_y\ast \b \hspace{1cm}
&\Rightarrow\hspace{1cm}D\v _y= \partial_yP_y\ast D\b \\
&\Rightarrow\hspace{1cm} S\v _y=\partial_yP_y\ast S\b \\
&\Rightarrow\hspace{1cm}(S\v _y)_{\alpha,\beta}=\partial_yP_y\ast (S\b )_{\alpha, \beta}.\endaligned$$ Therefore one can combine Lemma \[poissonBMO\] and Lemma \[higherorder\] and obtain $$\label{sb2}
\|D((S\v _y)_{\alpha,\beta})\|_{L^\infty}=\|D(\partial_yP_y\ast(S\b )_{\alpha,\beta})\|_{L^\infty}\leq\frac{C(n)\,\|(S\b )_{\alpha,\beta}\|_\ast}{y^2}\leq \frac{C(n)\|S\b \|_\ast}{y^2}$$ It is worth mentioning here that both in and one could replace the constant $\|S\b \|_\ast$ by $\|\b\|_Z$ (note that $\|\b\|_Z\leq C\,\|Sb\|_\ast$). To do this, one only needs to use Lemma \[poissonZ\] instead of Lemma \[poissonBMO\]. We now plug the above bounds for $\|D((S\u _y)_{\alpha,\beta})\|_{L^\infty}$ and $\|D((S\v _y)_{\alpha,\beta})\|_{L^\infty}$ into , and then take $h=y$. This finishes the proof in the case $\b \in C_c(\R^n)$.\
In order to remove the assumption on the compact support, we use once more the auxilliary function $g=g_t$ introduced at . We have $$D(g\b )-D^t(g\b )=\b \otimes \nabla g-\nabla g\otimes \b + g\,(D\b -D^t\b )$$ so $$\aligned
\|D(g\b )-D^t(g\b )\|_{L^\infty}
&\leq \|D\b -D^t\b \|_{L^\infty} + \sup_{t\leq|x|\leq t^{e^t}}|\b (x)||\nabla g(x)|\\
&\leq \|D\b -D^t\b \|_{L^\infty} + \sup_{t\leq|x|\leq t^{e^t}}C\,|x|\log|x|\frac{1}{t|x|\,\log|x|}\\
&\leq \|D\b -D^t\b \|_{L^\infty} + \frac{C}t\\
\endaligned$$ Now the claim follows since for every $x\in\R^n$ we can pick $t>0$ large enough and such that $|x|< t^{e^t}$ so that $\b =g\b $ in a neighbourhood of $x$, and therefore $D\b -D^t\b =D(g\b )-D^t(g\b )$.
[bib]{}
, [*Exponential self-similar mixing and loss of regularity for continuity equations*]{}, C. R. Acad. Sci. Paris, Ser. I Vol. 352, no. 11 (2014), pp. 901-906. , [*[ Area distortion of quasiconformal mappings]{}*]{}, Acta Math. 173, 37 (1994). , [*Elliptic partial differential equations and quasiconformal mappings in the plane*]{}, Princeton Mathematical Series, 48. Princeton University Press, Princeton, NJ, 2009. , [*Equations de transport relatives à des champs de vecteurs non-lipschitziens et mécanique des fluides*]{}, Arch. Rat. Mech. Anal 127 (1994) 159-181. , [*[Sobolev estimates for solutions of the transport equation and ODE flows associated to non-Lipschitz drifts]{}*]{}, Math. Ann. (2020), https://doi.org/10.1007/s00208-020-01988-5 , [*[A note on transport equations in quasiconformally invariant spaces]{}*]{}, Adv. Calc. Var., Vol. 11 (2018), no. 2, pp 193–202. , [*[Sobolev regular flows of non-Lipschitz vector fields]{}*]{}, J. Differential Equations 266 (2019), no. 8, 4544–4567. , [*[Partial differential equations]{}*]{}, vol. 19, Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1998. , [*[Spaces of harmonic functions representable by Poisson integrals of functions in $BMO$ and ${\mathcal{L}}_{p,\lambda}$]{}*]{}, Indiana Univ. Math. J. 25 (1976), no. 2, 159–170. , [*[ Quasiharmonic fields]{}*]{}, Ann. Inst. H. Poincaré Anal. Non Linéaire 18 (2001), no. 5, 519–572. , [*[ Geometric function theory and non-linear analysis]{}*]{}, Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, New York, 2001. , [*[Critical non-Sobolev regularity for continuity equations with rough velocity fields]{}*]{}, J. Differential Equations 260 (2016), no. 5, 4739–4757. , [*[Sobolev Spaces]{}*]{}, Department of Mathematics and Systems Analysis, Aalto University (FI), 2017. , [*[ Ordinary differential equations and quasiconformal mappings]{}*]{}, Invent. Math. 33 (1976), 247-270. , [*[Funktionen beschränkter mittlerer Oszillation]{}*]{} Lecture Notes in Mathematics, Vol. 487. Springer-Verlag, Berlin-New York, 1975. , [*[Singular integrals and differentiability properties of functions]{}*]{}, Princeton University Press. 1970. , [*[Non stationary flow of an ideal and incompressible liquid]{}*]{}, Zh. Vych. Math, 3, 1963, pp. 1032-1066.
Albert Clop\
Department of Mathematics and Computer Science\
Universitat de Barcelona\
08007-Barcelona\
CATALONIA\
[email protected]\
\
\
Banhirup Sengupta\
Departament de Matemàtiques\
Unirersitat Autònoma de Barcelona\
08193-Bellaterra\
CATALONIA\
[email protected]\
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
**Purpose:** Different multichannel methods for film dosimetry have been proposed in the literature. Two of them are the weighted mean method and the method put forth by Micke *et al* and Mayer *et al*. The purpose of this work was to compare their results and to develop a generalized channel-independent perturbations framework in which both methods enter as special cases.
**Methods:** Four models of channel-independent perturbations were compared: weighted mean, Micke-Mayer method, uniform distribution and truncated normal distribution. A closed-form formula to calculate film doses and the associated Type B uncertainty for all four models was deduced.
To evaluate the models, film dose distributions were compared with planned and measured dose distributions. At the same time, several elements of the dosimetry process were compared: film type EBT2 versus EBT3, different waiting-time windows, reflection mode versus transmission mode scanning, and planned versus measured dose distribution for film calibration and for $\gamma$-index analysis.
The methods and the models described in this study are publicly accessible through IRISEU. Alpha 1.1 (http://www.iriseu.com). IRISEU. is a cloud computing web application for calibration and dosimetry of radiochromic films.
**Results:** The truncated normal distribution model provided the best agreement between film and reference doses, both for calibration and $\gamma$-index verification, and proved itself superior to both the weighted mean model, which neglects correlations between the channels, and the Micke-Mayer model, whose accuracy depends on the properties of the sensitometric curves.
With respect to the selection of dosimetry protocol, no significant differences were found between transmission and reflection mode scanning, between $75 \pm 5$ min and $20 \pm 1$ h waiting-time windows or between employing EBT2 or EBT3 films. Significantly better results were obtained when a measured dose distribution was used instead of a planned one as reference for the calibration, and when a planned dose distribution was used instead of a measured one as evaluation for the $\gamma$-analysis.
**Conclusions:** The truncated normal distribution model of channel-independent perturbations was found superior to the other three models under comparison and we propose its use for multichannel dosimetry.
author:
- 'I. M[é]{}ndez'
- 'P. Peterlin'
- 'R. Hudej'
- 'A. Strojnik'
- 'B. Casar'
title: 'On multichannel film dosimetry with channel-independent perturbations'
---
Introduction
============
Radiochromic film dosimetry with flatbed scanners and Gafchromic films (Ashland Inc., Wayne, NJ) has been extensively studied in the literature[@devic:2005; @Fuss:2007; @Paelinck:2007; @battum:2008; @Martisikova:2008; @Richley:2010; @Devic:2011]. High spatial resolution, near water equivalence[@crijns:2102; @aapm:55] and weak energy dependence[@rink:2007; @richter:2009; @arjomandy:2010; @lindsay:2010; @massillon] make radiochromic films convenient for measurements whenever sharp dose gradients, tissue heterogeneities or charged particle disequilibrium conditions exist. This opens up a wide range of applications for radiochromic films in the field of radiotherapy.
Recently, different multichannel dosimetry methods have been proposed to take into account the information conveyed by all three color channels delivered by the scanner. Micke *et al*[@AMicke:2011] proposed the use of channel-independent perturbations to compensate for variations in the thickness of the active layer, artifacts, nonuniform response of the scanner or other disturbances. They found a substantial gain in dosimetric accuracy using this method. Van Hoof *et al*[@vanHoof:2012] found that this method performs at least as well as the conventional single-red-channel dosimetry. Mayer *et al*[@mayer:2012] derived a closed-form solution to obtain the dose employing channel-independent perturbations. They also compared different single, dual and triple channel methods, and found better agreement between planned and calculated dose distributions using the average dose of all three channels in comparison to using the channel-independent perturbations method. In an earlier article[@mendez:2013], our group suggested calculating the film dose as the weighted mean dose of all three channels. For each channel, the inverse of the mean square error obtained during the film calibration was used as weight. With this method, triple-channel dosimetry was found to be substantially superior to single-red-channel dosimetry.
The purpose of this work is to compare both weighted mean and Micke-Mayer methods, considering them as special cases of a more general channel-independent perturbations method. Deficiencies and important problems associated with both methods will be explained. To overcome these problems, an improved multichannel film dosimetry method will be introduced. Its performance against the other methods will be verified by comparing film dose distributions with planned as well as with measured dose distributions. In addition, other elements of the dosimetry process will be compared: film types[@reinhardt:2012], scanning modes[@park:2012], scanning waiting-time windows[@devic:2010] and choices of reference dose distribution.
Methods and materials
=====================
Channel-independent perturbations
---------------------------------
Channel-independent perturbations are obtained by applying a first order Taylor expansion to the dose due to a small perturbation:
$$\label{eqsystem}
\begin{cases}
D(r) = D_{R}(r) + \dot{D}_{R}(r) \Delta(r) + \epsilon_{R}(r)\\
D(r) = D_{G}(r) + \dot{D}_{G}(r) \Delta(r) + \epsilon_{G}(r)\\
D(r) = D_{B}(r) + \dot{D}_{B}(r) \Delta(r) + \epsilon_{B}(r)
\end{cases},$$
Micke *et al*[@AMicke:2011] derived the dose from the optical density (OD) of the irradiated film. Mayer *et al*[@mayer:2012] used pixel values directly. In this study, better results were found in preliminary tests using net optical density[@devic:2005] (NOD) in comparison to using OD. Therefore, the channel-independent perturbation consists of a change in NOD and is represented by $\Delta(r)$. $D(r)$ represents the dose absorbed by the film at point $r$. $D_{k}$ is the absolute dose measured by the channel $k$, [*i.e.*]{}, red (R), green (G) or blue (B) channel, when no disturbance is present, and it is calculated directly from the calibration model. In this study, the calibration model includes the lateral correction[@Paelinck:2007; @fiandra:2006; @battum:2008; @devic:2006; @lynch:2006; @Martisikova:2008; @Fuss:2007; @menegotti:2008; @saur:2008] and the sensitometric curve. $\dot{D}_{k}(r)$ is the first derivative of the dose, with respect to the NOD, at point $r$. Finally, $\epsilon_{k}(r)$ is an error term accounting for the difference between the dose absorbed by the film and the dose measured in the channel $k$ after correction by the perturbation.
Both for reflection[@ohuchi:2007] and transmission mode scanning, the NOD, denoted by $z$, was defined as
$$z := \log_{10} \frac{v_{\mathrm{nonirr}}}{v_{\mathrm{irr}}},$$
where $v_{\rm{nonirr}}$ and $v_{\rm{irr}}$ represent pixel values of nonirradiated and irradiated films, respectively, after applying lateral corrections. Our previous results[@mendez:2013] found better fit when lateral corrections are absolute corrections independent of dose, and sensitometric curves are polynomial fits of order four. Hence, lateral corrections were calculated as
$$\label{lateral}
v_{k} = a_{k_1} (x-x_{c}) + a_{k_2} (x-x_{c})^{2} + \hat{v_{k}},$$
where $\hat{v_{k}}$ represents uncorrected pixel values, the $x$ axis is parallel to the CCD array, $x_c$ is the [*x*]{} coordinate of the center of the scanner, $v_{k}$ represents corrected pixel values, and $a_{k}$ are fitting parameters. Sensitometric curves followed
$$\label{sensitometric}
D_{k} = \sum_{j=1}^{4} b_{k_j} z_{k}^j ,$$
and $\dot{D}_{k}$ was
$$\dot{D}_{k} = \sum_{j=1}^{4} j b_{k_j} z_{k}^{j-1} ,$$
where $b_{k}$ are fitting parameters.
Solving the equation system
---------------------------
The values of $\Delta(r)$ and $\epsilon_{k}(r)$ for $k=R,G,B$ in Eq.(\[eqsystem\]) are unknown. As a result, the absorbed dose $D(r)$ cannot be obtained directly. However, one can examine different probability density functions (pdf) for $\Delta(r)$ and $\epsilon_{k}(r)$ and, if $D(r)$ is known, analyze how well these models reproduce the absorbed dose distribution.
### Probability density function of the dose
Given the pdfs of $\Delta$, symbolized by $f(\Delta)$, and of each $\epsilon_{k}$, symbolized by $g_{k}(\epsilon_{k})$, the joint pdf of $D$, symbolized by $P(D)$, is:
$$P(D) = \int f(\Delta) \prod_{k} g_{k}(D - D_{k} - \dot{D}_{k} \Delta) \:\mathrm{d}\Delta ,$$
taking into account that $\Delta$ and $\epsilon_{k}$ are not independent from each other: $$\epsilon_{k} = D - D_{k} - \dot{D}_{k} \Delta .$$
Let us consider that the error terms are distributed normally with zero mean and $\sigma_{k}^{2}$ variance: $$g_{k}(\epsilon_{k}) = \mathcal{N}(0,\sigma_{k}^{2})$$
The joint pdf of $D$ becomes: $$P(D) = \int f(\Delta) \prod_{k} \frac{1}{\sigma_{k} \sqrt{2\pi}} \:\mathrm{e}^{-\frac{1}{2} \left(\frac{D - D_{k} - \dot{D}_{k}\Delta}{\sigma_{k}}\right)^{2}} \:\mathrm{d}\Delta.$$
Three different models for $f(\Delta)$ will be considered:
a\) Normally distributed perturbation ([*i.e.*]{}, $f(\Delta) = \mathcal{N}(0,\sigma_{\Delta}^{2})$):
$$P(D) = \frac{1}{(2\pi)^{\frac{n}{2}} \sigma_{\Delta} \prod_{k=1}^{n} \sigma_{k}} \: \frac{1}{\sqrt{A}} \: \:\mathrm{e}^{-\frac{1}{2} \left(C - \frac{B^2}{4A}\right)},$$
where $n$ represents the number of color channels ([*i.e.*]{}, $n = 3$) and
$$\label{Adef}
A= \frac{1}{\sigma_{\Delta}^{2}} + \sum_{k} \left( \frac{\dot{D}_{k}}{\sigma_{k}} \right)^{2}$$
$$B=-2\sum_{k} \frac{(D - D_{k})\dot{D}_{k}}{\sigma_{k}^{2}}$$
$$C=\sum_{k} \left( \frac{D - D_{k}}{\sigma_{k}} \right)^{2}.$$
b\) Truncated normal distribution with $\Delta \in (-\theta, \theta)$:
$$\label{truncated}
P(D)\propto \:\mathrm{e}^{-\frac{1}{2} \left(C - \frac{B^2}{4A}\right)} \: \left( \mathrm{erf} \left(\frac{\theta + \frac{B}{2A}}{\sqrt{\frac{2}{A}}}\right) - \mathrm{erf} \left(\frac{-\theta + \frac{B}{2A}}{\sqrt{\frac{2}{A}}}\right)\right) ,$$
excluding a normalizing term independent of $D$.
c\) Uniform distribution with $\Delta \in (-\theta, \theta)$: is a special case of Eq.(\[truncated\]) where $\sigma_{\Delta}$ goes to infinity, therefore $A= \sum_{k} \left( \frac{\dot{D}_{k}}{\sigma_{k}} \right)^{2}$.
### Dose calculation
The most likely value of the absorbed dose $D$, symbolized by $d$, is the one that maximizes $P(D)$. The exponential term in $P(D)$, $P(D) \propto \:\mathrm{e}^{-\frac{1}{2} \left(C - \frac{B^2}{4A}\right)}$, can be expressed in terms of $D$ as a gaussian function:
$$P(D) \propto \:\mathrm{e}^{-\frac{1}{2} \left(\frac{D - \mu_{D}}{\sigma_{D}}\right)^{2}},$$
where
$$\label{Dose}
\mu_{D} = d = \frac{A\beta - \gamma\delta}{A\alpha-\gamma^{2}}$$
and
$$\label{uncertainty}
\sigma_{D} = \sqrt{\frac{A}{A\alpha-\gamma^{2}}},$$
A is defined in Eq.(\[Adef\]) and
$$\alpha=\sum_{k} \frac{1}{\sigma_{k}^{2}}$$
$$\beta=\sum_{k} \frac{D_{k}}{\sigma_{k}^{2}}$$
$$\gamma=\sum_{k} \frac{\dot{D}_{k}}{\sigma_{k}^{2}}$$
$$\delta=\sum_{k} \frac{D_{k}\dot{D}_{k}}{\sigma_{k}^{2}} .$$
Eq.(\[Dose\]) and Eq.(\[uncertainty\]) can be considered, respectively, as the estimated absolute dose and its type B uncertainty [@JCGM:2008]. This result is exact for normally distributed perturbations and an approximation for truncated normal and uniform distributions.
Models of channel-independent perturbations under comparison
------------------------------------------------------------
Four models of channel-independent perturbations were compared: weighted mean (WM), Micke-Mayer (MM) method, uniform distribution (UD) and truncated normal distribution (TN). They are summarized in Table \[tab:models\].
The weighted mean method is a limit case of Eq.(\[eqsystem\]) in which $\Delta(r) = 0$. Thus, all three channels are independent of each other, which implies that correlations between channels are neglected.
The method employed by Micke *et al*[@AMicke:2011] and Mayer *et al*[@mayer:2012] is a special case of Eq.(\[eqsystem\]) where all $\sigma_{k}$ are equal and $f(\Delta)$ is uniformly distributed. Under these premises, Eq.(\[Dose\]) becomes:
$$d = \frac{A\beta - \gamma\delta}{A\alpha-\gamma^{2}} = \frac{(\sum_{k=1}^{n} \dot{D}_{k}) (\sum_{k=1}^{n} D_{k} \dot{D}_{k}) - (\sum_{k=1}^{n} \dot{D}_{k}^{2}) (\sum_{k=1}^{n} D_{k})}{(\sum_{k=1}^{n} \dot{D}_{k})^{2} - n (\sum_{k=1}^{n} \dot{D}_{k}^{2})},$$
which coincides with the closed-form solution derived by Mayer *et al* [@mayer:2012]. The uncertainty in the dose associated to this model becomes:
$$\label{MMuncertainty}
\sigma_{D} = \sqrt{\frac{A}{A\alpha-\gamma^{2}}} = \sigma_{k}
\sqrt{\frac{\sum_{k=1}^{n} \dot{D}^{2}_{k}}
{n \sum_{k=1}^{n} \dot{D}^{2}_{k} - (\sum_{k=1}^{n} \dot{D}_{k})^2}}$$
The uniform distribution model is a more general and realistic model for the perturbation than the MM one. In this case, $f(\Delta)$ is uniformly distributed but the $\sigma_{k}$ can differ.
Finally, the truncated normal distribution model considers that $f(\Delta)$ follows a truncated normal distribution. The WM model is a limit case and the UD and MM models are particular cases of this model.
Model Abbreviation Assumptions
------------------------------- -------------- ----------------------------------------------------------
Weighted mean WM $\Delta(r) = 0$
Micke-Mayer method MM $f(\Delta)$ uniform distribution, $\sigma_{k}$ are equal
Uniform distribution UD $f(\Delta)$ uniform distribution
Truncated normal distribution TD $f(\Delta)$ truncated normal distribution
Measurement protocol
--------------------
Ten 8 ${\rm inch}$ $\times$ 10 ${\rm inch}$ EBT2 films from lot A03171101A and seventeen EBT3 films from lot A05151201 were employed. They were handled following recommendations from the AAPM TG-55 report [@aapm:55].
Films were scanned with an Epson Expression 10000XL flatbed scanner (Seiko Epson Corporation, Nagano, Japan) using Epson Scan v.3.0 software. Images were acquired in 48-bit RGB mode (16 bit per channel), the resolution was 72 dpi (0.35 mm/px) and the image processing tools were turned off.
Before acquisitions, the scanner was warmed up for at least 30 min. After the warm-up, and whenever long interruptions occurred, five empty scans were taken to stabilize the temperature of the scanner lamp. Films were centered on the scanner with a black opaque cardboard frame and scanned in portrait orientation. Five consecutive scans were made for each film. To avoid the warm-up effect of the lamp due to multiple scans[@Paelinck:2007; @Martisikova:2008] the first scan was discarded and the resulting image was the average of the remaining four.
Films were scanned before irradiation both in reflection and in transmission mode. After irradiation, two waiting-time windows were studied: films were first scanned after $75 \pm 5$ min in transmission mode, and again after $20 \pm 1$ h both in reflection and transmission mode.
Irradiation was delivered with a 6 MV photon beam from a Novalis Tx accelerator (Varian, Palo Alto, CA, USA). Three different phantoms were used: CIRS Thorax Phantom (Model 002LFC, Computerized Imaging Reference Systems Inc. Norfolk, VA, USA), CIRS Pelvic Phantom (Model 002PRA) and IBA MatriXX Evolution MULTICube (IBA Dosimetry GmbH, Germany). Source-axis distance (SAD) setup was used for all three phantoms. To avoid the films lying in the beam axis plane[@Kunzler:2009], the films were placed at an offset of 1.5 cm from the beam axis in the CIRS Thorax Phantom and of 1.3 cm in the CIRS Pelvic Phantom. The IBA MatriXX Evolution MULTICube was used jointly with the IBA MatriXX Evolution ionization chamber array, which measured the dose distribution delivered. The film was situated atop the detector.
The absolute dose distributions in the plane of the film were calculated with Eclipse v.10.0 (Varian Medical Systems, Palo Alto, CA, USA) treatment planning system (TPS) using the anisotropic analytical algorithm (AAA). The planned dose distributions were exported to dose matrices with a resolution of 0.49 mm/px. The dose values were scaled to correct for the daily output of the linac. Whenever MatriXX Evolution was used, the dose distribution was simultaneously measured. The dose values were scaled with a constant factor to correct for the distance (which was 3.5 mm) between the film and the plane at the effective depth of measurement. The MatriXX 2D array has a resolution of 7.62 mm/px. Planned and measured dose distributions were bicubically interpolated to the resolution of the scan and registered with the film.
Film scans, planned dose ditributions and measured dose distributions were uploaded and processed with IRISEU. Alpha 1.1 (http://www.iriseu.com). IRISEU. is a cloud computing web application for calibration and dosimetry of radiochromic films. It is developed by one of the authors (IM) and incorporates the methods and models described in this study. It was employed for the calibration, dosimetry and gamma index evaluation. Additional statistical analysis was performed with R statistical software[@R:software].
Calibration
-----------
To fit the calibration parameters, the plan-based method[@mendez:2013] was chosen. Besides being faster than the calibration method with fragments, the plan-based method provides a more representative sample of perturbations (since it can use every pixel of the film). This method requires one or more 2D dose distributions as reference doses for the calibration. In order to obtain them, films were placed in the MatriXX Evolution phantom and irradiated with a $60^{\circ}$ Enhanced Dynamic Wedge (EDW) field of dimensions 20$\times$20 ${\rm cm^2}$. To reduce intralot variations[@mendez:2013], three separate films from each lot were exposed. The range of doses relevant for this study and for posterior clinical use was estimated between 50 cGy and 400 cGy. To encompass the whole range, two different fields were used: the wedge dose spanned from approximately 75 cGy to approximately 400 cGy (535 MU) for two of the films from each lot and from approximately 50 cGy to approximately 300 cGy (401 MU) for the remaining one.
Following this procedure, one set with EBT2 and another with EBT3 films were irradiated. Posteriorly, the films were scanned following the three protocols previously mentioned: reflection mode with $20 \pm 1$ h time window, transmission mode with $20 \pm 1$ h time window and transmission mode with $75 \pm 5$ min time window. Each set of images (six sets in total) was employed to calibrate each of the four models of channel-independent perturbations. Each of the models was calibrated against planned dose distributions (calculated with the TPS) and against measured dose distributions (simultaneously measured with MatriXX during the irradiations). Altogether, a total of 48 calibrations were computed.
Pixel values of the films exposed were translated into doses, for each color channel independently, fitting the calibration parameters. A genetic algorithm was used to fit the parameters minimizing the root-mean-square error (RMSE) of the differences between film doses for each channel ($D_{k}(r)$) and reference doses ($D(r)$).
This optimization provides the parameters used in Eq.(\[lateral\]) and Eq.(\[sensitometric\]). This is enough for film dosimetry following WM or MM models. However, to obtain $d(r)$ using UD or TN models $\sigma_{k}$ are necessary, and also $\sigma_{\Delta}$ if using the TN model. Knowing lateral correction, sensitometric curve parameters and the standard deviation of $f(\Delta)$, which depends on $\sigma_{\Delta}$ and $\theta$ and will be symbolized by $\tilde{\sigma}_{\Delta}$, $\sigma_{k}$ can be estimated with
$$\label{eq_cal}
\tilde{\sigma}_{k}^{2} \simeq (E[\dot{D}_{k}] \: \tilde{\sigma}_{\Delta})^{2} + \sigma_{k}^{2}$$
where $\tilde{\sigma}_{k}$ is the RMSE of the channel and $E[\dot{D}_{k}]$ is the expected value ([*i.e.*]{}, mean) of $\dot{D}_{k}$.
The values of $\tilde{\sigma}_{\Delta}$ for UD and TN models, and of $\sigma_{\Delta}$ for the TN model, were obtained optimizing the RMSE of the differences between film doses ($d(r)$) and reference doses ($D(r)$).
Verification
------------
To evaluate the four models of channel-independent perturbations, film dose distributions were compared with planned and with measured dose distributions. Global gamma analysis was conducted. The tolerances were 4 %, 3 mm with 20% of the maximum dose as threshold. Fourteen different cases were tested (Table \[tab:Tests\]). The cases were chosen with the intention of compiling a representative sample of dose distributions: several simple geometries, tissue heterogeneities, three-dimensional conformal radiotherapy (3D-CRT) plans and intensity modulated radiation therapy (IMRT) plans, including volumetric modulated arc therapy (VMAT) plans, were selected. EBT3 films were irradiated with all the cases but only a subset (considered representative) of them was used with EBT2 films, as shown in Table \[tab:Tests\]. Appropriate phantoms were employed dependent on the test case.
At the same time, several elements of the dosimetry process were compared: film type EBT2 versus EBT3, different waiting-time windows ([*i.e.*]{}, $75 \pm 5$ min versus $20 \pm 1$ h), reflection versus transmission mode scanning and planned versus measured reference dose distribution for film calibration and for gamma index analysis.
As a result, seven EBT2 and fourteen EBT3 films were irradiated with the cases shown in Table \[tab:Tests\]. They were scanned following the three scanning protocols under study. Each image was translated into a dose distribution following each of the four models of channel-independent perturbations. The film dose distributions were compared with the planned dose distributions in the plane of the film. Whenever the test was irradiated in the MatriXX phantom, the film dose distributions were also compared with the measured dose distributions. When film dose distributions were compared with planned ones, the calibration parameters of the model had been fitted using planned reference dose distributions, and analogously with measured dose distributions. If both planned and measured reference dose distributions are accurate, they should provide similar sets of calibration parameters. Following this, and for the TN model only, film distributions obtained with calibration parameters fitted using measured reference dose distribution were also compared with planned dose distributions, and vice versa ([*i.e.*]{}, film distributions obtained with calibration parameters fitted using planned reference dose distribution were compared with dose distributions measured with MatriXX).
Test Description Film type Phantom
------ ------------------------------------------------------ ----------- -------------
A Square 15$\times$15 ${\rm cm^2}$ EBT2,EBT3 MatriXX
B Chair test [@vanesch:2002] EBT2,EBT3 MatriXX
C Pyramid shaped in both axis[@ju:2002] EBT2,EBT3 MatriXX
D EDW $30^{\circ}$ field EBT3 MatriXX
E EDW $45^{\circ}$ collimator 90 field EBT3 MatriXX
F Y-shaped 3D CRT field EBT2,EBT3 MatriXX
G Predominantly convex shaped 3D CRT field EBT3 MatriXX
H RapidArc prostate 1 EBT3 CIRS Pelvic
I RapidArc prostate 2 EBT3 CIRS Pelvic
J RapidArc prostate 3 EBT2,EBT3 CIRS Pelvic
K Square 10$\times$10 ${\rm cm^2}$, lung inhomogeneity EBT3 CIRS Thorax
L Lateral incidence, lung inhomogeneity EBT2,EBT3 CIRS Thorax
M Four field box, lung inhomogeneity EBT3 CIRS Thorax
N EDW and asymmetric fields, lung inhomogeneity EBT2,EBT3 CIRS Thorax
Results and discussion
======================
Twenty-four different dosimetry protocols were analyzed in this study. To represent each protocol in a clear and concise way, they will be named using four characters (Table \[tab:symbols\]). The characters stand for: gamma analysis with either planned (P) or measured (M) evaluation dose distributions, scanning in reflection mode with $20 \pm 1$ h time window (R), in transmission mode with $20 \pm 1$ h time window (T) or in transmission mode with $75 \pm 5$ min time window (t), film type either EBT2 (2) or EBT3 (3) and calibration with either planned (p) or measured (m) reference dose distributions.
Element of the protocol Alternative Abbreviation
----------------------------------------------------- ------------------------------ --------------
Evaluation dose distribution for the gamma analysis Planned P
Measured M
Scanning mode and time window Reflection, $20 \pm 1$ h R
Transmission, $20 \pm 1$ h T
Transmission, $75 \pm 5$ min t
Film type EBT2 2
EBT3 3
Reference dose distribution for the calibration Planned p
Measured m
Selection of model of channel-independent perturbations
-------------------------------------------------------
---------- ------- --------------------- --------------- ------- --------------------- --------------- ------- --------------------- --------------- ------- --------------------- ---------------
RMSE $\overline{\gamma}$ $\gamma_{<1}$ RMSE $\overline{\gamma}$ $\gamma_{<1}$ RMSE $\overline{\gamma}$ $\gamma_{<1}$ RMSE $\overline{\gamma}$ $\gamma_{<1}$
Protocol (cGy) (%) (cGy) (%) (cGy) (%) (cGy) (%)
PR2p 3.5 0.18 98.4 5.4 0.28 93.9 5.2 0.27 93.6 2.8 0.19 98.0
Pt2p 3.8 0.23 96.9 3.3 0.22 96.9 3.2 0.22 97.2 2.8 0.19 98.0
PT2p 5.1 0.32 93.2 3.7 0.22 98.1 3.5 0.22 97.5 3.1 0.21 97.1
PR3p 2.9 0.18 98.1 4.9 0.26 95.1 4.8 0.25 95.6 2.7 0.15 98.7
Pt3p 6.2 0.24 96.3 4.1 0.19 98.6 4.1 0.19 98.6 4.1 0.17 99.4
PT3p 3.3 0.21 97.1 3.6 0.17 99.1 3.3 0.14 99.2 2.6 0.14 99.4
MR2m 3.6 0.31 98.1 5.1 0.39 93.1 4.9 0.39 92.6 2.9 0.31 97.3
Mt2m 4.4 0.35 94.6 3.2 0.34 95.3 3.1 0.33 96.1 2.9 0.32 96.7
MT2m 5.4 0.36 91.5 3.5 0.31 97.3 3.2 0.35 96.8 3.2 0.33 97.5
MR3m 3.1 0.28 97.7 5.0 0.44 88.0 4.8 0.29 98.1 2.9 0.29 98.1
Mt3m 6.2 0.34 94.2 4.2 0.30 96.9 4.2 0.28 97.8 4.2 0.28 97.8
MT3m 4.1 0.38 92.1 5.1 0.28 97.1 4.8 0.28 97.2 3.8 0.32 95.8
---------- ------- --------------------- --------------- ------- --------------------- --------------- ------- --------------------- --------------- ------- --------------------- ---------------
Table \[tab:Main\] compares film doses ($d(r)$) with planned or measured doses ($D(r)$), data are aggregated by model of channel-independent perturbations and dosimetry protocol. It contains RMSEs from the calibrations as well as gamma mean ($\overline{\gamma}$) and percentage of points with $\gamma_{<1}$ from the verification gamma analysis.
{width="\linewidth"}\
(a)
{width="\linewidth"}\
(b)
Considering the size of the sample and calculating likelihood from RMSE, Akaike Information Criterion (AIC) values are equivalent to calibration RMSEs. Since the TN model provided as good or better calibration RMSEs than the other models in all protocols, according to the AIC the TN model should be selected for multichannel dosimetry.
Protocol group WM - MM WM - UD WM - TN MM - UD MM - TN UD - TN
---------------- ------------------ ----------------- ----------------- ------------- ----------------- -----------------
R + T -0.03 (0.40) 0.01 (0.61) 0.03 (<0.05) 0.02 (0.26) 0.06 (<0.05) 0.04 (<0.05)
R -0.11 (<0.05) -0.03 (0.51) 0.00 (0.81) 0.04 (0.32) 0.11 (<0.05) 0.07 (0.06)
T 0.05 (<0.05) 0.06 (<0.05) 0.05 (<0.05) 0.00 (0.73) 0.01 (0.32) 0.01 (0.43)
If the calibration data is well-suited to the problem of model selection, lower calibration RMSEs result in better agreement between film doses and planned or measured doses and, consequently, lower values of $\overline{\gamma}$ and higher $\gamma_{<1}$ in gamma analysis. In Fig. \[fig:correlation\] it is shown that calibration RMSEs and $\overline{\gamma}$ or $\gamma_{<1}$ from Table \[tab:Main\] are significantly correlated. A consequence of this is shown in Table \[tab:Pair\_models\]. It compares the models of channel-independent perturbations employing paired difference test. Models are paired for each dosimetry protocol. Differences are in $\overline{\gamma}$ values from Table \[tab:Main\]. Mean differences between models are shown with dosimetry protocols grouped into protocols R and T. Protocols t were not used in this analysis in order to have both transmission mode and reflection mode scanning protocols equally weighted. The TN model provided significantly ($p < 0.05$) better results than the rest of models bringing together protocols R and T. Observing R protocols alone, the MM model was found significantly worse than WM and TN. Observing T protocols alone, the WM model was found significantly worse than the rest of models. Including in the analysis the rest of results, the TN model provided the best results both for R and T protocols, the UD model provided better results than the MM model, for R protocols the WM model provided better results than MM and UD models, however, for T protocols the WM model provided the worst results.
The WM model neglects correlations between channels. Poor performance of the WM model with T protocols can signify that these correlations ([*e.g.*]{}, due to variations in the thickness of the active layer) are important and can not be neglected. Still, this outcome does not mean that MM or UD models are preferable to WM in transmission mode scanning, it depends on the uncertainty $\sigma_{D}$ (Eq.(\[uncertainty\])), [*i.e.*]{}, on the properties of the dosimetry system under study. In fact, after analysing a sample of points from different test cases, $\sigma_{D}$ was found to be the reason why MM and UD models provided worse results than the WM model with R protocols. To illustrate the importance of $\sigma_{D}$, the protocol PR2p was calibrated with MM using only red and green color channels. In our dosimetry system the sensitometric curves of both channels were very similar and this caused the RMSE of the calibration to be 3300 cGy. From Eq.(\[MMuncertainty\]), it follows that the accuracy of the MM model depends on the properties of the sensitometric curves, and can result in unacceptable uncertainties. Another flaw of the MM model is that all $\sigma_{k}$ are considered equal. This hypothesis is usually false. As an example, it was found for protocol PR3p: $\sigma_{R} = 3.5$ cGy, $\sigma_{G} = 2.8$ cGy and $\sigma_{B} = 6.8$ cGy. The UD model corrects this deficiency, which could explain why it provided better results than the MM model. However, the accuracy of the UD model still depends on the properties of the sensitometric curves and of $\sigma_{k}$. With respect to the TN model, even though it is also submitted to Eq.(\[uncertainty\]), it can be considered as a metamodel that minimizes $\sigma_{D}$ and can derive (as a limit) into the WM model, or into an intermediate case between models WM and UD. As a conclusion, we believe the superior performance of the TN model of channel-independent perturbations makes it the best choice for multichannel dosimetry.
Once the TN model was selected, gamma analysis was conducted for Pm and Mp protocols. Values of $\overline{\gamma}$ and $\gamma_{<1}$ for this protocols are shown in Table \[tab:Main2\]. RMSEs from the calibrations are not included since they are already present in Table \[tab:Main\] ([*i.e.*]{}, the RMSE from the calibration is independent of the evaluation dose distribution used for gamma analysis).
Selection of dosimetry protocol
-------------------------------
------ --------------------- ---------------
$\overline{\gamma}$ $\gamma_{<1}$
PR2m 0.15 99.4
Pt2m 0.17 99.2
PT2m 0.20 98.5
PR3m 0.15 98.8
Pt3m 0.19 98.3
PT3m 0.18 98.4
MR2p 0.47 88.6
Mt2p 0.35 95.5
MT2p 0.43 91.2
MR3p 0.41 89.6
Mt3p 0.39 91.1
MT3p 0.33 95.2
------ --------------------- ---------------
Comparisons of elements of the dosimetry process were made employing paired difference tests for the TN model. For each point of each test case the difference in $\gamma$ values between two protocols was calculated. Between both protocols, only one element of the dosimetry process was modified. Since some test cases were not present in some protocols ([*e.g.*]{}, Test H in protocols M), the numbers of test cases differ between the comparisons. Results of the comparisons are shown in Table \[tab:Pair\_main\].
P2p P3p M2m M3m P2m P3m M2p M3p Mean ($p$-value)
-------- ------- -------- -------- ------- -------- -------- ------- -------------------------------
-0.034 0.013 -0.002 -0.046 0.032 -0.007 -0.132 0.023 -0.02 $\pm$ 0.06 ($p = 0.33$)
P2p P3p M2m M3m P2m P3m M2p M3p Mean ($p$-value)
------- ------- ------- ------- -------- -------- ------- ------- ------------------------------
0.005 0.046 0.020 0.021 -0.003 -0.019 0.071 0.168 0.04 $\pm$ 0.06 ($p = 0.11$)
PRp P(t+T)p MRm M(t+T)m PRm P(t+T)m MRp M(t+T)p Mean ($p$-value)
------- --------- -------- --------- ------- --------- -------- --------- ---------------------------------
0.038 0.061 -0.066 -0.003 0.008 0.024 -0.064 -0.001 -0.001 $\pm$ 0.050 ($p = 0.97$)
PR2 P(t+T)2 PR3 P(t+T)3 MR2 M(t+T)2 MR3 M(t+T)3 Mean ($p$-value)
------- --------- ------- --------- ------- --------- ------- --------- ------------------------------
0.084 0.027 0.049 0.002 0.232 0.102 0.179 0.085 0.10 $\pm$ 0.08 ($p < 0.05$)
R2p (t+T)2p R3p (t+T)3p R2m (t+T)2m R3m (t+T)3m Mean ($p$-value)
-------- --------- -------- --------- -------- --------- -------- --------- -------------------------------
-0.370 -0.251 -0.370 -0.239 -0.193 -0.149 -0.226 -0.151 -0.24 $\pm$ 0.09 ($p < 0.05$)
Table \[tab:Pair\_tT\] and Table \[tab:Pair\_RT\] compare transmission mode scanning with $75 \pm 5$ min time window (t) versus transmission mode scanning with $20 \pm 1$ h time window (T), and reflection mode scanning with $20 \pm 1$ h time window (R) versus transmission mode scanning with $20 \pm 1$ h time window, respectively. Protocols t provided better results than T, and T better than R. However, the differences are not significant. In Table \[tab:Pair\_23\], there is almost no difference between employing film type EBT2 (2) or EBT3 (3). Table \[tab:Pair\_em\] shows significant ($p < 0.05$) differences between calibration with planned (p) or measured (m) reference dose distributions. This result could be explained assuming that, for the EDW plan used in the calibration, the dose distribution measured with MatriXX has less uncertainty than the dose planned with Eclipse 10. Table \[tab:Pair\_EM\] shows significantly ($p < 0.05$) better results for the gamma analysis with planned (P) than with measured (M) dose distributions. This is a consequence of the resolution of the evaluation dose distribution which is much worse for MatriXX. The resolution of the array affects negatively the value of the $\gamma$-index in spite of using bicubic interpolation. Swapping reference and evaluation dose distributions was discarded since it would induce noise artifacts[@Clasie:2012].
Taking into account these comparisons, we selected the following dosimetry protocol: calibration with measured reference dose distributions, using film type EBT3, scanning in transmission mode with $75 \pm 5$ min time window and comparing the results with gamma analysis using planned evaluation dose distributions ([*i.e.*]{}, protocol Pt3m). Following this protocol allowed us to improve our previous $\gamma$-index tolerances from 4 % 3 mm to 3 % 3 mm or even 2.5 % 2.5 mm, results are presented in Table \[tab:final\_gamma\].
[l cc cc cc]{} & & &\
Test & $\overline{\gamma}$ & $\gamma_{<1}$ (%) & $\overline{\gamma}$ & $\gamma_{<1}$ (%) & $\overline{\gamma}$ & $\gamma_{<1}$ (%)\
A & 0.15 & 99.6 & 0.17 & 99.5 & 0.22 & 97.9\
B & 0.23 & 97.7 & 0.26 & 96.0 & 0.34 & 93.6\
C & 0.13 & 100 & 0.22 & 99.4 & 0.30 & 97.1\
D & 0.19 & 96.7 & 0.13 & 99.9 & 0.18 & 99.3\
E & 0.15 & 97.7 & 0.22 & 96.8 & 0.31 & 95.8\
F & 0.14 & 100 & 0.14 & 99.8 & 0.18 & 99.3\
G & 0.14 & 99.9 & 0.18 & 99.2 & 0.23 & 97.9\
H & 0.18 & 98.6 & 0.22 & 97.7 & 0.31 & 94.8\
I & 0.10 & 99.8 & 0.17 & 98.8 & 0.23 & 97.0\
J & 0.15 & 99.5 & 0.20 & 98.3 & 0.27 & 96.1\
K & 0.15 & 99.7 & 0.24 & 98.4 & 0.30 & 95.9\
L & 0.17 & 99.7 & 0.24 & 98.0 & 0.33 & 93.5\
M & 0.21 & 98.4 & 0.32 & 94.0 & 0.42 & 88.6\
N & 0.39 & 93.5 & 0.52 & 83.5 & 0.66 & 71.9\
Summary and recommendations
---------------------------
With respect to the model of channel-independent perturbations:
1. We recommend using the truncated normal distribution model because it can be considered as a metamodel which minimizes the uncertainty in the dose inherent in the method of channel-independent perturbations. The weighted mean model neglects correlations between the channels, which can be important, and the accuracy of the Micke-Mayer model depends on the properties of the sensitometric curves, which can result in unacceptable uncertainties for particular dosimetry systems. Since the other models are either limit cases or particular cases of the TN model, the latter should provide at least as good results as them.
2. For film calibration using the TN model, it is recommended to calibrate each color channel first. After that, two parameters: $\tilde{\sigma}_{\Delta}$ and of $\sigma_{\Delta}$, are obtained optimizing the RMSE of the differences between film doses ($d(r)$) and reference doses ($D(r)$), according to Eq.(\[Dose\]) and Eq.(\[eq\_cal\]).
3. Film doses can be calculated following a closed-form formula (Eq.(\[Dose\])). In addition, the type B uncertainty in the dose implicit in the method can be calculated (Eq.(\[uncertainty\])).
With respect to the dosimetry protocol, and excluding the comparisons between the particular TPS and array dosimeter used in this study:
1. No significant differences were found between transmission and reflection mode scanning.
2. Short waiting-time windows can be employed without losing accuracy, as pointed out by Lewis *et al*[@lewis:2012].
3. No significant differences were found between using EBT2 or EBT3 films.
Conclusions
===========
Four models of channel-independent perturbations for multichannel film dosimetry were examined. Two of them based on the literature: a model which employs channel-independent perturbations as proposed by Micke *et al*[@AMicke:2011] and further developed by Mayer *et al*[@mayer:2012], and another one which uses the weighted mean of all three channels to obtain the dose[@mendez:2013]. In addition to these, two novel models were proposed, a more realistic extension to the Micke-Mayer model which uses uniform distributed perturbations but allows the error terms to differ from one channel to another, and a truncated normal distribution, which comprises the other models as particular or limit cases.
A closed-form formula for dose calculation was derived for all four models, and it coincides with the published one[@mayer:2012] in the case of the Micke-Mayer model. In addition, Type B uncertainties in film dose due to the channel-independent perturbations method were obtained.
In order to assess the performance of the models, a set of tests was devised in which the dose distributions obtained from films were compared to either planned, or measured dose distributions. In these tests, the truncated normal distribution model provided the best agreement between film and reference doses, both for calibration and $\gamma$-index verification, and proved itself superior to both the weighted mean model, which neglects correlations between the channels, and the Micke-Mayer and the uniform distribution models, whose accuracy depends on the properties of the sensitometric curves. As a conclusion, we feel confident to recommend the truncated normal distribution model of channel-independent perturbations for multichannel dosimetry.
Along with the models, other factors which could influence the dosimetry process were also evaluated. No significant differences were found between transmission mode scanning and reflection mode scanning, between $75 \pm 5$ min versus $20 \pm 1$ h waiting-time window or between employing EBT2 or EBT3 films. However, significantly better results were obtained when a measured dose distribution was used instead of a planned one as reference for the calibration, and when a planned dose distribution was used instead of a measured one as evaluation for the $\gamma$-analysis.
The authors would like to thank Denis Brojan, Víctor Hernández and Sašo Pulko for their contributions to this work.
[34]{}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 10.1118/1.1929253) [****, ()](http://stacks.iop.org/0031-9155/52/i=14/a=013) [****, ()](http://stacks.iop.org/0031-9155/52/i=1/a=015) [****, ()](\doibase 10.1118/1.2828196) [****, ()](http://stacks.iop.org/0031-9155/53/i=24/a=001) [****, ()](http://stacks.iop.org/0031-9155/55/i=9/a=012) [****, ()](\doibase
10.1016/j.ejmp.2010.10.001) [****, ()](\doibase 10.1118/1.4771960) [****, ()](\doibase 10.1118/1.598407) [****, ()](\doibase 10.1118/1.2431425) [****, ()](\doibase 10.1118/1.3253902) [****, ()](\doibase
10.1118/1.3373523) [****, ()](\doibase 10.1118/1.3291622) [****, ()](\doibase 10.4236/ijmpcero.2012.12008) [****, ()](\doibase 10.1118/1.3576105) [****, ()](http://stacks.iop.org/0031-9155/57/i=13/a=4353) [****, ()](\doibase 10.1118/1.3694100) [****, ()](\doibase
10.1118/1.4772075) [****, ()](\doibase 10.1118/1.4737890) [****, ()](\doibase 10.1118/1.3700731) [****, ()](\doibase 10.1118/1.3378675) [****, ()](\doibase 10.1118/1.2362876) [****, ()](\doibase 10.1118/1.2357836) [****, ()](\doibase 10.1118/1.2370505) [****, ()](\doibase
10.1118/1.2936334) [****, ()](\doibase 10.1118/1.2938522) [****, ()](\doibase 10.1118/1.2795828) @noop [**]{}, ed. () [****, ()](http://stacks.iop.org/0031-9155/54/i=24/a=008) [**](http://www.R-project.org/), , (), @noop [****, ()]{} [****, ()](\doibase
10.1118/1.1449493) [****, ()](http://stacks.iop.org/0031-9155/57/i=21/a=6981) [****, ()](\doibase 10.1118/1.4754797)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The resonance scattering of ${\rm {Ly{\alpha}}}$ photons with neutral hydrogen atoms in the intergalactic medium not only couples the spin temperature to the kinetic temperature but also leads to a heating of the gas. We investigate the impact of this heating on the average brightness temperature of the 21-cm signal from the Cosmic Dawn in the context of the claimed detection by the EDGES low-band experiment. We model the evolution of the global signal taking into account the ${\rm {Ly{\alpha}}}$ coupling and heating and a cooling which can be stronger than the Hubble cooling. Using the claimed detection of a strong absorption signal at $z\approx 17$ as a constraint, we find that a strong ${\rm {Ly{\alpha}}}$ background is ruled out. Instead the results favour a weak ${\rm {Ly{\alpha}}}$ background combined with an excess cooling mechanism which is substantially stronger than previously considered. We also show that the cooling mechanism driven by the interaction between millicharged baryons and dark matter particles no longer provides a viable explanation for the EDGES result when ${\rm {Ly{\alpha}}}$ heating is taken into account.'
author:
- |
Raghunath Ghara$^{1,2,3}$[^1], Garrelt Mellema$^{1}$\
$^1$ The Oskar Klein Centre, Department of Astronomy, Stockholm University, AlbaNova, SE-10691 Stockholm, Sweden\
$^{2}$Department of Natural Sciences, The Open University of Israel, 1 University Road, PO Box 808, Ra’anana 4353701, Israel\
$^{3}$Department of Physics, Technion, Haifa 32000, Israel
bibliography:
- 'bibfile.bib'
date: 'Accepted XXX. Received YYY; in original form ZZZ'
title: 'Impact of ${\rm {Ly{\alpha}}}$ heating on the global 21-cm signal from the Cosmic Dawn'
---
\[firstpage\]
radiative transfer - galaxies: formation - intergalactic medium - cosmology: theory - dark ages, reionization, first stars
Introduction {#sec:intro}
============
The formation of the first sources of light is one of the milestone events in the history of our Universe. These primordial sources changed the ionization and thermal state of the gas in the intergalactic medium (IGM) and thus affected the further evolution of the Universe. The period when these very first sources formed is sometimes called the ‘Cosmic Dawn’ (CD). Details regarding these early sources, such as the time of their formation, their emission properties, etc. remain unknown. Models such as in @2006MNRAS.372.1093F [@Mesinger2013; @2017MNRAS.464.3498F; @2018MNRAS.478.2193C; @2019MNRAS.484..933P; @2019MNRAS.483.1980M], suggest that they formed around redshift 30 and their ultraviolet radiation first caused the spin temperature of the neutral hydrogen in the IGM to change due to the repetitive scattering of Lyman series photons, a process known as the Wouthuysen-Field effect [@wouth52; @field58; @hirata2006lya; @2006ApJ...651....1C]. The same models also predict that over time X-rays produced by these sources started to heat the IGM and only much later, in what usually is called the Epoch of Reionization (EoR) sufficient numbers of ionizing photons were produced to reionize the Universe.
The 21-cm signal produced by the neutral hydrogen in the IGM during these epochs can provide us with answers to many of the questions regarding the CD and the EoR. Therefore several efforts to detect this signal have been initiated. Two different types of experiments exist. The first type uses large interferometers to measure the spatial fluctuations of the neutral hydrogen ([H [i ]{}]{}) signal in terms of statistical quantities such as the power spectrum. Examples of these are the Low Frequency Array (LOFAR)[^2] [@van13; @2017ApJ...838...65P], the Giant Metrewave Radio Telescope (GMRT)[^3][@ghosh12; @paciga13], the Precision Array for Probing the Epoch of Reionization (PAPER)[^4] [@parsons13] and the Murchison Widefield Array (MWA)[^5] [@bowman13; @tingay13]. The future low-frequency component of the Square Kilometre Array (SKA-Low)[^6] will have the sensitivity to directly probe the spatial structure of the fluctuations by producing images of the signal [@2015aska.confE..10M; @ghara16].
The second type of experiment tries to detect the sky-averaged 21-cm signal, a quantity which the interferometers are unable to measure. Such a measurement only requires a single antenna. Examples of this type are EDGES [@2010Natur.468..796B], SARAS [@2015ApJ...801..138P], BigHorns [@2015PASA...32....4S], SciHi [@2014ApJ...782L...9V] and LEDA [@2012arXiv1201.1700G].
The detection of the redshifted 21-cm signal from the EoR and CD is very challenging for all types of experiments as it is several orders of magnitude weaker than the galactic and extra-galactic foreground signals at these frequencies. In addition, long integration times are needed to bring the system noise below the cosmological signal which makes calibration challenging, not only because of instrument stability but also because of the impact of time-dependent ionospheric effects. As a consequence, no undisputed detections of the signal have yet been made.
The strength of the redshifted 21-cm signal from the CD depends on the gas temperature and background ${\rm {Ly{\alpha}}}$ flux which are determined by the radiation sources and the heating/cooling processes. Several heating processes such as X-ray heating [@Pritchard07; @Mesinger2013; @ghara15b; @ghara15c; @2018arXiv180803287R; @2019MNRAS.487.2785I], shock heating [@2004ApJ...611..642F] and heating due to resonance scattering of ${\rm {Ly{\alpha}}}$ photons (hereafter ‘${\rm {Ly{\alpha}}}$ heating’) [@2004ApJ...602....1C; @2007ApJ...655..843C; @2006MNRAS.372.1093F] can increase the kinetic temperature of the gas in the IGM during these epochs. However, the relative contribution of these mechanisms is uncertain. In addition to these heating processes based on known physics, unknown physics such as dark matter decay may also convey energy to the IGM [[@2018PhRvD..98d3006C; @2018JCAP...05..069M; @2018arXiv180309739L]]{}. The gas cooling is expected to be dominated by the adiabatic cooling due to the expansion of the Universe (‘Hubble cooling’) with radiative cooling due to e.g. recombinations playing a subdominant role.
Recently, @EDGES2018 claimed a detection of a redshift-amplitude profile of the global 21-cm signal around redshift $z\sim 17$ from observations with the EDGES low-band instrument. However, the measured absorption signal was found to be stronger by several factors than the signal predicted by the previous theoretical studies such as @Pritchard07 [@Mesinger2013; @santos08; @ghara15a]. Explanations for the EDGES low-band results fall into two categories. The first kind assumes a lower than expected IGM temperature due to excess cooling caused by an unknown physical process such as the interaction between baryons and dark matter particles [@2018Natur.555...71B; @2018PhRvL.121a1101F; @2018arXiv180210094M; @PhysRevLett.121.011102]. The second type considers the presence of an excess radio background which can also enhance the measurement of the [H [i ]{}]{} signal, which is otherwise seen against the background of the Cosmic Microwave Background (CMB) [@2018ApJ...858L..17F; @2018ApJ...868...63E; @2018PhLB..785..159F]. Examples of sources which could cause such an excess radio background are supermassive black holes [@2018ApJ...868...63E] or supernova from first stars at $z\gtrsim 17$ [@2019MNRAS.483.1980M]. However, models relying on such astrophysical sources are unlikely as the time scale for generating a radio background is several orders of magnitude shorter than the duration of the EDGES signal centred at redshift $\sim 17$ [@2018MNRAS.481L...6S]. In addition the required excess background requires a $\sim 10^3$ times stronger flux of 1-2 GHz photons than observed from local galaxies [@2019MNRAS.483.1980M]. The viable alternative is an excess radio background of cosmological origin, e.g. decay of unstable particles into dark photons with non-zero mixing angle with electromagnetism [see e.g., @2018PhRvL.121c1103P].
Both of these explanations require the spin temperature to be strongly coupled to the gas temperature which in turn requires a strong ${\rm {Ly{\alpha}}}$ background. However, these ${\rm {Ly{\alpha}}}$ photons will also heat up the gas by resonance scattering. The question is whether this heating effect has an impact on the global signal. @madau1997 estimated the heating rate due to resonance scattering assuming that the scatterings occur with atoms at rest. For this estimate, the IGM temperature would exceed the CMB temperature in a fraction of Hubble time. A subsequent paper by @2004ApJ...602....1C included the effect of atomic thermal motions into the calculation and showed that the ${\rm {Ly{\alpha}}}$ heating rate is at least three orders of magnitude lower than estimated in @madau1997. Their calculation considered heating due to photons between ${\rm {Ly{\alpha}}}$ and ${\rm {Ly{\beta}}}$ (so-called ‘continuum photons’) as these redshift into the ${\rm {Ly{\alpha}}}$ resonance and cooling due to the cascade of higher resonance states into ${\rm {Ly{\alpha}}}$ (so-called ‘injected photons’). These authors showed that these two mechanism balance at a temperature $\sim 10$ K and thus the temperature would not increase beyond that. This low equilibrium value has prompted many works to neglect ${\rm {Ly{\alpha}}}$ heating as its effect would seem to be negligible compared to for example X-ray heating. However, both these works did not consider the forbidden transition from the 2s to the 1s level of hydrogen, something which was added to the calculation by @2007ApJ...655..843C who furthermore included the effect of deuterium. The result is a lower cooling contribution from the injected photons and which implies that the gas temperature can increase to an equilibrium value of $\sim 100$ K prior to the reionization.
Previous studies of the global 21-cm signal in the context of the EDGES results did either not consider ${\rm {Ly{\alpha}}}$ heating [[see e.g., @PhysRevLett.121.011103; @PhysRevD.98.023501; @PhysRevD.98.063021; @2018arXiv180303091B; @2018arXiv181209760N]]{} or used the erroneously low values from @2004ApJ...602....1C [[see e.g., @2019arXiv190202438F; @2018PhRvD..98j3513V]]{}. In this study, we for the first time adopt the calculation of the ${\rm {Ly{\alpha}}}$ heating rates from @2007ApJ...655..843C and investigate its impact on the global 21-cm signal from the CD. We include excess cooling so as to be able to reproduce EDGES low-band observations. We will consider models which use an excess radio background in a future work. We explore the parameter space of ${\rm {Ly{\alpha}}}$ heating and excess cooling to study the absorption profile of the global signal to find combinations of parameters that agree with the EDGES low-band results.
We have organised the paper in the following way. In Section \[sec:model\] we describe the analytical model we use to calculate the evolution of the global 21-cm signal, including the heating rates due to resonance scattering of the ${\rm {Ly{\alpha}}}$ photons. We first present results for a phenomenological excess cooling rate in Section \[sec:result\], followed by an investigation of a physically motivated excess cooling rate in Section \[sec:DMB\]. We conclude in Section \[sec:con\]. Throughout the paper we use the following set of cosmological parameters ${\Omega_{\rm m}}=0.32$, ${\Omega_{\rm B}}=0.049$, $\Omega_\Lambda=0.68$, $h=0.67$, $\sigma_8=0.83$ and $n_{\rm s}=0.96$ [@Planck2015].
Model for 21-cm signal {#sec:model}
======================
Analytical model {#sec:analytical}
----------------
The 21-cm signal from the [H [i ]{}]{}gas is measured as the differential brightness temperature against the CMB and can be written as $${\delta T_{\rm b}}= 27 ~x_{\rm HI} (1+\delta_{\rm B}) \left(\frac{{\Omega_{\rm B}}h^2}{0.023}\right) \sqrt[]{\frac{0.15}{{\Omega_{\rm m}}h^2}\frac{1+z}{10}} \left(1-\frac{{T_{\gamma}}}{{T_{\rm S}}} \right)\,\rm{mK},
\label{eq_tb}$$ where $x_{\rm HI}$, $\delta_{\rm B}$, ${T_{\rm S}}$ and ${T_{\gamma}}=2.73 \times (1+z)$ K denote the neutral fraction, density contrast, spin temperature of the hydrogen gas and CMB temperature at redshift $z$, respectively.
We adopt an analytic approach to model the expected 21-cm signal in the presence of spin temperature fluctuations. This approach follows previous works such as @Pritchard07 [@2005ApJ...630..643M]. It incorporates ${\rm {Ly{\alpha}}}$, UV and X-ray photons from the sources which are taken to be associated with dark matter halos. The number of dark matter halos at a given redshift is determined using the Press-Schechter halo mass function. We assume that only halos with virial temperatures above $10^4$ K contribute. The model estimates the volume averaged ionization fractions of the highly ionized [H [ii ]{}]{} regions ($x_i$) and of the mostly neutral gas in the IGM outside these [H [ii ]{}]{} regions ($x_e$). We assume the temperature of the ionized [H [ii ]{}]{} regions to be $\sim 10^4$ K. The gas temperature (${T_{\rm K}}$) of the largely neutral medium outside the [H [ii ]{}]{} regions is calculated using the various heating and cooling processes.
The heating rate due to resonance scattering as well as the spin temperature coupling depend critically on the number of ${\rm {Ly{\alpha}}}$ photons emitted from the sources. To estimate the average ${\rm {Ly{\alpha}}}$ photon flux, we follow the method from @2006MNRAS.372.1093F. We assume a power law spectrum $\epsilon_s(\nu) = f_{\alpha} A_{\alpha} \nu^{-\alpha_{s}-1}$ between ${\rm {Ly{\alpha}}}$ and ${\rm {Ly{\beta}}}$ and between ${\rm {Ly{\beta}}}$ and the Lyman limit, where the power law indices $\alpha_{s}$ can differ. The spectral index $\alpha_s$ between ${\rm {Ly{\alpha}}}$ and ${\rm {Ly{\beta}}}$ is taken to be 0.14 which corresponds to population II type sources. The normalization factor $A_{\alpha}$ is estimated such that the number of ${\rm {Ly{\alpha}}}$ photons per baryon in the range ${\rm {Ly{\alpha}}}$-${\rm {Ly{\beta}}}$ is 6520 for $f_{\alpha}=1$. The spectral index in the range ${\rm {Ly{\alpha}}}$-Lyman limit is adjusted so that the total number of photons per baryon for this wavelength regime is 9690. The parameter $f_\alpha$ determines the production rate of the ${\rm {Ly{\alpha}}}$ photons from the stars. The heating rate due to ${\rm {Ly{\alpha}}}$ scattering is described below in Section \[sec:heat\].
To model the X-ray heating, we follow @Pritchard07 and assume that the emissivity of X-ray photons from the sources follows the star formation rate density. We use an X-ray spectral distribution given by $$\epsilon_X(\nu) = \frac{L_0}{h\nu_0}\left(\frac{\nu}{\nu_0}\right)^{-\alpha_X-1}\,,$$ with $L_0=f_X\times 10^{41} ~\rm erg ~s^{-1} ~Mpc^{-3}$, $h\nu_0 = 1 ~\rm keV$. For our fiducial X-ray source we choose the X-ray efficiency parameter to be $f_X=1$ and the spectral index of the X-ray spectrum to be $\alpha_X=0.5$. Note that for most of our results we will set $f_X=0$ as we want to focus on the effect of ${\rm {Ly{\alpha}}}$ heating. This makes our results conservative as additional X-ray heating will only further increase the gas temperature.
Finally our model also includes the effect of ionizing UV radiation. The rate of emission of the UV photons per baryon is $$\Lambda_i = \zeta \frac{ {\rm d} f_{\rm coll}}{ {\rm d} t}.$$ The ionization efficiency parameter $\zeta = N_{\rm ion}\times f_{\rm esc}\times f_{\star}$ depends on the average number of ionizing photons per baryon produced in the stars ($N_{\rm ion}$), the star formation efficiency ($f_\star$) and the escape fraction of the UV photons ($f_{\rm esc}$). All these quantities are uncertain during the CD and EoR. In this study, we assume $N_{\rm ion}=4000$ which corresponds to population II types of stars, $f_\star=0.1$ and $f_{\rm esc}=0.1$ for modelling reionization. We note however that for most of our results ionization levels remain very low and do not impact the global 21-cm signal.
Heating due to resonance scattering {#sec:heat}
-----------------------------------
To estimate the heating rates due to the resonance scattering, we follow the calculations of @2007ApJ...655..843C. Photons emitted with frequencies between ${\rm {Ly{\alpha}}}$ and ${\rm {Ly{\beta}}}$ frequency (‘continuum photons’) will redshift to the ${\rm {Ly{\alpha}}}$ frequency at which point they suffer resonance scattering by [H [i ]{}]{}. This process will heat up the gas. [On the other hand, photons with a wavelength between ${\rm {Ly{\beta}}}$ and Lyman limit will be absorbed by the hydrogen atoms after redshifting to ${\rm {Ly{\beta}}}$ or other higher Lyman series lines. If higher resonance or excited states first decay to $2p$ state and then to the ground state, one ${\rm {Ly{\alpha}}}$ photon will be emitted. ]{} In contrast to the continuum photons, the emission of the ${\rm {Ly{\alpha}}}$ photons (‘injected photons’) due to the cascade of from the higher levels will cool the gas. The spectrum gets affected once the photons redshift through the ${\rm {Ly{\alpha}}}$ resonance. The intensity $J(\nu)$ at a frequency $\nu$ in the vicinity of the resonance frequency $\nu_\alpha$ can be written as [@2007ApJ...655..843C], $$J(x) = J(0) e^{-\frac{2\pi \gamma x^3}{3a}-2 \eta x},$$ for the injected photons. The above expression also hold for the continuum photons with $x>0$, otherwise $$J(x) = 2\pi J_0\gamma a^{-1} \int_{-\infty}^{x} e^{\frac{2\pi \gamma (z^3-x^3)}{3a}+2 \eta (z-x)} dz.$$ where $$\begin{aligned}
\label{eqn:eqlabel}
\begin{gathered}
x=(\nu/\nu_\alpha - 1)/(2k_{\rm B}{T_{\rm K}}/mc^2)^{1/2},
\\
a=A_{21}(2k_{\rm B}{T_{\rm K}}/mc^2)^{-1/2}/4\pi\nu_\alpha,
\\
\gamma = \tau_{\rm GP}^{-1}(1+0.4/{T_{\rm S}})^{-1},
\\
\eta = [h \nu_\alpha/(2k_{\rm B} {T_{\rm K}}m c^2)^{1/2}][(1+0.4/{T_{\rm S}})/(1+0.4/{T_{\rm K}})].
\end{gathered}\end{aligned}$$ Here $k_{\rm B}, m$, $c$ and $A_{21}$ are the Boltzmann constant, mass of hydrogen atom, speed of light and the Einstein spontaneous emission coefficient of ${\rm {Ly{\alpha}}}$ transition respectively. The quantities $\tau_{\rm GP}$ and $J_0$ are the Gunn-Peterson optical depth and the UV intensity at a frequency far away from $\nu_\alpha$, respectively.
The quantity $J(0)$ can be expressed as, $$\frac{J(0)}{J_0} = \frac{\pi\zeta\left(J_{1/3}(\zeta)-J_{-1/3}(\zeta)\right)}{\sqrt{3}}+ _1F_2\left(1; 1/3, 2/3, -\zeta^2/4 \right)$$ where $\zeta = \sqrt[]{16\eta^3a/9\pi\gamma}$, $_1F_2$ is hyper-geometric function, $J_{1/3}$ and $J_{-1/3}$ are the Bessel functions of first kind respectively.
The total heating/cooling rate due to the resonance scattering can be written as, $${\frac{d\log {T_{\rm K}}}{d\log t}} \bigg|_{\rm heating} = \frac{2 t}{3 k_{\rm B} {T_{\rm K}}} H_{\alpha},
\label{eq:heating}$$ where $t$ represents time, $H_{\alpha}$ is the rate of exchange of total energy by the photons due to resonance scattering. $$H_{\alpha} = \dot{N}_{\alpha} \left(\Delta E_c + \frac{J_i}{J_c} \Delta E_i \right)$$ where $\dot{N_{\alpha}}$ denotes the number of photons per hydrogen atom that pass through resonance scattering per unit time. The ratio of injected and continuum photons $J_i/J_c$ depends on the source’s surface temperature. We choose $J_i/J_c \approx 0.1$ which corresponds to a source with an effective temperature $\lesssim 5\times10^4$ K which corresponds approximately to population II type of sources [@2007ApJ...655..843C]. The quantities $\Delta E_c$ and $\Delta E_i$ are the total energy gain by the gas due to a resonance scattering by the continuum and injected photons respectively. This can be written as, $$\Delta E(x) =\frac{(h\nu)^2}{m c^2} \int \frac{J(x)}{J_0} \phi(x) dx$$ where $\phi(x)$ is the normalized scattering cross-section. Note that @2007ApJ...655..843C also considered the heating contribution from deuterium in their studies. Here we have not included this and thus our calculations somewhat underestimate the actual heating rates.
Cooling processes {#sec:ex-cool}
-----------------
The gas temperature of the IGM is one of the key quantities which determines the strength and nature of the 21-cm signal from the CD. As the heating and cooling processes during those epochs are uncertain, the gas temperature as well as the signal are poorly understood. The analytical method used in this study incorporates the adiabatic cooling due to the expansion of the Universe which dominates over radiative processes such as the collisional-ionization cooling, recombination cooling, collisional excitation cooling, free-free cooling, etc. After Compton scattering with CMB photons ceases to be important, this Hubble cooling causes the average gas temperature to evolve as ${T_{\rm K}}\propto (1+z)^2$. For standard physics, the post-recombination gas temperature is easily calculated, as can for example be done with the publicly available code [recfast]{} [@1999ApJ...523L...1S]. The results show that for our cosmological parameters the ${T_{\rm K}}\propto (1+z)^2$ relation is valid below $z_0\approx 138$. Expressed in the same form as the ${\rm {Ly{\alpha}}}$ heating rate in Equation \[eq:heating\], this adiabatic or Hubble cooling is given as $$\frac{d\log T_K}{d\log t}\bigg|_{\rm H} = -\frac{4}{3}.$$
However, as pointed out by @EDGES2018 this cooling process is unable to explain the strong absorption signal at redshift 17 found in the EDGES low-band results as it requires a lower temperature than can be achieved using standard cosmological models. In order to reproduce the EDGES results we therefore need to include an excess cooling rate in our calculations. Here we make two choices. In Section \[sec:result\] we use a simple phenomenological excess cooling model and in Section \[sec:DMB\] we use a physically motivated excess cooling rate based on interactions between dark matter particles and baryons.
Phenomenological cooling model {#sec:result}
==============================
In this section, we consider a simple phenomenological cooling rate inspired by @2019MNRAS.483.1980M, given by $$\frac{d\log T_K}{d\log t}\bigg|_{\rm cool} = \alpha \left[\frac{1+z}{1+z_0}\right]^\beta\,.
\label{eq:ex-cool}$$ The parameters $\alpha\le 0$ and $\beta$ determine the strength and redshift dependence of the excess cooling rate respectively. We only apply this excess cooling rate for redshifts $z\le z_0$. However, in principle, $z_0$ could be treated as a free parameter.
To gain insight into the impact of ${\rm {Ly{\alpha}}}$ heating on the volume averaged 21-cm signal from the CD, we first show the results for a number individual scenarios (Section \[sec:main\_scene\]). After this we will explore the parameter space made up of $\alpha$, $\beta$ and the average ${\rm {Ly{\alpha}}}$ flux (Section \[sec:parameter\]). Lastly, we will investigate scenarios that can explain the strong absorption signal as reported by the EDGES low-band observation (Section. \[sec:edges\_interpretation\]).
Scenarios $f_\alpha$ $\alpha$ $\beta$ ${{\delta T_{\rm b}}}_{,\rm min}$ $z({{\delta T_{\rm b}}}_{, \rm min})$ $\Delta z$
----------- ------------ ---------- --------- ----------------------------------- --------------------------------------- ------------
S$_1$ 1.0 0.0 0.0 -95.1 19.9 4.6
S$_2$ 1.0 -1.0 0.0 -367.1 20.9 3.7
S$_3$ 1.0 -1.0 -0.1 -404.6 20.9 3.6
S$_4$ 0.1 -1.0 -0.1 -294.4 17.9 4.2
: The ${\rm {Ly{\alpha}}}$ efficiency and excess cooling parameters for the four different scenarios considered in Section \[sec:main\_scene\]. Also shown are the quantities which describe the resulting absorption profile, ${{\delta T_{\rm b}}}_{,\rm min}$, $z({{\delta T_{\rm b}}}_{,\rm min})$ and $\Delta z$ which represent the minimum brightness temperature, its corresponding redshift and the FWHM of the absorption profiles, respectively.[]{data-label="tab2"}
Parameters Min range Max Range
------------ ----------- ----------- -- --
$f_\alpha$ 0.01 100.0
$\alpha$ -1.5 0.0
$\beta$ -0.5 0.5
: The range of the three parameters for the phenomenological cooling model explored in this study. []{data-label="tab1"}
Exploratory scenarios {#sec:main_scene}
---------------------
We choose four different sets of parameters to study the impact of different parameters/processes on the evolution of ${\delta T_{\rm b}}$. The parameters for these scenarios are listed in Table \[tab2\]. The fiducial model S$_1$ has $\alpha=0$ and therefore does not include any excess cooling. The left-hand panel of Fig. \[image\_global\] presents the redshift evolution of the average gas temperature of the neutral regions in the IGM for these four choices. For each, we consider two cases, namely without (thin lines) and with (thick lines) ${\rm {Ly{\alpha}}}$ heating. For the choice of no excess cooling S$_1$ we also consider a case without ${\rm {Ly{\alpha}}}$ heating but with heating by X-ray sources (thin double dot-dashed curve).
For the cases without ${\rm {Ly{\alpha}}}$ and X-ray heating, the temperature keeps decreasing over time as no other heating mechanisms are included in these scenarios. When we include ${\rm {Ly{\alpha}}}$ heating, it impacts the gas temperatures as early as redshift 22 in all these scenarios. For the scenario without excess cooling, the gas temperature increases to $\sim 60$ K at redshift $\sim 14$ which is roughly consistent with the results of @2007ApJ...655..843C. The small difference is due to ignoring the contribution from deuterium in our calculations. When we instead of ${\rm {Ly{\alpha}}}$ heating include X-ray heating according to the description in Section \[sec:analytical\], the gas temperature for S$_1$ increases more rapidly and reaches $\sim 300$ K by $z \sim 14$. This is why ${\rm {Ly{\alpha}}}$ heating is often ignored in simulations as X-ray heating will quickly dominate. However, if X-ray heating is inefficient or absent, ${\rm {Ly{\alpha}}}$ heating will have a non-negligible impact on the IGM temperature.
As scenarios S$_2$–S$_4$ include excess cooling, the Cosmic Dawn starts at lower gas temperatures than for S$_1$. In S$_2$ the excess cooling does not have a redshift dependence, in S$_3$ it increases with time. S$_4$ has the same excess cooling parameters as S$_3$ but a ten times lower ${\rm {Ly{\alpha}}}$ efficiency. When including the heating due to the scattering of ${\rm {Ly{\alpha}}}$ photons, it impacts the temperatures in S$_2$ and S$_3$ earlier compared to S$_1$, although the background ${\rm {Ly{\alpha}}}$ flux densities for these models are identical. This is due to the fact that the ${\rm {Ly{\alpha}}}$ heating rates increases as the kinetic temperature decreases (see Equation \[eq:heating\]). As expected the heating starts later for a lower ${\rm {Ly{\alpha}}}$ background (S$_4$).
Scenarios S$_1$, S$_2$ and S$_3$ each start with different temperatures. However, by $z\sim 16$ they all reach almost the same equilibrium temperature due to ${\rm {Ly{\alpha}}}$ heating. For $\alpha=1$ the excess cooling can thus not compete with ${\rm {Ly{\alpha}}}$ heating. For the case of a lower ${\rm {Ly{\alpha}}}$ flux (S$_4$), the heating is delayed and remains weaker compared to the other scenarios.
The right-hand panel of Fig. \[image\_global\] shows the redshift evolution of the global 21-cm signal corresponding to the nine scenarios (S$_1$ through S$_4$ with and without ${\rm {Ly{\alpha}}}$ heating and S$_1$ with X-ray heating). Note that we always include the ${\rm {Ly{\alpha}}}$ coupling for the spin temperature, even in those models where we ignore ${\rm {Ly{\alpha}}}$ heating. As for all these scenarios the IGM remains highly neutral at redshifts $>14$, the average brightness temperature is mostly determined by the gas temperature and the strength of the ${\rm {Ly{\alpha}}}$ coupling. As the background ${\rm {Ly{\alpha}}}$ flux is low at high redshift ($z\sim30$), the coupling between ${T_{\rm S}}$ and ${T_{\rm K}}$ is weak and ${T_{\rm S}}$ remains close to ${T_{\gamma}}$. This makes ${\delta T_{\rm b}}\approx 0$ at those redshifts. As more sources form with time, ${\rm {Ly{\alpha}}}$ coupling becomes stronger and the signal starts to appear in absorption, i.e., with a negative sign. However, different heating processes can increase the gas temperature and eventually ${\delta T_{\rm b}}$ transitions from absorption to emission. This produces a characteristic trough-like feature in the redshift evolution of ${\delta T_{\rm b}}$, which we refer to as the ‘absorption profile’.
In the absence of ${\rm {Ly{\alpha}}}$ or X-ray heating, ${\delta T_{\rm b}}$ decreases with redshift as ${T_{\rm K}}$ decreases with time and the signal remains in absorption until reionization ends. In such cases, ${\delta T_{\rm b}}$ slowly decreases to $\sim$ -250 mK at redshift $\sim 15$ for S$_1$ without excess cooling (thin solid line), while ${\delta T_{\rm b}}$ rapidly decreases to values below $\sim -500$ mK at redshift $\lesssim 20$ for models S$_2$ through S$_4$ which include excess cooling (thin long-dashed, short-dashed and dot-dashed lines).
In the presence of ${\rm {Ly{\alpha}}}$ heating, the increase of gas temperature as early as redshift $\sim 20$ resists the decrease of ${\delta T_{\rm b}}$ with time and produces prominent absorption profiles (thick lines). The minimum ${\delta T_{\rm b}}$ values of these profiles are much less deep than the corresponding signals from the no heating cases, demonstrating the large impact ${\rm {Ly{\alpha}}}$ heating has. For example in scenario S$_1$ the absorption profile does not reach below -100 mK and for S$_2$ and S$_3$ not below -400 mK.
The absorption profiles can be described by the minimum value of the brightness temperature (${{\delta T_{\rm b}}}_{,\rm min}$), the corresponding redshift $z({{\delta T_{\rm b}}}_\mathrm{, min})$ and the full width at half maximum (FWHM) of the absorption profile ($\Delta z$). We list the values for the cases with ${\rm {Ly{\alpha}}}$ heating in Table \[tab2\]. These numbers clearly depend on the excess cooling rate and ${\rm {Ly{\alpha}}}$ heating rates. We see the absorption profiles are much stronger and appear earlier for models S$_2$ and S$_3$ than for model S$_1$. This is due to the excess cooling in the former models which results in a lower initial gas temperature compared to S$_1$. The values for the $\Delta z$ are lower when excess cooling is present.
When the ${\rm {Ly{\alpha}}}$ background is lower ($f_\alpha=0.1$, scenario S$_4$) the absorption profile becomes less deep, widens and appears later compared to the scenario which has $f_\alpha=1$ (S$_3$), even though the gas temperature is actually lower. However, the profile can still reach a minimum of $\sim -300$ mK, below what can be achieved without excess cooling.
For completeness, the right-hand panel of Fig. \[image\_global\] also shows the differential brightness temperature evolution for the scenario without excess cooling and ${\rm {Ly{\alpha}}}$ heating but with X-ray heating (thin dot-dot-dashed curve). Due to the higher temperatures, this absorption profile is less deep and somewhat narrower than the corresponding case with ${\rm {Ly{\alpha}}}$ heating (thick solid curve).
Parameter space study {#sec:parameter}
---------------------
Now we will explore the parameter space of excess cooling ($\alpha$ and $\beta$) and ${\rm {Ly{\alpha}}}$ flux ($f_\alpha$) to find the impact on the absorption signal from the CD in terms of absorption profile parameters ${{\delta T_{\rm b}}}_{, \rm min}$, $z({{\delta T_{\rm b}}}_\mathrm{,\rm min})$ and $\Delta z$. The details of the parameter space are given in Table \[tab1\]. As the excess cooling is due to unknown processes, the parameter ranges for $\alpha$ and $\beta$ chosen here are somewhat arbitrary. However, as we will see this range covers the most interesting results in terms of the absorption feature and the EDGES low-band results.
We will study the global 21-cm signal around redshift $\sim 17$ which corresponds to $z({{\delta T_{\rm b}}}_\mathrm{, min})$ of the EDGES low-band detection. The different panels of Fig. \[image\_paramz\] represent the value of the differential brightness temperature at redshift 17 in 2D slices through the 3D parameter space. For these slices, the third parameter is chosen such that these slices contain the lowest brightness temperature at $z=17$ within the explored parameter space. The values are $\beta=-0.5$ (panel A$_1$), $\alpha=-1.5$ (panel A$_2$) and $f_\alpha=0.01$ (panel A$_3$).
The resonance photons impact the signal in two ways: ($i$) heating due to resonance scattering decreases for a lower background ${\rm {Ly{\alpha}}}$ flux, ($ii$) coupling of ${T_{\rm S}}$ with ${T_{\rm K}}$ decreases for a lower ${\rm {Ly{\alpha}}}$ background. These two effects create the vertical feature in ${\delta T_{\rm b}}$ around $f_\alpha \sim 0.1$ in panels A$_1$ and A$_2$. In the presence of significant ${\rm {Ly{\alpha}}}$ heating (e.g. for $f_\alpha >$1), the amplitude of ${\delta T_{\rm b}}$ at redshift 17 remains small for all values of $\alpha$ and $\beta$. As shown in panel A$_3$, strong excess cooling ($\alpha \sim -1.5$ and $\beta \sim -0.5$) can produce a deep absorption feature but only for a very weak ${\rm {Ly{\alpha}}}$ flux, reaching values as low as $-600$ mK for $f_\alpha \sim 0.01$.
We note that the color bar associated with the panels of Fig. \[image\_paramz\] represents ${\delta T_{\rm b}}$ at redshift 17, not ${{\delta T_{\rm b}}}_\mathrm{, min}$ for the choice of parameters. This figure shows that ${\delta T_{\rm b}}\sim -500$ mK at redshift 17 is only possible for a weak ${\rm {Ly{\alpha}}}$ background and strong excess cooling rates as shown by the contours in the panels. However, this does not mean that the values of ${\delta T_{\rm b}}$ in this figure are equal to ${{\delta T_{\rm b}}}_\mathrm{, \rm min}$, the minimum of the absorption profiles. Thus, we can not directly compare these with the EDGES low-band observations. However, we can see that a large part of the parameter space corresponds to ${\delta T_{\rm b}}$ values larger than -500 mK and thus, should be excluded by the EDGES observation. We will present a detailed investigation of this in Section \[sec:edges\_interpretation\].
First we will investigate the behaviour of absorption profiles over the parameter space. The top row of panels of Fig. \[image\_param\] show 2D slices of ${{\delta T_{\rm b}}}_{,\rm min}$ through the entire parameter space. As in Fig. \[image\_paramz\], the third parameter is chosen such that these slices contain the lowest value of ${{\delta T_{\rm b}}}_{,\rm min}$ obtained within the entire parameter space. In this case, the values for the third parameter are $\beta=-0.5$, $\alpha=-1.5$ and $f_\alpha=100$ which correspond to the left, middle and right panels, respectively. The middle row of panels shows the associated redshift $z({{\delta T_{\rm b}}}_{,\rm min})$ of the minimum of the absorption profiles and the bottom row the corresponding FWHM $\Delta z$.
Panels B$_1$ and B$_2$ show that ${{\delta T_{\rm b}}}_{,\rm min}$ decreases with increasing $f_\alpha$ as the coupling between ${T_{\rm S}}$ and ${T_{\rm K}}$ becomes stronger. However, this also implies that ${\rm {Ly{\alpha}}}$ heating becomes efficient earlier and thus the minima of the absorption profiles appear at higher redshifts when increasing $f_\alpha$ (see panels $C_1, C_2$). As shown in panel $B_3$, ${{\delta T_{\rm b}}}_{,\rm min}$ decreases for lower values of $\alpha$ and $\beta$ which corresponds to stronger excess cooling and also in this case $z({{\delta T_{\rm b}}}_\mathrm{,\rm min})$ shifts towards higher redshifts (panel $C_3$). The CD starts with a lower gas temperature for smaller values of $\alpha$ and $\beta$. As the ${\rm {Ly{\alpha}}}$ heating rate increases for lower temperatures, ${\rm {Ly{\alpha}}}$ heating become efficient earlier for a stronger excess cooling model. These results are consistent with our findings in Section \[sec:main\_scene\].
The bottom row of Fig. \[image\_param\] shows the corresponding FWHM $\Delta z$. The dependence of $\Delta z$ on the parameters is more complex compared to what we saw for ${{\delta T_{\rm b}}}_{,\rm min}$ and $z({{\delta T_{\rm b}}}_{,\rm min})$. Here we have to keep in mind a few aspects. One is that all models reach their equilibrium temperature due to ${\rm {Ly{\alpha}}}$ heating approximately at the same redshift for a fixed ${\rm {Ly{\alpha}}}$ background (as we have seen in Section \[sec:main\_scene\]). Secondly, the initial temperature (at $z=30$) of these models decreases rapidly with stronger cooling parameters. The absorption profile becomes deeper and shifts towards higher redshift for a larger excess cooling rate. On the other hand, the ${\rm {Ly{\alpha}}}$ heating starts earlier and ${\delta T_{\rm b}}$ of these profiles approaches zero at a similar redshift. These two facts make $\Delta z$ decrease initially with the increase of excess cooling rate for a fixed $f_\alpha$ as shown in panel $D_3$. However, $\Delta z$ starts increasing for $\alpha < -1$ and $\beta < 0$ as the initial temperature of these models becomes smaller and the minimum of the absorption profiles shifts towards higher redshifts. On the other hand, ${\rm {Ly{\alpha}}}$ heating becomes efficient earlier for a larger value of $f_\alpha$ which decrease $\Delta z$ for a fixed excess cooling (see panels $D_1$ and $D_2$).
The black lines in Fig. \[image\_param\] correspond to the absorption profile parameters estimated from the EDGES low-band observation. However, as these slices correspond to the minimum ${\delta T_{\rm b}}$ calculated by exploring the whole 3D parameter space, an interpretation of EDGES results from these contours is difficult. We therefore now turn our attention to the parts of the parameter space that are consistent with the EDGES absorption profile.
Interpretation of EDGES low-band results {#sec:edges_interpretation}
----------------------------------------
@EDGES2018 reported a measurement of a 21-cm absorption profile with ${{\delta T_{\rm b}}}_{, \rm min}=-500^{+200}_{-500}$ mK with $z({{\delta T_{\rm b}}}_\mathrm{, min})$ and $\Delta z$ equivalent to $78\pm 1$ MHz and $19^{+4}_{-2}$ MHz, respectively. We will investigate what part of our parameter space agrees with this observation. However, we do not consider the detailed shape of the absorption profile as reported in @EDGES2018, nor use parameter estimation techniques such as Markov chain Monte Carlo. Instead, we consider the values of ${{\delta T_{\rm b}}}_{, \rm min}, z({{\delta T_{\rm b}}}_\mathrm{, min})$ and $\Delta z$ corresponding to the profile to compare with the absorption profiles produced by our model. We would like to remind the reader that we have not included any X-ray heating in this parameter space study.
Fig. \[image\_edges\] presents which values for our parameters agree with the EDGES low band results. The left-hand panel shows a 2D plot for parameters $\alpha$ and $\beta$ where the colour of each point represents the average value of $f_\alpha$ for which the values of $\alpha, \beta$ are consistent with the EDGES observation. We see that two specific ranges of cooling parameters produce the desired profile, the broader of the two bands for a weak ${\rm {Ly{\alpha}}}$ background flux ($f_\alpha < 0.08$) and the narrower one for a very weak ${\rm {Ly{\alpha}}}$ background ($f_\alpha < 0.01$). The broader band is characterized by a strong but not too strong cooling around the redshift of the absorption profile ($-2\lesssim(\mathrm{d}\log T_K/\mathrm{d}\log t)_{\rm cool}\lesssim-1.2)$ and the narrower band by a stronger value of $(\mathrm{d}\log T_K/\mathrm{d}\log t)_{\rm cool}\sim -2.5$ This can also be characterized through the temperature which the IGM would achieve in the absence of ${\rm {Ly{\alpha}}}$ heating. For the cooling parameters in the broader of the two bands, this temperature is between 0.2 and 0.3 K and between 0.05 and 0.06 K for the narrower band. For stronger cooling than shown in the left-hand panel, ${{\delta T_{\rm b}}}_{, \rm min}$ will be lower and will shift towards higher redshifts. Similarly, $z({{\delta T_{\rm b}}}_\mathrm{, min})$ will shift towards higher redshifts for larger values of $f_\alpha$.
The right-hand panel of Fig. \[image\_edges\] shows a 3D representation of our parameter space where the colour indicates the $\chi^2$ value. We define $\chi^2$ error in this plot as $$\chi^2 =\sum_{i=1,3} \left(\frac{M_i-O_i}{\sigma_i}\right)^2,
\label{eq:chi2}$$ where $i$ represents there parameters to define the absorption profile used in this study, $M$ and $O$ are the model and observation parameters respectively and $\sigma_i$ represents the $1\sigma$ error on the measured parameters considered here. One can notice that for a certain choice of cooling parameters a range of $f_\alpha$ values can satisfy the agreement condition. However, all $f_\alpha$ values are low. The isolated region $f_\alpha=0.01$ corresponds to deeper absorption profiles with ${{\delta T_{\rm b}}}_{,\rm min} \lesssim -500 $ mK while the other region has absorption depths ${{\delta T_{\rm b}}}_{,\rm min} \sim -300 $ mK.
One thing to keep in mind that we have ignored all other heating processes such as X-ray heating, etc. If any other additional energy is added to the IGM, the excess cooling would have to compensate for this in order for the absorption profile to remain consistent with the EDGES result. In other words, the cooling rates derived here should be considered as lower limits. For example, the combination $f_\alpha \lesssim 0.1$, $\alpha \sim -1.5$ and $\beta \sim 0.1$ corresponds to the minimum excess cooling required to achieve the strong signal reported by EDGES. This minimum excess cooling rate is similar to the Hubble cooling rate at redshift 17.
Physically motivated cooling model {#sec:DMB}
==================================
So far we have considered a simple phenomenological form for the cooling rate as given by Equation \[eq:ex-cool\]. Now we will consider a physically motivated cooling model based on the interaction between cold dark matter and baryonic particles. Such interactions have the potential to cool the baryonic gas efficiently and explain the EDGES results [@2018Natur.555...71B; @2018arXiv180210094M; @2018arXiv180303091B; @2018PhRvL.121a1101F]. However, most of these interaction scenarios are highly constrained by limits from stellar cooling and fifth force experiments. This rules out scenarios in which the cooling of the gas occurs through Rutherford-like scattering with a dominant component of the dark matter. However, a scenario in which cooling is caused by interactions of electrons and protons with a small ($\sim 1\%$) fraction of millicharged dark matter particles is currently not entirely ruled out [see e.g., @2018arXiv180210094M] although only in a very small part of parameter space [@2018arXiv180303091B].
![[Parameter study for the millicharged dark matter cooling model. The parameters $\epsilon$ and $m_\chi$ represent the charge and the mass of the dark matter particles. The crosses indicate the values of the parameters which produce absorption profiles that agree with the EDGES low band results within a $1\sigma$ error. The colour bar represents the average values of $f_\alpha$ for these models.]{} []{data-label="image_edges_DMB"}](edges_DMB_MC.pdf)
The cooling rate for this scenario can be written as $$\frac{d\log {T_{\rm K}}}{d\log t}\bigg|_{\rm cool,DMB} = \frac{4 \dot{Q_b}}{9~H~{T_{\rm K}}},
\label{eq:ex-coolDMB}$$ where the cooling rate of the baryon $\dot{Q_b}$ can be expressed as the sum of the contributions due to collisions with electrons or protons as targets $t$ [@2018arXiv180210094M], $$\begin{aligned}
\dot{Q_b} \!\!\!\! & = & \!\!\!\! \frac{n_{\chi} x_e}{1+f_{\rm He}}\sum_{t=e,p}\frac{m_t m_{\chi}}{\left(m_t+m_\chi\right)^2}\frac{ \bar{\sigma}_t}{u_{{\rm th},t}}
\nonumber\\
&\times& \!\!\!\! \left[ \sqrt{\frac{2}{\upi}}\frac{ e^{-r_t^2/2} }{u^2_{{\rm th},t}}\left(T_\chi - {T_{\rm K}}\right) + m_{\chi}\frac{F(r_t)}{r_t} \right].
\label{eq:ex-coolDMB1}\end{aligned}$$ Here, $x_e$ is the residual electron fraction after recombination and $f_{\rm He}\approx0.08$ is the primordial helium fraction. The symbol $m$ stands for mass, where $e$, $p$ and $\chi$ stand for electron, proton and dark matter, respectively. The number density of millicharged dark matter is $n_{\chi} = f_{\rm dm}\times \rho_d/m_{\chi}$ where $\rho_d$ is the dark matter mass density and $f_{\rm dm}$ is the fraction of millicharged dark matter. ${T_{\rm K}}$ and $T_\chi$ represent the temperatures of the baryon gas and the dark matter respectively. The function $F(r_t)$ is defined as $$F(r_t) = {\rm Erf}(\frac{r_t}{\sqrt{2}})-\sqrt{\frac{2}{\upi}}r_t e^{-r^2_t/2}\,,$$ where $r_t = v_{\chi b}/u_{{\rm th},t}$, with $v_{\chi b}$ the relative velocity between the baryons and the dark matter and $u^2_{{\rm th},t} = {T_{\rm K}}/m_t + T_\chi/m_\chi$ the (iso)thermal sound speed of the DM-t fluid. Finally, the rate also depends on the interaction cross-section between the millicharged dark matter particles and a target $t$, $\bar{\sigma_t}$. We define a charge parameter $\epsilon=e_{\chi}/e$ where $e_{\chi}$ and $e$ are the dark matter and electron charge respectively. The interaction cross-section is assumed to scale with relative velocity as $\bar{\sigma_t} = \frac{2\upi \alpha^2_{f}\epsilon^2\xi}{\mu^2_{\chi t}v_{\chi b}^4}$ where $\alpha_{f}$ is the fine-structure constant, $\xi$ is the Debye logarithm and $\mu^2_{\chi t}$ is the reduced mass of the dark matter and target. We refer to @2018arXiv180210094M for more details on the various terms in Equation \[eq:ex-coolDMB1\].
We follow the approach as in @2018arXiv180210094M and solve the set of differential equations to track the evolution of the temperatures of the gas and dark matter. We initialize ${T_{\rm K}}= {T_{\gamma}}$ and $T_\chi = 0$ at $z=1010$. We assume that the initial distribution of $v_{\chi b, 0}$ is Gaussian with an root-mean-square value of 29 km s$^{-1}$. We solve the set of equations for many values $v_{\chi b, 0}$ taken from this distribution and in the end estimate the velocity averaged gas temperatures and brightness temperature.
We will vary two parameters for the cooling rate as described in Equation \[eq:ex-coolDMB\], namely $\epsilon$, the charge of the dark matter particles and their mass $m_{\chi}$. We keep the fraction of millicharged dark matter fixed at $f_{\rm dm}=0.01$ throughout this study. Figure \[image\_global\_DMB\] shows the redshift evolution of the gas temperature (left panel) and the brightness temperature (right panel) for three different combinations of $m_\chi$ and $\epsilon$. The thin lines show the results with ${\rm {Ly{\alpha}}}$ coupling but [*without*]{} ${\rm {Ly{\alpha}}}$ heating and the thick lines include ${\rm {Ly{\alpha}}}$ heating. We use $f_\alpha=1$, except for the dot-dashed line which has $f_\alpha=0.1$. These results illustrate the trends associated with the different parameters.
It should first of all be noted that these results are very similar to those shown in Fig. \[image\_global\]. In absence of ${\rm {Ly{\alpha}}}$ heating, a cooling rate with the combination of $m_\chi \sim 2$ MeV and $\epsilon\sim 10^{-6}$ cools the gas to a temperature $\sim 4$ K at redshift 17 which is sufficient to produce a signal which agrees with the EDGES low-band observations, consistent with the results of @2018arXiv180303091B [@2018arXiv180210094M]. However, as expected, the ${\rm {Ly{\alpha}}}$ heating prevents the gas temperature to reach such a low value, even for $f_\alpha=1$. The cooling, as well as the signal, becomes stronger for larger values of $\epsilon$ as this raises the interaction cross-section. On the other hand, increasing $m_{\chi}$ lowers the cooling rate and the signal. As above, we find that the absorption profile shifts towards lower redshifts for lower values of $f_\alpha$. These trends suggest that this cooling model might satisfy the EDGES results for the following two cases: (i) a higher cooling rate than estimated by @2018arXiv180303091B [@2018arXiv180210094M] which can arise due to either a larger $\epsilon$ or a smaller $m_{\chi}$, (ii) a lower ${\rm {Ly{\alpha}}}$ flux to keep the heating low at $z\sim 17$.
Next, we vary the parameter $\epsilon$ from $10^{-7}$ to $10^{-4}$ and $m_\chi$ from 0.5 MeV to 1 GeV while we keep the range of $f_\alpha$ the same as used previously. Equivalent to Fig. \[image\_edges\] for the phenomenological model, Fig. \[image\_edges\_DMB\] presents the parts of parameter space which agree with the EDGES measurements. Note that similar to the earlier case, we have not considered any heating mechanism other than heating due to scattering of the ${\rm {Ly{\alpha}}}$ photons. As expected, we find the required ${\rm {Ly{\alpha}}}$ flux in this scenario has to be less than 1 as shown by the color bar. Note however that the $f_\alpha$ values found are higher than what we obtained for the phenomenological cooling model. This suggests that the millicharged dark matter cooling process produces a larger cooling rate at redshift $\sim 17$ than the explored range in the previous cooling model.
While @2018arXiv180303091B [@2018arXiv180210094M] conclude that $\epsilon \gtrsim 10^{-6}$ will be required for $m_\chi=2$ MeV to reach agreement with the EDGES results, Fig. \[image\_edges\_DMB\] suggests a larger value of $\epsilon \gtrsim 1.5\times 10^{-5}$ for the same dark matter mass. The required ${\rm {Ly{\alpha}}}$ flux for these $m_\chi$ and $\epsilon$ values corresponds to $f_\alpha \sim 0.3$. Clearly, a larger dark matter-baryon interaction cross-section is required when ${\rm {Ly{\alpha}}}$ heating is taken into account.
However, the possible parameter space of the millicharge model in Fig. \[image\_edges\_DMB\] that can explain the EDGES result is disfavoured by the constraints from stellar and super- nova cooling, big bang nucleosynthesis and a range of particle physics experiments. Specifically, as shown in fig. 4 in @2018arXiv180303091B, these constraints require $m_\chi\gtrsim 10$ MeV and for these values we do not find any solutions that are consistent with the EDGES results. We therefore conclude that the millicharged dark matter model no longer offers a viable explanation for the absorption signal claimed by the EDGES team.
Discussions & Conclusions {#sec:con}
=========================
In this study we have considered the impact of the heating from resonance scattering of ${\rm {Ly{\alpha}}}$ photons in the IGM during the Cosmic Dawn on models with excess cooling constructed to explain the deep absorption feature around $z\sim 17$ reported by the EDGES team. This heating is an inevitable result of the resonance scattering which is needed to couple the spin temperature to the gas temperature, the only known process which can produce an observable 21-cm signal from the IGM at these redshifts. The required excess cooling requires new physics and thus its cause remains uncertain. We explored two possibilities, one simple phenomenological form of cooling and one physically motivated one relying on the interaction of putative millicharged dark matter particles with protons and electrons.
For these two scenarios we investigate the evolution of the average differential brightness temperature of the 21-cm signal. We explore a three-dimensional parameter space defined by two parameters describing the excess cooling ($\alpha$ and $\beta$ for the phenomenological model; $\epsilon$ and $m_\chi$ for the millicharged dark matter) and one parameter setting the strength of the ${\rm {Ly{\alpha}}}$ background ($f_\alpha$) to study the global 21-cm signal from the CD. The main findings of the paper are listed below.
Without any excess cooling, ${\rm {Ly{\alpha}}}$ heating can start heating the IGM as early as redshift 22 for a typical emissivity of $\sim 10000$ photons per baryon between ${\rm {Ly{\alpha}}}$ and the Lyman limit. Although this heating rate is smaller than the usually assumed X-ray heating rates, it can still increase the gas temperature to several tens of K which is the equilibrium temperature between the heating by the continuum photons and cooling by the injected photons. This is consistent with previous studies such as @2007ApJ...655..843C. For this case, we find an absorption signal of depth $\sim -100$ mK at redshift $\sim 20$.
When including excess cooling, the Cosmic Dawn starts with a very cold IGM. In such cases, ${\rm {Ly{\alpha}}}$ heating becomes efficient earlier and rapidly increases the IGM temperature to the equilibrium temperature. For these cases, we find absorption signals which can be factors 3 – 4 deeper than without excess cooling.
The exploration of the parameter space of the excess cooling rate parameters and the ${\rm {Ly{\alpha}}}$ background shows that the EDGES low-band results can only be reproduced for strong excess cooling combined with a weak ${\rm {Ly{\alpha}}}$ background. This puts an upper bound on the background ${\rm {Ly{\alpha}}}$ flux which is $\sim$15 times lower than our fiducial choice for the phenomenological cooling model, while for the millicharged dark matter model this upper limit is only a factor $\sim 2$ below the fiducial value. Thus the sources at redshift $\sim 17$ emit fewer ${\rm {Ly{\alpha}}}$ photons or the star formation efficiency is lower than expected. This result disagrees with the findings of @2019MNRAS.483.1980M who claim that the star formation efficiency should be higher than expected in order to produce the strong ${\rm {Ly{\alpha}}}$ background needed to achieve strong coupling between the spin and gas temperatures. However, these authors did not consider the effect of ${\rm {Ly{\alpha}}}$ heating.
Although we find that the millicharged dark matter model can reproduce the EDGES results for a relatively low ${\rm {Ly{\alpha}}}$ background and some combinations of DM charge and mass, these combinations are actually ruled out by constraints from stellar and supernova cooling, big bang nucleosynthesis and a range of particle physics experiments [@2018arXiv180303091B]. Including ${\rm {Ly{\alpha}}}$ heating therefore removes this model as a feasible explanation for the EDGES results.
In our exploration of the parameter space for the phenomenological model we frequently found interesting models at the edge of the parameter ranges that we considered. We did not explore a larger range of values as the trend is quite clear: only fairly strong cooling which without ${\rm {Ly{\alpha}}}$ heating would give gas temperatures below $\sim 0.3$ K around $z\approx 17$ combined with a weak ${\rm {Ly{\alpha}}}$ background ($f_\alpha \lesssim 0.06$) can reproduce the EDGES low-band results. Possibly even stronger cooling with an even weaker ${\rm {Ly{\alpha}}}$ background would also give consistent results but such models become increasingly unlikely.
In general, ${\rm {Ly{\alpha}}}$ heating works against all kind of excess cooling models that might explain the EDGES result and will potentially provide strong bounds on their parameters. The same is true for the alternative solutions which rely on a stronger radiation background at the Rayleigh-Jeans tail of the CMB [see e.g., @2018PhRvL.121c1103P]. However, we leave the study of the impact of ${\rm {Ly{\alpha}}}$ heating on those types of models to a future study.
We did not explore the impact of changing the source population. In our models, all halos with a virial temperature above $10^4$ K contribute to the ${\rm {Ly{\alpha}}}$ background. Obviously increasing this limit would also reduce the background and possibly lead to models in which fiducial values for $f_\alpha$ combined with strong excess cooling could reproduce the EDGES low-band absorption profile. Lowering the minimum virial mass would only increase the ${\rm {Ly{\alpha}}}$ background and thus require even lower values for $f_\alpha$. We also did not explore the impact of the star formation efficiency parameter $f_\star$ and the SED. However, for the redshift regime which we explore these parameters are degenerate with $f_\alpha$.
We thus find that heating due to resonance scattering with Lyman series photons may have a significant impact during the Cosmic Dawn and thus should be taken into account when modelling the 21-cm signal. Although we do find that for a fiducial value of X-ray heating ($f_\mathrm{X}=1$), the ${\rm {Ly{\alpha}}}$ heating is subdominant, many authors explore a wide range of values for $f_\mathrm{X}$ including low values for which ${\rm {Ly{\alpha}}}$ heating will dominate over X-ray heating [e.g @2017MNRAS.472.1915C; @2018MNRAS.477.3217G; @2019arXiv190110943M]. We note that none of these papers actually include the effect of ${\rm {Ly{\alpha}}}$ heating.
Models to explain the absorption feature seen in the EDGES results rely on ${\rm {Ly{\alpha}}}$ coupling to produce an observable signal and thus any excess cooling needs to overcome the heating caused by this coupling. As shown in this paper, this pushes for example the millicharged dark matter model into a regime ruled out by other constraints. It remains to be seen if there exist any physically motivated excess cooling processes which can explain the EDGES results.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors would like to thank Hannah Ross, Paul Shapiro, Anastasia Fialkov, Sambit Giri, Benedetta Ciardi, Avery Meiksin, Piero Madau, Tirthankar Roy Choudhury for useful discussions regarding this work. We also like to thank an anonymous referee whose comments have encouraged us to explore the millicharged dark matter model. We have used resources provided by the Swedish National Infrastructure for Computing (SNIC) (proposal number SNIC 2018/3-40) at PDC, Royal Institute of Technology, Stockholm.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: http://www.lofar.org/
[^3]: http://www.gmrt.tifr.res.in
[^4]: http://eor.berkeley.edu/
[^5]: http://www.mwatelescope.org/
[^6]: http://www.skatelescope.org/
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the derived categories of twisted supersingular K3 surfaces. We prove a derived crystalline Torelli theorem for twisted supersingular K3 surfaces, characterizing Fourier-Mukai equivalences in terms of the twisted K3 crystals introduced in [@BL17]. This is a positive characteristic analog of the Hodge-theoretic derived Torelli theorem of Orlov [@MR1465519] and its extension to twisted K3 surfaces by Huybrechts and Stellari [@HS04; @MR2310257]. We give applications to various questions concerning Fourier-Mukai partners, extending results of Căldăraru [@Cal01] and Huybrechts and Stellari [@HS04]. We also give an exact formula for the number of twisted Fourier-Mukai partners of a twisted supersingular K3 surface.'
author:
- Daniel Bragg
bibliography:
- 'biblio.bib'
title: Derived Equivalences of Twisted Supersingular K3 Surfaces
---
Introduction
============
If $X$ is a smooth projective variety, we write $D(X)$ for the bounded derived category of coherent sheaves on $X$. This object is studied as an invariant attached to $X$. The particular case of K3 surfaces over the complex numbers has received extensive attention, and has proven to be both interesting and tractable. A major result is Orlov’s Torelli theorem [@MR1465519] for the derived category, which characterizes derived equivalences between K3 surfaces over the complex numbers in terms of isomorphisms of certain associated Hodge structures. This result was extended to the case of twisted K3 surfaces by Căldăraru [@Cal01] and Huybrechts and Stellari [@HS04; @MR2310257]. These results allow difficult questions about derived equivalences to be rephrased in terms of Hodge structures, which are much easier to compute with.
A major obstacle to obtaining similar results for K3 surfaces in positive characteristic is that the classical Torelli theorem no longer applies (this result is a key input to both Orlov’s and Huybrechts and Stellari’s derived Torelli theorems). A partial replacement was found by Lieblich and Olsson in [@LO15], who introduced a certain filtration on the derived category. They used this to extend results on derived equivalences of K3 surfaces to positive characteristic by lifting to characteristic 0, for instance, the finiteness of the number of Fourier-Mukai partners. In this paper, we focus our attention on supersingular K3 surfaces, which are a particularly interesting class of K3 surfaces in positive characteristic. By work of Ogus [@Ogus78; @Ogus83] and Rudakov and Shafarevich [@RTS83], it is known that supersingular K3 surfaces satisfy a certain Torelli-type result, phrased in terms of crystalline cohomology. The form of this result (reviewed in Section \[sec:k3crystals\]) is very similar to that of the classical Hodge-theoretic Torelli theorem.
We use this crystalline Torelli theorem to investigate the derived categories of twisted supersingular K3 surfaces. We are able to obtain similarly strong results as are known for K3 surfaces over the complex numbers, although our methods are necessarily limited to the supersingular case. Our key technical tool is a certain crystalline analog of the twisted Hodge structures of Huybrechts and Stellari [@HS04], introduced in [@BL17]. This construction attaches to a twisted K3 surface $(X,\alpha)$ a certain K3 crystal $\tH(X/W,B)$. We show that a derived equivalence of twisted K3 surfaces induces an isomorphism of the corresponding K3 crystals, which preserves certain extra structure. We also show that the converse holds. Our main result is the following crystalline Torelli theorem for the derived category of a twisted supersingular K3 surface.
If $(X,\alpha)$ and $(Y,\beta)$ are twisted supersingular K3 surfaces, then there exists a Fourier-Mukai equivalence $D(X,\alpha)\cong D(Y,\beta)$ if and only if there exists an isomorphism $\tH(X/W,B)\cong\tH(Y/W,B')$ of K3 crystals.
That is, the twisted K3 crystal exactly captures the relationship of being Fourier-Mukai partners. We prove first a more refined statement (Theorem \[thm:Main1\]) characterizing which maps of K3 crystals are induced by derived equivalences.
To prove this result, we need to extend various related constructions and results to positive characteristic, such as the existence of certain twisted Chern characters and moduli spaces of twisted sheaves. Much of this has been already taken care of in [@BL17]. We review the needed definitions and results below. In two appendices, we show that our twisted Chern characters take integral values, and that derived equivalences of certain twisted K3 surfaces in positive characteristic are orientation preserving, extending results of Huybrechts, Macrì, and Stellari [@MR2553878] and Reinecke [@2017arXiv171100846R].
Our main theorem allows the translation of questions about derived equivalences into questions about isomorphisms of supersingular K3 crystals. By results of Ogus (reviewed in Section \[sec:characteristicsubspaces\]), supersingular K3 crystals are essentially objects of semi-linear algebra, and hence are quite easy to compute with. Using these techniques, we apply our main theorem in Sections \[sec:characteristicsubspaces\], \[sec:applications\], and \[sec:counting\] to prove various facts about derived equivalences of twisted supersingular K3 surfaces. For instance, we recover the result of Lieblich and Olsson [@LO15] that (non-twisted) supersingular K3 surfaces are derived equivalent if and only if they are isomorphic (see Theorem \[thm:Untwisted1\]). We also prove that if $(X,\alpha)$ is a twisted supersingular K3 surface then $D(X,\alpha)\cong D(X,\alpha^k)$ for any integer $k$ coprime to the order of $\alpha$ (Theorem \[thm:caldararusthing\]). Finally, in sections \[sec:orthogonal\] and \[sec:counting\] we use characteristic subspaces to count twisted Fourier-Mukai partners. The final result is Theorem \[Count2\], which gives the following formula for the number of twisted partners of a twisted supersingular K3 surface $(X,\alpha)$ in terms of two discrete invariants $\sigma_0,m$ attached to $(X,\alpha)$. $$\#\FM(X,\alpha)=
\begin{dcases}
\frac{p^{\sigma_0}+1}{p^m+1}(p^{\sigma_0-1}-1)+1 &\mbox{ if }1\leq\sigma_0\leq 10\\
\frac{p^{\sigma_0}+1}{p^m+1}(p^{\sigma_0-1}-1) &\mbox{ if }\sigma_0=11\\
\end{dcases}$$ We remark that Lieblich and Olsson [@LO15.2] have also used the crystalline Torelli thoerem to study the derived categories of K3 surfaces in positive characteristic. They obtain results for general K3 surfaces in positive characteristic by deformation to the supersingular locus, replacing the lifting arguments of [@LO15].
Notation
--------
We fix throughout an algebraically closed field $k$ of characteristic $p>0$. Unless otherwise state, we will assume that $p\neq 2$. If $X$ is a variety over $k$, the notation $\H^i(X,\m_{p^n})$ will always be used to denote the cohomology of $\m_{p^n}$ with respect to the flat (fppf) topology.
Acknowledgements
----------------
The results of this paper were obtained during the course of the author’s PhD work at the University of Washington. The author thanks Nicolas Addington, Ben Antieau, Max Lieblich, Martin Olsson, and Emanuel Reinecke for useful conversations and correspondences.
Supersingular K3 surfaces
=========================
Let $X$ be a K3 surface over $k$. We say that $X$ is *supersingular* if $X$ has Picard number 22.[^1] There is a beautiful theory of such surfaces, with many special results and techniques. Their systematic study was initiated by Artin [@Artin74]. It is known that the discriminant of the Néron-Severi group of a supersingular K3 surface is of the form $-p^{2\sigma_0}$ for some integer $1\leq\sigma_0\leq 10$, known as the *Artin invariant* of $X$. The moduli space of supersingular K3 surfaces is $9$ dimensional, and the locus of surfaces $X$ such that $\sigma_0(X)\leq\sigma_0$ is closed of dimension $\sigma_0-1$. In particular, the locus of supersingular K3 surfaces with $\sigma_0=1$ is discrete, and in fact there is up to isomorphism a unique such surface.
The Brauer group
----------------
We briefly describe the Brauer group of a supersingular K3 surface. If $X$ is a K3 surface over the complex numbers, or more generally over an algebraically closed field of characteristic 0, then there is an abstract isomorphism of groups $$\Br(X)\cong\left(\bQ/\bZ\right)^{\oplus 22-\rho(X)}$$ (see for instance Example 18.1.7ff of [@Huy16]). In particular, the Brauer group is divisible and countable. Suppose that $X$ is a K3 surface over an algebraically closed field $k$ of positive characteristic. If $X$ has finite height, then $\Br(X)$ is again divisible and countable, and has a similar form. If $X$ is supersingular, then the situation is completely different. Artin showed in [@Artin74] that there is an abstract isomorphism of groups $$\Br(X)\cong k$$ between the Brauer group and the additive group of the underlying field $k$. In particular, $\Br(X)$ is not divisible, is $p$-torsion, and has the same cardinality as $k$. The explanation for this distinctive behavior comes from flat cohomology. The flat Kummer sequence gives a short exact sequence $$0\to\Pic(X)/p\Pic(X)\to\H^2(X,\m_p)\to\Br(X)[p]\to 0$$ Artin showed that $\H^2(X,\m_p)$ is the $k$-points of a certain natural algebraic group, and defined the subgroup $\U^2(X,\m_p)\subset\H^2(X,\m_p)$ to be the $k$-points of the connected component of the identity (we record an alternative characterization of the subgroup $\U^2(X,\m_p)$ in Definition \[def:alternatedef\]). The completion of this algebraic group at the identity is identified with the formal Brauer group of $X$, which is isomorphic to $\widehat{\bG}_a$. This implies that we have an isomorphism $\U^2(X,\m_p)\cong\bG_a(k)$. Let $N(X)=\Pic(X)$ be the Néron-Severi group of $X$. Artin showed that the Kummer sequence induces a diagram $$\label{eq:artinssubgroup}
\begin{tikzcd}
0\arrow{r}&\dfrac{pN(X)^{\vee}}{pN(X)}\arrow{r}\arrow[hook]{d}&\U^2(X,\m_p)\arrow{r}\arrow[hook]{d}&\Br(X)\isor{d}{}\arrow{r}&0\\
0\arrow{r}&\dfrac{N(X)}{pN(X)}\arrow{r}&\H^2(X,\m_p)\arrow{r}&\Br(X)\arrow{r}&0
\end{tikzcd}$$ In particular, the map $\U^2(X,\m_p)\to\Br(X)$ is surjective, explaining the profusion of $p$-torsion Brauer classes in the supersingular case.
Despite the very non-classical behavior of the Brauer group in the supersingular case, we will see that many structures associated to twisted K3 surfaces over the complex numbers have analogs for supersingular K3 surfaces.
K3 crystals and Ogus’s crystalline Torelli theorem {#sec:k3crystals}
--------------------------------------------------
Let $W$ denote the ring of Witt vectors of $k$, let $K=W[p^{-1}]$ be its field of fractions, and let $\sigma\colon W\to W$ be its Frobenius. If $X$ is any K3 surface over $k$ (not necessarily supersingular), then the second crystalline cohomology group $\H^2(X/W)$ of $X$ is a free $W$-module of rank 22. It is equipped with a perfect pairing $$\H^2(X/W)\otimes_W\H^2(X/W)\to W$$ and an endomorphism $\Phi\colon \H^2(X/W)\to \H^2(X/W)$ which is $\sigma$-linear,[^2] and satisfies a certain compatibility with the pairing. This object may be viewed as a positive characteristic analog of the singular cohomology group $\H^2(X,\bZ)$ of a K3 surface over the complex numbers, equipped with its pairing and Hodge structure. Following Ogus (see [@Ogus83 Definition 1]), we abstract the above structure in the following definition.
\[def:K3crystal\] A *K3 crystal of rank $n$* is a $W$-module $H$ of rank $n$ equipped with a $\sigma$-linear endomorphism $\Phi\colon H\to H$ and a symmetric bilinear form $H\otimes_WH\to W$ such that
1. $p^2H\subset\Phi(H)$,
2. $\Phi\otimes k$ has rank 1,
3. the pairing $\langle\_,\_\rangle$ is perfect,
4. $\langle\Phi(x),\Phi(y)\rangle=p^2\sigma\langle x,y\rangle$ for all $x,y\in H$, and
5. the crystalline discriminant of $(H,\Phi)$ is $-1$.
We will use the notation $\varphi=p^{-1}\Phi$.
The second crystalline cohomology group $\H^2(X/W)$ of a K3 surface equipped with the cup product pairing and the endomorphism $\Phi$ is satisfies the above properties, and is the basic example of a K3 crystal (of rank 22) (see for instance [@Ogus83]).
The *Tate module* of a K3 crystal $H$ is $$T_H=\left\{h\in H|\Phi(h)=ph\right\}$$ This has a natural structure of $\bZ_p$-module, and is equipped with the restriction of the bilinear form on $H$.
We recall some abstract structural results. An $F$-*crystal* is a pair $(H,\Phi)$ where $H$ is a free $W$-module of finite rank and $\Phi\colon H\to H$ is a $\sigma$-linear endomorphism. A map $(H,\Phi)\to (H',\Phi')$ of $F$-crystals consists of a map $H\to H'$ of $W$-modules that commutes with the respective semilinear operators. Such a map is said to be an *isogeny* if it induces an isomorphism $H\otimes K\to H'\otimes K$. There is a well known classification of $F$-crystals up to isogeny in terms of slopes, due to Dieudonné-Manin. Note that any K3 crystal gives rise in particular to an $F$-crystal by forgetting the pairing. The isogeny type of this crystal is constrained by the extra structure imposed above, and gives rise to an invariant of the K3 crystal. We will mostly be interested in the following case in this paper.
A K3 crystal $H$ is *supersingular* if it has pure slope 1.
Equivalently, a K3 crystal $H$ is supersingular exactly when the inclusion $T_H\otimes W\to H$ is an isogeny (that is, the map $T_H\otimes K\to H\otimes K$ is an isomorphism). If $X$ is a K3 surface, then the K3 crystal $\H^2(X/W)$ is supersingular if and only if $X$ is supersingular.
If $H$ is supersingular, then by Proposition 3.13 of [@Ogus78] the discriminant of the pairing on $T_H$ is equal to $-p^{2\sigma_0}$ for some integer $\sigma_0\geq 1$.
\[def:artininvariant1\] The *Artin invariant* $\sigma_0(H)$ of a supersingular K3 crystal $H$ is the integer $\sigma_0$.
It is immediate that if $X$ is a supersingular K3 surface, then the Artin invariant of $X$ is equal to the Artin invariant of $\H^2(X/W)$. Suppose that $X$ and $Y$ are supersingular K3 surfaces. An obvious necessary condition for a map $\H^2(X/W)\iso\H^2(Y/W)$ of K3 crystals to be induced by a morphism is that it map an ample class to an ample class. It is easy to see that this need not always be the case. In fact, it need not even be the case that the Picard group of $X$ is mapped to the Picard group of $Y$ (unlike in the complex case, where this is automatic). We will therefore need to track the inclusion of the Picard group as a piece of extra data.
\[def:marking\] A *marked K3 crystal* is a pair $(H,N)$ where $H$ is a K3 crystal and $N$ is a lattice $N$ equipped with an isometric inclusion $N\to H$ that induces an isomorphism $N\otimes\bZ_p\iso T_H$.
An isomorphism of marked K3 crystals $(H,N)\to(H',N')$ consists of an isomorphism $H\to H'$ of K3 crystals that restricts to an isomorphism $N\to N'$.
The first Chern class map $\pic(X)\to\H^2(X/W)$ gives a marking of $\H^2(X/W)$. The following is Ogus’s crystalline Torelli theorem.[^3]
\[thm:Ogus1\][@Ogus83 Theorem II] Let $X$ and $Y$ be supersingular K3 surfaces. If $$f\colon (\H^2(X/W),\pic(X))\iso (\H^2(Y/W),\pic(Y))$$ is an isomorphism of marked supersingular K3 crystals such that the map $\pic(X)\to\pic(Y)$ sends an ample class to an ample class, then $f$ is induced by a unique isomorphism $X\iso Y$.
Somewhat surprisingly, it turns out that if we are only interested in the existence of an isomorphism, we may neglect the marking entirely.
\[thm:Ogus2\][@Ogus83 Theorem I] If $X$ and $Y$ are supersingular K3 surfaces, then $X$ is isomorphic to $Y$ if and only if there exists an isomorphism $\H^2(X/W)\iso\H^2(Y/W)$ of K3 crystals.
Twisted K3 crystals and derived equivalences
============================================
In the following sections we present the constructions and definitions needed to state our main theorem.
Crystalline B-fields
--------------------
We review the construction of crystalline B-fields defined in Section 3.4 of [@BL17]. Let $X$ be a K3 surface over $k$. We begin by relating the flat cohomology of $\m_{p}$ to the de Rham cohomology of $X$. Consider the Kummer sequence $$1\to\m_{p}\to\bG_m\xrightarrow{x\mapsto x^{p}}\bG_m\to 1$$ which is exact in the flat topology. Let $\varepsilon\colon X_{\fl}\to X_{\etale}$ be the natural map from the big fppf site of $X$ to the small étale site of $X$. By a theorem of Grothendieck, the cohomology of the complex $\R\varepsilon_*\bG_m$ vanishes in all positive degrees. Because $X$ is smooth over a perfect scheme, the restriction of $\m_{p}$ to the small étale site of $X$ is trivial. Applying $\varepsilon_*$ to the Kummer sequence, we obtain an exact sequence $$1\to\bG_m\xrightarrow{x\mapsto x^{p}}\bG_m\to\R^1\varepsilon_*\m_{p}\to 1$$ of sheaves on the small étale site of $X$. It follows that $$\R^1\varepsilon_*\m_{p}=\bG_m/\bG_m^{p}$$ where the quotient is taken in the étale topology. We therefore obtain isomorphisms $$\label{eq:flattoetale}
\H^i(X_{\fl},\m_{p})\iso\H^{i-1}(X_{\etale},\bG_m/\bG_m^{p})$$ We consider the map of étale sheaves $$\label{eq:dlog}
d\log\colon\bG_m\to\Omega^1_X$$ defined by $f\mapsto df/f$. This map sends any $p$-th power to 0, and hence descends to the quotient. Combined with the isomorphism (\[eq:flattoetale\]), we find a map $$\label{eq:dlog2}
d\log\colon\H^2(X,\m_p)\hto\H^2_{dR}(X/k)$$ (the injectivity follows for instance from Corollaire 2.1.18 of [@Ill79]). The essential property of this map that we will need is that it is compatible with the de Rham Chern character, in the sense that the diagram $$\begin{tikzcd}
\H^1(X,\G_m)\arrow{dr}{c_1^{dR}}\arrow{d}&\\
\H^2(X,\m_p)\arrow[hook]{r}{d\log}&\H^2_{dR}(X/k)
\end{tikzcd}$$ commutes, where the vertical arrow is the boundary map induced by the Kummer sequence. As the third crystalline cohomology group of $X$ is torsion free, the canonical map $$\H^2(X/W)\otimes k\to\H^2_{dR}(X/k)$$ is an isomorphism. Combined with (\[eq:dlog2\]), we arrive at a diagram $$\label{eq:crystallinediagram}
\begin{tikzcd}[row sep=small]
\H^2(X/W)\arrow[two heads]{r}{\pi}&\H^2_{dR}(X/k)&\\
&\H^2(X,\m_{p})\arrow[phantom]{u}{\rotatebox{90}{$\subset$}}\arrow[two heads]{r}&\Br(X)[p]
\end{tikzcd}$$ Write $\H^2(X/K)=\H^2(X/W)\otimes_WK$.
Let $\alpha\in\Br(X)[p]$ be a $p$-torsion Brauer class. A *B-field lift* of $\alpha$ is an element $B=\frac{a}{p}\in\H^2(X/K)$, where $a$ is an element of $\H^2(X/W)$ whose image in $\H^2_{dR}(X/k)$ is contained in the image of $d\log$, and maps to $\alpha$ under the lower horizontal map of (\[eq:crystallinediagram\]).
We note that, unlike in the Hodge-theoretic setting, it is no longer true that every class in the rational crystalline cohomology group is a B-field lift of some Brauer class. We give in [@BL17] the following characterization of B-fields.
[@BL17 Lemma 3.4.11]\[lem:Bfields!\] A class $B\in p^{-1}\H^2(X/W)$ is a B-field lift of a class in $\Br(X)[p]$ if and only if $$B-\varphi(B)\in\H^2(X/W)+\varphi(\H^2(X/W))$$ (recall that $\varphi=p^{-1}\Phi$)
It is possible to make a similar definition of a B-field lift associated to a class in $\Br(X)[p^n]$ by replacing de Rham cohomology with de Rham-Witt cohomology. One can show that a similar characterization to that of Lemma \[lem:Bfields!\] holds. As the Brauer group of a supersingular K3 surface is $p$-torsion, the above definition will suffice for our applications in this paper.
The twisted Mukai crystal
-------------------------
In this section we will show how to associate to a pair $(X,\alpha)$, where $X$ is a K3 surface and $\alpha\in\Br(X)[p]$, a certain K3 crystal of rank 24. The material in this section is essentially contained in Section 3.4 of [@BL17], although with slightly different notation. We begin by defining the Mukai crystal associated to a K3 surface $X$, as originally introduced in [@LO15]. Let $W(1)$ denote the $F$-crystal with underlying $W$-module $W$ and Frobenius action given by $\frac{1}{p}\sigma$. For any $F$-crystal $M$ and integer $n$, we set $M(n)=M\otimes W(1)^{\otimes n}$.
\[def:MukaiCrystal\] The *Mukai crystal* of $X$ is the $W$-module $$\tH(X/W)=\H^0(X/W)(\text{-}1)\oplus \H^2(X/W)\oplus \H^4(X/W)(1)$$ equipped with the twisted Frobenius $\tPhi:\tH(X/W)\to\tH(X/W)$. We define the *Mukai pairing* on $\tH(X/W)$ by $$\langle(a,b,c),(a',b',c')\rangle=bb'-ac'-a'c$$
Both $\H^0(X/W)$ and $\H^4(X/W)$ are canonically isomorphic (as $W$-modules) to $W$. Under these identifications, the twisted Frobenius action is given by $$\tPhi(a,b,c)=(p\sigma(a),\Phi(b),p\sigma(c))$$ It follows immediately from the definitions that $\tH(X/W)$ is a K3 crystal of rank 24.
We next explain how to twist the crystal $\tH(X/W)$ by a B-field. Given any element $B\in\H^2(X/K)$ (for instance, a B-field), cupping with the class $e^B=(1,B,B^2/2)$ defines an isometry $$e^B\colon \tH(X/K)\to\tH(X/K)$$ where $\tH(X/K)=\tH(X/W)\otimes_WK$. Explicitly, this map is given by $$e^B(a,b,c)=\left(a,b+aB,c+b.B+a\frac{B^2}{2}\right)$$
\[def:twistedK3crystal\] Let $\alpha$ be a class in $\Br(X)[p]$ and let $B$ be a B-field lift of $\alpha$. The *twisted Mukai crystal* associated to $(X,\alpha)$ and $B$ is $$\tH(X/W,B)=e^B\tH(X/W)\subset \tH(X/K)$$
In [@BL17], we show that $\tH(X/W,B)$ has a natural structure of K3 crystal.
[@BL17 Proposition 3.4.15]\[thm:itsacrystal\] The submodule $\tH(X/W,B)\subset\tH(X/K)$ is preserved by the Frobenius endomorphism $\tPhi$. The $W$-module $\tH(X/W,B)$ equipped with the endomorphism $\tPhi$ and the restriction of the Mukai pairing on $\tH(X/K)$ is a K3 crystal of rank 24, and is supersingular if and only if $X$ is supersingular.
The non-obvious part of this result is that $\tH(X/W,B)$ is preserved by $\tPhi$, which follows from the characterization of B-fields given in Lemma \[lem:Bfields!\].
Although the embedding $\tH(X/W,B)\subset\tH(X/K)$ does depend on the choice of B-field, we will show that the isomorphism class of the crystal $\tH(X/W,B)$ is independent of the choice of B-field.
\[lem:independent\] If $B$ and $B'$ are two different B-field lifts of $\alpha$, then there exists an isomorphism of K3 crystals $\tH(X/W,B)\cong\tH(X/W,B')$.
While cupping with $e^{B'-B}$ defines an isometry between $\tH(X/W,B)$ and $\tH(X/W,B')$, this map does not in general commute with the Frobenius operators, and so does not define an isomorphism of K3 crystals. Indeed, a direct computation yields the useful identity $$\label{eq:identity1}
\tPhi(e^A(a,b,c))=e^{\varphi(A)}\tPhi (a,b,c)$$ valid for any $A\in\H^2(X/K)$ and $(a,b,c)\in\tH(X/W)$ (recall that $\varphi=p^{-1}\Phi$). Instead, consider the diagram $$\begin{tikzcd}[row sep=small]
&&\H^2_{dR}(X/k)&&\\
0\arrow{r}&\Pic(X)\otimes\bZ/p\bZ\arrow{ur}{c_1^{dR}}\arrow{r}&\H^2(X,\m_p)\arrow[phantom]{u}{\rotatebox{90}{$\subset$}}\arrow{r}&\Br(X)[p]\arrow{r}&0
\end{tikzcd}$$ where the lower row is exact. The image of $pB'-pB$ in $\H^2_{dR}(X/k)$ maps to 0 in $\Br(X)[p]$, and hence is equal to $c_1^{dR}(\sL)$ for some line bundle $\sL$. Thus, writing $t=c_1^{\cry}(\sL)$, we have $$B'-B=\frac{t}{p}+h$$ for some $h\in\H^2(X/W)$. Note that if $h\in \H^2(X/W)$ then $e^h\in \H^*(X/W)$. Now, $\varphi(t/p)=t/p$, so by (\[eq:identity1\]) cupping with $e^{t/p}$ defines an isomorphism $$e^{t/p}\colon \tH(X/W,B)\iso e^{B'-h}\tH(X/W)=\tH(X/W,B')$$ of K3 crystals.
As a submodule of $\tH(X/K)$, the twisted crystal $\tH(X/W,B)$ depends only on the image of $B$ in $\H^2(X,\m_p)$. In other words, $\tH(X/W,B)$ may be viewed as a structure canonically attached to a class in $\H^2(X,\m_p)$. We take this viewpoint in [@BL17].
We can use the crystal $\tH(X/W,B)$ to extend the Artin invariant to twisted supersingular K3 surfaces.
The *Artin invariant* of a twisted supersingular K3 surface $(X,\alpha)$ is the Artin invariant of the supersingular K3 crystal $\tH(X/W,B)$ (in the sense of Definition \[def:artininvariant1\]).
In Section 3.4 of [@BL17], we make the following computation.
[@BL17 Corollary 3.4.23] If $(X,\alpha)$ is a twisted supersingular K3 surface, then $$\sigma_0(X,\alpha)=\begin{cases}
\sigma_0(X)+1&\mbox{ if }\alpha\neq 0\\
\sigma_0(X) &\mbox{ if }\alpha=0
\end{cases}$$
In particular, the Artin invariant of a twisted supersingular K3 surface is an integer $1\leq\sigma_0\leq 11$.
Twisted Chern characters and the twisted Néron-Severi group
-----------------------------------------------------------
Let $X$ be a K3 surface over an algebraically closed field $k$. Let $N(X)$ be the Néron-Severi group of $X$ (of course, $N(X)=\Pic(X)$ for a K3 surface).
We define the *extended Néron-Severi lattice* of $X$ by $$\tN(X)=\langle (1,0,0)\rangle\oplus N(X)\oplus\langle (0,0,1)\rangle$$ which we equip with the Mukai pairing $$\tN(X)\otimes\tN(X)\to\bZ$$ given by the formula $\langle (a,b,c),(a',b',c')\rangle=-a.c'+b.b'-a'.c$.
We define the *Mukai vector* of a coherent sheaf $\sE$ on $X$ by $$v(\sE)=\ch(\sE).\sqrt{\Td(X)}$$ Recall that we have $\sqrt{\Td(X)}=(1,0,1)$, and that the Chern characters of a coherent sheaf on a K3 surface are integral. The Mukai vector thus gives rise to a map $$v(\_):K(X)\to\tN(X)$$ which is an isomorphism of groups. If we equip $K(X)$ with the pairing induced by $$\label{eq:pairing}
\langle [\sE],[\sF]\rangle=-\sum_i(-1)^i\dim_k\Ext^i(\sE,\sF)$$ it is even an isometry.
We wish to replicate these structures in the twisted setting. Let $X$ be a smooth projective variety and $\alpha\in\Br(X)$ a Brauer class (we have in mind the case where $X$ is a K3 surface, or the product of two K3 surfaces). Let $\pi\colon\sX\to X$ be a $\bG_m$-gerbe whose associated cohomology class is $\alpha$. If $n\alpha=0$, then there exists an $n$-fold twisted invertible sheaf, say $\sL$, on $\sX$. Choosing such a sheaf allows us to compare $n$-fold twisted sheaves on $\sX$ and sheaves on $X$.
\[def:twistedcherncharacter\] Let $\sE$ be a locally free twisted sheaf of positive rank on $\sX$. The *twisted Chern character* of $\sE$ (with respect to $\sL$) is $$\ch^{\sL}(\sE)=\sqrt[\leftroot{-2}\uproot{2}n]{\ch(\pi_*(\sE^{\otimes n}\otimes\sL^{\vee}))}\in A^*(X)\otimes\bQ$$ where by convention we choose the $n$-th root so that $\rk(\ch^{\sL}(\sE))=\rk(\sE)$. We define the twisted Mukai vector of $\sE$ by $$v^{\sL}(\sE)=\ch^{\sL}(\sE).\sqrt{\Td(X)}$$
Any twisted sheaf on $\sX$ admits a resolution by locally free twisted sheaves of positive rank, so the ablove definition of the twisted Chern character and twisted Mukai vector extend by additivity to all of $K(X,\alpha)$. The choice of an $n$-fold twisted invertible sheaf $\sL$ on $\sX$ is essentially the same as a choice of preimage of $\alpha$ under the map $$\label{eq:map1}
\H^2(X,\m_n)\twoheadrightarrow\Br(X)[n],$$ Indeed, giving an $n$-fold twisted invertible sheaf $\sL$ on $\sX$ is the same as giving an isomorphism $\sX^n\iso B\bG_m$ of $\bG_m$-gerbes, which is the same as giving a $\m_n$-gerbe $\sX'$ together with an identification of its associated $\bG_m$-gerbe with $\sX$. The cohomology class of such a gerbe in $\H^2(X,\m_n)$ then maps to $\alpha$ under (\[eq:map1\]). Conversely, given a preimage $\alpha'$ of $\alpha$, choosing a representing $\m_n$-gerbe and an isomorphism between its associated $\bG_m$-gerbe and $\sX$ gives rise to an invertible $n$-fold twisted sheaf $\sL$ on $\sX$. Moreover, the resulting twisted Chern character $\ch^{\sL}$ depends only on $\alpha'$.
If $\alpha'$ is a preimage of $\alpha$ under (\[eq:map1\]), we write $\ch^{\alpha'}$ for the twisted Chern character determined by $\alpha'$, and set $v^{\alpha'}=\ch^{\alpha'}.\sqrt{\Td(X)}$.
Let us suppose now that $n=p$. Let $B$ be a B-field lift of $\alpha$. This gives rise to an element $\alpha'\in\H^2(X,\m_p)$ mapping to $\alpha$ by taking $pB\mod p$, and we denote by $\ch^{B}$ (resp. $v^B$) the resulting twisted Chern character (resp. twisted Mukai vector). Similarly, if $X\times Y$ is a product of two K3 surfaces, $\alpha\in\Br(X)$ and $\beta\in\Br(Y)$ are Brauer classes or order $p$ with B-field lifts $B,B'$, we define $\ch^{-B\boxplus B'}$ and $v^{-B\boxplus B'}$. We will identify a natural subgroup of $A^*(X)\otimes\bQ$ which will be a recipient for $\ch^B$ and $v^B$, and so provide a replacement for the twisted Néron-Severi group $\tN(X)$.
Let $X$ be a K3 surface, $\alpha\in\Br(X)[p]$ a $p$-torsion Brauer class, and $B$ a B-field lift of $\alpha$. The *twisted Néron-Severi lattice* associated to $(X,\alpha,B)$ is $$\tN(X,B)=(\tN(X)\otimes\bZ[p^{-1}])\cap \tH(X/W,B)$$ We equip $\tN(X,B)$ with the restriction of the Mukai pairing on $\tH(X/W,B)$.
We mention two results concerning this object.
\[prop:integrality1\] If $\sE$ is an $\alpha$-twisted sheaf, then $\ch^B(\sE)$ and $v^B(\sE)$ are contained in $\tN(X,B)$. The resulting map $$v^B(\_)\colon K(X,\alpha)\to\tN(X,B)$$ is an isomorphism of groups. If we equip the left hand side with the pairing (\[eq:pairing\]), it is an isometry.
We defer the proof to Appendix \[app:twistedChernCharacters\]. See also Proposition 4.1.9 of [@BL17] for a different proof.
[@BL17 Proposition 4.1.10] The inclusion $\tN(X,B)\hto\tH(X/W,B)$ induces an isomorphism $$\tN(X,B)\otimes\bZ_p\iso\tT(X,B)$$
Thus, the inclusion $\tN(X,B)\hto\tH(X/W,B)$ gives a marking of $\tH(X/W,B)$ in the sense of Definition \[def:marking\].
Action on cohomology
--------------------
Let $(X,\alpha)$ and $(Y,\beta)$ be twisted supersingular K3 surfaces, where $\alpha$ has order $n$ and $\beta$ has order $m$. Consider a complex $P^\bullet\in D(X\times Y,-\alpha\boxtimes\beta)$ of twisted sheaves such that the induced Fourier-Mukai functor $$\Phi_{P^\bullet}\colon D(X,\alpha)\to D(Y,\beta)$$ is an isomorphism. To study such objects, we will consider their action on the twisted Grothendieck groups and the twisted Mukai crystal. Choose B-field lifts $B$ of $\alpha$ and $B'$ of $\beta$. Using the standard formalism, we obtain maps $$\begin{aligned}
\Phi^{K}_{\left[P^\bullet\right]}(\_)&\colon K(X,\alpha)\to K(Y,\beta)\\
\Phi^{\cry}_{v^{-B\boxplus B'}(P^\bullet)}(\_)&\colon \tH(X/K)\to\tH(Y/K)\end{aligned}$$ It is not hard to show that these are compatible, in the sense that the diagram $$\begin{tikzcd}[column sep=large]
K(X,\alpha)\arrow{r}{\Phi^K_{[P^\bullet]}}\arrow{d}[swap]{v^B}&K(Y,\beta)\arrow{d}{v^{B'}}\\
\tH(X/K)\arrow{r}{\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}}&\tH(Y/K)
\end{tikzcd}$$ commutes (see for instance [@BL17 Lemma 4.1.6], or [@Huy06 Corollary 5.29] in the untwisted case). Similarly, it is straightforward to show that $\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}$ is compatible with the Mukai pairing and the Frobenius operators. As an immediate corollary, we obtain the following.
\[prop:FMpartner1\] Let $(X,\alpha)$ be a twisted supersingular K3 surface. If $Y$ is a smooth projective variety and there exists a derived equivalence $(X,\alpha)\cong (Y,\beta)$ for some $\beta\in\Br(Y)$, then $Y$ is a supersingular K3 surface.
It is well known that any twisted Fourier-Mukai partner of a twisted K3 surface is again a twisted K3 surface. As the isocrystal $\tH(X/K)$ has slope 1, the same is true for $\tH(Y/K)$. It follows that $Y$ is supersingular.
By analogy with the complex case, one expects that $\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}$ in fact preserves the integral structure. This is the content of the following result.
\[thm:Mainprelim\] Let $(X,\alpha)$ and $(Y,\beta)$ be twisted supersingular K3 surfaces with B-field lifts $B$ and $B'$. If $P^\bullet\in D(X\times Y,-\alpha\boxtimes\beta)$ is a complex of twisted sheaves such that the induced Fourier-Mukai functor $$\Phi_{P^\bullet}\colon D(X,\alpha)\to D(Y,\beta)$$ is an equivalence, then the induced cohomological transform gives an isomorphism $$\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}:(\tH(X/W,B),\tN(X,B))\iso(\tH(Y/W,B'),\tN(Y,B'))$$ of marked K3 crystals.
As it is somewhat unrelated to the rest of this paper, we defer the proof to Appendix \[app:twistedChernCharacters\] and Appendix \[app:deformation\]. An immediate consequence of Theorem \[thm:Mainprelim\] is that the Artin invariant is preserved under a derived equivalence.
\[cor:FMpartner2\] If $(X,\alpha)$ and $(Y,\beta)$ are twisted supersingular K3 surfaces that are derived equivalent, then $\sigma_0(X,\alpha)=\sigma_0(Y,\beta)$.
We mention a few consequences for supersingular K3 surfaces with small and large Artin invariants.
The unique supersingular K3 surface of Artin invariant 1 has no non-trivial twisted Fourier-Mukai partners.
Any twisted supersingular K3 surface $(X,\alpha)$ with $\alpha\neq 0$ must have Artin invariant $\sigma_0\geq 2$. So, every twisted partner is in fact untwisted, and has Artin invariant 1.
If $X$ is a supersingular K3 surface of Artin invariant 10 and $\alpha\in\Br(X)$ is a nontrivial Brauer class, then $(X,\alpha)$ is not derived equivalent to any non-twisted variety.
By Proposition \[prop:FMpartner1\], any potential Fourier-Mukai partner must be a supersingular K3 surface. Because $\alpha$ is nontrivial, $\sigma_0(X,\alpha)=11$. But, any non-twisted supersingular K3 surface has Artin invariant $\sigma_0\leq 10$.
Over the complex numbers, it is known that the action of a derived equivalence on the singular Mukai lattice preserves a certain orientation structure (see [@MR2553878], and [@2017arXiv171100846R] for the twisted case). As we do not have access to the singular cohomology lattice in positive characteristic, we instead formulate the orientation as a structure on the extended Néron-Severi group.
Let $N$ be a lattice of signature $(s_+,s_-)$.
\[def:orientation\] An *orientation* on $N$ is a choice of orientation (in the sense of vector spaces) on a positive definite subspace $P\subset N_{\bR}$ of dimension $s_+$. An isometry $g\colon N\to N'$ or $N_{\bQ}\to N'_{\bQ}$ of oriented lattices is *orientation preserving* if the composition $$P\to N_{\bR}\xrightarrow{g_{\bR}}N'_{\bR}\to P'$$ is orientation preserving in the usual sense, where $N'_{\bR}\to P'$ is the orthogonal projection.
We refer to [@2017arXiv171100846R] for further discussion. Let $X$ be a K3 surface. The extended Néron-Severi group $\tN(X)$ has signature $(2,22)$. Let $H\in\Pic(X)$ be an ample class. We consider the positive definite subspace $P_H=\langle (1,0,-1),H\rangle\subset\tN(X)\otimes\bR$. By the given ordering of the basis elements, $P_H$ is endowed with an orientation in the sense of vector spaces, and hence $\tN(X)$ acquires an orientation in the sense of Definition \[def:orientation\]. Note that this orientation does not depend on the choice of $H$, in the sense that the oriented lattices resulting from different choices are isometric by an orientation preserving isometry (namely the identity).
Suppose that $(X,\alpha)$ is a twisted K3 surface and $\alpha$ is killed by $p$. Given any B-field lift $B$ of $\alpha$, we obtain a natural orientation on $\tN(X,B)$ by declaring that the isometry $$e^B\colon \tN(X,B)_{\bQ}\iso\tN(X)_{\bQ}$$ is orientation preserving. If $P^{\bullet}\in D(X\times Y,-\alpha\boxtimes\beta)$ is a perfect complex of twisted sheaves that induces a Fourier-Mukai equivalence $\Phi_{P^{\bullet}}\colon D(X,\alpha)\to D(Y,\beta)$, then by Theorem \[thm:Mainprelim\] the cohomological transform $\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}$ restricts to an isometry $\tN(X,B)\iso\tN(Y,B')$. We say that $\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}$ is *orientation preserving* if this isometry is orientation preserving in the sense of Definition \[def:orientation\].[^4]
\[conj:mainPrelim2\] With the assumptions of Theorem \[thm:Mainprelim\], the cohomological transform $\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}$ is orientation preserving (in the sense that the induced map on twisted Néron-Severi groups is orientation preserving).
We resolve this conjecture in some special cases in Appendix \[app:deformation\].
Crystalline Torelli theorems for twisted supersingular K3 surfaces
------------------------------------------------------------------
In this section we formulate a crystalline Torelli theorem relating isomorphisms between the twisted Mukai crystals associated to a twisted supersingular K3 surface to Fourier-Mukai equivalences between derived categories of twisted sheaves. We will first prove a result relating the twisted Mukai crystal to equivalences of abelian categories of twisted sheaves. Suppose that $(X,\alpha)$ and $(Y,\beta)$ are twisted K3 surfaces. If $f\colon X\iso Y$ is an isomorphism such that $f^*\beta=\alpha$, then for any B-field lifts $B,B'$ of $\alpha$ and $\beta$, the induced map $f_*\colon\tH(X/K)\iso\tH(Y/K)$ restricts to an isomorphism $$\tH(X/W,B)\iso\tH(Y/W,B')$$ of K3 crystals. We wish to identify which such maps of crystals are induced by an isomorphism of surfaces. Consider the natural filtration on $\tH(X/K)$ by codimension. If $B$ is any B-field lift of a Brauer class $\alpha$, then there is an induced filtration $0\subset F^2\subset F^1\subset F^0=\tH(X/W,B)$ on the submodule $\tH(X/W,B)\subset\tH(X/K)$ defined by $$\begin{aligned}
F^2&=\tH(X/W,B)\cap\H^4(X/K)\\
F^1&=\tH(X/W,B)\cap(\H^2(X/K)\oplus\H^4(X/K))\\
F^0&=\tH(X/W,B)\cap(\H^0(X/K)\oplus\H^2(X/K)\oplus\H^4(X/K))\end{aligned}$$ Note that the $F^i$ are in fact subcrystals. There is a canonical isomorphism $$F^1/F^2\iso\H^2(X/W)$$ induced by $(0,h,h.B)\mapsto h$. This isomorphism is compatible with the Frobenius operators on both sides. Furthermore, as $F^2$ is isotropic, the Mukai pairing induces a pairing on the quotient $F^1/F^2$, and under the above isomorphism this pairing is taken to the cup product pairing.
We say that an isomorphism $g\colon\tH(X/W,B)\to\tH(Y/W,B')$ of K3 crystals *preserves the codimension filtrations* if $g(F^i)\subset F^i$ for $i=0,1,2$.
In fact, because $F^1$ is exactly the orthogonal complement of $F^2$, this is equivalent to $g(F^2)\subset F^2$. If $g$ preserves the filtrations, then it induces a map $$\frac{F^1\tH(X/W,B)}{F^2\tH(X/W,B)}\iso\frac{F^1\tH(Y/W,B')}{F^2\tH(Y/W,B')}$$ and hence an isomorphism $g_0\colon\H^2(X/W)\iso\H^2(Y/W)$ of K3 crystals.
\[thm:twistedOgus1\] Let $(X,\alpha)$ and $(Y,\beta)$ be twisted supersingular K3 surfaces. Let $B$ and $B'$ be B-field lifts of $\alpha$ and $\beta$. If $$g\colon (\tH(X/W,B),\tN(X,B))\to(\tH(Y/W,B'),\tN(Y,B'))$$ is an isomorphism of marked K3 crystals such that
1. $g$ sends $(0,0,1)$ to $(0,0,1)$ (and in particular preserves the codimension filtrations), and
2. $g_0$ maps an ample class to an ample class,
then there exists an isomorphism $f\colon X\to Y$ satisfying $f^*\beta=\alpha$ and a line bundle $\sL$ on $Y$ such that $\exp(c_1^{\cry}(\sL)/p)\circ f_*=g$.
As we have observed above, the assumption that $g$ is a map of K3 crystals implies that the induced isomorphism $g_0\colon\H^2(X/W)\iso\H^2(Y/W)$ is also a map of K3 crystals. Furthermore, as $g$ restricts to an isomorphism between the respective Néron-Severi groups, $g_0$ restricts to an isomorphism between the respective (non-twisted) Néron-Severi groups, and by assumption sends an ample class to an ample class. Thus, by Ogus’s crystalline Torelli theorem (in the formulation of Theorem \[thm:Ogus1\]), $g_0$ is induced by a unique isomorphism $f\colon X\iso Y$.
Consider the commuting diagram $$\begin{tikzcd}
\tH(X/W,B)\arrow{rr}{g}\arrow{dr}[swap]{f_*}&&\tH(Y/W,B')\\
&\tH(Y/W,g_0(B))\arrow{ur}[swap]{\exp(B'-g_0(B))}&
\end{tikzcd}$$ of $W$-modules. As both $g$ and $f_*$ are isomorphisms of marked K3 crystals, so is $\exp(B'-g_0(B))$. The inclusion $\tN(Y,g_0(B))\subset\tN(Y)\otimes\bQ$ becomes an isomorphism after tensoring with $\bQ$, so $\exp(B'-g_0(B))\in\tN(Y)\otimes\bQ$. It follows that $B'-g_0(B)=c_1^{\cry}(\sL)/p$ for some line bundle $\sL$ on $Y$. This completes the proof.
We next prove the main result of this paper: a Torelli theorem relating isomorphisms of the twisted Mukai crystals to equivalences of derived categories. This should be seen as a supersingular analog of the twisted derived Torelli theorem of Huybrechts and Stellari [@MR2310257].
\[thm:Main1\] Let $(X,\alpha)$ and $(Y,\beta)$ be twisted supersingular K3 surfaces with B-field lifts $B$ and $B'$.
1. If $\Phi_{P^\bullet}:D(X,\alpha)\to D(Y,\beta)$ is a Fourier-Mukai equivalence, the induced map $$\Phi^{\cry}_{v^{-B\boxplus B'}(P^\bullet)}:(\tH(X/W,B),\tN(X,B))\iso(\tH(Y/W,B'),\tN(Y,B'))$$ is an isomorphism of marked K3 crystals (which if Conjecture \[conj:mainPrelim2\] holds is orientation preserving).
2. If $$g:(\tH(X/W,B),\tN(X,B))\iso(\tH(Y/W,B'),\tN(Y,B'))$$ is an orientation preserving isomorphism of marked K3 crystals, then there exists a Fourier-Mukai equivalence $\Phi_{P^\bullet}:D(X,\alpha)\to D(Y,\beta)$ such that $\Phi^{\cry}_{v^{-B\boxplus B'}(P^\bullet)}=g$.
Part (1) is Theorem \[thm:Mainprelim\]. Let us show (2). The reader will note that the proof is essentially the same as the proof of the main theorem of [@MR2310257]. We start with an isomorphism $$g:(\tH(X/W,B),\tN(X,B))\iso(\tH(Y/W,B'),\tN(Y,B'))$$ of oriented marked crystals. By definition, this means that we have a diagram $$\begin{tikzcd}
\tH(X/W,B)\arrow{r}{g}&\tH(Y/W,B')\\
\tN(X,B)\arrow[hook]{u}\arrow{r}{g}&\tN(Y,B')\arrow[hook]{u}
\end{tikzcd}$$ where the upper horizontal arrow is an isomorphism of K3 crystals and the vertical arrows are the canonical inclusions. The lower horizontal arrow is then an isometry with respect to the Mukai pairings on the twisted Néron-Severi lattices, and preserves the canonical orientations. Let $w=(r,l,s)$ be the image $g(0,0,1)$ of the class of a point under $g$. We will first treat the case when $r\neq 0$. We may then assume that $r>0$ by composing with $g_0=-\id_{\tH(Y/W,B')}$, which lifts to the shift functor $F\mapsto F\left[1\right]$. As $g$ is an isometry we have $g^2=0$, so the moduli space $\sM_{(Y,\beta)}(w)$ of $\beta$-twisted sheaves on $Y$ with twisted Mukai vector $w$ is a $\bG_m$-gerbe over a K3 surface $M_{(Y,\beta)}(w)$, and there is a $B$-field $B''\in H^2(M_{(Y,\beta)}(w)/W)$ and an equivalence $D(Y,\beta)\cong D(M_{(Y,\beta)}(w),\alpha_{B''})$ inducing an orientation preserving isometry $\tH(Y/W,B')\to\tH(M_{(Y,\beta)}(w)/W,B'')$ (the relevant facts about moduli spaces of twisted sheaves have been extended to our setting in Section 4 of [@BL17]). The inverse of this equivalence sends $(0,0,1)$ to $w$, so composing we may assume without loss of generality that $g(0,0,1)=(0,0,1)$.
To apply Theorem \[thm:twistedOgus1\], we need to ensure that the induced map $g_0\colon\H^2(X/W)\to\H^2(Y/W)$ maps an ample class to an ample class (see the discussion preceding the statement of Theorem \[thm:twistedOgus1\]). By results of [@Ogus83] there exists a sequence of $(-2)$ classes $C_1,\dots,C_n$ such that composing $g_0$ with the composition of reflections $C_n\circ\dots\circ C_1$ takes the ample cone $C_X$ of $X$ to $\pm C_Y$. But $g$ was assumed to be orientation preserving, so in fact $C_X\mapsto C_Y$. As explained in [@MR2310257], each $C_i$ is induced by a spherical twist functor $T_{C_i}\colon D(Y,\beta)\to D(Y,\beta)$ on the derived category, and moreover the action of such a functor on the twisted Mukai crystal sends $(0,0,1)$ to $(0,0,1)$. We have reduced to the case when the assumptions of Theorem \[thm:twistedOgus1\] apply. We therefore find an isomorphism $f\colon X\to Y$ satisfying $f^*\beta=\alpha$ and a line bundle $\sL$ on $Y$ such that $$\exp(c_1^{\cry}(\sL)/p)\circ f_*=g$$ Both of these maps lift to Fourier-Mukai equivalences, so this completes the proof.
If we are only interested in the existence of a derived equivalence, it turns out that we may forget the marking and orientation. This will make our computations more manageable.
\[prop:forgetmarking\] Let $(H,N)$ and $(H',N')$ be marked oriented supersingular K3 crystals. If there exists an isomorphism $H\iso H'$ of K3 crystals (not necessarily preserving the marking or orientation), then there exists an isomorphism $(H,N)\iso (H',N')$ of marked K3 crystals that is orientation preserving.
By results of Nikulin (see for instance Theorem 15.1.5 of [@Huy06]), there exists an isometry $N\iso N'$. The remainder of the proof follows that of [@Ogus78 Theorem 7.4] exactly.
\[thm:main2\] Let $(X,\alpha)$ and $(Y,\beta)$ be twisted supersingular K3 surfaces, with B-field lifts $B,B'$ of $\alpha$ and $\beta$. The following are equivalent.
1. There exists a derived equivalence $D(X,\alpha)\cong D(Y,\beta)$.
2. There exists an isomorphism $\tH(X/W,B)\cong\tH(Y/W,B')$ of K3 crystals (not necessarily preserving any marking or orientation).
This follows from Theorem \[thm:Main1\] combined with Proposition \[prop:forgetmarking\].
Applications to Fourier-Mukai equivalences
==========================================
Characteristic subspaces {#sec:characteristicsubspaces}
------------------------
It is a remarkable discovery of Ogus that supersingular K3 crystals are determined by certain semi-linear algebraic data, called characteristic subspaces. These are very accessible for making explicit computations. This correspondence resembles the equivalence between Hodge structures and periods. In this section we will explain this connection, and reformulate the theorems of the previous section in terms of characteristic subspaces.
Let $V$ be a vector space over $\bF_p$ of even dimension $2\sigma_0$ equipped with a non-degenerate, non-neutral bilinear form. Let $\sigma\colon k\to k$ be the Frobenius map $\lambda\mapsto\lambda^p$, and define $$\varphi=\id\otimes\,\sigma\colon V\otimes k\to V\otimes k$$
A *characteristic subspace* of $V$ is a subspace $K\subset V\otimes k$ such that
1. $K$ is totally isotropic of dimension $\sigma_0$ and
2. $K+\varphi(K)$ has dimension $\sigma_0+1$.
If in addition $\sum_{i=0}^\infty\varphi^i(K)=V\otimes k$, we say that $K$ is *strictly characteristic*.
A *characteristic subspace datum* is a pair $(K,V)$, where $V$ is a vector space over $\bF_p$ of even dimension equipped with a non-degenerate, non-neutral bilinear form, and $K\subset V\otimes k$ is a strictly characteristic subspace.
An isomorphism $(K,V)\iso (K',V')$ of characteristic subspace data is an isometry $g:V\to V'$ such that $g\otimes k$ maps $K$ to $K'$.
The *Artin invariant* of a characteristic subspace datum is the dimension of $K$ as a $k$-vector space. Let $H$ be a supersingular K3 crystal with Tate module $T$ and Artin invariant $\sigma_0$. The quotient $T_0=T^\vee/T$ is an $\bF_p$-vector space of dimension $2\sigma_0$, which inherits a non-degenerate and non-neutral quadratic form from $T$, defined by $$\overline{v}.\overline{w}=p^{-1}\langle pv,pw\rangle$$ for $v,w\in T^{\vee}$. There are natural inclusions $$T\otimes W\subset H\subset T^{\vee}\otimes W$$ By Theorem 3.20 of [@Ogus78], the image $K$ of $H$ in $(T^{\vee}\otimes W)/(T\otimes W)\cong T_0\otimes k$ is a strictly characteristic subspace.
The characteristic subspace datum associated to a supersingular K3 crystal $H$ is the pair $(K,T_0)$.
It is perhaps more correct to replace $K$ with $\varphi^{-1}(K)$, which is also strictly characteristic (this is the convention taken in [@Ogus78]). This distinction is important when working over non-perfect bases.
Any isomorphism $H\iso H'$ of supersingular K3 crystals restricts to an isometry $T\iso T'$ of Tate modules, and hence induces an isomorphism $(K,T_0)\iso (K',T'_0)$ of characteristic subspace data. By results of Ogus, the converse is also true.
Suppose that $H,H'$ are supersingular K3 crystals of rank $n$. Let $T\subset H$ and $T'\subset H'$ be their Tate modules, and $(K,T_0)$ and $(K',T'_0)$ the corresponding characteristic subspace data. Any isomorphism $(K,T_0)\iso (K',T'_0)$ of characteristic subspace data is induced by an isomorphism $H\iso H'$ of K3 crystals.
Consider an isomorphism $(K,T_0)\to (K',T'_0)$ of characteristic subspace data. By Corollary 3.18 of [@Ogus78], there exists an isometry $T_0\iso T'_0$. Moreover, by results of Nikulin (see for instance Theorem 14.2.4 of [@Huy06]), every isometry $T_0\iso T'_0$ lifts to an isometry $T\iso T'$. The result then follows from Theorem 3.20 of [@Ogus78].
\[not:subspacedatum\] Let $X$ be a supersingular K3 surface. We write $T(X)$ for the Tate module of $\H^2(X/W)$ and write $K(X)$ for its associated characteristic subspace. Similarly, we write $\tT(X)$ and $\tK(X)$ for the Tate module and characteristic subspace associated to $\tH(X/W)$.
If $\alpha\in\Br(X)$ is a Brauer class and $B$ is a B-field lift of $\alpha$, then $\tH(X/W,B)$ is a supersingular K3 crystal of rank 24, and we let $\tT(X,B)$ be its Tate module and $\tK(X,B)$ its associated characteristic subspace.
Suppose that $X$ is a supersingular K3 surface. We can give an alternate description of the characteristic subspace datum associated to $\H^2(X/W)$ that does not use crystalline cohomology. Consider the first Chern class map $$c_1^{dR}\colon\Pic(X)\to\H^2_{dR}(X/k)$$ Tensoring with $k$, we get a map $$\label{eq:chern}
N(X)\otimes k\to\H^2_{dR}(X/k)$$ As the discriminant group of $N(X)$ is killed by $p$, there is a natural inclusion $pN(X)^{\vee}\subset N(X)$. This induces an inclusion $$\dfrac{pN(X)^{\vee}}{pN(X)}\subset\dfrac{N(X)}{pN(X)}$$ of $\bF_p$-vector spaces. Of course, multiplication by $p$ induces an isomorphism $$N(X)_0\defeq\dfrac{N(X)^{\vee}}{N(X)}\iso\dfrac{pN(X)^{\vee}}{pN(X)}$$ This gives a natural inclusion $$N(X)_0\otimes k\hto N(X)\otimes k$$ Unwinding definitions, we see that the composition $$K(X)\hto T(X)_0\otimes k=N(X)_0\otimes k\hto N(X)\otimes k$$ identifies $K(X)$ with the kernel of (\[eq:chern\]).
We now translate the crystalline Torelli theorems in previous sections into the language of characteristic subspaces. Let $X$ be a supersingular K3 surface. We write $(K(X),T(X)_0)$ for the characteristic subspace datum associated to the supersingular K3 crystal $\H^2(X/W)$ (see Notation \[not:subspacedatum\]). In terms of characteristic subspaces, Ogus’s Torelli theorem takes the following form.
\[thm:characteristictorelli1\] If $X$ and $Y$ are supersingular K3 surfaces, then $X$ is isomorphic to $Y$ if and only if there is an isomorphism $(K(X),T(X)_0)\iso (K(Y),T(Y)_0)$ of characteristic subspace data.
Suppose that $\alpha\in\Br(X)$ is a Brauer class with B-field lift $B$. The supersingular K3 crystal $\tH(X/W,B)$ gives rise to the characteristic subspace datum $(\tK(X,B),\tT(X,B)_0)$. If $\alpha$ is non-trivial, then the functional on $\tT(X,B)$ given by pairing with $(0,0,1)$ is divisible by $p$. Hence, $p^{-1}(0,0,1)$ gives rise to a non-trivial element of $\tT(X,B)_0$. Anticipating future notation, let us write $f_X$ for the element $p^{-1}(0,0,1)\in\tT(X,B)_0$. The twisted crystalline Torelli theorem then assumes the following form.
\[thm:characteristictorelli2\] Let $X$ and $Y$ be supersingular K3 surfaces and $\alpha\in\Br(X),\beta\in\Br(Y)$ non-trivial Brauer classes with B-field lifts $B,B'$. There exists an isomorphism $(X,\alpha)\cong (Y,\beta)$ if and only if there exists an isomorphism of characteristic subspace data $(\tK(X,B),\tT(X,B)_0)\iso (\tK(Y,B'),\tT(Y,B')_0)$ sending $f_X$ to $f_Y$.
Finally, we translate our main theorem on derived equivalences into characteristic subspaces.
\[thm:characteristictorelli3\] Let $X$ and $Y$ be supersingular K3 surfaces and $\alpha\in\Br(X),\beta\in\Br(Y)$ Brauer classes with B-field lifts $B,B'$. There exists a Fourier-Mukai equivalence $D(X,\alpha)\cong D(Y,\beta)$ if and only if there exists an isomorphism $(\tK(X,B),\tT(X,B)_0)\iso(\tK(Y,B'),\tT(Y,B')_0)$ of characteristic subspace data.
We summarize our results in the following table. In each row, the object in the left column is classified by the indicated K3 crystal, and by the indicated characteristic subspace datum. We continue the notation $f_X=p^{-1}(0,0,1)$.
1
[ c|c|c ]{} **[Geometric Object]{} & **[K3 Crystal]{} & **[Characteristic Subspace Datum]{}******
------------------------------------------------------------------------
\
$X$ & $H^2(X/W)$ & $(K(X),T(X)_0)$
------------------------------------------------------------------------
\
$(X,\alpha)$ & $\tH(X/W,B)$ + $(0,0,1)\in\tH$& $(\tK(X,B),\tT(X,B)_0)$ + $f_X\in \tT(X,B)_0$\
$D(X)$ & $\tH(X/W)$ & $(\tK(X),\tT(X)_0)$\
$D(X,\alpha)$ & $\tH(X/W,B)$ & $(\tK(X,B),\tT(X,B)_0)$
1
Let $X$ be a supersingular K3 surface, and $B$ a B-field on $X$. We will examine the relationship between the characteristic subspace data associated to $\H^2(X/W)$ and $\tH(X/W,B)$. This relationship is particularly simple if we make a special choice of B-field lift. We may characterize Artin’s subgroup $\U^2(X,\m_p)\subset\H^2(X,\m_p)$ as follows.
\[def:alternatedef\] The subgroup $\U^2(X,\m_p)\subset\H^2(X,\m_p)$ consists of exactly those classes in $\H^2(X,\m_p)$ that, under the inclusion $\H^2(X,\m_p)\hto\H^2_{dR}(X/k)$, are orthogonal to the image of the first Chern class map.
A B-field $B=\frac{a}{p}\in\H^2(X/K)$ is *transcendental* if the image of $a$ in $\H^2_{dR}(X/k)$ is contained in $\U^2(X,\m_p)$.
By the above characterization of $\U^2(X,\m_p)$, we see that a B-field $B=\frac{a}{p}$ is transcendental if and only if $\langle a,D\rangle$ is divisible by $p$ for every $D\in N(X)$, or equivalently if and only if $B\in T(X)^{\vee}\otimes W\subset p^{-1}\H^2(X/W)$.
Now, it also follows from the above that the image of $\U^2(X,\m_p)\hto\H^2_{dR}(X/k)$ is contained in the subspace $V\otimes k/K\subset\H^2_{dR}(X/k)$, where $V=N(X)_0$. We may identify the image of this map as follows.
\[lem:coolbijection\] The inclusion $\U^2(X,\m_p)\hto\H^2_{dR}(X/k)$ induces an isomorphism $$\U^2(X,\mu_p)\cong\dfrac{\left\{B\in V\otimes k|B-\varphi(B)\in K+\varphi(K)\right\}}{K}\subset \dfrac{V\otimes k}{K}$$
This is proved in [@BL17]. See Remark 3.3.9ff and Lemma 3.3.15.
Equivalently, this map may be described in the following way: given a class $\alpha\in\U^2(X,\m_p)$, we choose a B-field $B=\frac{a}{p}$ such that $\overline{a}=d\log\alpha$. Such a $B$ is transcendental, so it is contained in $T(X)^{\vee}\otimes W$. We send $\alpha$ to the image of $B$ under the composition $$T(X)^{\vee}\otimes W\to\dfrac{T(X)^\vee\otimes W}{T(X)\otimes W}=T(X)_0\otimes k=V\otimes k\to \dfrac{V\otimes k}{K}$$ Under the isomorphism of Lemma \[lem:coolbijection\], the subgroup $pN(X)^{\vee}/pN(X)$ described by (\[eq:artinssubgroup\]) is identified with the fixed points of $\varphi$ acting on $V\otimes k$.
Suppose $\alpha\in\Br(X)$ is a non-trivial Brauer class. Let $B=\frac{a}{p}$ be a transcendental B-field lift of $\alpha$. It is not hard to show that $\tT(X,B)$ is then the free $W$-module $$\tT(X,B)=\langle (p,0,0),(0,D,0),(0,0,1)\rangle$$ where $D$ ranges over all classes in $N(X)$ (see Proposition 3.4.21 of [@BL17]). There is an induced decomposition $$\tT(X,B)_0=T(X)_0\oplus U_2$$ where $U_2$ is the copy of the hyperbolic plane generated by the vectors $e$ and $f$, given respectively by pairing with $p^{-1}(p,0,0)$ and $p^{-1}(0,0,1)$. If $x_1,\dots,x_{\sigma_0}$ is a basis for $K(X)$, then the characteristic subspace $\tK(X,B)$ associated to $\tH(X/W,B)$ is given by $$\label{eq:decomposition222}
\tK(X,B)=\left\langle x_1+(x_1.B)f,\dots,x_{\sigma_0}+(x_{\sigma_0}.B)f,e+B+\dfrac{B^2}{2}f\right\rangle$$
With the above descriptions as motivation, we record the following result on abstract characteristic subspaces. We fix an $\bF_p$-vector space $V$ of dimension $2\sigma_0$, equipped with a non-degenerate, non-neutral bilinear form, and a characteristic subspace $K\subset V\otimes k$. Let $U_2$ be a copy of the hyperbolic plane over $\bF_p$, equipped with the standard basis $v,w$ satisfying $v^2=w^2=0$ and $v.w=-1$. Set $\tV=V\oplus U_2$. This is a vector space of dimension $2\sigma_0+2$, whose natural bilinear form is again non-degenerate and non-neutral.
\[prop:radbijection\] Let $x_1,\dots,x_{\sigma_0}$ be a basis for $K$. The map $$\dfrac{\left\{B\in V\otimes k|B-\varphi(B)\in K+\varphi(K)\right\}}{K}\to\left\{
\begin{tabular}{@{}c@{}}
characteristic subspaces $\tK\subset\tV\otimes k$ \\
such that $v\notin\tK$ and $\tK\cap v^\perp/v=K$\\
\end{tabular}
\right\}$$ given by $$B\mapsto \left\langle x_1+(x_1.B)v,\dots,x_{\sigma_0}+(x_{\sigma_0}.B)v,w+B+\frac{B^2}{2}v\right\rangle$$ is a bijection.
This is prove in Proposition 3.1.11 of [@BL17] (see also Remark 3.2.3).
Fourier-Mukai partners of twisted supersingular K3 surfaces {#sec:applications}
-----------------------------------------------------------
We will now deduce some consequences for derived equivalences of twisted supersingular K3 surfaces. We have attached to a supersingular K3 surface two supersingular K3 crystals: the second crystalline cohomology $\H^2(X/W)$, which has rank 22, and the Mukai crystal $\tH(X/W)$, which has rank 24. Note that there is an inclusion $\H^2(X/W)\subset\tH(X/W)$, which induces an equality $T(X)_0=\tT(X)_0$ identifying the corresponding characteristic subspaces. Thus, the characteristic subspace data of the two crystals are *equal* (of course, the crystals have different ranks, and thus are not isomorphic). Combining this observation with Theorem \[thm:characteristictorelli1\] and Theorem \[thm:characteristictorelli3\], we obtain another proof of the following result of Lieblich and Olsson [@LO15].
\[thm:Untwisted1\] If $X$ and $Y$ are supersingular K3 surfaces, then $D(X)\cong D(Y)$ if and only if $X\cong Y$.
If $D(X)\cong D(Y)$, then by Theorem \[thm:Main1\] we have in particular an isomorphism $\tH(X/W)\cong\tH(Y/W)$ of K3 crystals. Thus, there is an isomorphism $(\tK(X),\tT(X)_0)\cong(\tK(Y),\tT(Y)_0)$ of characteristic subspace data. But by the above observation, this gives an isomorphism $(K(X),T(X)_0)\cong (K(Y),T(Y)_0)$ between the characteristic subspace data associated to $\H^2(X/W)$ and $\H^2(Y/W)$. By Theorem \[thm:characteristictorelli2\], this implies $X\cong Y$.
The converse is immediate.
The characteristic subspace associated to a K3 crystal might be viewed as an analog of the transcendental lattice. Theorem \[thm:Untwisted1\] is then analogus to the classical result that two complex K3 surfaces with Picard number $\geq 12$ are derived equivalent if and only if they are isomorphic (see Corollary 16.3.7 of [@Huy06]).
\[cor:Untwisted2\] A twisted supersingular K3 surface $(X,\alpha)$ is derived equivalent to a non-twisted K3 surface if and only if $\sigma_0(X,\alpha)\leq 10$.
A non-twisted supersingular K3 surface has $\sigma_0\leq 10$. So, if $\sigma_0(X,\alpha)=11$ then $(X,\alpha)$ cannot be derived equivalent to a non-twisted K3 surface.
Conversely, suppose that $\sigma_0(X,\alpha)\leq 10$. Consider the characteristic subspace datum $(\tK(X,B),\tT(X,B)_0)$ associated to $\tH(X/W,B)$. By the surjectivity of the period morphism, we find a supersingular K3 surface $Y$ and an isomorphism $(K(Y),T(Y)_0)\cong (\tK(X,B),\tT(X,B)_0)$ of characteristic subspace data. But $(K(Y),T(Y)_0)=(\tK(Y),\tT(Y)_0)$, so Theorem \[thm:characteristictorelli3\] implies that $D(Y)\cong D(X,\alpha)$.
Let $E$ be a supersingular elliptic curve, and let $X=\Km(E\times E)$ be the unique supersingular K3 surface of Artin invariant $\sigma_0=1$. Every supersingular K3 surface of Artin invariant 2 is derived equivalent to $D(X,\alpha)$ for some Brauer class $\alpha\in\Br(X)$.
Let $Y$ be a supersingular K3 surface of Artin invariant 2. Let $(K(Y),T(Y)_0)$ be the characteristic subspace datum associated to $\H^2(Y/W)$. Pick a non-zero isotropic vector $v\in T(Y)_0$ (such a vector exists because the dimension of $T(Y)_0$ is $\geq 4$). Set $V=v^\perp$. We may then find an isotropic vector $w\in T(Y)_0$ such that $v.w=-1$, and such that there is an orthogonal decomposition $T(Y)_0=V\oplus U_2$ where $U_2$ is the hyperbolic plane generated by $v,w$. As $K(Y)$ is strictly characteristic, $v$ is not in $K(Y)$. Let $K=K(Y)\cap v^{\perp}/v$. This is a strictly characteristic subspace of $V$. By the surjectivity of the period morphism, the pair $(K,V)$ is isomorphic to the characteristic subspace datum associated to $\H^2(X/W)$, where $X$ is the unique supersingular K3 surface of Artin invariant $1$. By Proposition \[prop:radbijection\] and Lemma \[lem:coolbijection\], there is a Brauer class $\alpha\in\Br(X)$ and a transcendental B-field lift $B$ of $\alpha$ such that $(K(X,B),T(X,B)_0)\cong(K(Y),T(Y)_0)$. As before, $(K(Y),T(Y)_0)$ is equal to the characteristic subspace datum associated to $\tH(Y/W)$, and Theorem \[thm:characteristictorelli3\] then gives the result.
We also prove the following (compare to Corollary 7.9 of [@HS04]).
\[thm:caldararusthing\] If $X$ is a supersingular K3 surface and $\alpha\in\Br(X)$ is a Brauer class, then for any integer $k$ such that $(k,p)=1$, there is a derived equivalence $$D(X,\alpha)\cong D(X,\alpha^k)$$
Suppose that $\alpha$ is non-trivial. Let $B$ be a transcendental B-field lift of $\alpha$. We have a canonical decomposition $$\tT(X,B)_0=T(X)_0\oplus U_2$$ where $U_2$ is a copy of the hyperbolic plane generated by $e,f$, corresponding to pairing with $p^{-1}(p,0,0)$ and $p^{-1}(0,0,1)$ respectively. If $x_1,\dots,x_{\sigma_0}$ is a basis for $K(X)$, then we have that $$\tK(X,B)=\left\langle x_1+(x_1.B)f,\dots,x_{\sigma_0}+(x_{\sigma_0}.B)f,e+B+\dfrac{B^2}{2}f\right\rangle$$ (see the notation and discussion preceding (\[eq:decomposition222\])). Let $\lambda\in\bF_p^{\times}$ be the image of $k$ modulo $p$. Consider the isometry $m_{\lambda}\colon\tT(X,B)_0\iso\tT(X,B)_0$ defined by $v\mapsto v$ for $v\in T(X)_0$, $e\mapsto \lambda^{-1}e$, and $f\mapsto \lambda f$. The image of $\tK(X,B)$ under $m_{\lambda}$ is $$\left\langle x_1+(x_1.\lambda B)f,\dots,x_{\sigma_0}+(x_{\sigma_0}.\lambda B)f,e+\lambda B+\dfrac{(\lambda B)^2}{2}f\right\rangle$$ The element $\lambda B$ is again in $\U^2(X,\m_p)$, and its image in the Brauer group is $\alpha^k$ (here, we are abusing notation by identifying $B$ with its image in $\U^2(X,\m_p)\subset\H^2_{dR}(X/k)$, and with its image under the isomorphism of Lemma \[lem:coolbijection\]ff). If $\lambda'\in\bZ_p^{\times}$ is a lift of $\lambda$, then $\lambda'B$ is a (transcendental) B-field lift of $\alpha^k$, and $m_{\lambda}$ gives an isomorphism $(\tK(X,B),\tT(X,B)_0)\iso (\tK(X,\lambda' B),\tT(X,\lambda' B)_0)$ of characteristic subspace data. We conclude the result by Theorem \[thm:characteristictorelli3\].
As a further application of our main theorem, we will count the number of twisted Fourier-Mukai partners of a twisted supersingular K3 surface. This will require some more involved calculations with characteristic subspaces.
Interlude: the orthogonal group of a strictly characteristic subspace {#sec:orthogonal}
---------------------------------------------------------------------
Let $V$ be a vector space over $\bF_p$ of dimension $2\sigma_0$, equipped with a non-degenerate, non-neutral bilinear form, and let $K\subset V\otimes k$ be a strictly characteristic subspace.
We define $O_{K}(V)$ to be the automorphism group of the characteristic subspace datum $(K,V)$. That is, $O_K(V)$ is the group of isometries $g:V\to V$ such that $g\otimes k$ maps $K$ to itself.
In this section we will study the orthogonal group $O_K(V)$ using certain coordinates introduced by Ogus. Because $K$ is strictly characteristic, the subspace $$l_K=K\cap\varphi(K)\cap\dots\varphi^{\sigma_0-1}(K)$$ is one-dimensional, and $$\varphi^{\sigma_0-1}(K)=l_K+\varphi(l_K)+\dots+\varphi^{\sigma_0-1}(l_K)$$ Pick a generator $e$ for $L$, and scale $e$ so that $e.\varphi^{\sigma_0}(e)=1$. Such a choice of $e$ is unique up to multiplication by a $p^{\sigma_0}+1$-th root of unity. Set $e_i=\varphi^{i-1}(e)$, so that $\left\{e_1,\dots,e_{\sigma_0}\right\}$ is a basis for $\varphi^{\sigma_0-1}(K)$ and $\left\{e_1,\dots,e_{2\sigma_0}\right\}$ is a basis for $V\otimes k$. In this basis both the bilinear form and the Frobenius take a particularly simple form. Define $$\label{eq:structureconstants}
a_i=e_1.e_{\sigma_0+i+1}\quad 1\leq i\leq\sigma_0-1$$ By Lemma 3.22 of [@Ogus78], the bilinear form on $V\otimes k$ is given by the matrix $$\label{eqn:form}
\begin{pmatrix}
0 & A \\
A^T & 0
\end{pmatrix}$$ where $A$ is the $\sigma_0\times\sigma_0$-matrix $$\begin{pmatrix}
1 & a_1 & a_2 & a_3 & \dots & a_{\sigma_0-1} \\
0 & 1 & F(a_1) & F(a_2) & \dots & F(a_{\sigma_0-2}) \\
0 & 0 & 1 & F^2(a_1) & \dots & F^2(a_{\sigma_0-3})\\
0 & 0 & 0 & 1 & \dots & F^3(a_{\sigma_0-4})\\
\vdots & \vdots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & 0 & 0 & \dots & 1
\end{pmatrix}$$ The Frobenius is given by $\varphi(e_i)=e_{i+1}$ for $1\leq i\leq 2\sigma_0-1$, and $$\varphi(e_{2\sigma_0})=\lambda_1e_1+\dots+\lambda_{\sigma_0}e_{\sigma_0}+\mu_1e_{\sigma_0+1}+\dots+\mu_{\sigma_0}e_{2\sigma_0}$$ for some particular scalars $\lambda_i,\mu_i$ which satisfy $\lambda_1=1$ and $\mu_1=0$; the rest can be described in terms of the $a_i$. In this basis, every element of $O_K(V)$ is diagonalized. Indeed, suppose that $g\in O_K(V)$ is an isometry that preserves $K$. Because $g$ commutes with $\varphi$, it must also preserve $l_K$, and therefore $$g(e)=\zeta e$$ for some scalar $\zeta$. Because of the uniqueness of $e$, $\zeta$ must be a $(p^{\sigma_0}+1)$-th root of unity. We have $g(e_i)=\zeta^{p^{i-1}}(e_1)$, so $g$ is the diagonal matrix $$\label{eqn:diagonal}
\begin{pmatrix}
\zeta & & & & \\
& \zeta^p & & & \\
& & \zeta^{p^2} & & \\
& & & \ddots & \\
& & & &\zeta^{p^{2\sigma_0-1}}\\
\end{pmatrix}$$ In particular, $g$ is determined by $\zeta$. The map $g\mapsto\zeta$ defines an injective homomorphism $$\label{eqn:inclusion}
O_K(V)\hto \mu_{p^{\sigma_0}+1}(k)$$
We will determine the image of (\[eqn:inclusion\]).
\[prop:subgroup\] Let $m$ be a divisor of $\sigma_0$ such that $\sigma_0/m$ is odd. If $\zeta$ is a primitive $p^m+1$-th root of unity, then $\zeta$ is contained in the image of (\[eqn:inclusion\]) if and only if $a_i=0$ for every $1\leq i\leq\sigma_0-1$ such that $2m$ does not divide $i$.
Let $m$ be a divisor of $\sigma_0$ with $\sigma_0/m$ odd, let $\zeta$ be a primitive $(p^m+1)$-th root of unity, and let $g$ be the corresponding diagonal element (\[eqn:diagonal\]) of $\GL(V\otimes k)$.
Suppose that $\zeta$ is contained in the image of (\[eqn:inclusion\]). As $g$ is an isometry, we have for each $1\leq i\leq\sigma_0-1$ that $$a_i=e_1.e_{\sigma_0+i+1}=g(e_1).g(e_{\sigma_0+i+1})=\zeta^{p^{\sigma_0+i}+1}a_i$$ If $a_i\neq 0$ for some $1\leq i\leq\sigma_0-1$, then $\zeta^{p^{\sigma_0+i}+1}=1$, so $p^m+1$ divides $p^{\sigma_0+i}+1$. Thus, $p^{\sigma_0+i}\equiv -1$ modulo $p^m+1$. The order of $p$ in $(\bZ/(p^m+1)\bZ)^\times$ is $2m$, and because $\sigma_0/m$ is odd, we have $p^{\sigma_0}\equiv -1$ modulo $p^m+1$. It follows that $p^i\equiv 1$ modulo $p^m+1$, and therefore $2m$ divides $i$.
Conversely, suppose that $a_i=0$ for every $1\leq i\leq\sigma_0-1$ such that $2m$ does not divide $i$. We will show that $g\in O_K(V)$. It is clear that any matrix of the form (\[eqn:diagonal\]) preserves $K$. We next check that $g\in O(V\otimes k)$, that is, that $g$ preserves the bilinear form given by the matrix (\[eqn:form\]). We will show that $$g(e_k).g(e_j)=e_k.e_j$$ for all $1\leq k,j\leq 2\sigma_0$. It is clear that $\varphi(g(e_i))=g(\varphi(e_i))$ for $1\leq i\leq 2\sigma_0-1$, so it will suffice to check that $$\label{eq:misceq2}
g(e_1).g(e_{\sigma_0+i+1})=e_1.e_{\sigma_0+i+1}=a_i$$ for all $1\leq i\leq\sigma_0-1$. We have $$g(e_1).g(e_{\sigma_0+i+1})=\zeta^{p^{\sigma_0+i}+1}a_i$$ If $2m$ does not divide $i$, then by assumption $a_i=0$, so (\[eq:misceq2\]) holds. Suppose that $2m$ divides $i$. It follows that $p^i\equiv 1$ modulo $p^m+1$. We know that $p^{\sigma_0}\equiv -1$ modulo $p^m+1$, so $p^{\sigma_0+i}+1\equiv 0$ modulo $p^m+1$. We conclude that $\zeta^{p^{\sigma_0+i}+1}=1$. It follows that $g$ is an isometry.
Finally, we will show that $g$ is defined over $\bF_p$, that is, that it is in the image of the map $O(V)\to O(V\otimes k)$. Equivalently, we will show that $g$ commutes with the Frobenius $\varphi$. As previously observed, $\varphi(g(e_i))=g(\varphi(e_i))$ for $1\leq i\leq 2\sigma_0-1$, so it remains only to verify that $\varphi(g(e_{2\sigma_0}))=g(\varphi(e_{2\sigma_0}))$. The pairing is non-degenerate, so it will suffice to show that $$\label{eq:misceq1}
\varphi(g(e_{2\sigma_0})).e_k=g(\varphi(e_{2\sigma_0})).e_k$$ for all $2\leq k\leq 2\sigma_0+1$ (the vectors $e_2,\dots,e_{2\sigma_0+1}$ are a basis for $V\otimes k$). Using that $g$ is an isometry, (\[eq:misceq1\]) becomes $$\zeta^{p^{2\sigma_0}+p^{k-1}}\varphi(e_{2\sigma_0}).e_k=\varphi(e_{2\sigma_0}).e_k$$ We compute that $$\varphi(e_{2\sigma_0}).e_k=
\begin{cases}
F^{k-1}(a_{\sigma_0-k+1})&\mbox{ if }2\leq k\leq\sigma_0\\
1&\mbox{ if }k=\sigma_0+1\\
0&\mbox{ if }\sigma_0+2\leq k\leq 2\sigma_0+1
\end{cases}$$ It follows that (\[eq:misceq1\]) is an equality for $\sigma_0+1\leq k\leq 2\sigma_0+1$. Suppose that $2\leq k\leq\sigma_0$. If $2m$ does not divide $\sigma_0-k+1$, then by assumption $a_{\sigma_0-k+1}=0$, and so (\[eq:misceq1\]) is an equality. Suppose that $2m$ divides $\sigma_0-k+1$. Because $m$ divides $\sigma_0$, this implies that $m$ divides $2\sigma_0-k+1$. Furthermore, as $\sigma_0/m$ is odd, so is $(2\sigma_0-k+1)/m$. It follows that $p^m+1$ divides $p^{2\sigma_0-k+1}+1$, and therefore $p^m+1$ divides $p^{2\sigma_0}+p^{k-1}$. Thus, (\[eq:misceq1\]) is an equality in this case as well. We conclude that $g\in O_K(V)$, and hence $\zeta$ is in the image of (\[eqn:inclusion\]).
The image of (\[eqn:inclusion\]) is equal to the subgroup $\mu_{p^m+1}(k)\subset\mu_{p^{\sigma_0}+1}(k)$, where $m$ is the largest divisor of $\sigma_0$ such that
1. $\sigma_0/m$ is odd, and
2. if $1\leq i\leq\sigma_0-1$ and $2m$ does not divide $i$ then $a_i=0$.
or $m=0$ if no such divisor exists.
The inclusion (\[eqn:inclusion\]) realizes $O_K(V)$ as a subgroup of the cyclic group $\mu_{p^{\sigma_0}+1}(k)$. Thus, $O_K(V)$ is itself cyclic, say of order $k$. Let $g$ be a generator, and let $\zeta$ be the image of $g$ in $\mu_{p^{\sigma_0}+1}(k)$. Let $m'$ be the smallest non-negative integer such that $k$ divides $p^{m'}+1$. The order of $p$ modulo $k$ is then $2m'$. We also know that $p^{\sigma_0}+1$ is divisible by $k$, so therefore $m'$ divides $\sigma_0$, and $\sigma_0/m'$ is odd. By Proposition \[prop:subgroup\] applied to $m'$, we deduce that $a_i=0$ for every $1\leq i\leq\sigma_0-1$ such that $2m'$ does not divide $i$. That is, $m'$ is a divisor of $\sigma_0$ satisfying properties (1) and (2). Furthermore, as $g$ was a generator, we have that the image of (\[eqn:inclusion\]) is contained in $\mu_{p^{m'}+1}(k)$. On the other hand, it follows from Proposition \[prop:subgroup\] that $\mu_{p^m+1}(k)$ is contained in the image of (\[eqn:inclusion\]). We conclude that $\mu_{p^m+1}(k)\subset\mu_{p^{m'}+1}(k)$. By maximality of $m$, we deduce that $m=m'$, and the result follows.
Note that if the structure constants $a_i$ associated to $K$ are all non-zero then $m=0$. Thus, if $K$ is generic then $\#O_K(V)=2$. We remark that a similar calculation is carried out in Theorem 3.1 of [@2017arXiv170202290J]. There is also given a table of the dimensions of the locus of characteristic subspaces with different values of $m$ for each Artin invariant $1\leq \sigma_0\leq 10$.
Counting twisted Fourier-Mukai partners {#sec:counting}
---------------------------------------
We apply our techniques to count twisted partners of supersingular K3 surfaces. The analogous computation over the complex numbers is carried out in [@MR2550163]. We fix a twisted supersingular K3 surface $(X,\alpha)$ with Artin invariant $\sigma_0=\sigma_0(X,\alpha)$. We define $$\FM(X,\alpha)=\left\{
\begin{tabular}{@{}c@{}}
twisted supersingular K3 surfaces $(Y,\beta)$ \\
such that $D(X,\alpha)\cong D(Y,\beta)$\\
\end{tabular}
\right\}/\,\mbox{isomorphism}$$ Note that by Proposition \[prop:FMpartner1\], any twisted partner of $(X,\alpha)$ must be a twisted supersingular K3 surface. Our goal is to enumerate this set. We will do this by using characteristic subspaces and the computations of Section \[sec:orthogonal\].
If $V$ is an $\bF_p$-vector space with a bilinear form, we let $I(V)$ denote the set of non-zero vectors $v\in V$ such that $v^2=0$.
There is an obvious action of $O_K(V)$ on $I(V)$. Fix a B-field lift $B$ of $\alpha$. Let $(K,V)=(\tK(X,B),\tT(X,B)_0)$ be the characteristic subspace datum associated to $\tH(X/W,B)$.
\[thm:Count1\] $$\label{eq:count}
\#\FM(X,\alpha)=
\begin{cases}
\# O_{K}(V)\setminus I(V)+1&\mbox{ if }\sigma_0\leq 10\\
\# O_{K}(V)\setminus I(V)&\mbox{ if }\sigma_0=11\\
\end{cases}
$$
We let $\FM^d(X,\alpha)$ be the subset of $\FM(X,\alpha)$ consisting of those Fourier-Mukai partners $(X_1,\alpha_1)$ such that $\alpha_1$ has order $d$. Of course, in our situation $\FM^d(X,\alpha)$ is empty unless $d=1,p$. We can dispense with the $d=1$ case quickly.
If $\sigma_0\leq 10$, then $\#\FM^1(X,\alpha)=1$. If $\sigma_0=11$, then $\#\FM^1(X,\alpha)=0$.
This follows immediately from Corollary \[cor:Untwisted2\].
We next turn to $\FM^p(X,\alpha)$. Note that if $\sigma_0=1$, then this set is empty, so let us assume that $\sigma_0\geq 2$. If $(Y,\beta)\in \FM^p(X,\alpha)$, then there exists a $B$ field lift $B'$ of $\beta$ and an isometry $$g:\tT(Y,B')_0\iso V=\tT(X,B)_0$$ that sends $\tK(Y,B')$ to $K=\tK(X,B)$. Because the order of $\beta$ is $p$, the functional given by pairing with the vector $(0,0,1)\in\tT(Y,B')$ is divisible by $p$. As before, we denote by $f_Y$ the resulting (non-trivial) element $p^{-1}(0,0,1)\in\tT(Y,B')_0$. The image of this element under $g$ gives an element of $I(V)$. Note that this element does not depend on our choice of $B'$, although it does depend on $g$. We define a map of sets $$\mu:\FM^p(X,\alpha)\to O_K(V)\setminus I(V)$$ by $\mu(\left[(Y,\beta)\right])=\left[g(f_Y)\right]$, where $O_K(V)$ acts on $I(V)$ on the left in the obvious way. This is well defined, independent of the choices of $B'$ and $g$.
The map $\mu$ is a bijection.
We first show injectivity. Consider two Fourier-Mukai partners $$(X_1,\alpha_1),(X_2,\alpha_2)\in\FM^p(X,\alpha)$$ such that $\mu(X_1,\alpha_1)=\mu(X_2,\alpha_2)$. Pick $B$-field lifts $B_1,B_2$, and choose isometries $g_i:\tT(X_i,B_i)_0\iso\tT(X,B)_0$, $i=1,2$ that preserve the respective characteristic subspaces. By assumption, there exists $g\in O_K(V)$ such that $$g(g_1(f_{X_1}))=g_2(f_{X_2})$$ Thus, the isometry $$g_2^{-1}\circ g\circ g_1:\tT(X_1,B_1)_0\to\tT(X_2,B_2)_0$$ preserves the characteristic subspaces and satisfies $f_{X_1}\mapsto f_{X_2}$. By Theorem \[thm:characteristictorelli2\], we find an isomorphism $(X_1,\alpha_1)\cong(X_2,\alpha_2)$.
We next show surjectivity. Pick a non-zero isotropic vector $v\in V$. The vector space $V^-=v^\perp/v$ has dimension $2\sigma_0-2$, and because we assumed $\sigma_0\geq 2$, its induced bilinear form is again non-degenerate and non-neutral. Consider the subspace $$K^-=(K\cap v^\perp)/v\subset V^-\otimes k$$ Pick an isotropic vector $w\in V$ such that $v.w=-1$ and a decomposition $$V=V^-\oplus\langle v,w\rangle$$ Let $\left\{x_1,\dots,x_{\sigma_0-1}\right\}$ be a basis for $K^-$. By Proposition \[prop:radbijection\], there is an element $B'\in V^-\otimes k$ such that $$K=\left\langle x_1+(x_1.B')v,\dots,x_{\sigma_0-1}+(x_{\sigma_0-1}.B')v,w+B'+\frac{B'^2}{2}w\right\rangle$$ By the surjectivity of the period morphism, the pair $(K^-,V^-)$ is isomorphic to the characteristic subspace datum associated to $\H^2(Y/W)$ for some supersingular K3 surface $Y$. Under this isomorphism, $B'$ corresponds via Lemma \[lem:coolbijection\] to some class in $\U^2(Y,\m_p)$. Let $\beta$ be its image in $\Br(Y)$. By abuse of notation, let $B'$ denote a B-field lift of $\beta$ that maps to the given class in $\U^2(Y,\m_p)$. By construction, we have an isomorphism $(\tK(Y,B'),\tT(Y,B')_0)\iso (K,V)$ sending $f_Y$ to $v$. By Theorem \[thm:characteristictorelli3\], $(Y,\beta)$ is a Fourier-Mukai partner of $(X,\alpha)$, and its image under $\mu$ is $v$, as desired.
Using the results of Section \[sec:orthogonal\], we can explicitly determine the size of the set $$O_K(V)\setminus I(V)$$ in terms of the structure constants $a_i$ associated to $K$.
\[lem:free\] The action of $O_K(V)$ on $I(V)$ is free. That is, each $v\in I(V)$ is fixed only by the identity.
If $g\in O_K(V)$ fixes $v$, then $$v.e=g(v).g(e)=v.(\zeta e)=\zeta(v.e)$$ Because $v$ is non-zero and fixed by $\varphi$, we must have $v.e\neq 0$. It follows that $\zeta=1$, and hence $g$ is the identity.
Let $(X,\alpha)$ be a twisted supersingular K3 surface with Artin invariant $\sigma_0$, and let $a_i$ be be the constants associated to the strictly characteristic subspace $\tK(X,\alpha)$ corresponding to $(X,\alpha)$ (see \[eq:structureconstants\]). Let $m$ be the largest divisor of $\sigma_0$ such that $\sigma_0/m$ is odd and $a_i=0$ for all $1\leq i\leq\sigma_0-1$ such that $2m$ does not divide $i$, or set $m=0$ if no such divisor exists.
\[Count2\] The number of twisted Fourier-Mukai partners of $(X,\alpha)$ is given by the formula $$\#\FM(X,\alpha)=
\begin{dcases}
\frac{p^{\sigma_0}+1}{p^m+1}(p^{\sigma_0-1}-1)+1 &\mbox{ if }1\leq\sigma_0\leq 10\\
\frac{p^{\sigma_0}+1}{p^m+1}(p^{\sigma_0-1}-1) &\mbox{ if }\sigma_0=11\\
\end{dcases}$$
By eg. [@Ogus78 Lemma 4.12ff], if $V$ is a vector space over $\bF_p$ of dimension $2\sigma_0$ equipped with a non-degenerate, non-neutral bilinear form, then $$\# I(V)=(p^{\sigma_0}+1)(p^{\sigma_0-1}-1)$$ By Lemma \[lem:free\], the action of $O_K(V)$ on $I(V)$ is free. The result follows from Theorem \[thm:Count1\].
The integer $m$ associated to $(X,\alpha)$ is equal to 0 if all of the structure constants $a_i$ are non-zero. This is the case for instance if $X$ is a generic supersingular K3 surface, where by generic we mean that the corresponding point in Ogus’s moduli space of marked supersingular K3 surfaces [@Ogus83] is contained in the complement of an explicit proper closed subset.
Twisted Chern characters {#app:twistedChernCharacters}
========================
In this section we will discuss the twisted Chern characters introduced in Definition \[def:twistedcherncharacter\], with the goal of proving Proposition \[prop:integrality1\] and Theorem \[thm:Mainprelim\]. Although we have phrased the results in this section for $p$-torsion Brauer classes and crystalline B-fields, it will be clear that the proofs apply essentially unchanged to the singular cohomology constructions of [@HS04], the $l$-adic étale theory of [@LMS11], and the de Rham-Witt theory for $p^n$-torsion Brauer classes.
We will begin by comparing twisted sheaves on $\bG_m$-gerbes, $\m_n$-gerbes, and Brauer-Severi varieties. Let $X$ be a smooth projective variety over $k$. We assume that the characteristic of $k$ is a prime $p$, possibly equal to $2$. Let $\pi\colon\sX\to X$ be a $\m_n$-gerbe. If $\sE$ is a locally free coherent sheaf on $\sX$, we define $$\ch_{\sX}(\sE)=\sqrt[\leftroot{-2}\uproot{2}n]{\ch(\pi_*(\sE^{\otimes n}))}\in A^*(X)\otimes\bQ$$ We compare this twisted Chern character to the notion introduced in Definition \[def:twistedcherncharacter\]. Let $\sX_{\bG_m}$ be the $\bG_m$-gerbe obtained from $\sX$. The inclusion $\m_n\hto\bG_m$ induces a natural map $\iota\colon \sX\to\sX_{\bG_m}$, and $\iota_*$ induces an equivalence between the respective categories of twisted sheaves. The gerbe $\sX$ gives a $\m_n$-gerbe lift of $\sX_{\bG_m}$, and hence gives rise to a choice of invertible $n$-fold twisted sheaf $\sL$ on $\sX_{\bG_m}$ (see Definition \[def:twistedcherncharacter\]ff). This sheaf satisfies $\iota^*\sL\cong\cO_{\sX}$. It follows that for any locally free twisted sheaf $\sE$ on $\sX$, we have $$\ch_{\sX}(\sE)=\ch^{\sL}(\iota_*\sE)$$ The pushforward $\sA=\pi_*\sEnd(\sE)$ is an Azumaya algebra on $X$ whose associated cohomology class in $\H^1(X,\PGL_n)$ maps to the Brauer class of $\sX$ under the boundary map $\H^1(X,\PGL_n)\to\H^2(X,\bG_m)$. Let $f\colon P\to X$ be the Brauer-Severi scheme associated to $\sA$. We have a Cartesian diagram $$\begin{tikzcd}
\bP_{\sX}(\sE)\arrow{r}{\pi'}\arrow{d}[swap]{g}&P\arrow{d}{f}\\
\sX\arrow{r}{\pi}&X
\end{tikzcd}$$ The projective bundle $\bP_{\sX}(\sE)\to\sX$ comes with a universal quotient $g^*\sE\twoheadrightarrow\cO(1)$. We may also view $\bP_{\sX}(\sE)$ as a $\m_n$-gerbe via $\pi'$, and the invertible sheaf $\cO(1)$ is 1-twisted with respect to this gerbe structure. Define an invertible sheaf $\cO(n)=\pi'_*\cO(1)^{\otimes n}$ on $P$. The $\m_n$-gerbe $\bP_{\sX}(\sE)\to P$ is essentially trivial (that is, has trivial Brauer class), and is isomorphic to the gerbe of $n$-th roots of $\cO(n)$. Under this isomorphism, the universal quotient $\cO(1)$ acquires a second universal property: it is the universal (twisted) invertible sheaf equipped with an isomorphism $\cO(1)^{\otimes n}\iso \pi'^*\cO(n)$.
\[prop:cohomology\] There is a natural isomorphism of graded $W$-algebras $\H^*(P/W)\cong\H^*(X/W)[x]/f(x)$ where $x\in\H^2(P/W)$ is a class restricting to $c_1^{\cry}(\cO(1))$ on geometric fibers, and $f(x)$ is a monic polynomial of degree $n$. The map $f^*\colon \H^*(X/W)\otimes k\to\H^*(P/W)\otimes k$ is injective.
This follows from the Leray spectral sequence associated to $P\to X$ (see for instance [@Yos06 Lemma 1.6]).
\[prop:integral1\] Suppose that $X$ is a K3 surface, $\alpha\in\Br(X)$ is a Brauer class of order $p$, and $\sE$ is an $\alpha$-twisted locally free sheaf. If $B$ is a B-field lift of $\alpha$, then $\ch^{B}(\sE)\in\tN(X,B)$.
Let $\pi\colon\sX\to X$ be a $\m_p$-gerbe whose associated cohomology class maps to $\alpha$. We regard $\sE$ as a twisted sheaf on $\sX$. It will suffice to show the result for a particular choice of B-field lift. Consider a B-field lift $B=\frac{a}{p}$ of $\alpha$ where the image of $a$ modulo $p$ is equal to $d\log$ of the cohomology class $[\sX]\in\H^2(X,\m_p)$. It follows that $f^*B=c_1^{\cry}(\cO(p))/p+h$ for some $h\in\H^2(P/W)$. By the above discussion, we have $\ch_{\sX}(\sE)=\ch^B(\iota_*\sE)$. The Chern character of a coherent sheaf on a K3 surface is integral, so $\ch_{\sX}(\sE)^p\in\tN(X)$. Thus, $\ch_{\sX}(\sE)\in\tN(X)\otimes\bZ[p^{-1}]$. To prove the result, it will therefore suffice to show that $e^{-B}\ch_{\sX}(\sE)\in\H^*(X/W)$. By Proposition \[prop:cohomology\], this is true if and only if $f^*(e^{-B}\ch_{\sX}(\sE))\in\H^*(P/W)$. We compute $$\begin{aligned}
f^*(e^{-B}\ch_{\sX}(\sE))&=f^*(e^{-B}).f^*\ch_{\sX}(\sE)\\
&=e^{-h}.\sqrt[\leftroot{-2}\uproot{2}p]{\ch(-c_1^{\cry}(\cO(p)).\ch(f^*\pi_*(\sE^{\otimes p}))}\\
&=e^{-h}.\sqrt[\leftroot{-2}\uproot{2}p]{\ch(-c_1^{\cry}(\cO(p)).\ch(\pi'_*g^*(\sE^{\otimes p}))}\\
&=e^{-h}.\sqrt[\leftroot{-2}\uproot{2}p]{\ch(\pi'_*(\pi'^*(\cO(-p))\otimes g^*(\sE^{\otimes p})))}\\
&=e^{-h}.\sqrt[\leftroot{-2}\uproot{2}p]{\ch(\pi'_*((g^*(\sE)\otimes\cO(-1))^{\otimes p})))}\\
&=e^{-h}.\ch(\pi'_*(g^*(\sE)\otimes\cO(-1)))
\end{aligned}$$ Write $G=\pi'_*(g^*(\sE)\otimes\cO(-1))$. We have $$e^{-h}\ch(G)=(r,c_1(G)-rh,c_1(G)^2/2+c_2(G)-h.c_1(G)+rh^2/2,0,\dots,0)$$ where $r=\rk(G)$. The Chern classes $c_i(G)$ are all integral. In particular, this already gives the result if $p\neq 2$. Suppose $p=2$. We compute $$c_1(G)=c_1(\pi_*\det(\sE))-f^*(a)+ph$$ where $a=pB$. It is well known that the pairing on the Néron-Severi group of a K3 surface is even. By [@Ogus83 Theorem 4.7], the pairing on $\H^2(X/W)$ is also even, in the sense that $2$ divides $x.x$ for any $x\in\H^2(X/W)$ (this is a non-trivial statement only if $p=2$). We obtain that $c_1(G)^2$ is divisible by 2, as is $r$, which gives the result.
Proposition \[prop:integrality1\] follows, as the Todd class of a K3 surface is integral. Let $(X,\alpha)$, $(Y,\beta)$ be twisted K3 surfaces such that $\alpha$ and $\beta$ are killed by $p$, and fix B-field lifts $B,B'$. Define $$\label{eq:MoLatticesMoProblems}
\tN(X\times Y,-B\boxplus B')=(N^*(X\times Y)\otimes\bZ[p^{-1}])\cap(e^{-B\boxplus B'}H^*(X\times Y))$$ Using the method of Proposition \[prop:integral1\], we can show that the twisted Chern characters of a twisted sheaf on $X\times Y$ lie in this lattice. Unfortunately, we must assume here that $p\geq 5$, although we would be surprised if this assumption was necessary.[^5] In any case, under the additional assumptions that $X$ and $Y$ are supersingular and that $\Phi_{\sE}$ is an equivalence we will prove the result for all $p$ in Appendix \[app:deformation\].
\[prop:integral2\] If $\sE\in D(X\times Y,-\alpha\boxtimes\beta)$ is a locally free twisted sheaf and $p\geq 5$, then $$\ch^{-B\boxplus B'}(\sE)\in\tN(X\times Y,-B\boxplus B')$$
Let $\sX\to X$ be a $\m_p$-gerbe with cohomology class $\alpha$, and let $\sY\to Y$ be a $\m_p$-gerbe with cohomology class $\beta$. We regard $\sE$ as a $(-1,1)$-twisted sheaf on $\sX\times\sY$. As before, we consider the Cartesian diagram $$\begin{tikzcd}
\bP_{\sX\times\sY}(\sE)\arrow{r}{\pi'}\arrow{d}[swap]{g}&P\arrow{d}{f}\\
\sX\times\sY\arrow{r}{\pi}&X\times Y
\end{tikzcd}$$ There is a universal quotient map $g^*\sE\twoheadrightarrow\cO(1)$, and $\cO(1)$ is $(-1,1)$-twisted. We set $\cO(p)=\pi'_*\cO(1)^{\otimes p}$. As in the proof of Proposition \[prop:integral1\], we may assume that $B$ and $B'$ have the property that $f^*(-B+B')=c_1^{\cry}(\cO(p))/p+h$ for some $h\in\H^2(P/W)$. By the corresponding result in the untwisted case (eg. [@Huy06 Lemma 10.6]), it will suffice to show that $f^*(e^{-B\boxplus B'}\ch_{\sX\times\sY}(\sE))\in\H^*(P/W)$. The same computation as in the proof of Proposition \[prop:integral1\] shows that $$f^*(e^{-B\boxplus B'}\ch_{\sX\times\sY}(\sE))=e^{-h}\ch(\pi'_*(g^*(\sE)\otimes\cO(-1)))$$ The only primes dividing the denominators in right hand side of the above expression are $2$ and $3$, so this gives the result if $p\geq 5$.
Deformations of kernels of Fourier-Mukai transforms {#app:deformation}
===================================================
In this appendix we record some results concerning deformations of kernels of Fourier-Mukai equivalences. The relevant deformation theory has been worked out by Reinecke in [@2017arXiv171100846R], following [@LO15] in the non-twisted case. We apply this to complete the proof of Theorem \[thm:Mainprelim\], and to prove certain cases of Conjecture \[conj:mainPrelim2\].
Let $(X,\alpha)$, $(Y,\beta)$ be twisted supersingular K3 surfaces with B-field lifts $B,B'$, and suppose that $P^{\bullet}\in D(X\times Y,-\alpha\boxtimes\beta)$ is a perfect complex of twisted sheaves inducing a Fourier-Mukai equivalence $$\Phi_{P^\bullet}\colon D(X,\alpha)\to D(Y,\beta)$$ We will show that $P^{\bullet}$ deforms over a family of twisted surfaces connecting $(X\times Y,-\alpha\boxtimes\beta)$ to a product of twisted surfaces with trivial Brauer class. A special feature of the supersingular case is that there is a canonical choice of such a family. Specifically, we show in [@BL17 Lemma 2.2.5ff] that if $X$ is a twisted supersingular K3 surface, then there exists a class $$\alpha_X\in\H^2(X\times\bA^1,\bG_m)$$ with the property that for every $\alpha\in\Br(X)$, there exists a $k$-point $t\in\bA^1$ such that the restriction of $\alpha_X$ to $X\times t=X_{t}$ is equal to $\alpha$, and in particular the restriction of $\alpha_X$ to $X_0$ vanishes (there will be more than one such $t$ for a given $\alpha$).[^6] In particular, all twisted supersingular K3 surfaces with underlying surface $X$ are deformation equivalent.
\[prop:deformations!!\] Consider the universal class $\beta_Y\in\Br(Y\times\bA^1)$, and let $t\in\bA^1$ be a point such that the restriction of $\beta_Y$ to $Y_t$ is equal to $\beta$. There exists a connected étale neighborhood $(U,u)\to (\bA^1,t)$ and
1. a family of K3 surfaces $X_U\to U$ such that $(X_U)_u\cong X$,
2. a class $\alpha_U\in\Br(X_U)$ whose restriction to $(X_U)_u$ is equal to $\alpha$, and
3. a kernel $\widetilde{P}^\bullet\in D(X_U\times Y_U,\alpha_U^{-1}\boxtimes\beta_U)$ inducing a Fourier-Mukai equivalence on each fiber, and such that $\widetilde{P}^\bullet_u\cong P^\bullet$.
Let $\widetilde{Y}\to Y\times\bA^1$ be a $\bG_m$-gerbe with cohomology class $\beta_Y$. Its underlying surface is the constant family $Y\times\bA^1\to\bA^1$, so determinant of any twisted sheaf on a fiber $\widetilde{Y}_x$ is unobstructed. Hence, the stack of simple perfect complexes $s\mathscr{T}w_{\widetilde{Y}/\bA^1}$ is smooth over $\bA^1$ (see for instance Definition 3.7 of [@2017arXiv171100846R]). The result then follows by the proof of Proposition 4.2 of [@2017arXiv171100846R].
To apply this, we record the following lemma.
\[lem:itsalemma\] Let $S$ be the spectrum of a DVR over $k$ with residue field $k$ and field of fractions $L$, $\sX\to S$ and $\sY\to S$ two families of supersingular K3 surfaces, $\alpha\in\Br(X)$ and $\beta\in\Br(Y)$ two Brauer classes, and $\sP^\bullet\in D(\sX\times \sY,\alpha^{-1}\boxtimes\beta)$ a perfect complex of twisted sheaves. The following are equivalent.
1. There exists a geometric point $t\in S$ and B-field lifts $B,B'$ of $\alpha_t$ and $\beta_t$ such that $$v^{-B\boxplus B'}(\sP^{\bullet}_{t})\in\tN(\sX_{t}\times_{t}\sY_{t},-B\boxplus B')$$
2. For every geometric point $t\in S$ there exist B-field lifts $B,B'$ of $\alpha_t$ and $\beta_t$ such that $$v^{-B\boxplus B'}(\sP^{\bullet}_t)\in\tN(\sX_t\times \sY_t,-B\boxplus B')$$
The above notation was introduced in (\[eq:MoLatticesMoProblems\]). Note that if either of the above is true for some choice of B-field lifts, it is in fact true for every choice of B-field lifts. After possibly replacing $S$ with a connected étale cover, we may find a complete lift $S'$ of $S$ over $W$ and elements $\sB\in\H^2(\sX/S'_K)$, $\sB'\in\H^2(\sY/S'_K)$ such that the restriction of $\sB$ to the special fiber is a B-field lift of $\alpha_0$, and the restriction of $\sB$ to any geometric generic fiber $\sX_{\overline{L}}$ is a B-field lift of $\alpha_{\overline{L}}$, and similarly, the restrictions of $\sB'$ give B-field lifts of the restrictions of $\beta$ in both special and geometric fibers. We consider the class $$\exp(\sB\boxplus-\sB') v^{-\sB\boxplus\sB'}(\sP^\bullet)\in\H^*(\sX\times_S\sY/S'_K)$$ (the definition of the twisted Mukai vector makes sense for a twisted sheaf on a gerbe over any scheme). As $S$ is connected, this class is contained in the integral submodule $\H^*(\sX\times_S\sY/S')$ if and only if its restriction to any geometric point is. This gives the result.
It remains to show that the cohomological transform $\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}$ induces an isomorphism $\tH(X/W,B)\iso\tH(Y/W,B')$, and restricts to an isomorphism $\tN(X,B)\iso\tN(Y,B')$. To show this, it will suffice to prove that $$v^{-B\boxplus B'}(P^{\bullet})\in\tN(X\times Y,-B\boxplus B')$$ As the square root of the Todd class of a K3 surface is integral, Proposition \[prop:integral2\] already implies this when $p\geq 5$ (without using the assumptions that $X$ and $Y$ are supersingular, or that $\Phi_{P^{\bullet}}$ is an equivalence).
We apply Proposition \[prop:deformations!!\] to find a connected étale neighborhood $(U,u)\to (\bA^1,t)$, a family of K3 surfaces $X_U\to U$, a class $\alpha_U\in\Br(X_U)$, and a kernel $\widetilde{P}^\bullet\in D(X_U\times Y_U,\alpha_U^{-1}\boxtimes\beta_U)$ extending our given data. Let $U'\to\bA^1$ be the normalization of $U$ in the function field of $\bA^1$. Let $R$ be the hensalization of $U'$ at a closed point mapping to $0\in\bA^1$. Let $L$ be the function field of $R$. Pulling back the data produced by Proposition \[prop:deformations!!\], we find a family of K3 surfaces $X_L$ over $L$, a class $\alpha_L\in\Br(X_L)$, and a kernel $P^\bullet_L\in D(X_L\times_L Y_L,\alpha_L^{-1}\boxtimes\beta_L)$, inducing a Fourier-Mukai equivalence on geometric fibers. In particular, this implies that $X_{\overline{L}}$ is supersingular. After replacing $R$ with a finite extension, we may therefore assume that $X_L$ is the generic fiber of a family $X_R$ of supersingular K3 surfaces over $R$ by [@RS83 Theorem 3] (see also [@BL17 Theorem 5.2.1] for $p=3$). Consider a flat extension $P^\bullet_R$ of $P^{\bullet}_L$ to $X_R\times_RY_R$. By Lemma \[lem:itsalemma\], we reduce to the case when $\beta=0$. Applying the same reasoning with the roles of $X$ and $Y$ swapped, we furthermore reduce to the case where $\alpha=0$. The result then follows from the well known fact that the Mukai vector of any complex on the product of two K3 surfaces is integral (see eg. [@Huy06 Lemma 10.6]).
We now discuss orientation.
\[lem:this is a lemma\] Let $S$ be the spectrum of a DVR with residue field $k$ and field of fractions $L$, $\sX\to S$ and $\sY\to S$ two families of K3 surfaces, and $P\in\tN(\sX\times \sY)_{\bQ}$ a class that induces an isomorphism $\Phi_{P_t}\colon\tN(\sX_t)_{\bQ}\iso\tN(\sY_t)_{\bQ}$ on each geometric fiber $r\in S$. The following are equivalent.
1. There exists a geometric point $t\in S$ such that $\Phi_{P_t}$ is orientation preserving.
2. For every geometric point $t\in S$ the isometry $\Phi_{P_t}$ is orientation preserving.
As explained in [@2017arXiv171100846R], an orientation on a lattice $N$ gives rise to a choice of connected component of the associated positive definite Grassmannian, and an isometry of lattices is orientation preserving if and only if it preserves the choice of connected components. The result follows.
Using Lemma \[lem:this is a lemma\] and arguments of [@MR2553878] and [@LO15], we prove Conjecture \[conj:mainPrelim2\] in certain special cases.
\[thm:nontwisted\] Let $X$ and $Y$ be K3 surfaces over $k$, and suppose that $P^{\bullet}\in D(X\times Y)$ is a complex inducing a Fourier-Mukai equivalence $$\Phi_{P^{\bullet}}\colon D(X)\iso D(Y)$$ The induced cohomological transform $\Phi^{\cry}_{v(P^{\bullet})}$ is orientation preserving.
By composing with standard derived equivalences (which are known to be orientation preserving), we reduce to the case where $\Phi_{P^\bullet}$ is filtered, in the sense of [@LO15]. The lifting results of [@LO15] combined with Lemma \[lem:this is a lemma\] then reduce us to the case of K3 surfaces over the complex numbers, which follows from work of Huybrechts, Macrí, and Stellari [@MR2553878].
Suppose that $(X,\alpha)$ and $(Y,\beta)$ are twisted supersingular K3 surface such that $\sigma_0(X,\alpha),\sigma_0(Y,\beta)\leq 10$. If $P^{\bullet}\in D(X\times Y,\alpha^{-1}\boxtimes\beta)$ is a perfect complex of twisted sheaves that induces a Fourier-Mukai equivalence $$\Phi_{P^{\bullet}}\colon D(X,\alpha)\to D(Y,\beta)$$ then for any B-field lifts $B,B'$ of $\alpha$ and $\beta$ the cohomological correspondence $\Phi^{\cry}_{v^{-B\boxplus B'}(P^{\bullet})}$ is orientation preserving.
Under our assumptions on the Artin invariant, we may find primitive isotropic vectors $v,w\in\tN(X,B)$ such that $v.w=-1$, and $v$ is effective. By results of [@BL17] (extending results of Mukai and Yoshioka over the complex numbers) the moduli space $X'=M_{(X,-\alpha)}(v)$ of $-\alpha$-twisted sheaves on $X$ with Mukai vector $v$ is a supersingular K3 surface, and the universal sheaf induces a derived equivalence $D(X,\alpha)\to D(X')$. As in [@HS04], one shows that the induced cohomological correspondence is orientation preserving. Let $Y'$ be the same for $(Y,\beta)$. Now, given a Fourier-Mukai equivalence $D(X,\alpha)\to D(Y,\beta)$ that is not orientation preserving, we obtain by composition a Fourier-Mukai equivalence $D(X')\to D(Y')$ that is not orientation preserving, in contradiction to Theorem \[thm:nontwisted\].
One can show that if $X$ is a K3 surface that is not supersingular and $\alpha\in\Br(X)$ is any Brauer class, then $(X,\alpha)$ lifts to characteristic 0. The argument of Theorem \[thm:nontwisted\] then extends to prove Conjecture \[conj:mainPrelim2\] for all twisted K3 surfaces in positive characteristic that are not supersingular. Thus, the only case of Conjecture \[conj:mainPrelim2\] left open is that of non-trivial Brauer classes on supersingular K3 surfaces of Artin invariant $10$.
[^1]: This is the strongest possible notion of supersingularity, and is sometimes known as *Shioda supersingular*. This condition is equivalent to all of the various weaker notions appearing in the literature (see eg. Chapter 18 of [@Huy06]) by work of Charles, Kim, Madapusi Pera, and Maulik [@Charles13; @Charles14; @KMP16; @MP15; @Mau14].
[^2]: That is, $\Phi(\lambda h)=\sigma(\lambda)\Phi(h)$ for all $h\in\H^2(X/W)$ and $\lambda\in W$.
[^3]: Ogus proved this result under the assumption that $p\geq 5$. It has been extended to $p\geq 3$ in [@BL17]
[^4]: We do not really need Theorem \[thm:Mainprelim\] to make this definition, as it is immediate that the cohomological transform gives an isometry after tensoring with $\bQ$, which is enough to define the property of being orientation preserving.
[^5]: In fact, we would imagine that the same method should give this result for all $p$, although we have not been able to make it work.
[^6]: Roughly speaking, the class $\alpha_X$ realizes $\bA^1$ as the moduli space of twisted supersingular K3 surfaces with underlying surface $X$ equipped with a certain level structure. This viewpoint is developed in detail in [@BL17]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we motivate the need for a publicly available, generic software framework for question-answering (QA) systems. We present an open-source QA framework [**<span style="font-variant:small-caps;">Qanus</span>**]{} which researchers can leverage on to build new QA systems easily and rapidly. The framework implements much of the code that will otherwise have been repeated across different QA systems. To demonstrate the utility and practicality of the framework, we further present a fully functioning factoid QA system [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} built on top of [**<span style="font-variant:small-caps;">Qanus</span>**]{}.'
author:
- |
Jun-Ping Ng\
Department of Computer Science\
National University of Singapore\
[[email protected]]{} Min-Yen Kan\
Department of Computer Science\
National University of Singapore\
[[email protected]]{}
bibliography:
- 'Bibliography.bib'
title: 'QANUS: An Open-source Question-Answering Platform'
---
Introduction
============
There has been much research into question-answering (QA) over the past decades. However the community is still lacking QA systems which are readily available for use. This translates into a high barrier of entry for researchers who are new to the field. The absence of easily accessible systems also means that there is a lack of credible, reproducible baseline systems against which new QA systems can be evaluated.
To address the highlighted limitations, we are releasing an open-source, Java-based, QA framework [**<span style="font-variant:small-caps;">Qanus</span>**]{} (pronounced KAY-NESS). [**<span style="font-variant:small-caps;">Qanus</span>**]{} is a framework on which new QA systems can be easily and rapidly developed. [**<span style="font-variant:small-caps;">Qanus</span>**]{} makes it easy to build new QA systems as only a minimal set of components needs to be implemented on top of the provided framework. To demonstrate the utility and practicality of [**<span style="font-variant:small-caps;">Qanus</span>**]{}, a reference implementation of a QA system [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} has also been developed using the framework. [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} is also made available to the community. When it matures, it can serve as an accessible, reproducible baseline system for evaluations.
To ensure the availability of the system to the community, as well as to maximise the benefits of any derivative projects for everyone, [**<span style="font-variant:small-caps;">Qanus</span>**]{} is released under the Open Software License (OSL) v3.0.
Related Work
============
There has been previous efforts in generalising the architecture of QA systems. Hirschman and Gaizauskas for example described a pipelined approach to QA ([<span style="font-variant:small-caps;">HG-01</span>]{}), where different stages are combined serially into a QA system. Figure \[fig:hirschman-compare\] highlights the different stages in their pipeline vis-a-vis the stages found in [**<span style="font-variant:small-caps;">Qanus</span>**]{}. The informal correspondence between the various stages of the two pipelines are also shown in the figure.
![Comparing pipeline stages of [<span style="font-variant:small-caps;">HG-01</span>]{} and [**<span style="font-variant:small-caps;">Qanus</span>**]{}.[]{data-label="fig:hirschman-compare"}](hirschman-compare.jpg){width="0.95\linewidth"}
The architecture of [<span style="font-variant:small-caps;">HG-01</span>]{} is slanted towards QA systems based on current state-of-the-art information retrieval (IR) techniques. These techniques typically involve manipulating the lexical and syntactic form of natural language text and do not attempt to comprehend the semantics expressed by the text. Systems which make use of these techniques [@Hickl2007; @Chali2007] have been able to perform ahead of their peers in the Text Retrieval Conference (TREC) QA tracks [@Dang2007].
In IR-based systems, answer processing revolves around units of information stored in documents. To reflect the importance of this organisation two separate stages *(c) candidate document selection* and *(d) candidate document analysis* are described in Hirschman’s architecture. Further, *(f) answer generation* is included as they considered interactive QA systems which could participate in a dialogue with end-users.
Not all QA systems are IR-centric however, and interactive QA systems are likely not imminent given the limitations of natural language understanding and generation. [**<span style="font-variant:small-caps;">Qanus</span>**]{} thus generalises stages *(c)*, *(d)* and *(e)* into one to avoid over-committing to any particular architecture or paradigm, and leaves out *(f)*.
Another important point of comparison is that [**<span style="font-variant:small-caps;">Qanus</span>**]{} is an implemented, functional QA architecture whereas [<span style="font-variant:small-caps;">HG-01</span>]{} serves mainly as a general discussion and introduction to the architecture of QA systems.
Though few in numbers, some QA systems have previously been made available to the community. One such system is <span style="font-variant:small-caps;">Aranea</span>[^1] [@Lin2007]. <span style="font-variant:small-caps;">Aranea</span> is a factoid QA system which seeks to exploit the redundancy of data on the web and has achieved credible performances at past TREC evaluations. <span style="font-variant:small-caps;">Aranea</span> is not designed however as a generic QA platform. We argue that a framework such as [**<span style="font-variant:small-caps;">Qanus</span>**]{} which is designed from the start with extensibility and flexibility in mind will greatly reduce the effort needed for any such customisation.
<span style="font-variant:small-caps;">Qanda</span> by <span style="font-variant:small-caps;">Mitre</span>[^2] is another QA system which has featured in the TREC QA track. It has a project page on SourceForge. However currently only one module of the system is made available for download. We are at the time of writing unable to verify if there are plans for the release of the rest of the system in the near future.
[**<span style="font-variant:small-caps;">Qanus</span>**]{} Framework
=====================================================================
{width="0.90\linewidth"}
The [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework adopts a pipelined approach to QA. The pipeline consists of four stages executed serially.. The stages include (1) *information source preparation*, (2) *question processing*, (3) *answer retrieval* and (4) *evaluation*. Within the framework we have implemented much of the programming code that will otherwise have been repeated across different QA systems. The framework can thus be likened to a foundation on top of which components can be added to obtain a complete QA system.
Figure \[fig:extensions\] illustrates a complete QA system built with the framework. The upper-half of the figure delineates clearly the key classes that constitute the four stages of the framework listed earlier. The bottom-half of the figure shows additional components that can be added to the framework to complete the QA system. For completeness, the input and output to the various stages of the system are also depicted as shaded boxes at the bottom of the figure.
The top half of Figure \[fig:extensions\] shows that each of the stages share a common architecture, composed of two main classes. The `FrameworkController` is responsible for directing the program flow and managing any input and output required or produced by the stage. It also invokes appropriate methods in the latter to process any input sent to the stage. The `FrameworkEngine` class provides the required processing that is needed on the various pieces of input to the stage. The processing that is required in each stage differs. For example, in the *information source preparation* stage, processing may involve part-of-speech tagging an input corpus, while in *question processing*, processing may instead be classifying the expected answer type of the posed questions.
Due to space constraints, the individual interfaces and function calls presented by [**<span style="font-variant:small-caps;">Qanus</span>**]{} are not explained in detail here. The full documentation together with the source code for the framework are available at the [**<span style="font-variant:small-caps;">Qanus</span>**]{} download site[^3].
We briefly explain the operations that may be carried out in each stage. Note that this description serves merely as a guide, and users of the framework have full flexibility in deciding the operations to be carried out at each stage.
**Information Source Preparation**. In this stage, an information source from which answers are to be obtained is set up. The framework is not restricted to any particular type of information source. Depending on the required needs and specifications, the eventual information source can be as varied as a <span style="font-variant:small-caps;">Lucene</span>[^4] index of the source documents, a full-fledged ontology or the Internet. Any necessary pre-processing to set up the information source is done here. Note that this stage prepares static information sources. Using the Web dynamically as an information source is done in the subsequent answer retrieval stage. **Question Processing**. Typically, questions posed to the system need to be parsed and understood before answers can be found. Necessary question processing is carried out here. Typical operations here can include forming a query suitable for the information source from the posed questions, question classification to determine the expected answer type, as well as part-of-speech tagging and parsing. The outputs of these various operations are stored so that they can subsequently be used by the next stage in the QA pipeline.
**Answer Retrieval**. The answer retrieval stage makes use of the annotations from the question processing stage, and looks up the information source for suitable answers to the posed questions. Incorporating candidate answers from dynamic sources, such as the Web or online databases, can also be incorporated here. Proper answer strings that can answer the questions are extracted in this stage. If desired, answer validation can be performed as well.
**Evaluation**. With the three stages above, [**<span style="font-variant:small-caps;">Qanus</span>**]{} already provides the support necessary for a fully functional QA system. The *evaluation* stage is introduced to complement the earlier stages and ease the verification of the performance of the developed QA system. It is optional and may be omitted if desired. The evaluation stage cross-checks the answers computed previously by the answer retrieval stage with a set of *gold-standard* answers. The results of the evaluation are then output for easy review.
Additional Components
---------------------
The four stages of the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework establish the flow of data through the entire QA pipeline, and form the backbone of any instantiated QA system. To realise the framework and obtain a fully functional QA system, additional components such as those shown in the bottom half of Figure \[fig:extensions\] must be coupled to the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework.
The classes in the framework enforce the required interfaces that need to be adhered to by these additional components. By following the specified interfaces, any desired functionality can be plugged into the framework.
To give a better picture of how these components can be easily added to the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework to complete a QA system, let us walk through an example for the *question processing* (QP) stage. From Figure \[fig:extensions\], the minimum set of components that need to be implemented for QP include the `QPController`, `QuestionInputHandler`, and `QPEngine`.
**QPController**. `QPController` inherits from the `QPFrameworkController` component of the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework. This component is responsible for initializing and integrating any text processing modules that will be used to process input questions with the framework. Suppose we want to perform part-of-speech tagging on the input questions, a part-of-speech component module needs to be created in `QPController`. `QPController` next notifies the `QPEngine` component about this part-of-speech tagger component.
**QuestionInputHandler**. This component is responsible for reading in provided input questions. The implementation is thus dependent on how the input questions are formatted and presented. **QPEngine**. This component is derived from the `QPFrameworkEngine` component of the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework. It makes use of the earlier `QuestionInputHandler` component to read in input questions, and invokes any text processing modules registered with it by the `QPController` to annotate the question text.
It is useful to emphasise here the ease and flexibility provided by the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework: (1) The abstraction provided by the framework greatly reduces the amount of code that needs to be written for a QA system. Only a minimal set of customisation needs to be carried out to complete the implementation of the QP stage. (2) The framework is sufficiently flexible to allow for a range of QA systems to be built. In the explanation here, only a part-of-speech tagger is described. Depending on requirements, other text processing algorithms and techniques can also be incorporated.
Implementation of [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{}
==============================================================================
To demonstrate the utility and practicality of the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework, we have developed a QA system, referenced to as [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} on top of the framework. The implementation of [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} is included when downloading [**<span style="font-variant:small-caps;">Qanus</span>**]{} to serve as an effective reference implementation and help reduce the learning curve for researchers in using the framework.
[**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} is a fully functioning QA system developed to run on the well-known dataset from the TREC 2007 QA track [@Dang2007]. [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} makes use of IR-based techniques to perform the QA task. As can be seen later, this includes making use of a text search engine to perform document lookup, as well as lexicon-based techniques including named entity recognition for answer retrieval. An IR-based approach is adopted because it has been shown to turn in credible performances as explained earlier [@Hickl2007; @Chali2007].
Conforming to the description of the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework, Figure \[fig:actual-implementation\] shows the various classes that have been implemented as part of [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{}. This figure is similar to Figure \[fig:extensions\], which shows possible components needed to obtain a complete QA system.
{width="90.00000%"}
**Information Source Preparation**. Similar to the participating machines of the TREC 2007 QA track, [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} makes use of the <span style="font-variant:small-caps;">Aquaint-2</span> corpus[^5] which is stored in XML format. A XML parser `AQUAINTXMLParser` is written to interface the corpus with [**<span style="font-variant:small-caps;">Qanus</span>**]{}. `LuceneWriter` makes use of <span style="font-variant:small-caps;">Lucene</span> to build an index of the input corpus. We will subsequently make use of this index to retrieve documents relevant to posed questions in the later stages of the QA pipeline.
**Question Processing**. In this stage, [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} attempts to classify the expected answer type of the input questions based on the taxonomy described in Li and Roth with `QuestionClassifier`. We built the classifier used by training the Stanford Classifier [@Manning2003] on the data described in Li and Roth . The classification assigned to each question is stored and passed on to the *answer retrieval* stage.
**Answer Retrieval**. To look up answers to the posed questions, [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} form a query out of the question by dropping stop-words found in the question. `LuceneQuery` uses this query to search through the <span style="font-variant:small-caps;">Lucene</span> index built earlier in the *information source preparation* stage. Documents retrieved by the <span style="font-variant:small-caps;">Lucene</span> search engine are then broken down into individual passages. `AnswerRetrieval` scores each of these passages using a variety of heuristics such as by tabulating the occurrences of the query terms within the passages.
From the ranked passages, answer candidates are extracted depending on the expected answer type previously determined in *question processing*. For a question seeking a person name for example, a named entity recogniser [@Finkel2005] is used to extract candidate people names from the ranked passages. For other expected answer types such as dates, hand-written regular expressions are used to aid in the extraction of answer candidates.
Finally, the answer candidates are ranked based again on a set of heuristics which include the proximity of the candidates within the ranked passages to the query terms for example. The highest ranked candidate is returned as the preferred answer.
**Evaluation**. The evaluation stage provided by the [**<span style="font-variant:small-caps;">Qanus</span>**]{} framework makes it possible to easily test the performance of [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{}. Currently [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} supports only factoid questions, and so the evaluation metric used here is factoid *accuracy* [@Dang2007], defined as: $$\begin{aligned}
\textit{accuracy} = \frac{\textrm{no. of correctly answer questions}}{\textrm{total no. of test factoid questions}}\end{aligned}$$
which is implemented in `FactoidAccuracyEvaluator`.
The top system in the TREC 2007 QA track <span style="font-variant:small-caps;">LymbaPA07</span> and the tenth-placed system <span style="font-variant:small-caps;">Quanta</span> achieved accuracy scores of 0.706 and 0.206 respectively. [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} currently obtains an accuracy of 0.119. There is room for improvement before [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} can catch up with the state-of-the-art. The current implementation is simplistic and does not do much processing of the input questions, nor does it perform elaborate ranking of retrieved documents. As work on the system progresses and more sophisticated components are included into the system, [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} should be able to achieve better results.
Future Work
===========
[**<span style="font-variant:small-caps;">Qanus</span>**]{} and [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} are currently under development. [**<span style="font-variant:small-caps;">Qanus</span>**]{} is relatively mature, having undergone several iterations of improvements and our work is now focused on improving the performance and functionalities of [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{}.
**Performance**. Conventionally, QA systems have been benchmarked against the systems participating in the TREC QA track. However recently the QA track has been dropped from both TREC and the Text Analysis Conference (TAC). As the years go by, the results from the QA track will age and become irrelevant. There is also a trend towards the use of the Web as an aid for QA. The Web is dynamic and any such QA system will likely not generate the same results in different instances of time. For useful benchmarking, it is thus important to be able to use a baseline system which makes use of the Internet at the same time instance as the QA system being compared to. Having access to such a baseline system is thus critical and essential. This is the niche that [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} serves to address.. When the performance of [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} catches up with the state-of-the-art, it will be a useful baseline system against which other QA systems can be evaluated against.
To boost performance, more work needs to be done for the *question processing* and *answer retrieval* stages. There are plans to include a query expansion component which will be helpful in boosting the precision of the documents retrieved by <span style="font-variant:small-caps;">Lucene</span>. To improve on answer retrieval, *soft* patterns as described in Cui [*et al.*]{} can replace the current *hard* hand-written patterns used in the system. More advanced measures like the use of dependency relations [@Cui2005] can also be adopted to improve on the current passage ranking implementation.
**List questions**. Besides performance, it will also be useful to expand the functionalities of [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{}. It does not handle list questions for the moment. An implementation based on the use of redundancies found within the source text [@Banko2002; @Lin2007] is being considered.
**Internet front-end**. An online demonstration of [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} is currently hosted online[^6] and supports querying over a pre-indexed <span style="font-variant:small-caps;">Aquaint-2</span> corpus or the Internet. The answer retrieval component working with data from the Internet is rudimentary and lacks techniques to process the noise that accompanies data downloaded from the Internet. It will be useful to improve on this Internet-querying component by adding better post-processing over the retrieved data.
Conclusion
==========
The lack of community-available QA systems has made it difficult to create new QA systems and perform comparisons across published studies. This motivated our work on an open-source QA framework [**<span style="font-variant:small-caps;">Qanus</span>**]{}. The framework implements much of the code needed for a QA system and reduces the development effort needed to build new systems. It is carefully designed to be flexible and supports the use of a wide range of QA techniques.
As a demonstration of the utility and practicality of [**<span style="font-variant:small-caps;">Qanus</span>**]{}, we have also implemented a fully functional factoid QA system [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} on top of the framework. Our goal is to improve [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} so that it will serve as a useful and accessible baseline to benchmark future QA systems and technologies against. Through this work, we hope to lower the high barriers of entry facing new QA researchers and reduce the time needed for them to begin productive research in this area.
[^1]: Available for download at http://www.umiacs.umd.edu/$\sim$jimmylin/downloads/index.html
[^2]: http://www.openchannelsoftware.org/projects/Qanda
[^3]: http://junbin.com/qanus
[^4]: Open-source text search engine written in Java
[^5]: The corpus is not included with the download for [**<span style="font-variant:small-caps;">Qa-Sys</span>**]{} as it is the intellectual property of the <span style="font-variant:small-caps;">Linguistic Data Consortium</span>.
[^6]: http://wing.comp.nus.edu.sg/$\sim$junping/qanus/online/main.php
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Upon introducing a one-parameter quadratic deformation of the $q$-boson algebra and a diagonal perturbation at the end point, we arrive at a semi-infinite $q$-boson system with a two-parameter boundary interaction. The eigenfunctions are shown to be given by Macdonald’s hyperoctahedral Hall-Littlewood functions of type $BC$. It follows that the $n$-particle spectrum is bounded and absolutely continuous and that the corresponding scattering matrix factorizes as a product of two-particle bulk and one-particle boundary scattering matrices.'
address: ' Facultad de Matemáticas, Pontificia Universidad Católica de Chile, Casilla 306, Correo 22, Santiago, Chile'
author:
- 'J.F. van Diejen'
- 'E. Emsiz'
date: May 2013
title: |
The semi-infinite $q$-boson system\
with boundary interaction
---
[^1]
Introduction {#sec1}
============
The $q$-boson system [@bog-ize-kit:correlation] is a lattice discretization of the one-dimensional quantum nonlinear Schrödinger equation [@gau:fonction; @gut:quantum; @kor-bog-ize:quantum; @mat:many-body; @sut:beautiful] built of particle creation and annihilation operators representing the $q$-oscillator algebra [@kli-sch:quantum 5]. Its $n$-particle eigenfunctions are given by Hall-Littlewood functions [@tsi:quantum; @kor:cylindric; @die-ems:diagonalization]. In the present paper we study a system of $q$-bosons on the semi-infinite lattice with boundary interactions, in the spirit of previous works concerned with the quantum nonlinear Schrödinger equation on the half-line [@gau:boundary; @gat-lig-min:nonlinear; @hal-lan:exact; @cau-cra:exact; @tra-wid:bose].
Specifically, by introducing at the end point creation and annihilation operators representing a quadratic deformation of the $q$-oscillator algebra together with a diagonal perturbation, we arrive at the hamiltonian of a $q$-boson system on the semi-infinite integer lattice endowed with a two-parameter boundary interaction. By means of an explicit formula for the action of the hamiltonian in the $n$-particle subspace, it is deduced that the Bethe Ansatz eigenfunctions are given by Macdonald’s three-parameter Hall-Littlewood functions with hyperoctahedral symmetry associated with the $BC$-type root system [@mac:orthogonal §10].
It follows that the $q$-boson system fits within a large class of discrete quantum models with bounded absolutely continous spectrum for which the scattering behaviour was determined in great detail by means of stationary phase techniques [@rui:factorized; @die:scattering]. In particular, the $n$-particle scattering matrix is seen to factorize as a product of explicitly computed two-particle bulk and one-particle boundary scattering matrices.
Semi-infinite $q$-boson system {#sec2}
==============================
Let $$\label{AFock}
\mathcal{F}:=\bigoplus_{n\in\mathbb{N}} \mathcal{F}(\Lambda_n)$$ denote the algebraic Fock space consisting of finite linear combinations of $f_n\in\mathcal{F}(\Lambda_n)$, $n\in\mathbb{N}:=\{0,1,2,\ldots \}$, where $\mathcal{F}(\Lambda_n)$ stands for the space of functions $f:\Lambda_n\to\mathbb{C}$ on the set of partitions of length at most $n$: $$\label{dominant}
\Lambda_n:=\{\lambda=(\lambda_1,\dots, \lambda_n)\in\mathbb{N}^n \mid \lambda_1\geq \lambda_2 \geq \cdots \geq \lambda_n\} ,$$ with the additional convention that $\Lambda_0:=\{ 0\}$ and $\mathcal{F}(\Lambda_0):=\mathbb{C}$. For $l\in\mathbb{N}$, we introduce the following actions on $f\in\mathcal{F}(\Lambda_n)\subset \mathcal{F}$: $$(\beta_l f)(\lambda):=
f(\beta_l^*\lambda)
\qquad\qquad\ (\lambda\in \Lambda_{n-1})$$ if $n>0$ and $\beta_l f:=0$ if $n=0$, $$\begin{split}
(\beta^*_l f)(\lambda)&:=
\begin{cases}
[m_l(\lambda)](1-c\delta_{l}q^{m_0(\lambda)-1})f(\beta_l\lambda)&\text{if}\ m_l(\lambda)>0 \\
0&\text{otherwise}
\end{cases}
\quad ( \lambda\in \Lambda_{n+1}),\\
(q^{N_l+k} f)(\lambda)&:=q^{m_l(\lambda)+k}f(\lambda) \qquad\quad (\lambda\in \Lambda_n),
\end{split}$$ with $q,c\in\mathbb{R}$ such that $|q|\neq 0,1$ and $k\in\mathbb{Z}$. Here $$\delta_l:= \begin{cases} 1 & \text{for}\ l=0,\\
0 & \text{otherwise}
\end{cases},
\qquad
[m]:=\frac{1-q^{m}}{1-q}=\begin{cases} 0&\text{for}\ m=0\\
1+q+\cdots +q^{m-1}&\text{for}\ m>0
\end{cases},$$ and the multiplicity $m_l(\lambda)$ counts the number of parts $\lambda_j$, $1\leq j\leq n$ of size $\lambda_j=l$ (so $m_0(\lambda)$, $\lambda\in\Lambda_n$ is equal to $n$ minus the number of nonzero parts), while $\beta^*_l\lambda\in\Lambda_{n+1}$ and $\beta_l\lambda\in \Lambda_{n-1}$ stand for the partitions obtained from $\lambda\in\Lambda_n$ by inserting/deleting a part of size $l$, respectively (where it is assumed in the latter situation that $m_l(\lambda)>0$). It is clear from these definitions that $\beta_l$, $\beta_l^*$ and $q^{N_l+k}$ map $\mathcal{F}(\Lambda_n)$ into $ \mathcal{F}(\Lambda_{n-1})$, $ \mathcal{F}(\Lambda_{n+1})$ and $ \mathcal{F}(\Lambda_{n})$, respectively (with the convention that $ \mathcal{F}(\Lambda_{-1})$ is the null space).
The operators in question represent a quadratic deformation of the $q$-boson field algebra at the boundary site $l=0$ parametrized by the constant $c$:
$$\begin{aligned}
\beta_l q^{N_l} = q^{N_l+1}\beta_l,\ & \
\beta_l^*q^{N_l} =q^{N_l-1} \beta_l^* ,\nonumber \\
\beta_l\beta_l^*=[N_l+1](1-c\delta_{l}q^{N_0}),\ & \ [\beta_l,\beta_l^*]_q =1-c\delta_{l}q^{2N_0}\end{aligned}$$
and preserving the ultralocality: $$[\beta_l,\beta_k]=[\beta^*_l,\beta^*_k]=[N_l,N_k]=[N_l,\beta_k]=[N_l,\beta^*_k]=[\beta_l,\beta^*_k]= 0$$
for $l\neq k$ (where $[A,B]:=AB-BA$, $[A,B]_q:=AB-qBA$, and $[N_l+r]:=(1-q^{N_l+r})/(1-q)$).
When interpreting the characteristic function $|\lambda\rangle\in\mathcal{F}(\Lambda_n)$ supported on $\lambda\in\Lambda_n$ as a state representing a configuration of $n$ particles on $\mathbb{N}$ such that $m_l(\lambda)$ particles are siting on the site $l\in\mathbb{N}$, it is clear that the operators $\beta_l$ and $\beta_l^*$ act as particle annihilation and creation operators: $$\beta_l |\lambda\rangle = \begin{cases}
|\beta_l\lambda\rangle &\text{if}\ m_l(\lambda)>0 \\
0&\text{otherwise}
\end{cases},\quad
\beta_l^* |\lambda\rangle =[m_l(\lambda)+1](1-c\delta_{l}q^{m_0(\lambda)}) | \beta_l^*\lambda\rangle,$$ while $q^{N_l}$ counts the number of particles at the site $l$ (as a power of $q$): $$q^{N_l}|\lambda\rangle=q^{m_l(\lambda)}|\lambda\rangle .$$
The dynamics of our $q$-boson system is governed by a hamiltonian built of left and right hopping operators together with a diagonal boundary term: $$\label{qbH}
\text{H}_q =a[N_0] +\sum_{l\in \mathbb{N}} (\beta_{l+1}\beta^*_l+\beta_{l+1}^*\beta_l ),$$ $a\in\mathbb{R}$. This hamiltonian constitutes a well-defined operator on $\mathcal{F}$ as for any $f\in\mathcal{F}(\Lambda_n)$ and $\lambda\in\Lambda_n$ the infinite sum $(\text{H}_q f)(\lambda)$ contains only a finite number of nonvanishing terms.
The $n$-Particle hamiltonian and its eigenfunctions {#sec3}
===================================================
By construction $\text{H}_q$ preserves the $n$-particle subspace $\mathcal{F}(\Lambda_n)$. The following proposition describes the action of the hamiltonian in this subspace explicitly.
\[action-Hq:prp\] For any $f\in \mathcal{F}(\Lambda_n)$ and $\lambda\in\Lambda_n$, one has that $$\begin{aligned}
&(\text{H}_q f)(\lambda)
= a[m_0(\lambda)] f(\lambda)\ +\\
&\sum_{\substack{1\leq j \leq n\\ \lambda+e_j\in\Lambda_n}} (1-c\delta_{\lambda_j}q^{m_0(\lambda)-1})[m_{\lambda_j}(\lambda)] f(\lambda+e_j)
+\sum_{\substack{1\leq j \leq n\\ \lambda-e_j\in\Lambda_n}} [m_{\lambda_j}(\lambda)] f(\lambda-e_j) ,\end{aligned}$$ where $e_1,\ldots ,e_n$ refer to the unit vectors comprising the standard basis of $\mathbb{Z}^n$.
It is clear from the definitions that $([N_0]f)(\lambda)=[m_0(\lambda)]f(\lambda)$, and that for any $l\in\mathbb{N}$: $$(\beta_{l+1} \beta_l^* f)(\lambda)=
\begin{cases}
[m_l(\lambda)](1-c\delta_lq^{m_0(\lambda)-1})f(\beta_{l+1}^* \beta_l\lambda) &\text{if}\ m_l(\lambda)>0 ,\\
0&\text{otherwise},
\end{cases}$$ where $\beta_{l+1}^* \beta_l\lambda=\lambda+e_j$ with $j=\min\{k\mid \lambda_k=l\}$ (so $l=\lambda_j$), and $$(\beta_{l+1}^*\beta_l f)(\lambda)=
\begin{cases}
[m_{l+1}(\lambda)]f(\beta_{l+1}\beta_l^*\lambda)&\text{if}\ m_{l+1}(\lambda)>0 ,\\
0&\text{otherwise},
\end{cases}$$ where $\beta_{l+1}\beta_l^*\lambda=\lambda-e_j$ with $j=\max\{k\mid \lambda_k=l+1\}$ (so $l=\lambda_j-1$).
The $n$-particle hamiltonian has Bethe Ansatz eigenfunctions given by the following plane wave expansion
$$\label{HLf}
\phi_\xi (\lambda) :=
\sum_{\substack{\sigma\in S_n\\ \epsilon\in \{\pm 1\}^n}}
C(\epsilon \xi_{\sigma}) e^{i \langle \lambda, \epsilon \xi_{\sigma}\rangle} ,$$
with expansion coefficients of the form $$\begin{aligned}
\label{Cf}
\lefteqn{C(\xi) :=
\prod_{1\leq j\leq n} \frac{1-ae^{-i\xi_j}+ce^{-2i\xi_j}}{1-e^{-2i\xi_j}}} && \\
&& \times \prod_{1\leq j<k \leq n} \Bigl(\frac{1-q e^{-i(\xi_{j}-\xi_k)}}{1-e^{-i(\xi_{j}-\xi_k)}}\Bigr)\Bigl( \frac{1-q e^{-i(\xi_{j}+\xi_k)}}{1-e^{-i(\xi_{j}+\xi_k)}} \Bigr) . \nonumber\end{aligned}$$
Here $\langle\cdot,\cdot\rangle$ denotes the standard inner product on $\mathbb{R}^n$, $\epsilon\xi_\sigma:=(\epsilon_1\xi_{\sigma_1},\epsilon_2\xi_{\sigma_2},\ldots ,\epsilon_n\xi_{\sigma_n})$, and the summation is meant over all permutations $\sigma$ in the symmetric group $S_n$ and all sign configurations $\epsilon=(\epsilon_1,\ldots,\epsilon_n)\in \{ 1,-1\}^n$. Viewed as a function of the spectral parameter $\xi=(\xi_1,\ldots ,\xi_n)$ in the fundamental alcove $$\label{alcove}
A:=\{ (\xi_1,\xi_2,\ldots,\xi_n)\in\mathbb{R}^n\mid \pi>\xi_1>\xi_2>\cdots >\xi_n>0\} ,$$ the expression $\phi_\xi(\lambda)$, $\lambda\in\Lambda_n$ amounts to Macdonald’s three-parameter Hall-Littlewood polynomial with hyperoctahedral symmetry associated with the root system $BC_n$ [@mac:orthogonal §10].
\[BAE:prp\] The $n$-particle Bethe Ansatz wave function $\phi_\xi$, $\xi\in A$ solves the eigenvalue equation $$\label{ev:eq}
\text{H}_q \phi_\xi = E_n(\xi) \phi_\xi,\qquad E_n(\xi):=2\sum_{j=1}^n \cos (\xi_j).$$
It follows from Proposition \[action-Hq:prp\] that the stated eigenvalue equation boils down to the following identity $$\begin{aligned}
a[m_0(\lambda)] \phi_\xi(\lambda)\ &+\
\sum_{\substack{1\leq j \leq n\\ \lambda+e_j\in\Lambda_n}} (1-c\delta_{\lambda_j}q^{m_0(\lambda)-1})[m_{\lambda_j}(\lambda)] \phi_\xi (\lambda+e_j) \\
&+\sum_{\substack{1\leq j \leq n\\ \lambda-e_j\in\Lambda_n}} [m_{\lambda_j}(\lambda)] \phi_\xi(\lambda-e_j) =
2\phi_\xi(\lambda ) \sum_{j=1}^n \cos (\xi_j) ,\end{aligned}$$ which is in turn equivalent to the Pieri formula for the hyperoctahedral Hall-Littlewood function in Eq. of Appendix \[appA\].
Diagonalization {#sec4}
===============
From now on it will be assumed unless stated otherwise that $0<|q|<1$ and that the boundary parameters $a$ and $c$ are chosen such that the roots $r_1$, $r_2$ of the quadratic polynomial $r^2-ar+c$ belong to the interval $(-1,1)$: $$\label{od}
a=r_1+r_2\ \text{and}\ c=r_1r_2\ \text{with}\ r_1,r_2\in (-1,1) .$$
Let $ L^2(A,\Delta\text{d}\xi)$ be the Hilbert space of functions $\hat{f}:A\to\mathbb{C}$ characterized by the inner product $$\label{ip}
\langle \hat{f},\hat{g}\rangle_\Delta=\frac{1}{(2\pi)^n}\int_A \hat{f}(\xi)\overline{\hat{g}(\xi)}\Delta(\xi)\text{d}\xi ,
\quad \text{where}\quad \Delta (\xi):= \frac{1}{|C(\xi)|^2}$$ with $C(\xi)$ given by Eq. . It is well-known that for the parameter regime in question Macdonald’s hyperoctahedral Hall-Littlewood functions form an orthogonal basis of $ L^2(A,\Delta\text{d}\xi)$ [@mac:orthogonal §10]:
$$\label{orthogonality}
\langle \phi(\lambda) ,\phi (\mu) \rangle_\Delta =\begin{cases}
\mathcal{N}(\lambda)&\text{if}\ \lambda =\mu ,\\
0&\text{otherwise},
\end{cases}$$
where $$\label{norm}
\mathcal{N}(\lambda):= (c;q)_{m_0(\lambda)} \prod_{\ell\in\mathbb{N}} [m_\ell (\lambda)]!$$
with $(c;q)_m:=(1-c)(1-cq)\cdots (1-cq^{m-1})$ (and the convention that $(c;q)_0:=1$) and $[m]!:=(q;q)_m/(q;q)_1^m=[m][m-1]\cdots [2][1]$. By combining the orthogonality in Eqs. , with Proposition \[BAE:prp\], the spectral decomposition of $\text{H}_{q}$ in the $n$-particle Hilbert space $\ell^2(\Lambda_n,\mathcal{N}^{-1})\subset\mathcal{F}(\Lambda_n)$ characterized by the inner product $$\langle f,g\rangle_n:=\sum_{\lambda\in\Lambda_n} f(\lambda) \overline{g(\lambda)} \mathcal{N}^{-1}(\lambda)$$ becomes immediate.
\[diagonal:thm\] For $0< |q|<1$ and values of the boundary parameters $a$ and $c$ in the orthogonality domain , the $q$-boson Hamiltonian $\text{H}_{q}$ restricts to a bounded self-adjoint operator in $\ell^2(\Lambda_n,\mathcal{N}^{-1})$ with purely absolutely continuous spectrum. More specifically, its spectral decomposition reads explicitly $$\label{s-d}
\text{H}_{q}=\boldsymbol{F_q}^{-1} \circ \hat{{E}} \circ\boldsymbol{F_q},$$ where $\boldsymbol{F_q}:\ell^2(\Lambda_n,\mathcal{N}^{-1})\to L^2(A,\Delta\text{d}\xi)$ denotes the unitary Fourier transform associated with the hyperoctahedral Macdonald-Hall-Littlewood basis:
$$\label{ft1}
(\boldsymbol{F_q}f)(\xi):= \langle f,\phi_\xi \rangle_n=\sum_{\lambda\in\Lambda_n}f(\lambda)
\overline{\phi_\xi (\lambda)}\mathcal{N}^{-1}(\lambda)$$
($f\in \ell^2(\Lambda_n,\mathcal{N}^{-1})$) with the inversion formula given by $$\label{ft2}
(\boldsymbol{F_q}^{-1}\hat{f})(\lambda) = \langle \hat{f},\overline{\phi ( \lambda)}\rangle_\Delta=
\frac{1}{(2\pi)^n}\int_A \hat{f}(\xi) \phi_\xi (\lambda)\Delta(\xi)\text{d}\xi$$
$(\hat{f}\in L^2(A,\Delta\text{d}\xi))$, and $(\hat{{E}}\hat{f})(\xi):=E_n(\xi) \hat{f}(\xi)$ stands for the bounded real multiplication operator in $L^2(A,\Delta\text{d}\xi)$ associated with the $n$-particle eigenvalue $E_n(\xi)$ .
In the Fock space $
\mathcal{H}:=\bigoplus_{n\geq 0} \ell^2(\Lambda_n,\mathcal{N}^{-1})
$, built of all linear combinations $\sum_{n\geq 0} c_n f_n$ with $c_n\in\mathbb{C}$ and $f_n\in \ell^2(\Lambda_n,\mathcal{N}^{-1})$ such that $\sum_{n\geq 0} |c_n|^2 \langle f_n,f_n\rangle_n<\infty$, the $q$-boson hamiltonian $\text{H}_q$ constitutes an unbounded operator that is essentially self-adjoint on the dense domain $\mathcal{D}:=\mathcal{F}\cap\mathcal{H}$ (because for $z\in\mathbb{C}\setminus\mathbb{R}$ the range $(\text{H}_{q}-z)\mathcal{D}$ is dense in $\mathcal{H}$ and $\lim_{n\to\infty} \sup_{\xi\in A} |E_n(\xi)|=\infty$). The representation of the deformed $q$-boson field algebra in Section \[sec2\] on the other hand gives rise to a bounded representation on $\mathcal{H}$: $$\begin{aligned}
\langle \beta_l f,\beta_l f\rangle_{n-1}&\leq \frac{1+|c|\delta_{l}}{1-q} \langle f,f\rangle_{n} ,\nonumber\\
\langle \beta_l^* f,\beta_l^*f\rangle_{n+1}&\leq \frac{1+|c|\delta_{l}}{1-q} \langle f,f\rangle_{n},\\
\langle q^{N_l} f,q^{N_l}f\rangle_{n}&\leq \langle f,f\rangle_{n} , \nonumber\end{aligned}$$ preserving the $*$-structure: $$\langle \beta_l^*f,g\rangle_{n+1}= \langle f, \beta_l g\rangle_{n}\quad\text{and}\quad
\langle q^{N_l}f,g\rangle_{n}= \langle f, q^{N_l} g\rangle_{n} .$$
Upon rescaling the lattice $\Lambda_n$ and performing an appropriate continuum limit [@die:plancherel Sec. 5], Macdonald’s hyperoctahedral Hall-Littlewood functions tend to the eigenfunctions of the quantum nonlinear Schrödinger equation on the half-line with a boundary interaction [@gau:boundary; @gat-lig-min:nonlinear; @hal-lan:exact; @cau-cra:exact; @tra-wid:bose]. In particular, it follows from [@die:plancherel Sec. 5.3] that for $a=0$ (which corresponds to a reduction from type $BC$ to type $C$ root systems) a renormalized version of the $q$-boson hamiltonian $\text{H}_q$ then converges in the $n$-particle subspace in the strong resolvent sense to a hamiltonian that can be written formally as: $$-\sum_{j=1}^n \frac{\partial^2}{\partial x_j^2}+
g \sum_{1\leq j< k\leq n} \bigl( \delta (x_j-x_k)+\delta (x_j+x_k)\bigr)
+g_0 \sum_{1\leq j\leq n}\delta(x_j)$$ with $g, g_0> 0$ (where $\delta (\cdot )$ stands for the ‘delta potential’).
Factorized scattering
=====================
The similarity transformation $$\label{Hq-lebesgue}
H :=\mathcal{N}^{-1/2}\,\text{H}_{q}\,\mathcal{N}^{1/2}$$ turns the $n$-particle $q$-boson hamiltonian in Proposition \[action-Hq:prp\] into a self-adjoint operator in $\ell^2(\Lambda_n)$ diagonalized by the normalized wave function
$$\begin{split}
\Psi_\xi(\lambda)&:= e^{\frac{\pi i}{2}n^2} |C (\xi)|^{-1} \mathcal{N}(\lambda )^{-1/2} \phi_\xi (\lambda ) \\
&=\mathcal{N}(\lambda )^{-1/2}\sum_{\substack{\sigma \in S_n\\ \epsilon\in \{ \pm 1\}^n}}\text{sign} (\epsilon \sigma )
\hat{\mathcal S}(\epsilon \xi_\sigma )^{1/2} e^{i\langle \rho+\lambda, \epsilon\xi_\sigma\rangle} ,
\end{split}$$
with $\xi\in A$ , $\text{sign} (\epsilon \sigma ):=\epsilon_1\cdots\epsilon_n\text{sign} ( \sigma )$, $\rho:=(n-1,n-2,\ldots ,2,1,0)$, and $$\hat{ {\mathcal S}} (\xi)
:= \prod_{1\leq j<k\leq n} s(\xi_j-\xi_k)s(\xi_j+\xi_k)\prod_{1\leq j\leq n} s_0(\xi_j) ,$$ where $$\begin{aligned}
s(x):=\frac{1-qe^{-ix}}{1-qe^{ix}}\quad&\text{with}\quad s(x)^{1/2}=\frac{1-qe^{-ix}}{|1-q e^{ix}|}
\intertext{and}
s_0(x):=\frac{1-ae^{-ix}+ce^{-2ix}}{1-ae^{ix}+ce^{2ix}}
\quad&\text{with}\quad
s_0(x)^{1/2}=\frac{1-ae^{-ix}+ce^{-2ix}}{|1-ae^{ix}+ce^{2ix}|}.\end{aligned}$$
Specifically, one has that $
H=\boldsymbol{F}^{-1} \circ \hat{{E}} \circ\boldsymbol{F}
$ where $\boldsymbol{F}:\ell^2(\Lambda_n)\to L^2(A,\text{d}\xi)$ denotes the unitary Fourier transformation determined by the kernel $\Psi_\xi(\lambda)$ (and $\hat{E}$ is now interpreted as a bounded multiplication operator in $ L^2(A,\text{d}\xi)$). For $q,a,c\to 0$ the $n$-particle $q$-boson hamiltonian $H$ simplifies to a hamiltonian modeling impenetrable bosons on $\mathbb{N}$: $$(H_0 f)(\lambda)
=
\sum_{\substack{1\leq j \leq n\\ \lambda+e_j\in\Lambda_n}} f(\lambda+e_j)
+\sum_{\substack{1\leq j \leq n\\ \lambda-e_j\in\Lambda_n}} f(\lambda-e_j)$$ ($f\in\ell^2(\Lambda_n)$), which is diagonalized by the conventional Fourier transform $\boldsymbol{F_0}:\ell^2(\Lambda_n)\to L^2(A,\text{d}\xi)$ obtained from $\boldsymbol{F}$ by setting $\hat{\mathcal{S}}(\xi)\equiv 1$, $\mathcal{N}(\lambda)\equiv 1$.
As a very special case of the results in [@die:scattering 4], it now follows that the wave- and scattering operators comparing the $q$-boson dynamics $$\label{dynamics}
(e^{itH }f)(\lambda)=
\frac{1}{(2\pi)^n}\int_A e^{itE_n (\xi )}\hat{f}(\xi) \Psi_\xi (\lambda)\text{d}\xi\qquad
\hat{f}= \boldsymbol{F} f$$ with the corresponding impenetrable boson dynamics generated by $H_0$ are governed by a unitary $S$-matrix $ \hat{\mathcal S} : L^2(A,\text{d}\xi)\to L^2(A,\text{d}\xi)$ of the form $$( \hat{\mathcal S}\hat{f})(\xi):= \hat{\mathcal S}(\epsilon_\xi \xi_{\sigma_\xi})\hat{f}(\xi)\qquad
(\hat{f}\in C_0(A_r).$$ Here $C_0(A_r)$ denotes the dense subspace of $L^2(A,\text{d}\xi)$ consisting of smooth test functions with compact support in the open dense subset $A_r\subset A$ for which the components of $\nabla E_n(\xi)=(-2\sin(\xi_1),\ldots,-2\sin(\xi_n))$ do not vanish and are all distinct in absolute value, and the sign-configuration $\epsilon_\xi$ and the permutation $\sigma_\xi$ are such that the components of $\nabla E_n(\epsilon_\xi \xi_{\sigma_\xi})$ are all positive and ordered from large to small. Specifically, by comparing the large-time asymptotics of oscillatory integrals of the form in Eq. for the dynamics generated by $H$ and $H_0$ one concludes that [@die:scattering 4.2 4.3]:
\[scattering:thm\] The operator limits
$$\Omega^{\pm} :=s-\lim_{t\to \pm \infty} e^{i t H}e^{-it H_{0}}$$
converge in the strong $\ell^2(\Lambda_n)$-norm topology and the corresponding wave operators $\Omega^\pm_r$ are given by unitary operators in $\ell^2(\Lambda_n)$ of the form $$\Omega_r^\pm = \boldsymbol{F}^{-1} \circ \hat{\mathcal S}^{\mp 1/2} \circ \boldsymbol{F_0}.$$ Hence, the scattering operator comparing the dynamics of $H$ and $H_{0}$ is given by the unitary operator $$\mathcal{S}:=(\Omega_r^+)^{-1} \Omega_r^- = \boldsymbol{F_0}^{-1} \circ \hat{\mathcal S} \circ \boldsymbol{F_0} .$$
Pieri formula for Macdonald’s hyperoctahedral Hall-Littlewood function {#appA}
======================================================================
Let $x:=(x_1,\ldots ,x_n)=(e^{i\xi_1},\ldots,e^{i\xi_n})$ and $\tau:=(\tau_1,\ldots,\tau_n)$, where $\tau_j=rq^{n-j}$ ($j=1,\ldots ,n$) with $r=\frac{a}{2}+\sqrt{(\frac{a}{2})^2-c}$ (cf. Eq. ). Upon setting $$P_\lambda (x):= \frac{\tau_1^{\lambda_1}\cdots \tau_n^{\lambda_n}}{\mathcal{N}(0)}\phi_\xi(\lambda)\qquad
(\lambda\in\Lambda_n),$$ where $\mathcal{N}(0)$ is given by Eq. with $\lambda=0$, the hyperoctahedral Hall-Littlewood function is renormalized to have unital principal specialization values: $P_\lambda(\tau)=1$ ($\forall\lambda\in\Lambda_n)$ [@mac:orthogonal §10]. With this normalization, the following Pieri formula holds: $$\begin{aligned}
\label{pieri}
\lefteqn{P_\lambda (x) \sum_{j=1}^n (x_j+x_j^{-1}-\tau_j-\tau_j^{-1})=} &&\\
&&\sum_{\substack{1\leq j\leq n\\ \lambda +e_j\in\Lambda_n}}
V_j^+(\lambda) \left( P_{\lambda +e_j}(x)-P_\lambda(x) \right)
+\sum_{\substack{1\leq j\leq n\\ \lambda -e_j\in\Lambda_n}}
V_j^-(\lambda) \left( P_{\lambda -e_j}(x)-P_\lambda(x) \right) ,\nonumber\end{aligned}$$ where $$\begin{aligned}
V_j^+(\lambda) &= \tau_j^{-1}\Bigl(\frac{1-c^2\delta_{\lambda_j}q^{2(n-j)}}{1+c\delta_{\lambda_j}q^{2(n-j)}}\Bigr)
\prod_{\substack{j<k\leq n\\ \lambda_k=\lambda_j}}
\Bigl(\frac{1-q^{1+k-j}}{1-q^{k-j}}\Bigr)\Bigl( \frac{1+c\delta_{\lambda_j}q^{1+2n-k-j}}{1+c\delta_{\lambda_j}q^{2n-k-j}}\Bigr),\\
V_j^-(\lambda) &= \tau_j \prod_{\substack{1\leq k<j\\ \lambda_k=\lambda_j}}
\Bigl(\frac{1-q^{1+j-k}}{1-q^{j-k}}\Bigr) .\end{aligned}$$
The formula in question is readily obtained through degeneration from an analogous Pieri formula for a $BC_n$-type Macdonald function that arises as a special case of the Pieri formulas in [@die:properties . 6.1]. Specifically, by substituting $t_2=q^{1/2}$, $t_3=-q^{1/2}$ (which amounts to a reduction from the Macdonald-Koornwinder function to the $BC_n$-type Macdonald function) in the Pieri formula of [@die:properties . (6.4), (6.5)] with coefficients taken from [@die:properties . (6.12), (6.13)], the relation in Eq. is retrieved for $q\to 0$ (which corresponds to a transition from Macdonald type functions to Hall-Littlewood type functions). Notice in this connection that the parameters $q$, $a$, $c$ (and $r$) of the present paper are related to the parameters $t$, $t_0$, $t_1$ of Ref. [@die:properties] via $q=t$, $a=t_0+t_1$, $c=t_0t_1$ (and $r=t_0$).
Since $$V_j^+(\lambda)=\tau_j^{-1}(1-c\delta_{\lambda_j}q^{m_0(\lambda)-1})[m_{\lambda_j}(\lambda)] ,
\qquad V_j^-(\lambda)= \tau_j [m_{\lambda_j}(\lambda)] ,$$ and $$\sum_{j=1}^n( \tau_j+\tau_j^{-1})
-\sum_{\substack{1\leq j \leq n\\ \lambda-e_j\in\Lambda_n}} \tau_j [m_{\lambda_j}(\lambda)]
-\sum_{\substack{1\leq j \leq n\\ \lambda+e_j\in\Lambda_n}}
\tau_j^{-1} [m_{\lambda_j}(\lambda)]
=r[m_0(\lambda)] ,$$ the Pieri formula can be condensed into the more compact form $$\begin{aligned}
\label{pieri2}
P_\lambda (x) \sum_{j=1}^n (x_j+x_j^{-1})&=a[m_0(\lambda)] +\sum_{\substack{1\leq j\leq n\\ \lambda -e_j\in\Lambda_n}}
\tau_j [m_{\lambda_j}(\lambda)] P_{\lambda -e_j}(x) \\
&
+\sum_{\substack{1\leq j\leq n\\ \lambda +e_j\in\Lambda_n}}
\tau_j^{-1} (1-c\delta_{\lambda_j}q^{m_0(\lambda)-1})[m_{\lambda_j}(\lambda)] P_{\lambda +e_j}(x)
.\nonumber\end{aligned}$$
[000000]{}
N.M. Bogoliubov, A.G. Izergin, and A.N. Kitanine, Correlation functions for a strongly correlated boson system, Nuclear Phys. B [**516**]{} (1998), 501–528.
V. Caudrelier and N. Crampé, Exact results for the one-dimensional many-body problem with contact interaction: including a tunable impurity, Rev. Math. Phys. [**19**]{} (2007), 349–370.
J.F. van Diejen, Properties of some families of hypergeometric orthogonal polynomials in several variables, Trans. Amer. Math. Soc. [**351**]{} (1999), 233–270.
, On the Plancherel formula for the (discrete) Laplacian in a Weyl chamber with repulsive boundary conditions at the walls, Ann. Henri Poincaré [**5**]{} (2004), 135–168.
, Scattering theory of discrete (pseudo) Laplacians on a Weyl chamber, Amer. J. Math. [**127**]{} (2005), 421–458.
J.F. van Diejen and E. Emsiz, Diagonalization of the infinite $q$-boson system, preprint 2013.
M. Gattobigio, A. Liguori, and M. Mintchev, The nonlinear Schrödinger equation on the half line, J. Math. Phys. [**40**]{} (1999), 2949–2970.
M. Gaudin, Boundary energy of a Bose gas in one dimension. Phys. Rev. A [**4**]{} (1971), 386–394.
, [*La fonction d’onde de Bethe*]{}, Masson, Paris, 1983.
E. Gutkin, Quantum nonlinear Schrödinger equation: two solutions, Phys. Rep. [**167**]{} (1988), 1–-131.
M. Hallnäs and E. Langmann, Exact solutions of two complementary one-dimensional quantum many-body systems on the half-line, J. Math. Phys. [**46**]{} (2005), no. 5, 052101.
A. Klimyk and K. Schmüdgen, [*Quantum Groups and their Representations*]{}, Springer-Verlag, Berlin, 1997.
V.E. Korepin, N.M. Bogoliubov, and A.G. Izergin, [*Quantum Inverse Scattering Method and Correlation Functions*]{}, Cambridge University Press, Cambridge, 1993.
C. Korff, Cylindric versions of specialised Macdonald functions and a deformed Verlinde algebra, Comm. Math. Phys. [**318**]{} (2013), 173–246.
I.G. Macdonald, Orthogonal polynomials associated with root systems, Sém. Lothar. Combin. [**45**]{} (2000/01), Art. B45a, 40 pp.
D.C. Mattis, [*The Many-Body Problem: An Encyclopedia of Exactly Solved Models in One Dimension*]{}, World Scientific, Singapore, 1994.
S.N.M. Ruijsenaars, Factorized weight functions vs. factorized scattering, Comm. Math. Phys. [**228**]{} (2002), 467–494.
B. Sutherland, [*Beautiful Models: 70 Years of Exactly Solved Quantum Many-Body Problems*]{}, Singapore: World Scientific, 2004.
C.A. Tracy and H. Widom, The Bose gas and asymmetric simple exclusion process on the half-line, J. Stat. Phys. [**150**]{} (2013), 1–12.
N.V. Tsilevich, The quantum inverse scattering method for the $q$-boson model and symmetric functions, Funct. Anal. Appl. [**40**]{} (2006), 207–217.
[^1]: Work was supported in part by the [*Fondo Nacional de Desarrollo Científico y Tecnológico (FONDECYT)*]{} Grants \# 1130226 and \# 11100315, and by the [*Anillo ACT56 ‘Reticulados y Simetrías’*]{} financed by the [*Comisión Nacional de Investigación Científica y Tecnológica (CONICYT)*]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Tuberculosis (TB) and human immunodeficiency virus (HIV) can be considered a deadly human syndemic. In this article, we formulate a model for TB and HIV transmission dynamics. The model considers both TB and acquired immune deficiency syndrome (AIDS) treatment for individuals with only one of the infectious diseases or both. The basic reproduction number and equilibrium points are determined and stability is analyzed. Through simulations, we show that TB treatment for individuals with only TB infection reduces the number of individuals that become co-infected with TB and HIV/AIDS, and reduces the diseases (TB and AIDS) induced deaths. Analogously, the treatment of individuals with only AIDS also reduces the number of co-infected individuals. Further, TB-treatment for co-infected individuals in the active and latent stage of TB disease, implies a decrease of the number of individuals that passes from HIV-positive to AIDS.'
author:
- |
Cristiana J. Silva\
`[email protected]`
- |
Delfim F. M. Torres[^1]\
`[email protected]`
date: |
Center for Research and Development in Mathematics and Applications (CIDMA)\
Department of Mathematics, University of Aveiro, 3810-193 Aveiro, Portugal
title: 'Modeling TB-HIV syndemic and treatment[^2]'
---
#### Keywords:
Tuberculosis, Human immunodeficiency virus, Syndemic, Treatment, Equilibrium, Stability.
#### Mathematics Subject Classification 2010:
34D30; 92D30; 93A30.
Introduction
============
Tuberculosis (TB) and human immunodeficiency virus/acquired immune deficiency syndrome (HIV/AIDS) are the leading causes of death from an infectious disease worldwide [@TB_WHO_report_2013]. Individuals infected with HIV are more likely to develop TB disease because of their immunodeficiency, and HIV infection is the most powerful risk factor for progression from TB infection to disease [@Getahun:etall:CID2010]. This interaction justifies the fact that HIV and TB can be considered a deadly human *syndemic*, where syndemic refers to the convergence of two or more diseases that act synergistically to magnify the burden of disease [@Kwan_Ernst_HIV_TB_Syndemic].
Following UNAIDS global report on AIDS epidemic 2013 [@UNAIDS_report_2013], globally, an estimated 35.3 million people were living with HIV in 2012. An increase from previous years, as more people are receiving the life-saving antiretroviral therapy (ART). There were approximately 2.3 million new HIV infections globally, showing a 33% decline in the number of new infections with respect to 2001. At the same time, the number of AIDS deaths is also declining with around 1.6 million AIDS deaths in 2012, down from about 2.3 million in 2005. In 2012, 1.1 million of 8.6 million people who developed TB worldwide were HIV-positive. The number of people dying from HIV-associated TB has been falling since 2003. However, there were still 320 000 deaths from HIV-associated TB in 2012 and further efforts are needed to reduce this burden [@TB_WHO_report_2013]. ART is a critical intervention for reducing the risk of TB morbidity and mortality among people living with HIV and, when combined with isoniazid preventive therapy, it can have a significant impact on TB prevention [@TB_WHO_report_2013].
Collaborative TB/HIV activities (including HIV testing, ART therapy and TB preventive measures) are crucial for the reduction of TB-HIV coinfected individuals. The World Health Organization (WHO) estimates that these collaborative activities prevented 1.3 million people from dying, from 2005 to 2012. However, significant challenges remain: the reduction of tuberculosis related deaths among people living with HIV has slowed in recent years; the ART therapy is not being delivered to TB-HIV coinfected patients in the majority of the countries with the largest number of TB/HIV patients; the pace of treatment scale-up for TB/HIV patients has slowed; less than half of notified TB patients were tested for HIV in 2012; and only a small fraction of TB/HIV infected individuals received TB preventive therapy [@UNAIDS_report_2013].
The study of the joint dynamics of TB and HIV present formidable mathematical challenges due to the fact that the models of transmission are quite distinct [@CChavez_TB_HIV_2009]. Few mathematical models have been proposed for TB-HIV coinfection (see, for example, [@Bhunu:BMB:2009:HIV:TB; @Kirschner:TB:HIV:1999; @Naresh:TB:HIV:2005; @CChavez_TB_HIV_2009; @Song:TB:HIV:2008]). Kirschner [@Kirschner:TB:HIV:1999] developed a cellular model for HIV-1 and TB coinfection inside a host. Roeger et al. [@CChavez_TB_HIV_2009] proposed a population model for TB-HIV/AIDS coinfection transmission dynamics, assuming that TB-infected individuals in the active stage of the disease are too ill to remain sexually active and therefore they are unable to transmit HIV. In this work we assume that active TB-infected individuals are susceptible to HIV-infection. Naresh and Tripathi [@Naresh:TB:HIV:2005] proposed a model for TB-HIV coinfection in a variable size population with only TB treatment. Here we consider TB and HIV treatment in different stages of the disease. Bhunu et al. [@Bhunu:BMB:2009:HIV:TB] studied a TB-HIV coinfection model with both TB and HIV treatment. The authors did not take into account that an individual co-infected with TB and HIV can effectively recover from TB infection. We assume that TB can be cured, even in HIV-positive individuals [@TB_WHO_report_2013]. Sharomi et al. [@Song:TB:HIV:2008] also considered these assumptions, subdividing the total population into 15 classes. It is our aim in this work to develop a model that balances two goals: simplicity and useful information.
The paper is organized as follows. Section \[sec:2\] describes our model for TB-HIV syndemic with TB and HIV treatment. In Section \[sec:3\] the positivity and boundedness of solutions of the model are proved and in Section \[sec:4\] equilibrium points and respective stability are analyzed. Section \[sect:NumSim\] is devoted to numerical simulations and discussion of results.
TB-HIV/AIDS model {#sec:2}
=================
The model subdivides the human population into 10 mutually-exclusive compartments, namely susceptible individuals ($S$), TB-latently infected individuals, who have no symptoms of TB disease and are not infectious ($L_T$), TB-infected individuals, who have active TB disease and are infectious ($I_T$), TB-recovered individuals ($R_T$), HIV-infected individuals with no clinical symptoms of AIDS ($I_H$), HIV-infected individuals with AIDS clinical symptoms ($A$), TB-latent individuals co-infected with HIV (pre-AIDS) ($L_{TH}$), HIV-infected individuals (pre-AIDS) co-infected with active TB disease ($I_{TH}$), TB-recovered individuals with HIV-infection without AIDS symptoms ($R_{TH}$), HIV-infected individuals with AIDS symptoms co-infected with TB ($A_T$). The total population at time $t$, denoted by $N(t)$, is given by $$N(t) = S(t) + L_T(t) + I_T(t) + R_T(t) + I_H(t) + A(t)
+ I_{TH}(t) + L_{TH}(t) + R_{TH}(t) + A_T(t) \, .$$ The susceptible population is increased by the recruitment of individuals (assumed susceptible) into the population, at a rate $\Lambda$. All individuals suffer from natural death, at a constant rate $\mu$. Susceptible individuals acquire TB infection from individuals with active TB at a rate $\lambda_T$, given by $$\lambda_T = \frac{\beta_1}{N} \left(I_T + I_{TH} + A_T\right) \, ,$$ where $\beta_1$ is the effective contact rate for TB infection. Similarly, susceptible individuals acquire HIV infection, following effective contact with people infected with HIV at a rate $\lambda_H$, given by $$\lambda_H = \frac{\beta_2}{N} \left[ I_H + I_{TH} + L_{TH}
+ R_{TH} + \eta \left(A + A_T\right) \right] \, ,$$ where $\beta_2$ is the effective contact rate for HIV transmission and the modification parameter $\eta {\geqslant}1$ accounts for the relative infectiousness of individuals with AIDS symptoms, in comparison to those infected with HIV with no AIDS symptoms. Individuals with AIDS symptoms are more infectious than HIV-infected individuals (pre-AIDS) because they have a higher viral load and there is a positive correlation between viral load and infectiousness [@art:viral:load].
Individuals leave the latent-TB class $L_T$ by becoming infectious, at a rate $k_1$, or recovered, with a treatment rate $\tau_1$. The treatment rate for active TB-infected individuals is $\tau_2$. We assume that TB-recovered individuals $R_T$ acquire partial immunity and the transmission rate for this class is given by $\beta'_1 \lambda_T$ with $\beta'_1 {\leqslant}1$. Individuals with active TB disease suffer induced death at a rate $d_T$. We assume that individuals in the class $R_T$ are susceptible to HIV infection at a rate $\lambda_H$. On the other hand, TB-active infected individuals $I_T$ are susceptible to HIV infection, at a rate $\delta \lambda_H$, where the modification parameter $\delta {\geqslant}1$ accounts for higher probability of individuals in class $I_T$ to become HIV-positive.
HIV-infected individuals (with no AIDS symptoms) progress to the AIDS class $A$, at a rate $\rho_1$. HIV-infected individuals with AIDS symptoms are treated for HIV at the rate $\alpha_1$ and suffer induced death at a rate $d_A$. Individuals in the class $I_H$ are susceptible to TB infection at a rate $\psi \lambda_T$, where $\psi {\geqslant}1$ is a modification parameter traducing the fact that HIV infection is a driver of TB epidemic [@Kwan_Ernst_HIV_TB_Syndemic].
HIV-infected individuals (pre-AIDS) co-infected with TB-disease, in the active stage $I_{TH}$, are treated for TB at the rate $\tau_3$ and progress to the AIDS-TB co-infection class $A_T$ at a rate $\rho_2$. Individuals in the class $I_{TH}$ suffer TB induced death at a rate $d_T$. The anti-TB drugs can prevent or decrease the likelihood of TB infection progression to active TB disease in individuals in the class $L_{TH}$ [@USAID:TB:HIV]. The treatment rate for individuals in this class is given by $\tau_4$. However, individuals in the class $L_{TH}$ are more likely to progress to active TB disease than individuals infected only with latent TB. In our model, this progression rate is given by $k_2$. Similarly, HIV infection makes individuals more susceptible to TB reinfection when compared with non HIV-positive patients. The modification parameter associated to the TB reinfection rate, for individuals in the class $R_{TH}$, is given by $\beta'_2$, where $\beta'_2 {\geqslant}1$. Individuals in this class progress to class $A_T$, at a rate $\rho_3$.
HIV-infected individuals (with AIDS symptoms), co-infected with TB, are treated for HIV, at a rate $\alpha_2$. Individuals in the class $A_T$ suffer from AIDS-TB coinfection induced death rate, at a rate $d_{TA}$.
The aforementioned assumptions result in the following system of differential equations that describes the transmission dynamics of TB and HIV disease: $$\label{model:TB:HIV}
\begin{cases}
\dot{S}(t) = \Lambda - \lambda_T S(t) - \lambda_H S(t) - \mu S(t),\\[0.2 cm]
\dot{L}_T(t) = \lambda_T S(t) + \beta^{'}_1 \lambda_T R_T(t) - (k_1 + \tau_1 + \mu)L_T(t),\\[0.2 cm]
\dot{I}_T(t) = k_1 L_T(t) - (\tau_2 +d_T +\mu + \delta \lambda_H)I_T(t), \\[0.2 cm]
\dot{R}_T(t) = \tau_1 L_T(t) + \tau_2 I_T(t) - (\beta^{'}_1 \lambda_T + \lambda_H + \mu) R_T(t),\\[0.2 cm]
\dot{I}_H(t) = \lambda_H S(t) - (\rho_1 + \psi \lambda_T + \mu)I_H(t) + \alpha_1 A(t) + \lambda_H R_T(t), \\[0.2 cm]
\dot{A}(t) = \rho_1 I_H(t) - \alpha_1 A(t) - (\mu + d_A) A(t),\\[0.2 cm]
\dot{L}_{TH}(t) = \beta^{'}_2 \lambda_T R_{TH}(t) - (k_2 + \tau_4 + \mu) L_{TH}(t),\\[0.2 cm]
\dot{I}_{TH}(t) = \delta \lambda_H I_T(t) + \psi \lambda_T I_H(t) + \alpha_2 A_T(t)+ k_2 L_{TH}(t)
- (\tau_3 + \rho_2 + \mu + d_T)I_{TH}(t),\\[0.2 cm]
\dot{R}_{TH}(t) = \tau_3 I_{TH}(t) + \tau_4 L_{TH}(t) - (\beta^{'}_2 \lambda_T + \rho_3 + \mu)R_{TH}, \\[0.2 cm]
\dot{A}_T(t) = \rho_2 I_{TH}(t) + \rho_3 R_{TH} -(\alpha_2 + \mu + d_{TA})A_T(t) \, .
\end{cases}$$ The model flow is described in Figure \[fig:model:flow\].
The initial conditions of model satisfy $$\label{eq:init:cond:geral}
\begin{split}
&S(0)= S_0 {\geqslant}0 \, , \quad \quad L_T(0) = L_{T0} {\geqslant}0 \, ,
\quad \quad I_T(0) = I_{T0} {\geqslant}0 \, ,
\quad \quad R_T(0) = R_{T0} {\geqslant}0 \, ,\\
&I_H(0) = I_{H0} {\geqslant}0 \, , \quad \quad A(0) = A_0 {\geqslant}0 \, ,
\quad \quad L_{TH}(0) = L_{TH0} {\geqslant}0 \, ,\\
&I_{TH}(0) = I_{TH0} {\geqslant}0 \, , \quad \quad R_{TH}(0) = R_{TH0} {\geqslant}0
\, , \quad \quad A_T(0) = A_{T0} {\geqslant}0 \, .
\end{split}$$ Note that if we consider the sub-model of with no HIV/AIDS disease, that is, $I_H = A = L_{TH} = I_{TH}
= R_{TH} = A_T = 0$, then we obtain the TB model from [@Castillo_Chavez_1997]. On the other hand, if we consider the sub-model with no TB, that is, $L_T = I_T = R_T = L_{TH} = I_{TH} = R_{TH} = A_T = 0$, then we obtain an HIV/AIDS model based on the models proposed in [@Bhunu:BMB:2009:HIV:TB; @Hyman:MathBio:1999].
Positivity and boundedness of solutions {#sec:3}
=======================================
Let $(S, L_T, I_T, R_T, I_H, A, L_{TH}, I_{TH}, R_{TH}, A_T) \in \R^{10}_{+}$ be any solution of with initial conditions . Consider the biologically feasible region given by $$\label{eq:feasible:region}
\Omega = \left\{ \left(S, L_T, I_T, R_T, I_H, A, L_{TH}, I_{TH}, R_{TH}, A_T\right) \,
\in \R^{10}_{+} \, : \, 0 {\leqslant}N(t) {\leqslant}\frac{\Lambda}{\mu} \right\} \, .$$ For the model system to be epidemiologically meaningful, it is important to prove that all its state variables are nonnegative for all time $t > 0$. Suppose, for example, that at some $\bar{t} > 0$ the variable $L_T$ becomes zero, i.e., $L_T(\bar{t}) = 0$, while all other variables are positive. Then, from the $L_T$ equation we have $d L_T(\bar{t})/dt > 0$. Thus, $L_T(t) {\geqslant}0$ for all $t > 0$. Analogously, we can prove that all variables remain nonnegative for all time $t > 0$.
Adding all equations in model gives $$\frac{d N}{dt}(t) = \Lambda - \mu N(t) - d_T I_T(t)
- d_A A(t) - d_T I_{TH}(t) - d_{TA} A_T(t) \, .$$ Since $N(t) {\geqslant}I_T(t) + A(t) + I_{TH}(t) + A_T(t)$, then $$\Lambda - (\mu + d_T + d_A + d_{TA})N(t) {\leqslant}\frac{d N}{dt}(t)
{\leqslant}\Lambda - \mu N(t) \, .$$ Therefore, we conclude that $N(t)$ is bounded for all $t > 0$ and every solution of system with initial condition in $\Omega$ remains in $\Omega$. This result is summarized below.
The region $\Omega$ is positively invariant for the model with non-negative initial conditions in $\R^{10}_{+}$.
Stability analysis {#sec:4}
==================
The model has four non-negative equilibria, namely
- The disease-free equilibrium (no disease) $$\label{eq:DFE:model:TBHIV}
\Sigma_0 = \left(S_0, L_{T_0}, I_{T_0}, R_{T_0}, I_{H_0}, A_0, L_{TH_0},
I_{TH_0}, R_{TH_0}, A_{T_0}\right) = \left(\frac{\Lambda}{\mu},
0, 0, 0, 0, 0, 0, 0, 0, 0 \right) \, .$$
- The HIV-AIDS free equilibrium $$\Sigma_T = \left(S^\diamond, L_T^\diamond, I_{T}^\diamond, R_{T}^\diamond,
I_{H}^\diamond, A^\diamond, L_{TH}^\diamond, I_{TH}^\diamond,
R_{TH}^\diamond, A_{T}^\diamond\right)$$ with $I_{T}^\diamond > 0$ and $I_{H}^\diamond = A^\diamond = L_{TH}^\diamond
= I_{TH}^\diamond = R_{TH}^\diamond = A_{T}^\diamond = 0$ for $R_1 > 1$, where $R_1$ is the basic reproduction number of the model with $I_H = A = L_{TH} = I_{TH} = R_{TH} = A_T = 0$ (only TB model) that is given by $$\label{eq:R1}
R_1 = \frac{\Lambda}{N\mu}\left(\frac{\beta_1}{d_T+\mu+\tau_2} \right)
\left(\frac{k_1}{k_1+\tau_1+\mu } \right)$$ (see [@Castillo_Chavez_1997]).
- The TB-free equilibrium $$\Sigma_H = \left(S^\star, L_T^\star, I_{T}^\star, R_{T}^\star, I_{H}^\star,
A^\star, L_{TH}^\star, I_{TH}^\star, R_{TH}^\star, A_{T}^\star\right)$$ with $L_T^\star = I_{T}^\star = R_{T}^\star = L_{TH}^\star
= I_{TH}^\star = R_{TH}^\star = A_{T}^\star = 0$ and $$S^\star = \frac{\Lambda}{\mu R_2}\, , \quad I_H^\star
= (R_2 - 1)\frac{\mu N_H (\alpha_1 + d_A + \mu)}{\beta_2 (\alpha_1
+ d_A + \mu + \eta \rho_1)}\, , \, \, \quad A^\star
= (R_2 - 1) \frac{\rho_1 \mu N_H}{\beta_2 (\alpha_1 + d_A + \mu + \eta \rho_1)},$$ for $R_2 > 1$, where $R_2$ is the basic reproduction number of model with $L_T = I_{T} = R_{T} = L_{TH} = I_{TH} = R_{TH} = A_{T} = 0$ (only HIV-AIDS model), that is, $$\label{eq:R2}
R_2 =\frac{\Lambda}{N\mu} \beta_2 \left(\frac{\mu+\alpha_1+d_A
+\eta\,\rho_1 }{\mu\,\alpha_1+ (\mu + \rho_1)(\mu + d_A)}\right)\, .$$
- The syndemic equilibrium $$\Sigma^* = (S^*, L_T^*, I_T^*, R_T^*, I_H^*, A^*, L^*_{TH}, I_{TH}^*, R_{TH}^*, A_T^*)$$ with $I_T^* > 0$, $I_H^* > 0$, $A^* > 0$, $L_{TH}^* > 0$, $I_{TH}^* > 0$, $R_{TH}^*>0$ and $A_T^* > 0$, for $R_0 > 1$, where $R_0$ is the basic reproduction number of the model , that is, $$R_0 = \max \{ R_1, R_2 \} \, .$$
The details of the computation of the basic reproduction number $R_0$ are given in Appendix \[A.1\].
The following theorem states the stability of the equilibrium points.
\[theo:stab:equil\] The disease free equilibrium $\Sigma_0$ is locally asymptotically stable if $R_0 < 1$, and unstable if either $R_i > 1$ with $i=1, 2$. The HIV-AIDS free equilibrium $\Sigma_T$ is locally asymptotically stable if $R_1 > 1$, and the TB-free equilibrium $\Sigma_H$ is locally asymptotically stable for $R_2$ near 1.
Details of the proof of Theorem \[theo:stab:equil\] are given in Appendix \[A.2\].
Explicit expressions for the coinfection endemic equilibrium $\Sigma^*$ are very difficult to compute analytically. In Section \[sect:NumSim\], we consider an example, with $R_0 > 1$, for which there exists a syndemic equilibrium, and analyze, numerically, the local asymptotical stability of the syndemic equilibrium $\Sigma^*$.
Numerical analysis and discussion {#sect:NumSim}
=================================
For numerical simulations, we consider the following initial conditions for system : $$\begin{gathered}
\label{init:cond}
\left(S(0), L_T(0), I_T(0), R_T(0), I_H(0), A(0), L_{TH}(0), I_{TH}(0), R_{TH}(0), A_T(0)\right)\\
= \left(\frac{60N}{100}, \frac{14 N}{100}, \frac{3N}{100}, 0, \frac{4N}{100},
\frac{N}{100}, \frac{12N}{100}, \frac{5N}{100}, 0, \frac{N}{100}\right)\end{gathered}$$ with $N = 50000$. The parameters of model take the values of Table \[table:parameters:TB-HIV\].
-- -- ----------------- -- -- --
[@USAID:TB:HIV]
-- -- ----------------- -- -- --
: Parameters of the TB-HIV/AIDS model .[]{data-label="table:parameters:TB-HIV"}
Equilibrium points and stability analysis
-----------------------------------------
In Table \[table:effect:beta1\] we show the effect of the transmission coefficient $\beta_1$ on the state $I_T^\diamond$ of the HIV-free equilibrium $\Sigma_T$ and on the basic reproduction number $R_1$. Table \[table:effect:beta2\] shows the effect of the transmission coefficient $\beta_2$ on the states $I_H^\star$ and $A^\star$ of the TB-free equilibrium $\Sigma_H$ and on the basic reproduction number $R_2$. We conclude that the equilibrium states $I_T^\diamond$ and $(I_H^\star, A^\star)$ increase with the transmission coefficients $\beta_1$ and $\beta_2$, respectively.
-- -- -- -- -- --
-- -- -- -- -- --
: Effect of $\beta_1$ on $I_T^\diamond$ and $R_1$.[]{data-label="table:effect:beta1"}
-- -- -- -- -- --
-- -- -- -- -- --
: Effect of $\beta_2$ on $I_H^\star$, $A^\star$ and $R_2$.[]{data-label="table:effect:beta2"}
In Figure \[fig:DFE:stab\] we considered different initial conditions in a neighborhood of the initial conditions given by and $R_0 < 1$ ($R_1 < 1$ and $R_2 < 1$) to illustrate the stability of the disease-free equilibrium $\Sigma_0$ given by . In these numerical simulations we considered $\beta_1 = 2.7$ and $\beta_2 = 0.03$, corresponding to $R_1 = 0.62632$ and $R_2 = 0.55077$, while the rest of the parameters take the values in Table \[table:parameters:TB-HIV\].
$\begin{array}{cc}
\hspace*{-0.4in}
\includegraphics[scale=0.6]{S_DFE_stability_V4.eps} &
\hspace*{-0.6in}\includegraphics[scale=0.6]{NmenosS_DFE_stab_V4.eps}\\
\end{array}$
Figure \[fig:EE:stab\] shows that, for $R_0 > 1$, the syndemic equilibrium $\Sigma^*$ exists. We considered different initial conditions for the state variables of system in a neighborhood of , $\beta_1 = 6$ and $\beta_2 = 0.1$, corresponding to $R_1 = 1.39239$ and $R_2 = 1.83593$, and the rest of the parameters take the values in Table \[table:parameters:TB-HIV\]. We observe that the state variables converge to $\Sigma^*$ when $t \rightarrow \infty$. In this case, $\Sigma^*$ is given by $$\begin{split}
\Sigma^* &= \left(S^*, L_T^*, I_T^*, R_T^*, I_H^*, A^*, L^*_{TH}, I_{TH}^*, R_{TH}^*, A_T^*\right)\\
&= (4766.84, 2019.66, 943.06, 28621.89, 362.66, 56.29, 31.39, 55.15, 495.68, 112.33) \, .
\end{split}$$
![Stability of the syndemic equilibrium $\Sigma^*$.[]{data-label="fig:EE:stab"}](S_LT_EE_stability.eps "fig:") ![Stability of the syndemic equilibrium $\Sigma^*$.[]{data-label="fig:EE:stab"}](IT_EE_stability.eps "fig:") ![Stability of the syndemic equilibrium $\Sigma^*$.[]{data-label="fig:EE:stab"}](RT_EE_stability.eps "fig:") ![Stability of the syndemic equilibrium $\Sigma^*$.[]{data-label="fig:EE:stab"}](IH_A_EE_stability.eps "fig:") ![Stability of the syndemic equilibrium $\Sigma^*$.[]{data-label="fig:EE:stab"}](LTH_ITH_RTH_EE_stability.eps "fig:") ![Stability of the syndemic equilibrium $\Sigma^*$.[]{data-label="fig:EE:stab"}](AT_EE_stability.eps "fig:")
Treatment impact on TB-HIV/AIDS coinfection
-------------------------------------------
Consider $\beta_1=13$ and $\beta_2 = 0.06$, while the rest of the parameters take the values of Table \[table:parameters:TB-HIV\]. Figure \[fig:treatTBonly:death\] shows the impact of treating the individuals with active and latent TB on the number of individuals co-infected with TB-HIV/AIDS. The treatment of individuals with only-TB, $I_T$ and $L_T$, has a positive impact on the reduction of the number of individuals co-infected with TB-HIV/AIDS. Moreover, the number of individuals that suffered from disease (TB and AIDS) induced death is higher when individuals with TB-single infection are not treated. In this case the total population at the end of 20 years is around 10509 and, in the case where individuals with only TB are treated, the total population at the end of 20 years is around 29758 individuals. In Figure \[fig:treatTBonly:no:death\], we assume that there are no disease induced deaths, that is, $d_T = d_A = d_{TA} = 0$. The impact of treating individuals with only TB on the reduction of the number of individuals co-infected is more evident.
![Impact of TB treatment on single-infected individuals with disease induced death.[]{data-label="fig:treatTBonly:death"}](LTH_ITH_treat_TB.eps "fig:") ![Impact of TB treatment on single-infected individuals with disease induced death.[]{data-label="fig:treatTBonly:death"}](RTH_treat_TB.eps "fig:") ![Impact of TB treatment on single-infected individuals with disease induced death.[]{data-label="fig:treatTBonly:death"}](AT_treat_TB.eps "fig:")
![Impact of TB treatment on single-infected individuals with no disease induced death.[]{data-label="fig:treatTBonly:no:death"}](LTH_ITH_treat_TB_noDeath.eps "fig:") ![Impact of TB treatment on single-infected individuals with no disease induced death.[]{data-label="fig:treatTBonly:no:death"}](RTH_treat_TB_noDeath.eps "fig:") ![Impact of TB treatment on single-infected individuals with no disease induced death.[]{data-label="fig:treatTBonly:no:death"}](AT_treat_TB_noDeath.eps "fig:")
Figure \[fig:treatHIVonly:death\] illustrates the case where we compare the number of individuals co-infected with TB-HIV/AIDS when individuals with only AIDS symptoms $A_T$ are or not treated. We observe that treating this class of individuals is important for the reduction of the number of individuals that become co-infected, with special attention to the individuals that have AIDS symptoms and TB infection. In figure \[fig:treatHIVonly:no:death\], we considered that there is no disease induced deaths ($d_T = d_A = d_{TA} = 0$).
![Impact of AIDS treatment on single-infected individuals with disease induced death.[]{data-label="fig:treatHIVonly:death"}](LTH_ITH_treat_HIV.eps "fig:") ![Impact of AIDS treatment on single-infected individuals with disease induced death.[]{data-label="fig:treatHIVonly:death"}](RTH_treat_HIV.eps "fig:") ![Impact of AIDS treatment on single-infected individuals with disease induced death.[]{data-label="fig:treatHIVonly:death"}](AT_treat_HIV.eps "fig:")
![Impact of AIDS treatment on single-infected individuals with no disease induced death.[]{data-label="fig:treatHIVonly:no:death"}](LTH_ITH_treat_HIV_nodeath.eps "fig:") ![Impact of AIDS treatment on single-infected individuals with no disease induced death.[]{data-label="fig:treatHIVonly:no:death"}](RTH_treat_HIV_nodeath.eps "fig:") ![Impact of AIDS treatment on single-infected individuals with no disease induced death.[]{data-label="fig:treatHIVonly:no:death"}](AT_treat_HIV_nodeath.eps "fig:")
It is crucial that TB-infected individuals (in the latent and active stage), which are also HIV-positive, take anti-TB drugs, since they can recover from TB. We analyze the impact of treating TB-HIV/AIDS co-infected individuals $L_{TH}$, $I_{TH}$ and $A_T$ on the reduction of the number of individuals coinfection. If anti-TB drugs are supplied, then latent and active-TB individuals with HIV can recover and pass to the class $R_{TH}$ (the number of individuals in the class $R_{TH}$ tends to zero when TB is not treated). In Figure \[fig:treatTBHIV:no:death\], we observe that after 7 years the number of individuals infected with active-TB and HIV, in the case without treatment, becomes lower than in the case with treatment. This is due to the fact that coinfection precipitates AIDS symptoms.
![Impact of TB and AIDS treatment on co-infected individuals with no disease induced death.[]{data-label="fig:treatTBHIV:no:death"}](LTH_tratarTBHIV_nodeath.eps "fig:") ![Impact of TB and AIDS treatment on co-infected individuals with no disease induced death.[]{data-label="fig:treatTBHIV:no:death"}](ITH_tratarTBHIV_nodeath.eps "fig:") ![Impact of TB and AIDS treatment on co-infected individuals with no disease induced death.[]{data-label="fig:treatTBHIV:no:death"}](AT_tratarTBHIV_nodeath.eps "fig:")
Appendix: Computation of $R_0$ {#A.1}
==============================
The basic reproduction number represents the expected average number of new infections produced by a single infectious individual when in contact with a completely susceptible population [@van:den:Driessche:2002]. Following [@van:den:Driessche:2002], the basic reproduction number $R_{0}$ is obtained as the spectral radius of the matrix $F \cdot V^{-1}$ at the disease-free equilibrium $\Sigma_0$, given by , with $F = \left[ \begin{array}{cc} F_1 \quad F_2 \end{array} \right]$ and and $V = \left[ \begin{array}{cc} V_1 \quad V_2 \end{array} \right]$ with The dominant eigenvalues of the matrix $F \cdot V^{-1}$ are $$\begin{split}
R_1&=\frac{\Lambda}{N\mu}\left(\frac{\beta_1}{d_T+\mu+\tau_2} \right)
\left(\frac{k_1}{k_1+\tau_1+\mu } \right) \, ,\\
R_2&=\frac{\Lambda}{N\mu} \beta_2 \left(\frac{\mu+\alpha_1+d_A
+\eta\,\rho_1 }{\mu\,\alpha_1+ (\mu + \rho_1)(\mu + d_A)}\right)\, .
\end{split}$$ Thus, the basic reproduction number $R_0$ of the model is given by $$R_0 = \max \{ R_1, R_2 \} \, .$$ Note that $R_1$ is the basic reproduction number of the model with $I_T = A = L_{TH} = I_{TH} = R_{TH} = A_T = 0$ (only TB model), and $R_2$ is the basic reproduction number of the model with $L_T = I_T = R_T = L_{TH} = I_{TH} = R_{TH} = A_T = 0$ (only HIV-AIDS model).
Appendix: Proof of Theorem \[theo:stab:equil\] {#A.2}
==============================================
In this Appendix we provide details of the proof of Theorem \[theo:stab:equil\].
#### Local asymptotical stability of the disease-free equilibrium $\Sigma_0$.
Following Theorem 2 of [@van:den:Driessche:2002], the disease-free equilibrium, $\Sigma_0$, is locally asymptotically stable if all the eigenvalues of the Jacobian matrix of the system , here denoted by $M_T\left(\Sigma_0\right)$, computed at the disease free equilibrium $\Sigma_0$, given by , have negative real parts.
The Jacobian matrix of the system at disease free equilibrium $\Sigma_0$ is given by $$M_T\left(\Sigma_0\right) = \left[
\begin{array}{cc}
M_{T1}\left(\Sigma_0\right) \quad M_{T_2}\left(\Sigma_0\right)
\end{array}
\right]$$ with $$M_{T1}\left(\Sigma_0\right)=
{\footnotesize{
\left[ \begin {array}{ccccc}
-\mu&0&-{\frac{\beta_1 \Lambda}{\mu N}}&0&
-{\frac{\beta_2 \Lambda}{\mu N}}\\ \noalign{\medskip}
0&-d_1&{\frac{\beta_1 \Lambda}{\mu N}}&0&0\\ \noalign{\medskip}
0&k_1&-d_2&0&0\\ \noalign{\medskip}
0&\tau_1&\tau_2&-\mu&0\\ \noalign{\medskip}
0&0&0&0&{\frac{\beta_2 \Lambda}{\mu N}}-d_3\\ \noalign{\medskip}
0&0&0&0&\rho_1\\ \noalign{\medskip}
0&0&0&0&0\\ \noalign{\medskip}
0&0&0&0&0\\ \noalign{\medskip}
0&0&0&0&0\\ \noalign{\medskip}
0&0&0&0&0
\end {array} \right]
}}$$ and $$M_{T2}\left(\Sigma_0\right)=
{\tiny{
\left[ \begin {array}{cccccccccc}
-{\frac{\beta_2 \eta \Lambda}{\mu N}}&
-{\frac{\beta_2 \Lambda}{\mu N}}&
-{\frac {\beta_1 \Lambda}{\mu N}}-{\frac{\beta_2 \Lambda}{\mu N}}&
-{\frac {\beta_2 \Lambda}{\mu N}}&
-{\frac{\beta_1 \Lambda}{\mu N}}-{\frac{\beta_2 \eta \Lambda}{\mu N}}\\
\noalign{\medskip}
0&0
&{\frac{\beta_1 \Lambda}{\mu N}}&0&{\frac{\beta_1 \Lambda}{\mu N}}\\
\noalign{\medskip}
0&0&0&0&0\\
\noalign{\medskip}
0&0&0&0&0\\
\noalign{\medskip}
{\frac{\beta_2 \eta \Lambda}{\mu N}}+\alpha_1
&{\frac {\beta_2 \Lambda}{\mu N}}
&{\frac{\beta_2 \Lambda}{\mu N}}
&{\frac{\beta_2 \Lambda}{\mu N}}
&{\frac {\beta_2 \eta \Lambda}{\mu N}}\\
\noalign{\medskip}
-d_4&0&0&0&0\\ \noalign{\medskip}
0&-d_5&0&0&0\\ \noalign{\medskip}
0&k_2&-d_6&0&\alpha_2\\ \noalign{\medskip}
0&\tau_4&\tau_3&-d_7&0\\ \noalign{\medskip}
0&0&\rho_2&\rho_3&-d_8
\end {array} \right] ,
}}$$ where $d_1 = k_1 +\tau_1+\mu$; $d_2 = \tau_2+\mu+d_T$; $d_3 = \rho_1+\mu$; $d_4 = \alpha_1+\mu+d_A$; $d_5 = k_2+\mu + \tau_4$; $d_6 = \rho_2+\tau_3+\mu+d_T$; $d_7 = \rho_3+\mu$; $d_8 = \alpha_2+ d_{TA}+\mu$. One has $$trace\left[ M_T\left(\Sigma_0\right) \right] =
-2 \mu - (d_1 + d_2 + d_3 + d_4 + d_5 + d_6 + d_7 + d_8) < 0$$ and $$\begin{split}
det \left[ M_T\left(\Sigma_0\right) \right] &
= \frac{1}{N^2} (
d_5\, ( d_6 \, d_7 + d_T(\alpha_2 + \mu)\,d_7 + \alpha_2 \mu \,d_6 + d_T d_{TA}\, d_7 )\\
&\times (N \mu (\alpha_1 \mu + (\mu + \rho_1)(\mu + d_A))
-\beta_2 \Lambda(\alpha_1 + \mu + d_A + \rho_1 \eta))\\
&\times (N \mu (d_T + \mu + \tau_2)
(k_1 + \tau_1 + \mu)-k_1 \beta_1 \Lambda) > 0
\end{split}$$ for $$R_1 = \frac{\Lambda}{N\mu}\left(\frac{\beta_1}{d_T+\mu+\tau_2} \right)
\left(\frac{k_1}{k_1+\tau_1+\mu } \right) < 1$$ and $$R_2 =\frac{\Lambda}{N\mu} \beta_2 \left(\frac{\mu+\alpha_1+d_A
+\eta\,\rho_1 }{\mu\,\alpha_1+ (\mu + \rho_1)(\mu + d_A)}\right) < 1 \, .$$ We have just proved that the disease free equilibrium $\Sigma_0$ of model is locally asymptotically stable if $R_0 < 1$, and unstable if either $R_i > 1$, $i=1, 2$.
#### Global asymptotical stability of the disease-free equilibrium $\Sigma_0$.
For convenience, let us rewrite the model system as $$\label{mod:append}
\begin{split}
&\frac{dX}{dt} = F(X, Z) \, , \\
&\frac{dZ}{dt} = G(X, Z)\, , \quad G(X, 0) = 0 \, ,
\end{split}$$ where $X = (S, R_T)$ and $Z = (L_T, I_T, I_H, A, L_{TH}, I_{TH}, R_{TH}, A_T)$, with $X \in \R^2_+$ denoting (its components) the number of uninfected individuals and $Z \in \R^8_+$ denoting (its components) the number of infected individuals including the latent and infectious.
The disease-free equilibrium is denoted by $$E_0 = (X_0, 0)\, , \quad \text{where} \, \, X_0
= \left(\frac{\Lambda}{\mu}, 0 \right) \, .$$ Following [@Bhunu:BMB:2009:HIV:TB], if
- $E_0$ is globally asymptotically stable for $\frac{dX}{dt} = F(X, 0)$,
- $\hat{G}(X, Z) {\geqslant}0$ for $(X, Z) \in \Omega$, where $G(X, Z)
= AZ - \hat{G}(X, Z)$, $A = D_Z G(E_0, 0)$ is a Metzler matrix and $\Omega$ is given by ,
then the fixed point $E_0 = (X_0, 0)$ is a globally asymptotically stable equilibrium of system . We have $$\frac{dX}{dt} = F(X, Z)
= \left[
\begin{array}{c}
\Lambda - \lambda_T S - \lambda_H S - \mu S \\[0.2 cm]
\tau_1 L_T + \tau_2 I_T - (\beta^{'}_1 \lambda_T + \lambda_H + \mu) R_T
\end{array} \right] \, ,$$ $$F(X, 0) = \left[ \begin{array}{c}
\Lambda - \mu S \\[0.2 cm]
-\mu R_T
\end{array} \right] \, ,$$ $$\frac{dZ}{dt} = G(X, Z) = \left[
\begin{array}{c}
\lambda_T S+ \beta^{'}_1 \lambda_T R_T- (k_1 + \tau_1 + \mu)L_T,\\[0.2 cm]
k_1 L_T - (\tau_2 +d_T +\mu + \delta \lambda_H)I_T, \\[0.2 cm]
\lambda_H S - (\rho_1 + \psi \lambda_T + \mu)I_H + \alpha_1 A + \lambda_H R_T \\[0.2 cm]
\rho_1 I_H - \alpha_1 A - (\mu + d_A) A\\[0.2 cm]
\beta^{'}_2 \lambda_T R_{TH} - (k_2 + \tau_4 + \mu) L_{TH}\\[0.2 cm]
\delta \lambda_H I_T + \psi \lambda_T I_H + \alpha_2 A_T+ k_2 L_{TH}
- (\tau_3 + \rho_2 + \mu + d_T)I_{TH}\\[0.2 cm]
\tau_3 I_{TH} + \tau_4 L_{TH} - (\beta^{'}_2 \lambda_T + \rho_3 + \mu)R_{TH} \\[0.2 cm]
\rho_2 I_{TH} + \rho_3 R_{TH} -(\alpha_2 + \mu + d_{TA})A_T \, ,
\end{array}
\right],$$ and $G(X, 0) = 0$. Thus, $$\frac{dX}{dt} = F(X, 0) = \left[
\begin{array}{c}
\Lambda - \mu S \\[0.2 cm]
- \mu R_T
\end{array} \right] \, ,$$ $$A = D_Z G(X_0, 0) = \left[ \begin {array}{cc} D_1 & D_2
\end{array} \right]$$ with $$D_1 = {\footnotesize{ \left[ \begin {array}{cccc}
-k_1-\tau_1-\mu & \frac{\beta_1 \Lambda}{\mu N}&0&0
\\ \noalign{\medskip}
k_1&-\tau_2-\mu- d_T&0&0\\ \noalign{\medskip}
0&0&\frac {\beta_2 \Lambda}{\mu N}-\rho_1-\mu&
\frac{\beta_2 \eta \Lambda}{\mu N}+\alpha_1\\ \noalign{\medskip}
0&0&\rho_1&-\alpha_1-\mu- d_A\\ \noalign{\medskip}
0&0&0&0\\ \noalign{\medskip}
0&0&0&0\\ \noalign{\medskip}
0&0&0&0\\ \noalign{\medskip}
0&0&0&0
\end {array} \right] }} \, ,$$ $$D_2 = {\footnotesize{ \left[ \begin {array}{cccccccc}
0&{\frac{\beta_1 \Lambda}{\mu N}}&0&{\frac{\beta_1 \Lambda}{\mu N}}\\
\noalign{\medskip}
0&0&0&0\\ \noalign{\medskip}
{\frac{\beta_2 \Lambda}{\mu N}}&
{\frac{\beta_2 \Lambda}{\mu N}}&{\frac{\beta_2 \Lambda}{\mu N}}
&{\frac{\beta_2 \eta \Lambda}{\mu N}}\\ \noalign{\medskip}
0&0&0&0\\ \noalign{\medskip}
-k_2-\tau_4 -\mu &0&0&0\\ \noalign{\medskip}
k_2&-\rho_2-\tau_3-\mu - d_T&0&\alpha_2\\ \noalign{\medskip}
\tau_4&\tau_3&-\rho_3-\mu&0\\ \noalign{\medskip}
0&\rho_2&\rho_3&-\alpha_2-d_{TA}-\mu
\end {array} \right] }}$$ and $$\label{eq:hat:G}
\hat{G}(X, Z)
= {\footnotesize{ \left[ \begin {array}{c}
\lambda_T \left(\frac{\Lambda}{\mu} - S - \beta_1^{'} R_T \right)\\
\noalign{\medskip}
-\delta \lambda_H I_T\\
\noalign{\medskip}
\lambda_H \left( \frac{\Lambda}{\mu} - S - R_T - \psi I_H \right)\\
\noalign{\medskip}
0\\ \noalign{\medskip}
-\beta_2^{'} \lambda_T R_{TH}\\ \noalign{\medskip}
-\left(\delta \lambda_H I_T + \psi \lambda_T I_H \right)\\
\noalign{\medskip}
\beta_2^{'} \lambda_T R_{TH}\\
\noalign{\medskip}
0
\end {array} \right] }} \, .$$ From the condition (H2) is not satisfied, since $\hat{G}(X, Z) {\geqslant}0$ is not true. Therefore, the disease-free equilibrium $E_0$ may not be globally asymptotically stable. Following [@CChavez_TB_Exog_TPB_2000], the backward bifurcation occurs at $R_0 = 1$ and the double endemic equilibria can be supported for $R_c < R_0 < 1$, where $R_c$ is a positive constant.
#### Existence and stability of HIV-AIDS free equilibrium $\Sigma_T$.
The expressions for $S^\diamond$, $L_T^\diamond$, $I_T^\diamond$ and $R_T^\diamond$ are obtained if we consider a sub-model of for which $I_H = A = L_{TH} = I_{TH} = R_{TH} = A_T = 0$ and the total population $N$ is given by $N_T = S + L_T + I_T + R_T$. The basic reproduction number of this submodel is given by $R_1$ . The existence, uniqueness and local asymptotic stability of $\Sigma_T$ is proven in [@Castillo_Chavez_1997 Theorem 1].
#### Existence and stability of TB free equilibrium $\Sigma_H$.
To prove the existence of $\Sigma_T$, consider the sub-model of for which $L_T = I_T = R_T = L_{TH} = I_{TH} = R_{TH} = A_T = 0$ and the total population $N_H$ is given by $N_H = S + I_H + A$. The equations of this submodel are $$\label{submodel:HIV}
\begin{cases}
\dot{S}(t) = \Lambda - \lambda_H S(t) - \mu S(t)\\[0.2 cm]
\dot{I}_H(t) = \lambda_H S(t) - (\rho_1 + \mu)I_H(t) + \alpha_1 A(t)\\[0.2 cm]
\dot{A}(t) = \rho_1 I_H(t) - \alpha_1 A(t) - (\mu + d_A)A \, ,
\end{cases}$$ where $\lambda_H = \beta_2 \frac{I_H + \eta A}{N_H}$. Setting the right hand sides of submodel to zero, we obtain the endemic equilibrium $\Sigma_H^\star = (S^*, I_H^*, A^*)$ given by $$S^\star = \frac{\Lambda}{\mu R_2}\, , \quad I_H^\star
= (R_2 - 1)\frac{\mu N_H (\alpha_1 + d_A + \mu)}{\beta_2 (\alpha_1
+ d_A + \mu + \eta \rho_1)}\, , \, \, \quad A^\star
= (R_2 - 1) \frac{\rho_1 \mu N_H}{\beta_2 (\alpha_1 + d_A + \mu + \eta \rho_1)},$$ where $I_H^\star > 0$ and $A^\star > 0$, whenever $R_2 > 1$.
In what follows we prove the local asymptotic stability of the endemic equilibrium $\Sigma_H^\star$, using the center manifold theory [@Carr:1981], as described in [@CChavez_Song_2004 Theorem 4.1] (see also [@van:den:Driessche:2002]), considering ART treatment. The basic reproduction number of this sub-model $R_2$ is given by . Chose as bifurcation parameter, $\beta^*$, by solving for $\beta_2$ from $R_2 = 1$: $$\beta^* = {\frac {\mu \alpha_1 + (\mu + \rho_1)(\mu + d_A)}{
\alpha+ d_A +\mu+\eta \rho}} \, .$$ The submodel has a disease free equilibrium given by $\Sigma_{H0}^*=(x_{10}, x_{20}, x_{30})= \left(\frac{\Lambda}{\mu}, 0, 0 \right)$.
The Jacobian of the system , evaluated at $\Sigma_{H0}^*$ and with $\beta_2 = \beta^*$, is given by $$\label{Jacob:betaast}
J(\Sigma_{H0}^*) = \left[ \begin {array}{ccc}
-\mu&- \beta_2&-\beta_2 \eta\\
\noalign{\medskip}
0&\beta_2-\rho-\mu & \beta_2 \eta + \alpha\\
\noalign{\medskip}
0&\rho&-\alpha- d_A - \mu
\end {array} \right] \, .$$ The eigenvalues of the linearized system are $$\lambda_1 = 0, \quad \lambda_2 = -\mu \quad \text{and} \quad \lambda_3
= -\frac{ \eta \rho (2\mu^2 + \rho + d_A + \alpha) + d_A(2 \alpha + 2 \mu +d_A)
+\rho \alpha+ (\mu + \alpha)^2}{\alpha+ d_A+\mu+ \eta \rho} \, .$$ We observe that there is a simple eigenvalue with zero real part and the other two eigenvalues have negative real part. Thus, the system , with $\beta_2 = \beta^*$, has a hyperbolic equilibrium point and the center manifold theory [@Carr:1981] can be used to analyze the dynamics of the submodel near $\beta_2 = \beta^*$.
The Jacobian $J(\Sigma_{H0}^*)$ at $\beta_2 = \beta^*$ has a right eigenvector (associated with the zero eigenvalue) given by $w = [w_1, w_2, w_3]^T$, where $$\begin{split}
w_1&=-\frac{\left( \mu \alpha_1+ (\mu + \rho_1)(\mu + d_A) \right) w_3}{\rho_1 \mu}\, ,\\
w_2&=\frac { \left( \alpha_1 + d_A+\mu \right) w_3}{\rho_1}\, ,\\
w_3&= w_3 > 0 \, .
\end{split}$$ Further, $J(\Sigma_{H0}^*)$ for $\beta_2 = \beta^*$ has a left eigenvector $v = [v_1, v_2, v_3]$ (associated with the zero eigenvalue), where $$\begin{split}
v_1&=0\, ,\\
v_2&=\frac{v_3 \left( \alpha_1+ d_A +\mu+\eta \rho_1 \right)}{
\alpha_1 +\eta \rho_1 +\mu \eta}\, ,\\
v_3&= v_3 > 0 \, .
\end{split}$$ To apply Theorem 4.1 in [@CChavez_Song_2004] it is convenient to let $f_k$ represent the right-hand side of the $k$th equation of the system and let $x_k$ be the state variable whose derivative is given by the $k$th equation for $k = 1, 2, 3$. The local stability near the bifurcation point $\beta_2 = \beta^*$ is then determined by the signs of two associated constants, denoted by $a$ and $b$, defined (respectively) by $$a = \sum_{k, i, j=1}^3 \, v_k w_i w_j \frac{\partial^2 f_k}{\partial x_i \partial x_j}(0, 0)
\quad \text{and} \quad b = \sum_{k, i =1}^3 \, v_k w_i
\frac{\partial^2 f_k}{\partial x_i \partial \phi}(0, 0)$$ with $\phi = \beta_2 - \beta^*$.
For the system , the associated partial derivatives at the disease free equilibrium $\Sigma_{H0}$ are given by $$\begin{split}
\frac{\partial^2 f_1}{\partial x_2^2} &= \frac{2\beta^*\mu}{\Lambda}\, , \quad
\frac{\partial^2 f_1}{\partial x_2 \partial x_3} = \frac{\beta^*\mu(1+\eta)}{\Lambda}\, , \quad
\frac{\partial^2 f_1}{\partial x_3^2} = \frac{2\beta^*\mu \eta}{\Lambda} \, ,\\
\frac{\partial^2 f_2}{\partial x_2^2} &= \frac{-2\beta^*\mu}{\Lambda}\, , \quad
\frac{\partial^2 f_2}{\partial x_2 \partial x_3} = \frac{-\beta^*\mu(1+\eta)}{\Lambda}\, , \quad
\frac{\partial^2 f_2}{\partial x_3^2} = \frac{-2\beta^*\mu \eta}{\Lambda} \, .
\end{split}$$ It follows from the above expressions that $$a=-\frac {v_3 w_3^2 \beta^* \mu \left(k_1 +\mu+\eta \rho_1 \right)
\left( 2 k_1^2 + 4 \mu k_1 + 2 \mu^2 + \rho_1 (\alpha_1
+ \eta (\alpha_1 + \mu + 2 \rho_1) + d_A(1 + \eta) + \mu)
\right)}{\rho_1^2\Lambda \left( \alpha_1
+\eta \rho_1 +\mu \eta \right)} < 0$$ with $k_1 = \alpha_1+ d_A $.
For the sign of $b$, it can be shown that the associated non-vanishing partial derivatives are $$\begin{split}
\frac{\partial^2 f_1}{\partial x_2 \partial \beta^*} &= -1\, , \quad
\frac{\partial^2 f_1}{\partial x_3 \partial \beta^*} = -\eta \, , \\
\frac{\partial^2 f_2}{\partial x_2 \partial \beta^*} &= 1 \, , \quad
\frac{\partial^2 f_2}{\partial x_3 \partial \beta^*} = \eta \, .
\end{split}$$ It also follows from the above expressions that $$b=\frac{ v_3 w_3 \left( k_1 +\mu+\eta \rho_1 \right)
\left( k_1 +\mu \right) }{ \left( \alpha_1 +\eta
\rho_1 +\mu \eta \right) \rho_1}+ \frac {\eta v_3 w_3
\left( k_1 +\mu+\eta \rho_1 \right)}{\alpha_1 +\eta \rho_1 + \mu \eta} > 0 \, .$$ Thus, $a< 0$ and $b > 0$. Using Theorem 4.1 of [@CChavez_Song_2004], the endemic equilibrium $\Sigma_H^\star$ is locally asymptotically stable for $R_2$ near 1.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by Portuguese funds through the *Center for Research and Development in Mathematics and Applications* (CIDMA, University of Aveiro), and *The Portuguese Foundation for Science and Technology* (“FCT — Fundação para a Ciência e a Tecnologia”), within project PEst-OE/MAT/UI4106/2014. Silva was also supported by FCT through the post-doc fellowship SFRH/BPD/72061/2010; Torres by the FCT project PTDC/EEI-AUT/1450/2012, co-financed by FEDER under POFC-QREN with COMPETE reference FCOMP-01-0124-FEDER-028894.
[99]{}
C. P. Bhunu, W. Garira and Z. Mukandavire, *Modeling HIV/AIDS and tuberculosis coinfection*, Bul. Math. Biol. 71, 1745–1780 (2009).
J. Carr, *Applications centre manifold theory*, Springer-Verlag, New-York (1981).
C. Castillo-Chavez and Z. Feng, *To treat or not to treat: The case of tuberculosis*, J. Math. Biol. 35, no. 6, 629–656 (1997).
C. Castillo-Chavez and B. Song, *Dynamical models of tuberculosis and their applications*, Math. Biosc. Engrg. 1, no. 2, 361–404 (2004).
P. W. David, G. L. Matthew, E. G. Andrew, A. C. David and M. K. John, *Relation between HIV viral load and infectiousness: A model-based analysis*, The Lancet 372, no. 9635, 314–320 (2008).
Z. Feng, C. Castillo-Chavez and A. F. Capurro, *A model for tuberculosis with exogenous reinfection*, Theor. Pop. Biology 57, 235–247 (2000).
H. Getahun, C. Gunneberg, R. Granich and P. Nunn, *HIV infection-associated tuberculosis: The epidemiology and the response*, Clin. Infect. Dis. 50 (Suppl 3), S201–S207 (2010).
J. M. Hyman, J. Li and E. A. Stanley, *The differential infectivity and staged progression models for the transmission of HIV*, Math. Biosci. 155, 77–109 (1999).
D. Kirschner, *Dynamics of co-infection with M. tuberculosis and HIV-1*, Theor. Pop. Biol. 55, 94–109 (1999).
C. K. Kwan and J. D. Ernst, *HIV and tuberculosis: A deadly human syndemic*, Clin. Microbiol. Rev. 24, no. 2, 351–376 (2011).
R. Naresh and A. Tripathi, *Modelling and analysis of HIV-TB co-infection in a variable size population*, Math. Model. Anal. 10, 275–286 (2005).
L. W. Roeger, Z. Feng and C. Castillo-Chavez, *Modeling TB and HIV co-infections*, Math. Biosc. and Eng. 6, no. 4, 815–837 (2009).
O. Sharomi, C.N. Podder, A.B. Gumel and B. Song, *Mathematical analysis of the transmission dynamics of HIV/TB coinfection in the presence of treatment*, Math. Biosc. Eng. 5, no. 1, 145–174 (2008).
UNAIDS, *Global report: UNAIDS report on the global AIDS epidemic 2013*, Geneva, World Health Organization (2013).
P. van den Driessche and J. Watmough, *Reproduction numbers and subthreshold endemic equilibria for compartmental models of disease transmission*, Math. Biosc. 180, 29–48 (2002).
WHO, *Global tuberculosis report 2013*, Geneva, World Health Organization (2013).
<http://www.usaid.gov/news-information/fact-sheets/twin-epidemics-hiv-and-tb-co-infection>
<http://en.wikipedia.org/wiki/HIV_disease_progression_rates>
<http://hivinsite.ucsf.edu/InSite?page=kb-03-01-04>
[^1]: Corresponding author.
[^2]: This is a preprint of a paper whose final and definite form is: Journal of Applied Mathematics (ISSN 1110-757X) 2014, Article ID 248407, [http://dx.doi.org/10.1155/2014/248407]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper provides a comprehensive review of the domain of physical layer security for wireless communications. The essential premise of physical-layer security is to enable the exchange of confidential messages over a wireless medium in the presence of unauthorized eavesdroppers. This can be achieved primarily in two ways: without the need for a secret key by intelligently designing transmit coding strategies, or by exploiting the wireless communication medium to develop secret keys over public channels. We begin with an overview of the foundations dating back to the pioneering work of Shannon and Wyner on information-theoretic security. We then describe the evolution of secure transmission strategies from point-to-point channels to multiple-antenna systems, followed by generalizations to multiuser broadcast, multiple-access, interference, and relay networks. Subsequently, we evaluate secret-key establishment protocols based on physical layer mechanisms, along with an overview of practical secrecy-preserving code design.'
author:
- 'Amitav Mukherjee, S. Ali Fakoorian, Jing Huang, and A. Lee Swindlehurst, *Fellow, IEEE*[^1]'
title: 'Principles of Physical Layer Security in Multiuser Wireless Networks: A Survey'
---
INTRODUCTION
============
The two fundamental characteristics of the wireless medium, namely *broadcast* and *superposition*, present different challenges in ensuring reliable and/or secure communications in the presence of adversarial users. The broadcast nature of wireless communications makes it difficult to shield transmitted signals from unintended recipients, while superposition can lead to the overlapping of multiple signals at the receiver. As a consequence, adversarial users are commonly modeled either as (1) an unauthorized receiver that tries to extract information from an ongoing transmission without being detected, or (2) a malicious transmitter (*jammer*) that tries to degrade the signal at the intended receiver [@Stark88]-[@Basar04].
While jamming and counter-jamming physical layer strategies have been of long-standing interest especially in military networks, the security of data transmission has traditionally been entrusted to key-based enciphering (cryptographic) techniques at the network layer [@Massey88]. However, in dynamic wireless networks this raises issues such as key distribution for symmetric cryptosystems, and high computational complexity of asymmetric cryptosystems. More importantly, all cryptographic measures are based on the premise that it is computationally infeasible for them to be deciphered without knowledge of the secret key, which remains mathematically unproven. The information-theoretic aspects of secrecy at the *physical layer* have experienced a resurgence of interest only in the past decade or so. Therefore, the remainder of this paper is devoted to surveying and reviewing the various aspects of physical-layer security in modern wireless networks.
The fundamental principle behind physical-layer security is to limit the amount of information that can be extracted at the ‘bit’ level by an unauthorized receiver. With appropriately designed coding and transmit precoding schemes in addition to the exploitation of any available channel state information, physical-layer security schemes enable secret communication over a wireless medium without the aid of an encryption key. However, if it is desirable to use a secret key for encryption, then information-theoretic security also describes techniques that allow for the evolution of such a key over wireless channels that are observable by the adversary. Alternatively, since they can operate essentially independently of the higher layers, physical layer techniques can be used to augment already existing security measures. The vast majority of information-theoretic security research reviewed in this survey contains the premise that the eavesdropper is passive, i.e., does not transmit in order to conceal its presence.
The survey does not proceed in a strictly chronological order, nor is the list of references intended to be exhaustive. Furthermore, due to inaccessibility we are forced to omit contributions to the field published within the former Soviet Union [@Yakovlev81] and other international forums. Instead, we aim to provide a high-level overview of the historical development of the field with references that are easily accessible, juxtaposed with recent and ongoing research efforts.
The remainder of the paper is organized as follows. In the next section, the fundamental mathematical precepts of secrecy are presented, along with a description of the most elementary secrecy problem: the wiretap channel. The state-of-the-art in the burgeoning area of multi-antenna wiretap channels is described in Section \[sec:MIMOWiretap\]. The extension to more than three terminals for broadcast, multiple-access, and interference channels is described in Section \[sec:BC\_MAC\_IC\]. The development of secrecy in relay channels and miscellaneous systems such as sensor and cognitive radio networks is carried out in \[sec:Relay\]. The important issue of secret-key agreement in wireless networks is studied in Section \[sec:Key\]. Section \[sec:RelatedTopics\] highlights the emerging areas of practical wiretap code design and cross-disciplinary approaches to secrecy. Finally, in Section \[sec:concl\] we summarize our discussion.
Fundamentals {#sec:Fundamentals}
============
The simplest network where problems of secrecy and confidentiality arise is a three-terminal system comprising a transmitter, the intended (legitimate) receiver, and an unauthorized receiver, wherein the transmitter wishes to communicate a private message to the receiver. In the sequel, the unauthorized receiver is referred to interchangeably as an *eavesdropper* or *wiretapper*. Encryption of messages via a secret key known only to the transmitter and intended receiver has been the traditional route to ensuring confidentiality. Up until the 20th century, the design of cryptographic methods was based on the notion of computational security, without a solid mathematical basis for secrecy. A classical example was Vernam’s one-time pad cipher [@Vernam26], where the binary message or plaintext is XOR’ed with a random binary key of the same length.
Performance Metrics
-------------------
Shannon postulated the information-theoretic foundations of modern cryptography in his ground-breaking treatise of 1949 [@Shannon49]. Shannon’s model assumed that a non-reusable private key $K$ is used to encrypt the confidential message $M$ to generate the cryptogram $C$, which is then transmitted over a noiseless channel. The eavesdropper has unbounded computational power, has knowledge of the transmit coding scheme, and has access to an identical copy of the signal at the intended receiver. The notion of perfect secrecy was introduced, which requires that the *a posteriori* probability of the secret message computed by the eavesdropper based on her received signal be equal to the *a priori* probability of the message. In other words, $$\label{eq:Shannon_perfsecrecy}
I(M;C)=0,$$ where $I(\cdot;\cdot)$ denotes mutual information. A by-product of this analysis was that perfect secrecy [@Geffe65] can be guaranteed only if the secret key has at least as much entropy as the message to be encrypted, $H(K) \geq H(M)$, which validated Vernam’s one-time pad cipher system.
![The degraded wiretap channel [@Wyner75].[]{data-label="fig_Wyner_Wiretap"}](Wyner){width="\linewidth"}
Wyner ushered in a new era in information-theoretic security when he introduced the wiretap channel in [@Wyner75], which considered the imperfections introduced by the channel. In the wiretap channel, the information signal $X$ is transmitted to the intended receiver over the ‘main channel’ which is modeled as a discrete memoryless channel. The receiver observes $Y$, which subsequently passes through an additional ‘wiretap channel’ before being received by the eavesdropper as $Z$, as shown in Fig. 1. Under the assumption that the source-wiretapper channel is a probabilistically degraded version of the main channel [@Hellman77], Wyner sought to maximize the transmission rate $R$ in the main channel while making negligible the amount of information leaked to the wiretapper channel.
In his development Wyner defined the *equivocation* rate of the wiretapper, where the wiretapper equivocation or ‘ambiguity’ is the conditional entropy $H(S^k\mid Z^n)$, and therefore the equivocation rate is $$R_e \leq \frac{H(S^k\mid Z^n)}{n}.$$ If the equivocation rate $R_e$ is arbitrarily close to the information rate $R$, then $R$ is the *secrecy capacity*[^2] of the wiretap channel. He constructed a randomized coding scheme which sought to hide the information stream in the additional noise impairing the wiretapper by mapping each message to many codewords according to an appropriate probability distribution. This way, one induces maximal equivocation at the wiretapper, and Wyner was able to show that secure communication was possible *without* the use of a secret key. In 1993, Maurer [@Maurer93] presented a strategy that allowed a positive rate even when the wiretapper observes a “better" channel than the one used by the legitimate users. The essence of Maurer’s scheme is the joint development of a secret key by the transmitter and receiver via communication over a public (insecure) and error-free feedback channel. Thereafter, research in information-theoretic secrecy developed along two main branches: secret key-based secrecy as in the work by Shannon and Maurer, and keyless security as in the work by Wyner. In Sections III-V we trace the evolution of keyless security over the past decade. We revisit the topic of key-based security for wireless channels in Section \[sec:Key\].
Single-Antenna Wiretap Channels Since Wyner {#subsec:SISOWiretap}
-------------------------------------------
Carleial and Hellman [@Hellman77] considered a special case of Wyner’s model where the main channel is noiseless and the wiretap channel is a binary symmetric channel, and analyzed the applicability of systematic linear codes for preserving the secrecy of an arbitrarily portion of the transmitted message. For the degraded wiretap channel with additive Gaussian noise [@LeungH78], the essential result for the secrecy capacity $C_S$ was the following: $$\label{eq:degradedGaussianSecCap}
C_S=C_M-C_W,$$ where $C_M$ and $C_W$ are the Shannon capacities of the main and wiretap channels, respectively. Ultimately, it was established that a non-zero secrecy capacity can only be obtained if the eavesdropper’s channel is of lower quality than that of the intended recipient.
Csizar and Korner considered a more general (non-degraded) version of Wyner’s wiretap channel in [@Csiszar78], where they obtained a single-letter characterization of the achievable (private message rate, equivocation rate, common message rate)-triple for a two-receiver broadcast channel. For the special case of no common messages, the secrecy capacity was defined as $$\label{eq:Korner1978}
C_S = \mathop {\max }\limits_{V \to X \to YZ} I\left( {V;Y} \right) - I\left( {V;Z} \right),$$ which is achieved by maximizing over all joint probability distributions such that a Markov chain $V,X,YZ$ is formed, where $V$ is an auxiliary input variable.
In [@Ozarow84], Ozarow and Wyner studied the type-II wiretap channel, where the main communication channel is noiseless but the wiretapper has access to an arbitrary subset $\mu$ of the $N$ coded bits, and optimal tradeoffs between code rate $k/N$ and $\mu$ that guaranteed secrecy were characterized.
The consideration of channel fading in wiretap channels has recently opened new avenues of research. Barros and Rodrigues *et al* [@Barros06]-[@Bloch06] analyzed the outage secrecy capacity of slow fading channels and have showed that in the presence of fading information-theoretic security is achievable even when the eavesdropper has a better average SNR than the legitimate receiver.
Li *et al*. [@Trappe07] examined an achievable secrecy rate for an AWGN main channel, while the eavesdropper’s channel is Rayleigh fading with additive Gaussian noise, and its realizations are unknown to Alice and Bob. The main result of this paper is that with Gaussian random codes, artificial noise injection and power bursting, positive secrecy rate is achievable even when the main channel is arbitrarily worse than the eavesdropper’s average channel.
Relatively fewer studies consider the case of a complete absence of eavesdropper CSI at the transmitter in fading wiretap channels. In [@ElGamal08], the authors consider a block-fading scalar wiretap channel where the number of channel uses within each coherence interval is large enough to invoke random coding arguments. This assumption is critical for their achievable coding scheme which attempts to “hide" the secure message across different fading states. A recent approach towards understanding the information-theoretic limits of wiretap channels with no eavesdropper CSI has been taken by studying the compound wiretap channel [@Kramer09]. The compound wiretap channel captures the situation in which there is no or incomplete CSI at the transmitter by characterizing the eavesdropper’s channel with a finite set of states, and guarantees secure communication under any state that may occur. The schemes designed for the compound channel are robust to various communication environments.
Multi-antenna Channels {#sec:MIMOWiretap}
======================
The explosion of interest in multiple-input multiple-output (MIMO) systems soon led to the realization that exploiting the available spatial dimensions could also enhance the secrecy capabilities of wireless channels. The work by Hero [@Hero03] is arguably the first to consider secret communication in a MIMO setting, and sparked a concerted effort to apply and extend the single-antenna wiretap theory to this new problem. In a fading MIMO channel where the transmitter, receiver, and eavesdropper are equipped with $N_T,N_R,N_E$ antennas respectively, a general representation for the signal received by the legitimate receiver is $$\mathbf{y}_b = \mathbf{H}_{b} \mathbf{x}_a + \mathbf{n}_b, \label{eq:yHa}$$ while the received signal at the eavesdropper is $$\mathbf{y}_e = \mathbf{H}_{e} \mathbf{x}_a + \mathbf{n}_e \label{eq:yHe},$$ where $\mathbf{x}_a \in \mathbb{C}^{N_T\times 1}$ is the transmit signal with covariance $E\left\{ {{\mathbf{x}}_a {\mathbf{x}}_a^H } \right\} = {\mathbf{Q}}_x, \operatorname{Tr} \left( {{\mathbf{Q}}_x } \right) \leqslant P$, $\mathbf{H}_{ba} \in \mathbb{C}^{N_R\times N_T}, \mathbf{H}_{ba} \in \mathbb{C}^{N_E\times N_T}$ are the MIMO complex Gaussian channel matrices, and $\mathbf{n}_b, \mathbf{n}_e$ are the respective zero-mean complex Gaussian additive noise vectors.
![General MIMO wiretap channel.[]{data-label="fig_MIMOwiretap"}](MIMOwiretap){width="\linewidth"}
Hero examined the utility of space-time block coding for covert communications in [@Hero03], and designed CSI-informed transmission strategies to achieve either a low probability of intercept (equivalent to secrecy rate), or a low probability of detection for various assumptions about the CSI available to the eavesdropper. One of the main results was that if the eavesdropper is completely unaware of its receive CSI, then a secrecy capacity-achieving (i.e., equivocation-maximizing) strategy is to employ a space-time constellation with a constant spatial inner product.
Parada and Blahut analyzed a degraded single-input multiple-output ($N_T=1, N_R,N_E>1$) wiretap channel in [@Blahut05], and obtained a single-letter characterization of its secrecy capacity via transformation into a scalar Gaussian wiretap channel, and then re-applying (\[eq:degradedGaussianSecCap\]). The authors also proposed a secrecy rate outage metric for the SIMO wiretap channel with slow fading, and observed a secrecy diversity gain of order proportional to the number of receiver antennas. The corresponding MISO case was studied in [@Yates07; @Shafiee07], who noted that the MIMO wiretap channel is not degraded in general. Since this renders a direct computation of (\[eq:Korner1978\]) difficult, they therefore restricted attention to Gaussian input signals. For the special case of $N_T=2, N_R=2, N_E=1$ analyzed by Shafiee and coworkers in [@Shafiee09], a beamforming transmission strategy was shown to be optimal.
The next steps toward understanding the full-fledged MIMO wiretap channel were taken in [@Khisti07]-[@GoelN08], which considered the case of multiple antennas at all nodes and termed it the MIMOME (multiple-input multiple-output multiple-eavesdropper) channel. Khisti *et al*. [@Khisti07] developed a genie-aided upper bound for the MIMO secrecy capacity for which Gaussian inputs are optimal. When the eavesdropper’s instantaneous channel state is known at the transmitter, it was shown that an asymptotically optimal (high SNR) scheme is to apply a transmit precoder based upon the generalized singular value decomposition (GSVD) of the pencil $(\mathbf{H}_b,\mathbf{H}_e)$, which decomposes the system into parallel channels and leads to a closed-form secrecy rate expression. For the so-called MISOME special case where $N_R=1, N_T,N_E>1$, the optimal transmit beamformer is obtained as the generalized eigenvector ${\bm{\psi }}_m$ corresponding to the largest generalized eigenvector $\lambda _m$ of $${\mathbf{h}}_b^H {\mathbf{h}}_b {\bm{\psi }}_m = \lambda _m {\mathbf{H}}_e^H {\mathbf{H}}_e {\bm{\psi }}_m.$$
If only the statistics of $\mathbf{H}_e$ are known to the transmitter, then the authors proposed an *artificial noise* injection strategy as first suggested by Goel and Negi [@Negi05; @GoelN08]. The artificial noise is transmitted in conjunction with the information signal, and is designed to be orthogonal to the intended receiver, such that only the eavesdropper suffers a degradation in channel quality [@Khisti10; @Wornell09].
![The MIMO secrecy rates of GSVD-beamforming [@Ali10; @Khisti10], artificial noise [@GoelN08], and waterfilling over the main channel, $N_T=N_E=3,N_R=2$. []{data-label="fig_MIMOSecrecy"}](comp2){width="\linewidth"}
An example of the secrecy rate performance of various transmission strategies for the MIMO wiretap channel is shown in Fig. \[fig\_MIMOSecrecy\]. The GSVD schemes require instantaneous knowledge of eavesdropper channel $\mathbf{H}_e$, the artificial noise scheme requires the statistics of $\mathbf{H}_e$, and if no information is available regarding $\mathbf{H}_e$, then the relatively poor performance of waterfilling on the main channel is also shown.
The MIMO wiretap channel was studied independently by Oggier and Hassibi [@OggierH08], who computed a similar upper bound on the MIMO secrecy capacity, and showed after a matrix optimization analysis that $$\label{eq:MIMOsecCap}
C_S=\mathop {\max }\limits_{{\mathbf{Q}}_x \succeq 0} \log \det \left( {{\mathbf{I}} + {\mathbf{H}}_b {\mathbf{Q}}_x {\mathbf{H}}_b^H } \right) - \log \det \left( {{\mathbf{I}} + {\mathbf{H}}_e {\mathbf{Q}}_x {\mathbf{H}}_e^H } \right).$$ In [@LiuShitz09], Liu and Shamai reexamined the MIMO wiretap channel with a more general matrix input power-covariance constraint ${\mathbf{Q}}_x \preceq {\mathbf{S}}$, and showed that the conjecture of a Gaussian input $U = X$ without prefix coding is indeed an optimal secrecy capacity-achieving choice.
Zhang *et al*. attempted to bypass the non-convex optimization of the optimal input covariance matrix by drawing connections to a sequence of convex cognitive radio transmission problems, and obtained upper and lower bounds on the MIMO secrecy capacity. Li and Petropulu [@Petropulu10] computed the optimal input covariance matrix for a MISO wiretap channel, and presented a set of equations characterizing the general MIMO solution. Bustin and coauthors [@Bustin09] exploited the fundamental relationship between mean-squared error and mutual information to provide a closed-form expression for the optimal input covariance $\mathbf{Q}_x$ that achieves the MIMO wiretap channel secrecy capacity, again under an input power-covariance constraint. More precisely, it was shown in [@Bustin09] that, under the matrix power constraint $\mathbf{Q}_x\preceq \mathbf{S}$, the solution of (6) is given by $$\label{eq:MIMOsecCapunderS}
C_{sec}(\mathbf{S})=\sum_{i=1}^{\lambda}\log\alpha_i$$ where $\alpha_i$, $i=1$, . . .$\lambda$, are the generalized eigenvalues of the pencil $$\label{eq:r7}
(\mathbf{S}^{\frac{1}{2}}{\mathbf{H}}_b^H{\mathbf{H}}_b\mathbf{S}^{\frac{1}{2}}+\textbf{I},\quad \mathbf{S}^{\frac{1}{2}}{\mathbf{H}}_e^H{\mathbf{H}}_e\mathbf{S}^{\frac{1}{2}}+\textbf{I})$$ that are greater than 1.
Note that, since both elements of the pencil (\[eq:r7\]) are strictly positive definite, all the generalized eigenvalues of the pencil (\[eq:r7\]) have real positive values [@LiuLiu10; @Horn]. In (\[eq:MIMOsecCapunderS\]), a total of $\lambda$ of them are assumed to be greater than 1. Clearly, if there are no such eigenvalues, then the information signal received at the intended receiver is a degraded version of that of the eavesdropper and in this case, the secrecy capacity is zero.
It should be noted that, under the average power constraint $\operatorname{Tr} \left( {{\mathbf{Q}}_x } \right) \leqslant P$, there is not a computable secrecy capacity expression for the general MIMO case. In fact, for the average power constraint, the secrecy capacity is found through an exhaustive search over the set $\{\mathbf{S}: \mathbf{S}\succeq 0, \text{Tr}(\mathbf{S})\leq P\}$. More precisely, we have [@LiuLiu10], [@BCWeingarten06 Lemma 1] $$\begin{aligned}
\label{eq:12}
C_{sec}(P)=\max_{\mathbf{S}\succeq 0, \text{Tr}(\mathbf{S})\leq P}C_{sec}(\mathbf{S})\end{aligned}$$ where, for any given semidefinite $\mathbf{S}$, $C_{sec}(\mathbf{S})$ can be computed as given by (\[eq:MIMOsecCapunderS\]).
Subsequently, numerous research contributions emerged that considered a number of practical issues regarding the MISO/MIMO wiretap channel, of which we enumerate a few below:
- Optimal power allocation methods for the artificial noise strategy were presented in [@Zhou09], and for the GSVD-based precoding scheme in [@Ali10].
- If even statistical information regarding the eavesdropper’s channel is unavailable, then Swindlehurst *et al*. [@Swindlehurst09; @Mukherjee09] suggested an approach where just enough power is allocated to meet a target performance criterion (SNR or rate) at the receiver, and any remaining power is used for broadcasting artificial noise, since the secrecy rate cannot be computed at the transmitter. A compound wiretap channel approach and a resultant universal coding scheme that guarantees a positive secrecy rate is presented in [@He10].
- The effects of imperfect and quantized main channel state information at the transmitter upon the secrecy rate were examined in [@Swindle10] and [@ISIT09], respectively.
- MIMO secrecy capacity has also been studied for frequency-selective [@Debbah08] and ergodic [@Bhargava09] channel fading processes.
A summary of transmission strategies in the MIMO wiretap channel for various assumptions regarding eavesdropper channel state information at the transmitter (ECSIT) is presented in Table \[table\_MIMOstrats\].
----------------------------------------------- --
**Parameters & **Strategy\
MIMOME, no ECSIT & Artificial noise\
MIMOME, statistical ECSIT & Artificial noise\
MISOME, complete ECSIT & GEVD beamforming\
MIMOME, complete ECSIT & GSVD precoding\
****
----------------------------------------------- --
: Comparison of MIMO wiretap transmission strategies for various ECSIT assumptions[]{data-label="table_MIMOstrats"}
Broadcast, Multiple-Access, and Interference Channels {#sec:BC_MAC_IC}
=====================================================
The concept of information-theoretic security is easily extended to larger multi-user networks with more than 2 receivers and/or transmitters. The original wiretap channel as proposed by Wyner [@Wyner75], is a form of broadcast channel (BC) where the source sends confidential messages to the destination while the messages should be kept as secret as possible from the other receiver(s)/ eavesdropper(s). Csiszàr and Körner, extended this work to the case where the source sends common information to both the destination and the eavesdropper, and confidential messages are sent only to the destination [@Csiszar78]. The secrecy capacity region of this scenario, for the case of a BC with parallel independent subchannels is considered in [@Liang08] and the optimal source power allocation that achieve the boundary of the secrecy capacity region is derived. The secrecy capacity region of the MIMO Gaussian broadcast channel with common and confidential messages, is characterized in [@BCLiang09] using a channel enhancement approach [@BCWeingarten06] and under matrix input power-covariance constraint ${\mathbf{Q}}_x \preceq {\mathbf{S}}$. The notion of an enhanced broadcast channel was firstly introduced in [@BC; @Weingarten06] and was used jointly with entropy power inequality to characterize the capacity region of the conventional Gaussian MIMO broadcast channel (without secrecy constraint). But, as we will show, most of the current works in the literature, studying different examples of MIMO broadcast channel with secrecy, use this notion. Moreover, instead of the average total power constraint $\operatorname{Tr} \left( {{\mathbf{Q}}_x } \right) \leqslant P$, they consider the more general matrix input power-covariance constraint ${\mathbf{Q}}_x \preceq {\mathbf{S}}$.
The discrete memoryless broadcast channel with two confidential messages sent to two receivers, where each receiver acts as an eavesdropper for the other one, was studied in [@Yates08], where inner and outer bounds for the secrecy capacity region were established. This problem was studied in [@Liu09] for the multiple-input single-output (MISO) Gaussian case and in [@LiuLiu10] for general MIMO Gaussian case. Rather surprisingly, it was shown in [@LiuLiu10] that, under the matrix input power-covariance constraint, both confidential messages can be simultaneously communicated at their respected maximum secrecy rates, where the achievability is obtained using the dirty-paper coding. To prove this result, Liu et al. revisited the MIMO Gaussian wiretap channel and showed that a coding scheme that uses artificial noise and random binning achieves the secrecy capacity of the MIMO Gaussian wiretap channel as well [@LiuLiu10].
Consider the broadcast channel represented by (3) and (4), but this time the transmitter has independent confidential messages $W_1$ (intended for receiver 1 but needing to be kept secret from receiver 2) and $W_2$ (intended for receiver 2 but needing to be kept secret from receiver 1). From [@LiuLiu10 Corollary 2], under the matrix constraint $\mathbf{S}$, the secrecy capacity region is given by the set of nonnegative rate pairs $(R_1,R_2)$ such that $$\label{eq:MIMOBCregS1}
R_1\leq\sum_{i=1}^{\lambda}\log\alpha_i$$ $$\label{eq:MIMOBCregS2}
R_2\leq\sum_{j=1}^{N_T-\lambda}\log\frac{1}{\beta_j}$$ where $\alpha_i$, $i=1$, . . .$\lambda$, are the generalized eigenvalues of the pencil (\[eq:r7\]) that are bigger than 1, and $\beta_j$ $j=1$, . . .$(N_T-\lambda)$ are the generalized eigenvalues of the pencil (\[eq:r7\]) that are less than or equal to 1.
The secrecy capacity region of the MIMO Gaussian broadcast channels with confidential and common messages, where the transmitter has two independent confidential messages and a common message, is characterized in [@BCLiuISIT10]. The achievability is obtained using secret dirty-paper coding, while the converse is proved by using the notion of channel splitting [@BCLiuISIT10].
Secure broadcasting with more than two receivers are considered in [@Wornell08],[@BCKhandani09], [@BCUlukus09], and [@BCLiuLiu10] (and reference therein). More precisely, there is one transmitter which wants to communicate with several legitimate users in the presence of an external eavesdropper. The secrecy capacity region for the two-legitimate receiver case is characterized by Khandani *et al.* [@BCKhandani09] using the enhanced channels, and for an arbitrary number of legitimate receivers by Ekrem et. al [@BCUlukus09]. Ekrem et. al. use the relationships between the minimum-mean-square-error and the mutual information, and equivalently, the relationships between the Fisher information and the differential entropy to provide the converse proof.
In [@BCLiuLiu10], Liu *et al.* considered the secrecy capacity regions of the degraded vector Gaussian MIMO broadcast channel with layered confidential messages. They presented a vector generalization of Costa’s Entropy Power Inequality to provide their converse proof. The role of artificial noise for jamming unintended receivers in multiuser downlink channels was investigated in [@MukherjeeAllerton09; @ICASSP10].
Other recent works on secure multi user communications investigated the multiple-access channel (MAC) with confidential messages [@MACMaric06], [@LiangIT08], the MAC wiretap channel (MAC-WT) [@Tekin08a], [@Tekin08b], and the cognitive MAC with confidential messages [@MACLiu09]. In [@MACMaric06] and [@LiangIT08], two transmitters communicating with a common receiver try to keep their messages secret from each other. For this scenario, the achievable secrecy rate region, and the capacity region for some special cases, are considered.
In [@Tekin08a], the Gaussian multiple access wire-tap channel (GMAC-WT) is considered where multiple users are transmitting to a base station in the presence an eavesdropper which receives a noisy version of the signal received at the base station (degraded wiretapper). In [@Tekin08a], achievable rate regions were found for different secrecy constraints, and it was shown that the secrecy sum capacity can be achieved using Gaussian inputs and stochastic encoders, where the secrecy sum capacity is given by: In [@Tekin08b], a general, not necessarily degraded, Gaussian MAC-WT is considered, and the optimal transmit power allocation that achieves the maximum secrecy sum-rate is obtained. It is shown in [@Tekin08b] that, a user that is prevented from transmitting based on the obtained power allocation can help increase the secrecy rate for other users by transmitting artificial noise to the eavesdropper.
In [@MACLiu09], Liu et al. consider the fading cognitive multiple-access channel with confidential messages (CMAC-CM), where two users attempt to transmit common information to a destination while user 1 also has confidential information intended for the destination and tries to keep its confidential messages as secret as possible from user 2. The secrecy capacity region of the parallel CMAC-CM is established and the closed-form power allocation that achieves every boundary point of the secrecy capacity region is derived [@MACLiu09]. It should be noted that, all the above works in the field of MAC with confidential messages, assume single antenna nodes.
The interference channel (IFC) refers to the case where multiple communication links are simultaneously active in the same time and frequency slot, and hence potentially interfere with each other. A special application of the IFC with secrecy constraints is addressed in [@Shitz09], where the message from only one of the transmitters is considered confidential. The more general case, where each receiver acts as an eavesdropper for the other transmitter, was studied in [@Yates08] where, in the absence of a common message, the authors imposed a perfect secrecy constraint and obtained inner and outer bounds for the perfect secrecy capacity region.
For multiuser networks, a useful metric that captures the scaling behavior of sum secrecy rate $R_\Sigma$ as the transmit SNR, $\rho$, goes to infinity is the degrees of freedom (DoF), which can be defined as $$\eta \triangleq \mathop {\lim }\limits_{\rho \to \infty } \frac{{R_\Sigma \left( \rho \right)}}
{{\log \left( \rho \right)}}.$$ The number of secure DoF for $K$-user Gaussian IFCs ($K\geq 3$) has been addressed in [@Koylouglu08], [@Koylouglu09], [@IFCYener09], and it was shown that under very strong interference, positive secure DoFs are achievable for each user in the network. More precisely, for the case of $K$-user SISO Gaussian interference channel with confidential messages, where each node has 1 antenna and each transmitter needs to ensure the confidentiality of its message from all non-intended receivers, a secure DoF of $$\label{eq:DoF1}
\eta=\frac{K-2}{2K-2}$$ is almost surely achievable for each user [@Koylouglu09]. The achievablity is obtained by interference alignment and channel extension [@Jafar09]. Moreover, for the case of $K$-user SISO Gaussian interference channel with an external eavesdropper, each user can achieve $$\label{eq:DoF2}
\eta=\frac{K-2}{2K}$$ secure DoF in the ergodic setting.
It should be noted that all of the above references [@Shitz09]-[@IFCYener09] assume single antenna nodes. In fact, to the best of our knowledge, the only works considering the effect of multi-antenna nodes on secrecy in interference channel are [@IFCJorswieck09]-[@IFCAlij10]. In [@IFCJorswieck09], Jorswieck *et al*. study the achievable secrecy rates of a two-user MISO interference channel, where each receiver has single antenna. They model a non-cooperative game in MISO interference channel and obtain the Nash equilibrium point using an iterative algorithm.
In [@IFCAlic10] and [@IFCAlij10], Swindlehurst *et al*. investigate the two-user MIMO Gaussian interference channel with confidential messages, where each node has arbitrary number of antennas. Several cooperative and non-cooperative transmission schemes are described, and their achievable secrecy rate regions are derived. A game-theoretic formulation of the problem is adopted to allow the transmitters to find an operating point that balances network performance and fairness (the so called Kalai-Smorodinsky bargaining solution [@IFCAlij10]). It is shown in [@IFCAlij10] that, while ordinary jamming is near optimal for the standard wiretap channel [@Wornell09], its performance is far from optimal for the interference channel.
Relay Channels and Miscellaneous Networks {#sec:Relay}
=========================================
Relays and Helpers
------------------
The physical layer security in relay networks has drawn much attention recently, as a natural extension to the secure transmission in MIMO networks. The secrecy capacity and achievable secrecy rate bounds have been investigated for various types of relay-eavesdropper channels, and many cooperative strategies, based on the ones that serve for the conventional relay systems, have been proposed.
![Models of trusted and untrusted relay networks[]{data-label="fig_Wyner_Wiretap"}](trusted_relay){width="40.00000%"}
As an extreme case study, the relay itself can be considered to be an *untrusted* user, where the relay node acts both as an eavesdropper and a helper, i.e., the eavesdropper is co-located with the relay node. The source desires to use the relay to communicate with the destination, but at the same time intends to shield the message from the relay. This group of model was first studied in [@Oohama_Capacity07] for the general relay channel. Coding problems of the relay-wiretap channel are studied under the assumption that some of transmitted messages are confidential to the relay, and deterministic and stochastic rate regions are explicitly derived in [@He_Cooperation10; @He_Equivocation07; @He_role08], which show that the cooperation from the untrusted relay is still essential for achieving a non-zero secrecy rate. In [@He_Cooperation10], for the general untrusted relay channel which can be described as $p(Y,Y_r|X,X_r)$ with $X,X_r$ being the input from the source and the relay, respectively, and $Y,Y_r$ being the signals received by the relay and destination, respectively, the following achievable region of rate pairs $(R_1,R_e)$ is derived: $$\bigcup \left \{
\begin{array}{l}
R_e \le R_1 < I(X;Y,\hat{Y}_r|X_r) \\
0 \le R_e < [I(X;Y,\hat{Y}_r|X_r)-I(X;Y_r|X_r)]^+
\end{array}
\right \}.$$ Based on this region, the cooperation of an untrusted relay node is found to be beneficial for a specific model where there is an orthogonal link in the second hop. A more symmetric case is discussed in [@Ekrem_Secrecy08], where both the source and the relay send their own private messages while keeping them secret from the destination. Assuming a half-duplex amplify-and-forward protocol, another effective countermeasure in this case is to have the destination jam the relay while it is receiving data from the source. This intentional interference can then be subtracted out by the destination from the signal it receives via the relay.
However, unlike the aforementioned case, in a *trusted* relay scenario, the eavesdroppers and relays are separate network entities. The relays can play various roles with external eavesdroppers. They may act purely as traditional relays while utilizing help from other nodes to ensure security; they may also act as both relaying components as well as cooperative jamming partners to enhance the secure transmission; or they can assume the role of stand-alone *helpers* to facilitate jamming unintended receivers.
Helpers serve as friendly jammers that do not have any information of their own to transmit, but instead cooperate with authorized nodes to degrade the signals intercepted by eavesdroppers. For example, from an information-theoretic viewpoint, a helper can send a random codeword at a rate that ensures that it can be decoded and subtracted from the received signal by the intended receiver, but cannot be decoded by the eavesdropper. Alternatively, a helper can transmit a jamming signal that interferes with the ability of the eavesdropper to intercept and decode the desired signal. In both cases, there is either no or minimal impact on the mutual information of the desired link, but that of the eavesdropper’s link is reduced, and hence the secrecy rate is improved. For example, in a single-antenna wiretap channel with external helpers, an interesting approach is to split the transmission time into two phases. In the first phase, the transmitter and the intended receiver both transmit independent artificial noise signals to the helper nodes. The helper nodes and the eavesdropper receive different weighted versions of these two signals. In the second stage, the helper nodes simply replay a weighted version of the received signal, using a publicly available sequence of weights. At the same time, the transmitter transmits its secret message, while also canceling the artificial noise at the intended receiver [@GoelN08].
A typical model of a relay channel with an external eavesdropper is investigated in [@Lai_Relay--Eavesdropper08], where the four-terminal network is introduced and an outer-bound on the optimal rate-equivocation region is derived. Specifically, for the traditional decode-and-forward (DF) relaying strategy, a achievable perfect secrecy rate is derived as $$\begin{gathered}
R_s^{(DF)} = \sup_{p(u)p(v_1,v_2|u)p(x_1,x_2|v_1,v_2)} [\min\{I(V_1,V_2;Y|U),\\
I(V_1;Y_1|V_2,U)\}-I(V_1,V_2;Y_2|U)]^+\end{gathered}$$ for some random variables $U \rightarrow (V_1,V_2) \rightarrow (X_1,X_2) \rightarrow (Y,Y_1,Y_2)$ where $X_1,X_2$ are the channel inputs from the source and the relay respectively, while $Y,Y_1,Y_2$ are the channel outputs at the destination, relay and eavesdropper respectively. Lai et al. also propose a noise-forwarding strategy where the full-duplex relay sends codewords independent with the secrete message to confuse the eavesdropper. In [@Dong_Improving10], several cooperative schemes are proposed for a two-hop multiple-relay network, and the corresponding relay weights are derived aiming to maximize the achievable secrecy rate, under the constraint that the link between the source and the relay is not protected from eavesdropping.
![\[fig:sys\_mod\] Two-hop MIMO relay network with external eavesdropper.](figure_sys_mod2.eps){width="\linewidth"}
In [@Tang_Interference-assisted08], a wiretap channel with an independent helping jammer is considered. The interferer can send a random codeword at a rate that ensures that it can be decoded and subtracted from the received signal by the intended receiver but cannot be decoded by the eavesdropper. [@Wang_Cooperative09] consider the cooperative artificial interference approach for MIMO ad hoc networks. The model therein can also be regarded as the external helper category, since when one pair of nodes are communicating with each other, all the nodes surrounding the legitimate receiver cooperate to interfere with the eavesdropper by sending jamming signals. A general model is given by $$\begin{aligned}
\mathbf{y} &= \mathbf{H}_{B,0} \mathbf{x} + \Sigma_{i=1}^N \mathbf{H}_{B,i} \mathbf{q}_i + \mathbf{n}_B \\
\mathbf{z} &= \mathbf{H}_{E,0} \mathbf{x} + \Sigma_{i=1}^N \mathbf{H}_{E,i} \mathbf{q}_i + \mathbf{n}_E\end{aligned}$$ where $\mathbf{H}_{B,0}$ and $\mathbf{H}_{E,0}$ are the channels from the source to the destination and eavesdropper respectively and $\mathbf{H}_{B,i}$ and $\mathbf{H}_{E,i}$ are the channels from jammer $i$ to the destination and eavesdropper respectively. $\mathbf{x}$ is the information signals and $\mathbf{q}_i$ is the artificial interference, and $\mathbf{y}$ and $\mathbf{z}$ are received signals at the destination and eavesdropper respectively. For the proposed coordinated cooperative jamming scheme in [@Wang_Cooperative09], orthogonal information subspace and jamming subspace are broadcast across the network and $\mathbf{q}_i$ is chosen to lie in the publicized jamming subspace such that there will be no interference at the destination when an appropriate receive beamformer is used. An uncoordinated cooperative jamming strategy is also proposed for the case where the public jamming subspace is unavailable. In this case, $\mathbf{q}_i$ is simply the right singular vector of $\mathbf{H}_{B,i}$ corresponding to the smallest singular value. Both schemes have been shown to efficiently increase the secrecy capacity, even if the eavesdropper has knowledge of the associated subspaces.
A more general case where cooperative jamming strategies guarantee secure communication in both hops using without the need for external helpers is studied in [@Huang_Secure10]. In these approaches, the normally inactive node in the relay network can be used as cooperative jamming sources to confuse the eavesdropper and provide better performance in terms of secrecy rate. In the proposed cooperative jamming strategies, the source and the destination nodes act as *temporary helpers* to transmit jamming signals during transmission phases in which they are normally inactive. We define two types of cooperative jamming schemes, *full cooperative jamming* (FCJ) and *partial cooperative jamming* (PCJ), depending on whether or not both the transmitter and the temporary helper transmit jamming signals at the same time.
![\[fig:Jack\] Secrecy rate vs. transmit power, ECSIT unknown, $N_a=N_b=N_e=N_r=4,d_{ij}=800\text{m}$, $R_t=2$bps/Hz.](figure_multipleR_AN_Ne4){width="\linewidth"}
In [@Tekin08b], a two-way wiretap channel is considered, in which both the source and receiver transmit information over the channel to each other in the presence of a wiretapper. Achievable rates for the two-way Gaussian channel are derived. Besides, a cooperative jamming scheme that utilizing the potential jammers is shown to be able to further increase the secrecy sum rate. [@He_role08a] shows that using feedback for encoding is essential in Gaussian full-duplex two-way wiretap channels, while feedback can be ignored in the Gaussian half-duplex two-way relay channel with untrusted relays. More recently, secure transmission strategies are studied for multi-antenna two-way relay channel with network coding with the presence of eavesdroppers by [@Mukherjee_Securing10]-[@DebbahGC10]. By applying the analog network-coded relaying protocol, the end nodes exchange messages in two time slots. In this scenario, the eavesdropper has a significant advantage since it obtains two observations of the transmitted data compared to a single observation at each of the end nodes. As a countermeasure, in each of the two communication phases the transmitting nodes jam the eavesdropper, either by optimally using any available spatial degrees of freedom, or with the aid of external helpers.
Cognitive Radio and Sensor Networks
-----------------------------------
As a promising technique to alleviate spectrum scarcity, cognitive radio (CR) [@Mitola_Cognitive99] is capable of dynamically sensing and locating unused spectrum segments in a target spectrum pool and communicating using the unused spectrum segments in ways that cause no harmful interference to the primary users of the spectrum. Due to the vulnerability of physical layer spectrum sensing of CR, research attention on physical layer security issues, though limited, has emerged recently. In [@Clancy_Security08], several classes of physical layer attacks for dynamic spectrum access and adaptive radio scenarios are described, and corresponding mitigation techniques to these attacks are proposed. In [@Brown_Potential07], the denial-of-service vulnerabilities, from the perspectives of the network architecture employed, the spectrum access technique used and the spectrum awareness model, are examined and possible remedies are provided.
A so-called *primary user emulation* threat to spectrum sensing is identified in [@Chen_Defense08], and a transmitter verification scheme is proposed to verifies whether a given signal is that of an incumbent transmitter by estimating its location and observing its signal characteristics. In [@Wang_PHY-layer10], the security problem of collaborative sensing for spectrum occupation is formulated as $$Q_f = \left\{ {\begin{array}{*{20}c}
{\sum\nolimits_{i = q - k}^{n - k} {\left( {\begin{array}{*{20}c}
{n - k} \\
i \\
\end{array} } \right)P_f^i \left( {1 - P_f } \right)^{\left( {n - k - i} \right)} } } & {{\text{if }}k < q} \\
1 & {{\text{if }}k \geqslant q} \\
\end{array} } \right.$$ where $Q_f$ and $P_f$ are the overall and individual false alarm probabilities respectively. In the sensing network, $n$ users are collaboratively reporting sensing results and $k$ of which are malicious users who deliver false reports. The base station makes the final decision based on certain counting rules related to $q$. The work provides a numerical algorithm to optimize $q$ such that $Q_f$ is minimized. Thus the issue of improving efficiency of spectrum access on a non-interfering basis is formulated as a constrained parameter optimization problem, which provides a better understanding on the effectiveness of the attacks and their countermeasures.
Wireless sensor networks and corresponding distributed estimation algorithms have been at the forefront of signal processing research in the past decade. The downlink and uplink phases of communication between the sensors and a fusion center (FC) are inherently vulnerable to eavesdropping. Li, Chen, and Ratazzi [@RatazziSensor05] tackle downlink secrecy where the FC has multiple antennas by deliberately inducing rapid time-varying fluctuations in the eavesdropper’s channel. [@Ives05] proposed the use of artificial noise-like schemes on the uplink to ‘confuse’ eavesdroppers about the aggregate sensor observations sent to the FC. Kundur *et al.* examine cross-layer secrecy-preserving design methodologies for multimedia sensor networks in [@Kundur08].
Wireless Secret Key Agreement {#sec:Key}
=============================
We recall that the original secure communication system studied by Shannon was based on secret-key encryption. Shannon’s result that perfect secrecy required encryption with a random one-time pad cipher at least as long as the message was widely regarded as a pessimistic result, until it was reexamined in the context of noisy channels by Maurer [@Maurer93]. In his seminal work, Maurer decried Wyner’s degraded wiretap channel as being too unrealistic, and instead proposed a secret-key agreement protocol that could be implemented over a publicly observable two-way channel in the presence of a passive eavesdropper.
The key elements of Maurer’s strategy are the *information reconciliation* and *privacy amplification* procedures. The information reconciliation phase is aimed at generating an identical random sequence between the two terminals by exploiting the public discussion channel. The privacy amplification stage extracts a secret key from the identical random sequence agreed to by two terminals in the preceding information reconciliation phase. Less formally, after public discussion based on *correlated randomness* in the first stage, privacy amplification reduces a initial piece of random nature into a smaller entity (e.g., by linear mapping and universal hashing) which is known only by the legitimate users, even if the eavesdropper has a less noisy channel in certain cases.
![Secret key agreement by public discussion.[]{data-label="fig_Maurer"}](Maurer){width="2.02in"}
More precisely, it was assumed that the transmitter, receiver and adversary have access to repeated independent realizations of random variables $X, Y,$ and $Z$, respectively, with some joint probability distribution $P_{X,Y,Z}$ as in Fig. \[fig\_Maurer\]. The eavesdropper is completely ignorant of $X$ and $Y$. The secret-key rate $S(X; Y|| Z)$ was then defined as the maximal rate at which Alice and Bob can generate a secret key by communication over the noiseless and authentic but otherwise insecure channel in such a way that the opponent obtains information about this key only at an arbitrarily small rate (cf. (\[eq:Korner1978\])). The following upper and lower bounds on the secret key rate were presented: $$S\left( {X;Y|| Z} \right) \leqslant \min \left[ {I\left( {X;Y} \right),I\left( {X;Y|Z} \right)} \right],$$ $$S\left( {X;Y|| Z} \right) \geqslant \max \left[ {I\left( {X;Y} \right) - I\left( {X;Z} \right),I\left( {Y;X} \right) - I\left( {Y;Z} \right)} \right].$$
Closely related results were offered in the concurrent work by Ahlswede and Csizar [@Ahlswede93]. Csiszar and Narayan studied the augmentation of key-based secrecy capacity with the aid of a helper which supplies additional correlated information in [@Csizar00], and obtained a single-letter characterization of the key-based secrecy capacities with an arbitrary number of terminals in [@Csiszar04]. Maurer and Wolf subsequently extended the secret-key sharing analysis of [@Maurer93] to account for the presence of an active eavesdropper in [@Maurer03I]-[@Maurer03III], and showed that either a secret key can be generated at the same rate as in the passive-adversary case, or such secret-key agreement is infeasible.
The next evolution in secret-key sharing was the exploitation of the common randomness inherent in reciprocal wireless communication channels. Koorapaty *et al.* relied on the independence of the channels between transmitter/receiver and transmitter/eavesdropper to use the phase of the fading coefficients as a secret key [@Hassan00]. Other techniques include key generation via
- discretizing extracted coefficients of the multipath components [@Shah07],
- quantizing the channel phases for a multitone communication system such that multiple independent phases are used to generate longer keys [@Sayeed08],
- directly quantizing the complex channel coefficients [@Shah06],
- a purposely constructed random variable whose realizations are communicated between the legitimate nodes, with secrecy achieved when the eavesdropper lacks channel state information [@Mclaughlin08],
- exploiting the level crossing rates of the fading processes at the legitimate terminals [@Trappe10],
- utilizing appropriately timed one-bit feedback available in practical networks due to Automatic Repeat reQuest (ARQ) protocols [@Abdallah10].
Unsurprisingly, multiple-antenna channels have attracted considerable attention for their capabilities of increasing common randomness at the legitimate users. Li and Ratazzi [@Ratazzi05] design a MIMO precoder based on knowledge of the main channel that renders difficult blind channel estimation by the eavesdropper. Chen and Jensen developed practical key generation protocols for MIMO systems with temporally and spatially correlated channel coefficients in [@Chen09; @Jensen10]. One of the first experimental measurement campaigns on secret key generation in reciprocal MIMO channels was presented by Wallace and Sharma in [@Sharma10].
The role of a feedback channel in improving the secrecy rate of a wiretap channel has been revisited in recent work. In [@Lai08], the authors show that a noisy feedback channel that is observable by all parties can still be utilized to generate a secrecy rate equal to the main channel capacity, since the feedback from the (either full- or half-duplex) receiver can be used to jam the eavesdropper (assuming a modulo-additive channel model). Javidi *et al.* [@Javidi09] consider a secure but rate-limited feedback channel, and prove that it is optimal for the receiver to feedback a random secret key that is independent of its received channel output symbols.
Related Topics {#sec:RelatedTopics}
==============
Code Design for Secrecy
-----------------------
Once the groundwork had been laid for the limits of information-theoretic security, several researchers turned their attention to the development of practical channel codes for secrecy. Wyner [@Wyner75] and Csiszàr and Körner [@Csiszar78] had used a stochastic coding argument to provide a non-constructive proof of the existence of channel codes that guarantee both robustness to transmission errors and a prescribed degree of data confidentiality as the block length tends to infinity.
In Wyner’s stochastic encoding scheme, a mother codebook $C_0(n)$ of length n is randomly partitioned into “secret bins” or subcodes $\left\{C_1(n),C_2(n),\ldots,C_M(n)\right\}$. A message $w$ is associated with a sub-code $C_w(n)$ and the transmitted codeword is randomly selected within the sub-code. The mother code $C_0(n)$ provides enough redundancy so that the legitimate receiver can decode the message reliably, whereas each sub-code is sufficiently large and, hence, introduces enough randomness so that the eavesdropper’s uncertainty about the transmitted message can be guaranteed. However, the development of practical wiretap codes for general wiretap channels was not as rapid as that of classical error-correction codes, and several open problems remain till date.
Therefore, it was natural to turn to capacity-achieving channel codes and examine their applications for secrecy. In [@Thangaraj07], Thangaraj *et al.* advanced the idea of using graph-based codes such as low density parity check (LDPC) codes for binary erasure wiretap channels (noiseless main channel), and showed that both reliability and Wyner’s weak secrecy criterion could be satisfied simultaneously. Bloch and coauthors [@Mclaughlin08] adopted LDPC codes and multi-level coding for the information reconciliation phase of a practical secret key agreement protocol. For Gaussian wiretap channels, appropriately punctured LDPC codes were employed with the relative bit error rate at the receiver and eavesdropper as as a proxy security metric in [@Klinc09], where the authors showed that a ‘security gap’ was achievable. A turbo code-based scheme with the puncturing pattern determined by a pre-shared secret key was presented in [@ArefIEE].
Graph-based unstructured codes are not the only viable approach for wiretap coding. He and Yener [@He09StructCodes] show that an arbitrarily large secrecy rate is achievable for Gaussian wiretap channels with an external helper using structured integer and nested lattice codes. Nested lattice codes were also deployed over the binary symmetric wiretap channel in [@Liu08]. Arora and Sang presented the notion of dialog codes wherein the receiver aids the transmitter by jamming the eavesdropper while still being able to recover the transmitted symbol [@Arora09]. If the receiver is half-duplex, then this can be achieved using a rate-1/2 code with memory where the receiver jams either of the code bits but is able to recover the message from the remaining bit, whereas the equivocation at the eavesdropper is unity. The recently proposed polar coding scheme has been shown to achieve the secrecy capacity for binary symmetric wiretap channels [@Vardy].
While the emerging area of network coding is not directly related to traditional channel coding design, we briefly mention physical-layer security issues encountered in this field. Network coding is an paradigm for wireline and wireless networks that allows intermediate nodes to mix signals received from multiple paths, with the objective of improving throughput. Therefore, such networks are vulnerable to eavesdropping, as all other networks discussed thus far in this work.
The secure network coding problem was introduced in [@Cai02] for multicast wireline networks where each link has equal capacity, and a wiretapper can observe an unknown set of up to $k$ network links. For this scenario, the secrecy capacity is given by the cut set bound and is achieved by injecting $k$ random keys at the source which are decoded at the sink along with the message [@Cai02; @Stein04]. Silva and Kschischang [@Kschischang08] among others have drawn connections between the multicast problem and the type-II wiretap channel studied by Ozarow and Wyner, as described in Section \[subsec:SISOWiretap\]. Eavesdropping countermeasures for wireless network coding systems are described in [@Mukherjee_Securing10; @Zhang09], among others.
Cross-Disciplinary Tools
------------------------
The interactions between various agents (transmitters, receivers, helpers, and attackers) in multiuser wireless networks is accurately captured by inter-disciplinary analyses based on game theory and microeconomics, and this holds true for problems of secrecy as well. Cooperative game theory was applied in [@Saad09] to demonstrate the improvement in secrecy capacity of an ad hoc network, when users form coalitions to null the signals overheard by eavesdroppers via collaborative beamforming. Han *et al.* [@Debbah09] developed a two-stage Stackelberg game where a transmitter ‘pays’ a number of external helpers to jam an eavesdropper, and computed the corresponding equilibrium prices and convergence properties. The same authors examined a similar scenario in [@Marina09], where an auction game was used instead to model the transactions between transmitters and helping jammers. Anand and Chandramouli studied a $M$-user non-cooperative power control game with secrecy considerations in [@Anand], and applied pricing functions to improve the energy efficiency and sum secrecy capacity of the network. For the 2-user MIMO-IC with confidential messages, the so called Kalai-Smorodinsky bargaining solution is adopted to allow the transmitters to find an operating point that balances network performance and fairness [@IFCAlic10].
Utilizing secrecy rate as the payoff in a game-theoretic formulation is a relatively new concept. Yuksel, Liu, and Erkip studied a SISO wiretap network with an adversarial jammer helping the eavesdropper as a zero-sum game, and presented the Nash Equilibrium input and jammer cumulative distribution functions [@Erkip09]. Mukherjee and Swindlehurst posed the MIMO wiretap channel with an active eavesdropper as a zero-sum dynamic (sequential) game, and examined the equilibrium transmit/wiretapper strategies for games with and without perfect information [@MukherjeeICC].
CONCLUSION {#sec:concl}
==========
This paper provided a comprehensive survey of the field of physical-layer security in wireless networks based on information-theoretic principles. We commenced with an overview of the foundations dating back to Shannon’s pioneering work on information-theoretic security. We then describe the evolution of secure transmission strategies from point-to-point channels to multiple-antenna systems, followed by generalizations to multiuser networks. We also evaluate secret-key establishment protocols based on physical layer mechanisms, along with an overview of practical secrecy-preserving code design and inter-disciplinary models for security. Several recent monographs that provide a more rigorous and in-depth introduction to the topic of information-theoretic security are [@LiangBook09; @TrappeBook; @JorsweickBook10].
[200]{} W. E. Stark and R. J. McEliece, “On the capacity of channels with block memory," *IEEE Trans. Inf. Theory*, vol. 34, no. 3, pp. 322-324, Mar. 1988. M. Medard, “Capacity of correlated jamming channels," in *Proc. 35th Allerton Conf.*, pp. 1043-1052, 1997. A. Kashyap, T. Basar, and R. Srikant, “Correlated jamming on MIMO gaussian fading channels," *IEEE Trans. Inf. Theory*, vol. 50, no. 9, pp. 2119-2123, Sep. 2004. J. L. Massey, “An introduction to contemporary cryptology," *Proc. IEEE*, vol. 76, no. 5, pp. 533-549, May 1988. V. I. Korzhik and V. A. Yakovlev, “Nonasymptotic estimates for efficiency of code jamming in a wire-tap channel," *Problemy Peredachi Informatsii (USSR)*, vol. 17, no. 4, pp. 11-18, 1981. G. S. Vernam, “Cipher printing telegraph systems for secret wire and radio telegraphic communications," *Trans. American Institute Electrical Engineers*, vol. XLV, pp. 295-301, 1926. C. E. Shannon, “Communication theory of secrecy systems," *Bell Sys. Tech. Journ.*, vol. 28, pp. 656-715, 1949. P. R. Geffe, “Secrecy systems approximating perfect and ideal secrecy," *Proc. of the IEEE*, vol. 53 , no. 9, pp. 1229-1230, 1965. A. D. Wyner, “The wire-tap channel," *Bell Sys. Tech. Journ.*, vol. 54, pp. 1355-1387, 1975. A. B. Carleial and M. Hellman, “A note on Wyner’s wiretap channel," *IEEE Trans. Inf. Theory*, vol. 23, no. 5, pp. 625-627, May 1977. I. Csiszàr and J. Körner, “Broadcast channels with confidential messages," *IEEE Trans. Inf. Theory*, vol. 24, no. 3, pp. 339-348, May 1978. S. L. Y. Cheong and M. Hellman, “The Gaussian wire-tap channel," *IEEE Trans. Inf. Theory*, vol. 24, no. 4, pp. 451-456, July 1978. L. H. Ozarow and A. D. Wyner, “Wire-tap channel II," in *Proc. Eurocrypt, Workshop on Advances in Cryptology*, pp. 33–51, Paris, 1985.
U. Maurer, “Secret key agreement by public discussion from common information," *IEEE Trans. Inf. Theory*, vol. 39, no. 3, pp. 733-742, May 1993. C. Mitrpant, A. J. H Vinck, and Y. Luo, “An achievable region for the Gaussian wiretap channel with side information," *IEEE Trans. Inf. Theory*, vol. 52, no. 5, pp. 2181-2190, May 2006. J. Barros and M. R. D. Rodrigues, “Secrecy capacity of wireless channels," in *Proc. IEEE ISIT*, Seattle, July 2006. M. Bloch, J. Barros, M. R. D. Rodrigues, and S. W. McLaughlin, “An opportunistic physical-layer approach to secure wireless communications," *Proc. Allerton Conf*., Monticello, Sep. 2006. Z. Li, R. Yates, and W. Trappe, “Secret communication with a fading eavesdropper channel," in *Proc. IEEE ISIT*, Nice, July 2007. P. Gopala, L. Lai, and H. El Gamal, “On the secrecy capacity of fading channels", *IEEE Trans. Inf. Theory*, vol. 54, no. 10, pp. 4687-4698, Oct. 2008. Y. Liang, G. Kramer, H. V. Poor, and S. Shamai, “Compound wiretap channels," *EURASIP Journ. Wireless Commun. Network.*, 2009.
A. Hero, “Secure space-time communication," *IEEE Trans. Inf. Theory*, vol. 49, no. 12, pp. 3235-3249, Dec. 2003. P. Parada and R. Blahut, “Secrecy capacity of SIMO and slow fading channels," in *Proc. IEEE ISIT*, Adelaide, 2005. Z. Li, W. Trappe, and R. Yates, “Secret communication via multi-antenna transmission", in *Proc. CISS*, Mar. 2007. S. Shafiee and S. Ulukus, “Achievable rates in Gaussian MISO channels with secrecy constraints," in *Proc. IEEE ISIT*, Nice, France, June 2007. S. Shafiee, N. Liu, and S. Ulukus, “Towards the secrecy capacity of the Gaussian MIMO wire-tap channel: The 2-2-1 channel," *IEEE Trans. Inf. Theory*, vol. 55, no. 9, pp. 4033-4039, Sep. 2009. A. Khisti, G. Wornell, A. Wiesel, and Y. Eldar, “On the Gaussian MIMO wiretap channel," in *Proc. IEEE Int’l Symp. on Inf. Theory*, pp. 2471-2475, June 2007. F. Oggier and B. Hassibi, “The secrecy capacity of the MIMO wiretap channel," in *Proc. IEEE Int’l Symp. on Inf. Theory*, pp. 524-528, July 2008. R. Negi and S. Goel, “Secret communication using artificial noise," in *Proc. IEEE Veh. Tech. Conf.*, vol. 3, pp. 1906-1910, Dallas, Sept. 2005. S. Goel and R. Negi, “Guaranteeing secrecy using artificial noise," *IEEE Trans. Wireless Commun*., vol. 7, no. 6, pp. 2180-2189, June 2008. T. Liu and S. Shamai, “A note on the secrecy capacity of the multiple-antenna wiretap channel," *IEEE Trans. Inf. Theory*, vol. 55, no. 6, pp. 2547-2553, June 2009. L. Zhang, R. Zhang, Y. Liang, Y. Xin, and S. Cui, “On the relationship between the multi-antenna secrecy communications and cognitive radio communications," *IEEE Trans. Commun*. vol. 58, no. 6, pp. 1877-1886, June 2010. J. Li and A. Petropulu, “Transmitter optimization for achieving secrecy capacity in Gaussian MIMO wiretap channels,” submitted to *IEEE Trans. Inform. Theory*, 2010 \[Online\]. Available: http://arxiv.org/abs/0909.2622v1. R. Bustin, R. Liu, H. V. Poor, and S. Shamai, “An MMSE approach to the secrecy capacity of the MIMO Gaussian wiretap channel," *EURASIP Journ. Wireless Commun. Network.*, 2009. A. Khisti and G. Wornell, “Secure transmission with multiple antennas I: the MISOME wiretap channel", *IEEE Trans. Inf. Theory*, vol. 56, no. 7, pp. 3088-3104, July 2010. A. Khisti and G. Wornell, “Secure transmission with multiple antennas II: the MIMOME wiretap channel", to appear, *IEEE Trans. Inf. Theory*, 2010. \[Online\]. Available: http://www.rle.mit.edu/sia/. R. Liu, T. Liu, H. V. Poor, and S. Shamai, “Multiple-input multiple-output Gaussian broadcast channels with confidential messages," *IEEE Trans. Inf. Theory*, to appear, 2010. A. L. Swindlehurst, “Fixed SINR solutions for the MIMO wiretap channel," in *Proc. IEEE ICASSP*, pp. 2437-2440, Taipei, Apr. 2009. A. Mukherjee and A. L. Swindlehurst, “Fixed-rate power allocation strategies for enhanced secrecy in MIMO wiretap channels," in *Proc. IEEE SPAWC*, pp. 344-348, Perugia, June 2009. X. Zhou and M. R. McKay, “Physical layer security with artificial noise: Secrecy capacity and optimal power allocation", in *Proc. Int. Conf. on Sig. Proc. and Commun. Syst*., Omaha, NE, Sept. 2009. S. A. Fakoorian and A. L. Swindlehurst, “Optimal power allocation for GSVD-based beamforming in the MIMO wiretap channel," 2010 \[Online\]. Available: http://arxiv.org/abs/1006.1890. A. Mukherjee and A. L. Swindlehurst, “Robust beamforming for secrecy in MIMO wiretap channels with imperfect CSI," *IEEE Trans. Signal Proc*., Oct. 2010. Y.-L. Liang, Y. Wang, T. Chang, Y.-W. P. Hong, and C. Chi, “On the impact of quantized channel feedback in guaranteeing secrecy with artificial noise ," in *Proc. IEEE ISIT*, pp. 2351-2355, Seoul, 2009. X. He and A. Yener, “MIMO wiretap channels with arbitrarily varying eavesdropper channel states," 2010 \[Online\]. Available: http://arxiv.org/abs/1007.4801 M Kobayashi, M Debbah, “On the secrecy capacity of frequency-selective fading channels: A practical Vandermonde precoding," in Proc. *IEEE 19th PIMRC*, 2008. Z. Rezki, F. Gagnon, and V. Bhargava, “The ergodic capacity of the MIMO wire-tap channel," \[Online\]. Available: http://arxiv.org/abs/0902.0189v1, Feb. 2009.
Y. Liang, H. V. Poor, and S. Shamai, “Secure communication over fading channels," *IEEE Trans. Inf. Theory*, vol. 54, no. 6, pp. 32470-2492, June 2008. H. D. Ly, T. Liu, and Y. Liang, “Multiple-input multiple-output Gaussian broadcast channels with common and confidential messages," *IEEE Trans. Inf. Theory*, submitted July 2009. H.Weingarten, Y. Steinberg, and S. Shamai, “The capacity region of the Gaussian multiple-input multiple-output broadcast channel," *IEEE Trans. Inf. Theory*, vol. 52, no. 9, pp. 3936-3964, Sep. 2006. R. Liu, I. Maric, P. Spasojevic, and R. D. Yates, “Discrete memoryless interference and broadcast channels with confidential messages: Secrecy rate regions," *IEEE Trans. Inf. Theory*, vol. 54, no. 6, pp. 2493-2512, June 2008. R. Liu and H. V. Poor, “Secrecy capacity region of a multiple-antenna Gaussian broadcast channel with confidential messages," *IEEE Trans. Inf. Theory*, vol. 55, no. 3, pp. 1235-1249, Mar. 2009. R. Liu, T. Liu, H. V. Poor, and S. Shamai, “Multiple-input multiple-output Gaussian broadcast channels with confidential messages," *IEEE Trans. Inf. Theory*, to appear, 2010. R. Liu, T. Liu, H. V. Poor, and S. Shamai (Shitz), “MIMO Gaussian broadcast channels with confidential and common messages”, in *Proc. IEEE Int. Symp. Information Theory*, Texas, U.S.A., June 2010, pp. 2578-2582. A. Khisti, A. Tchamkerten, and G. Wornell, “Secure broadcasting over fading channels," *IEEE Trans. Inf. Theory*, vol. 54, no. 6, pp. 2453-2469, June 2008. G. Bagherikaram, A. S. Motahari and A. K. Khandani, “The secrecy capacity region of the Gaussian MIMO broadcast channel,” submitted to *IEEE Trans. Inform. Theory*, March 2009, available at http://arxiv.org/PScache/arxiv/pdf/0903/0903.3261v2.pdf E. Ekrem and S.Ulukus, “The secrecy capacity region of the Gaussian MIMO multi-receiver wiretap channel,” submitted to *IEEE Trans. Inform. Theory*, March 2009, available at http://arxiv.org/PS cache/arxiv/pdf/0903/0903.3096v1.pdf. R. Liu, T. Liu, H. V. Poor, and S. Shamai, “A vector generalization of Costa’s entropy-power inequality with applications,”*IEEE Trans. Inf. Theory*, vol. 56, no. 4, pp. 1865-1879, Apr. 2010. A. Mukherjee and A. L. Swindlehurst, “Utility of beamforming strategies for secrecy in multiuser MIMO wiretap channels," in *Proc. of Forty-Seventh Allerton Conf.*, Oct. 2009. W. Liao, T. Chang, W. Ma, and C. Chi, “Joint transmit beamforming and artificial noise design for QoS discrimination in wireless downlink," in *Proc. IEEE ICASSP*, pp. 256-2565, Dallas, Mar. 2010. R. Liu, I. Maric, R. D. Yates, and P. Spasojevic, “The discrete memoryless multiple access channel with confidential messages,” in *Proc. IEEE Int. Symp. Inf. Theory*, Seattle, WA, Jul. 9-14, 2006. Y. Liang and H. V. Poor, “Multiple-access channels with confidential messages," *IEEE Trans. Inf. Theory*, vol. 54, no. 3, pp. 976-1002, Mar. 2008. E. Tekin and A. Yener, “The Gaussian multiple access wire-tap channel," *IEEE Trans. Inf. Theory*, vol. 54, no. 12, pp. 5747-5755, Dec. 2008. E. Tekin and A. Yener, “The general Gaussian multiple access and two-way wire-tap channels: Achievable rates and cooperative jamming,” *IEEE Trans. Inf. Theory*, vol. 54, no. 6, pp. 2735 - 2751, June 2008. R. Liu, Y. Liang and H. V. Poor, “Fading cognitive multiple-access channels with confidential messages,” submitted to *IEEE Trans. Inform. Theory*, Dec. 2009, available at http://arxiv.org/PS-cache/arxiv/pdf/0910/0910.4613v2.pdf Y. Liang, A. Somekh-Baruch, H. V. Poor, S. Shamai, and S. Verdu, “Capacity of cognitive interference channels with and without secrecy," *IEEE Trans. Inf. Theory*, vol. 55, no. 2, pp. 604-619, Feb. 2009. O. Koyluoglu, H. El Gamal, L. Lai, and H. V. Poor, “On the secure degrees of freedom in the $K$-user Gaussian interference channel," in *Proc. IEEE ISIT*, pp. 384-388, July 2008. O. O. Koyluoglu, H. El Gamal, L. Lai, and H. V. Poor, “Interference alignment for secrecy," submitted to *IEEE Trans. Inf. Theory*, 2008 \[Online\]. Available: http://arxiv.org/pdf/0810.1187. X. He and A. Yener, “K-user interference channels: achievable secrecy Rate and degrees of freedom,” *ITW*, Greece, June 2009. V. R. Cadambe and S. A. Jafar, “Interference Alignment and Degrees of Freedom of the K-User Interference Channel,” *IEEE Trans. Inf. Theory*, vol. 54, no. 8, pp. 3425-3441, Aug. 2008. E. A. Jorswieck and R. Mochaoura, “Secrecy rate region of MISO interference channel: pareto boundary and non-cooperative games,” in *Proc. WSA*, 2009. S. A. A. Fakoorian and A. L. Swindlehurst, “MIMO interference channel with confidential messages: game theoretic beamforming designs,” in *Proc. Asilomar Conf. on Signals, Systems, and Computers*, Nov. 2010. S. A. A. Fakoorian and A. L. Swindlehurst, “MIMO interference channel with confidential messages: achievable secrecy rates and beamforming design,” submitted to *IEEE Trans. on Inf. Forensics and Security*, Aug. 2010.
Y. Oohama, “[C]{}apacity [t]{}heorems for [r]{}elay [c]{}hannels with [c]{}onfidential [m]{}essages,” in *Proc. IEEE ISIT*, pp. 926-930, Jun. 2007.
X. He and A. Yener, “[C]{}ooperation [w]{}ith an [u]{}ntrusted [r]{}elay: [A]{} [s]{}ecrecy [p]{}erspective,” *[IEEE]{} Trans. Inf. Theory*, vol. 56, no. 8, pp. 3807-3827, Aug. 2010.
——, “[O]{}n the [e]{}quivocation [r]{}egion of [r]{}elay [c]{}hannels with [o]{}rthogonal [c]{}omponents,” in *Proc. Forty-First Asilomar Conf.*, pp. 883-887, Nov. 2007.
——, “[T]{}he role of an untrusted relay in secret communication,” in *Proc. IEEE ISIT*, pp. 2212-2216, Jul. 2008.
E. Ekrem and S. Ulukus, “[S]{}ecrecy in cooperative relay broadcast channels,” in *Proc. IEEE ISIT*, pp. 2217-2221, Jul. 2008.
L. Lai and H. El Gamal, “[T]{}he [r]{}elay-[e]{}avesdropper [c]{}hannel: [c]{}ooperation for [s]{}ecrecy," *[IEEE]{} Trans. Inf. Theory*, vol. 54, no. 9, pp. 4005-4019, Sep. 2008.
L. Dong, Z. Han, A. P. Petropulu, and H. V. Poor, “[I]{}mproving [w]{}ireless [p]{}hysical [l]{}ayer [s]{}ecurity via [c]{}ooperating [r]{}elays,” *[IEEE]{} Trans. Signal Process.*, vol. 58, no. 3, pp. 1875-1888, Mar. 2010.
J. Li, A. P. Petropulu, and S. Weber, “Optimal cooperative relaying schemes for improving wireless physical layer security," \[Online\]. Available: http://arXiv.org/abs/1001.1389.
X. Tang, R. Liu, P. Spasojevic, and H. V. Poor, “[I]{}nterference-assisted secret communication,” in *Proc. IEEE ITW*, pp. 164-168, May 2008.
J. Wang and A. L. Swindlehurst, “[C]{}ooperative [j]{}amming in [MIMO]{} ad-hoc networks,” in *Proc. Forty-Third Asilomar Conf.*, pp. 1719-1723, Nov. 2009.
J. Huang and A. L. Swindlehurst, “[S]{}ecure communications via cooperative jamming in two-hop relay systems,” in *Proc. IEEE GLOBECOM*, Dec. 2010.
X. He and A. Yener, “[O]{}n the role of feedback in two-way secure communication,” in *Proc. 42nd Asilomar Conf. Signals, Systems and Computers*, pp. 1093-1097, Oct. 2008.
A. Mukherjee and A. L. Swindlehurst, “[S]{}ecuring multi-antenna two-way relay channels with analog network coding against eavesdroppers,” in *Proc. 11th IEEE SPAWC*, Jun. 2010.
S. Al-Sayed and A. Sezgin, “Secrecy in Gaussian MIMO bidirectional broadcast wiretap channels: Transmit strategies," in *Proc. 44th Asilomar Conf. Signals, Systems and Computers*, Nov. 2010.
R. Zhang, L. Song, Z. Han, B. Jiaa, and M. Debbah, “Physical layer security for two way relay communications with friendly jammers," in *Proc. IEEE GLOBECOM*, Miami, 2010.
I. Mitola, J., “[C]{}ognitive radio for flexible mobile multimedia communications,” in *Proc. IEEE Int Mobile Multimedia Commun. Work.*, pp. 3-10, Nov. 1999.
T. C. Clancy and N. Goergen, “[S]{}ecurity in [c]{}ognitive [r]{}adio [n]{}etworks: [T]{}hreats and [m]{}itigation,” in *Proc. 3rd Int. Conf. Cognitive Radio Oriented Wireless Networks and Communications CrownCom 2008*, pp. 1-8, May 2008.
T. X. Brown and A. Sethi, “[P]{}otential [c]{}ognitive [r]{}adio [d]{}enial-of-[s]{}ervice [v]{}ulnerailities and [p]{}rotection [c]{}ountermeasures: [a]{} [m]{}ulti-dimensional [a]{}nalysis and [a]{}ssessment,” in *Proc. IEEE CrownCom*, pp. 456-464, Aug. 2007.
R. Chen, J.-M. Park, and J. H. Reed, “[D]{}efense against [p]{}rimary [u]{}ser emulation attacks in cognitive radio [n]{}etworks,” *[IEEE]{} J. Sel. Areas Commun.*, vol. 26, no. 1, pp. 25–37, Jan. 2008.
H. Wang, L. Lightfoot, and T. Li, “[O]{}n [PHY]{}-layer security of cognitive radio: [C]{}ollaborative sensing under malicious attacks,” in *Proc. 44th Annual Conf. Information Sciences and Systems (CISS)*, pp. 1-6, Mar. 2010. S. Anand and R. Chandramouli, “On the secrecy capacity of fading cognitive wireless networks," in *Proc. IEEE CrownCom*, May 2008. Y. Pei, Y. Liang, L. Zhang, K. C. Teh, and K. H. Li, “Secure communication over MISO cognitive radio channels," *IEEE Trans. Wireless Commun.*, vol. 9, no. 4, pp. 1494-1592 , Apr. 2010.
X. Li, M. Chen, and E. P. Ratazzi, “Array-transmission based physical-layer security techniques for wireless sensor networks," in *Proc. IEEE ICMA*, pp. 1618-1623, 2005. M. Anand, Z. Ives, and I. Lee, “Quantifying eavesdropping vulnerability in sensor networks," in *Proc. 2nd International Workshop on Data Management For Sensor Networks*, pp. 3-9, Aug. 2005. D. Kundur, W. Luh, U. N. Okorafor, and T. Zourntos, “Security and privacy for distributed multimedia sensor networks," *Proc. of the IEEE Special Issue on Distributed Multimedia*, vol. 96, no. 1, pp. 112-130, Jan. 2008.
R. Ahlswede and I. Csiszár, “Common randomness in information theory and cryptography - part I: Secret sharing," *IEEE Trans. Inf. Theory*, vol. 39, no. 4, pp. 1121-1132, July 1993. I. Csiszár and P. Narayan, “Common randomness and secret key generation with a helper," *IEEE Trans. Inform. Theory*, vol. 46, no. 3, pp. 344-366, Mar. 2000. I. Csiszár and P. Narayan, “Secrecy capacities for multiple terminals," *IEEE Trans. Inf. Theory*, vol. 50, no. 12, pp. 3047-3061, Dec. 2004. U. M. Maurer and S. Wolf, “Secret key agreement over a nonauthenticated channel-Part I: Definitions and bounds," *IEEE Trans. Inf. Theory*, vol. 49, no. 4, pp. 822-831, Apr. 2003. U. M. Maurer and S. Wolf, “Secret key agreement over a nonauthenticated channel-Part II: The simulatability condition," *IEEE Trans. Inf. Theory*, vol. 49, no. 4, pp. 832-838, Apr. 2003. U. M. Maurer and S. Wolf, “Secret key agreement over a nonauthenticated channel-Part III: Privacy amplification," *IEEE Trans. Inf. Theory*, vol. 49, no. 4, pp. 839-851, Apr. 2003. H. Koorapaty, A. A. Hassan, and S. Chennakeshu, “Secure information transmission for mobile radio," *IEEE Commun. Lett.*, vol. 4, pp. 52–55, Feb. 2000. C. Ye, A. Reznik, G. Sternberg, and Y. Shah, “On the secrecy capabilities of ITU channels," in *Proc. IEEE 66nd Veh. Tech. Conf.*, pp. 2030–2034, Baltimore, MD, Oct. 2007. A. Sayeed and A. Perrig, “Secure wireless communications: Secret keys through multipath," in *Proc. IEEE ICASSP*, Las Vegas, pp. 3013-3016, Apr. 2008. C. Ye, A. Reznik, and Y. Shah, “Extracting secrecy from jointly Gaussian random variables," in *Proc. IEEE ISIT*, pp. 2593-2597, Seattle, July 2006. M. Bloch, J. Barros, M. R. D. Rodrigues, and S. W. McLaughlin, “Wireless information-theoretic security," *IEEE Trans. Inf. Theory*, vol. 54, no. 6, pp. 2515-2534, June 2008. Y. Chunxuan, S. Mathur, A. Reznik, Y. Shah, W. Trappe, and N.B. Mandayam, “Information-theoretically secret key generation for fading wireless channels," *IEEE Trans. Inform. Forensics and Security*, vol.5, no.2, pp. 240-254, June 2010. Y. Abdallah, M. A. Latif, M. Youssef, A. Sultan, and H. El Gamal, “Keys through ARQ: Theory and practice," submitted to *IEEE Trans. Inform. Forensics and Security*, May 2010 \[Online\]. Available: http://arxiv.org/pdf/1005.5063. X. Li and E. P. Ratazzi, “MIMO transmissions with information-theoretic secrecy for secret-key agreement in wireless networks," in *Proc. IEEE MILCOM*, Atlantic City, NJ, 2005. C. Chen and M. A. Jensen, “Secrecy extraction from increased randomness in a time-varying MIMO channel," in *Proc. IEEE GLOBECOM*, Honolulu, Dec. 2009. C. Chen and M. A. Jensen, “Secret key establishment using temporally and spatially correlated wireless channel coefficients,“ *IEEE Trans. Mobile Computing*, submitted Oct. 2009. J. W. Wallace and R. K. Sharma, ”Automatic secret keys from reciprocal MIMO wireless channels: Measurement and analysis,” *IEEE Trans. Inform. Forensics and Security*, vol. 5, no. 3, pp. 381-392, Sep. 2010. L. Lai, H. El Gamal, and H. V. Poor, “The wiretap channel with feedback: Encryption over the channel," *IEEE Trans. Inf. Theory*, vol. 54, no. 11, pp. 5059-5067, Nov. 2008. E Ardestanizadeh, M. Franceschetti, T. Javidi, and Y. Kim, “Wiretap channel with secure rate-limited feedback," *IEEE Trans. Inf. Theory*, vol. 55, no. 12, pp. 5353-5361, Dec. 2009.
A. Thangaraj, S. Dihidar, A. R. Calderbank, S. W. McLaughlin, and J. Merolla, “Applications of LDPC codes to the wiretap channel," *IEEE Trans. Inf. Theory*, vol. 53, no. 8, pp. 2933-2945, Aug. 2007. D. Klinc, J. Ha, S. McLaughlin, J. Barros, and B.-J. Kwak, “LDPC codes for physical layer security," in *Proc. IEEE GLOBECOM*, Honolulu, HI, Nov. 2009. A. Payandeh, M. Ahmadian, and M. Reza Aref, “Adaptive secure channel coding based on punctured turbo codes," *IEE Proc.-Commun*., vol. 153, no. 2, pp. 313-316, Apr. 2006 X. He and A. Yener, “Providing secrecy with structured codes: Tools and applications to Gaussian two-user channels," submitted to *IEEE Trans. Inf. Theory*, 2009 \[Online\]. Available: http://arxiv.org/abs/0907.5388. R. Liu, H. V. Poor, P. Spasojevic, and Y. Liang, “Nested codes for secure transmission," in *Proc. IEEE PIMRC*, pp. 1–5, Sep. 2008. A. Arora and L. Sang, “Dialog codes for secure wireless communications," in *Proc. IPSN*, 2009. H. Mahdavifar and A. Vardy, “Achieving the secrecy capacity of wiretap channels using polar codes," 2010, \[Online\]. Available: http://arxiv.org/abs/1001.0210. R. W. Yeung, *Information Theory and Network Coding*. Springer, August 2008. N. Cai and R. Yeung, “Secure network coding," in *Proc. IEEE ISIT*, p. 323, June 2002. J. Feldman, T. Malkin, R. Servedio, and C. Stein, “On the capacity of secure network coding," in *Proc. Allerton Conf.*, Sept. 2004. D. Silva and F. R. Kschischang, “Security for wiretap networks via rank-metric codes," in *Proc. IEEE ISIT*, Toronto, 2008. K. Lu, S. Fu, Y. Qian, Y., and T. Zhang, “On the security performance of physical-layer network coding," in *Proc. IEEE ICC*, Dresden, Germany, June 2009.
W. Saad, Z. Han, M. Debbah, A. Hjorungnes, and T. Basar, “Physical layer security: Coalitional games for distributed cooperation," in *Proc. 7th WiOpt*, 2009. Z. Han, N. Marina, M. Debbah, and A. Hj[ø]{}rungnes, “Physical layer security game: Interaction between source, eavesdropper and friendly jammer", *Eurasip Journ. Wireless Commun. and Network.*, special issue on physical layer security, 2009. Z. Han, N. Marina, M. Debbah, and A. Hjorungnes, “Improved wireless secrecy capacity using distributed auction theory", in *Proc. 5th ICMAS*, China, 2009. S. Anand and R. Chandramouli, “Secrecy capacity of multi-terminal networks with pricing," \[Online\]. Available: http://koala.ece.stevens-tech.edu/ mouli/IT02.pdf. M. Yuksel, X. Liu, and E. Erkip, “A secure communication game with a relay helping the eavesdropper," in *Proc. IEEE Inf. Theory Workshop*, Taormina, Italy, Oct. 2009. A. Mukherjee and A. L. Swindlehurst, “Equilibrium outcomes of dynamic games in MIMO channels with active eavesdroppers," in *Proc. IEEE ICC*, Cape Town, South Africa, May 2010.
Y. Liang, H. V. Poor, and S. Shamai, *Information Theoretic Security*. Foundations and Trends in Communications and Information Theory, vol. 5, nos. 4-5, pp. 355-580, Now Publishers, Hanover, MA, USA, 2008. R. Liu and W. Trappe, *Securing Wireless Communications at the Physical Layer*. Springer, Norwell, MA, USA, 2009. E. Jorswieck, A. Wolf, and S. Gerbracht, *Secrecy on the Physical Layer in Wireless Networks*. Telecommunications, In-Tech Publishers, 2010. R. A. Horn and C. R. Johnson, *Matrix Analysis*, University Press, Cambridge, UK, 1985.
[^1]: The authors are with the Dept. of Electrical Engineering & Computer Science, University of California, Irvine, CA 92697-2625, USA. [(e-mail: {amukherj; afakoori; jing.huang; swindle}@uci.edu)]{}
[^2]: Strictly speaking, Wyner’s definition of “perfect secrecy" as the scenario in which the block-length-normalized mutual information at the eavesdropper vanishes in the limit of long block lengths was weaker than the one proposed by Shannon \[cf. (\[eq:Shannon\_perfsecrecy\])\], which requires that the mutual information be zero regardless of the block length.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper concludes the series begun in \[M. Dafermos and I. Rodnianski *Decay for solutions of the wave equation on Kerr exterior spacetimes I-II: the cases $|a| \ll M$ or axisymmetry*, arXiv:1010.5132\], providing the complete proof of definitive boundedness and decay results for the scalar wave equation on Kerr backgrounds in the general subextremal $|a|<M$ case without symmetry assumptions. The essential ideas of the proof (together with explicit constructions of the most difficult multiplier currents) have been announced in our survey \[M. Dafermos and I. Rodnianski *The black hole stability problem for linear scalar perturbations*, in Proceedings of the 12th Marcel Grossmann Meeting on General Relativity, T. Damour et al (ed.), World Scientific, Singapore, 2011, pp. 132–189, arXiv:1010.5137\]. Our proof appeals also to the quantitative mode-stability proven in \[Y. Shlapentokh-Rothman *Quantitative Mode Stability for the Wave Equation on the Kerr Spacetime*, arXiv:1302.6902, to appear, Ann. Henri Poincaré\], together with a streamlined continuity argument in the parameter $a$, appearing here for the first time. While serving as Part III of a series, this paper repeats all necessary notations so that it can be read independently of previous work.'
address:
- 'Princeton University, Department of Mathematics, Fine Hall, Washington Road, Princeton, NJ 08544, United States, [[email protected]]{}'
- 'University of Cambridge, Department of Pure Mathematics and Mathematical Statistics, Wilberforce Road, Cambridge CB3 0WA, United Kingdom, [[email protected]]{} '
- 'Princeton University, Department of Mathematics, Fine Hall, Washington Road, Princeton, NJ 08544, United States, [[email protected]]{}'
- 'Massachusetts Institute of Technology, Department of Mathematics, 77 Massachusetts Avenue, Cambridge, MA 02139, United States, [[email protected]]{}'
- 'Massachusetts Institute of Technology, Department of Mathematics, 77 Massachusetts Avenue, Cambridge, MA 02139, United States, [[email protected]]{}'
author:
- Mihalis Dafermos
- Igor Rodnianski
- 'Yakov Shlapentokh-Rothman'
date: 'December 3, 2014'
title: |
Decay for solutions of the wave equation\
on Kerr exterior spacetimes *III*:\
The full subextremal case $|a| < M$
---
Introduction
============
The boundedness and decay properties of solutions to the scalar wave equation $$\label{WAVE}
\Box_{g_{a,M}}\psi = 0$$ on the exterior regions of Kerr black hole backgrounds $(\mathcal{M},g_{a,M})$ have been the subject of considerable recent activity, in view of the intimate relation of this problem to the stability of these spacetimes themselves in the context of Einstein’s theory of general relativity (cf. [@ck]). Following definitive results [@kw:lss; @2bachelots; @dr1; @BlueSof0; @dr3; @BlueSof; @BlueSter; @dr5] in the Schwarzschild case $a=0$, boundedness in the very slowly rotating Kerr case $|a|\ll M$ was first proven in our [@dr6], and subsequently, decay results have been established for $|a|\ll M$ in [@jnotes] and in the first parts of this series [@dr7], and independently by Tataru–Tohaneanu [@tattoh] and Andersson–Blue [@anblue]. See also [@luk2]. Our [@dr7] also obtained such decay results in the general subextremal case $|a|<M$, under the assumption that $\psi$ is itself axisymmetric. (Let us mention also the previous non-quantitative study [@fksy; @fksy2] of fixed azimuthal modes on Kerr.) The main significance of these restrictive assumptions is that the well-known difficulty of superradiance is controlled by a small parameter (the case $|a|\ll M$) or is in fact completely absent (the case of axisymmetric $\psi$). The present paper represents the culmination of this programme by dropping these restrictions, extending the above boundedness and decay results to the *general subextremal case $|a|<M$ without axisymmetry*:
1. General solutions $\psi$ of $(\ref{WAVE})$ on the exterior of a Kerr black hole background $(\mathcal{M}, g_{a,M})$ , arising from bounded initial energy on a suitable Cauchy surface $\Sigma_0$, have bounded energy flux through a global foliation $\Sigma_\tau$ of the exterior, bounded energy flux through the event horizon $\mathcal{H}^+$ and null infinity $\mathcal{I}^+$, and satisfy a suitable version of “local integrated energy decay”.
2. Similar statements hold for higher order energies involving time-translation invariant derivatives. This implies immediately uniform bounds on $\psi$ and translation-invariant derivatives to arbitrary order, up to and including $\mathcal{H}^+$, in terms of a sufficiently high order initial energy.
The precise statements of parts 1 and 2 of the above Theorem will be given in Section \[statementsections\] as Theorems \[theResult\] and \[h.o.s.\]. Note that these are exact analogues of Theorems 1.2 and 1.3 of [@dr7], dropping the assumption of axisymmetry (alternatively, Theorem 1.1 dropping the assumption $a_0\ll M$). The main results of the present paper have been previously announced in our survey paper [@stabi], which provided both an outline of the proof and many details of the crucial arguments, including all high frequency multiplier constructions. To complete the outline, one required a quantitative refinement of Whiting’s classical mode stability result [@whiting] and a continuity argument in the parameter $a$. The former refinement has very recently been obtained [@shlapRot] and will be indeed used in our proof. As for the latter, the proof presented here introduces a streamlined continuity argument which as an added benefit in fact *reproves* the theorems of the first parts of the series [@dr7] in the case $|a|\ll M$. We will only rely on [@dr7] for a detailed discussion of the background, the setup and several of its elementary propositions. All necessary notations and results from [@dr7] are reviewed and quoted explicitly, however, so that the present paper can be read independently. We will also repeat all constructions originally introduced in the survey [@stabi].
In view of our general “black box” decay result [@icmp], a corollary of the above Theorem is
Solutions $\psi$ of (\[WAVE\]) arising from sufficiently regular and initial data (i.e. whose initial suitably higher-order energy is finite) satisfy uniform of the energy flux through a hyperboloidal foliation $\widetilde{\Sigma}_\tau$ of the exterior region as well as uniform pointwise polynomial decay bounds.
As in the case of Minkowski space, there is a hierarchy of polynomial decay bounds that can be obtained, both for energy fluxes and pointwise, depending on the quantification of the initial localisation assumed on initial data. The precise statement we shall give (Corollary \[thecorol\] of Section \[statementsections\]) is motivated by applications to quasilinear problems; further refinements fail to be robust from this point of view. We remark explicitly that the decay estimates of the above Corollary are indeed sufficient for applications to quasilinear problems with quadratic non-linearities. See [@luk3; @shiwu; @shiwu2; @Yang]. We note also that the non-quantitative fixed-azimuthal mode statements of [@fksy; @fksy2] are of course implied *a fortiori* by the above Corollary. To obtain from our Theorem alternative more refined corollaries for compactly supported smooth initial data, see [@tatar].
As stated, the above Theorem and its Corollary concern the black hole exterior. Note that boundedness and polynomial decay statements in the Kerr exterior propagate easily to any fixed-$r$ hypersurface in the interior following [@luk] using the red-shift effect and stationarity, for $r$ *strictly greater* than its value on the Cauchy horizon. This insight goes back to [@cbh]. On the other hand, by a result of [@Sbierski], uniform non-degenerate energy boundedness cannot hold up to the Cauchy horizon in view of the blue-shift. Uniform $L^\infty$ bounds on $\psi$ and its *tangential* derivatives up to and including the Cauchy horizon have been obtained in the thesis of Franzen [@franzen].
Besides the Kerr family, there is an additional class of black hole spacetimes of interest in classical general relativity: the sub-extremal Kerr–Newman metrics. These now form a $3$-parameter family of metrics, with parameters $a$, $M$ and $Q$ (the latter representing *charge*), which, when coupled with a suitable Maxwell field associated to $Q$, satisfy the *Einstein–Maxwell* equations. See [@he:lssst]. (Setting $Q=0$, the Maxwell field vanishes and the family reduces to the Kerr case.) As shown in the thesis of Civin [@civin], all the miraculous properties of the Kerr family that allow for the results proven in the present paper in fact extend to the Kerr–Newman family in the full sub-extremal parameter range $\sqrt{a^2+Q^2}<M$, leading to a precise analogue of our Theorem and its Corollary in this more general class.
Though outside of the domain of astrophysically relevant black holes, it is interesting to consider the problem of boundedness and decay for scalar waves on the analogues of the Kerr family when a non-zero cosmological constant $\Lambda$ is added to the Einstein equations. These spacetimes are known as Kerr-de Sitter ($\Lambda>0$) and Kerr-anti de Sitter ($\Lambda<0$). See [@he:lssst]. It is in fact the negative case $\Lambda<0$ which presents more surprising new phenomena from the mathematical point of view and has been definitively treated in the work of Holzegel–Smulevici [@holz-smul; @holz-smul2]. See also [@gannot]. The $\Lambda>0$ case is from some points of view easier than $\Lambda=0$, and results in the Schwarzschild-de Sitter $(a=0)$ and very slowly rotating Kerr-de Sitter $(|a| \ll M, |a|\ll \Lambda)$ case followed soon after the analogous results for Schwarzschild and very slowly rotating Kerr had been obtained. See [@dr4; @bh; @sbvm; @vasy; @dyatlov1; @dyatlov2; @schlue2]. Let us note however, that Kerr-de Sitter is still not understood in its full subextremal range, in particular in view of the absence of an analogue of the mode stability statements [@whiting; @shlapRot]. The best results to date have been obtained by Dyatlov [@dyatlov-last].[^1]
Returning to the classical astrophysical domain, let us recall finally that the Kerr metrics $g_{a,M}$ represent black hole spacetimes in the full *closed* parameter range $|a|\le M$; the geometry of the extremal case $|a|=M$, however, exhibits several qualitative differences, most conspicuously, the degeneration of the celebrated *red-shift effect* at the horizon. In view of the recently discovered *Aretakis instability* [@aretakis; @aretakis2; @aretakisHor; @luciet; @murataetal], the precise analogue of the above Theorem *does not* in fact hold without qualification in the case of extremality $|a|=M$. In particular, in the extremal case, for generic solutions, certain higher order time-translation-invariant derivatives asymptotically blow up along the event horizon. This generic blow up is completely unrelated to superradiance and holds even for solutions $\psi$ restricted to be *axisymmetric*. For such axisymmetric $\psi$, restricted decay results of a definitive nature have been obtained by Aretakis [@aretakisKerr]. The fundamental remaining problem for scalar waves on black hole backgrounds of interest in classical general relativity is thus the precise understanding of the for general -axisymmetric solutions.
Overview of the main difficulties
---------------------------------
We begin with an overview of the difficulties of the problem and the basic elements of the proof. In this discussion, we will assume some familiarity with the Schwarzschild and Kerr families of metrics as well as basic aspects of the analysis of wave equations on Lorentzian manifolds. See our lecture notes [@jnotes].
### Review of the very slowly rotating case $|a|\ll M$ {#reviewsr}
We have discussed at length in the first parts of this series [@dr7] the various difficulties connected to showing energy boundedness and “integrated local energy decay” for solutions of the wave equation $(\ref{WAVE})$ in the very slowly rotating case. We review these briefly.
Already in the Schwarzschild case $a=0$, to show boundedness, one must face the difficulty that at the *event horizon* $\mathcal{H}^+$, the conserved $\partial_t$-energy (associated to stationarity of the metric) degenerates. To show integrated local energy decay, one must moreover understand both “low” and “high” frequency obstructions to dispersion, in particular, the high-frequency obstructions connected to the presence of *trapped null geodesics*.
The horizon difficulties were overcome by our introduction of the red-shift vector field [@dr3], while the difficulties concerning both excluding low frequency obstructions and quantifying the high frequency phenomena connected to trapped geodesics were overcome in one go by appeal to the energy identity of Morawetz-type (cf. [@mora2]) multipliers associated to a vectorfield $f(r)\partial_r$, where $f$ is a carefully chosen function vanishing at the so-called photon sphere $r=3M$, the timelike cylinder to which all trapped null geodesics asymptote. In the Schwarzschild context, such Morawetz estimates were pioneered by [@labasoffer; @BlueSof]. The boundedness of the nonnegative space-time integral given by the energy identity associated to this multiplier is precisely the statement of “integrated local energy decay”. See also [@BlueSter; @dr3]. The degeneracy of any such estimate at trapping is necessary in view of a general result of Sbierski [@Sbierski] in the spirit of the classical [@Ralston]. On the other hand, the fact that such a degenerate estimate indeed holds (and the precise nature of the degeneracy) means that the trapping is “good”; at the level of geodesic flow, this corresponds to the fact that dynamics is hyperbolic near the trapped set. This estimate also degenerates at the event horizon when only the initial conserved $\partial_t$-energy is included on the right hand side. On the other hand, as shown in [@dr3], again using the red-shift vector field, this degeneracy is removed by replacing the latter with the initial *non-degenerate* energy.
Turning from Schwarzschild to the very slowly rotating Kerr case $|a|\ll M$, the above difficulties are combined with a new one: *superradiance*. Now, the vector field $\partial_t$ is spacelike in a region outside the horizon known as the *ergoregion*, hence its energy identity gives no obvious *a priori* control over the solution. Moreover, it is clear that the high-frequency obstructions to decay cannot be captured from classical physical space vector field multipliers [@alinhac]. This can be seen at the level of geodesic flow as the projection of the trapped space to physical space is no longer a codimension-$1$ hypersurface.
The problem of superradiance was first overcome in our proof of boundedness [@dr6] mentioned previously, which concerned in fact the more general setting of the wave equation on backgrounds that are small axisymmetric stationary perturbations of Schwarzschild, a class including the very slowly rotating Kerr case $|a|\ll M$. In this class of spacetimes, one can analyse solutions with respect to frequencies $\omega\in \mathbb R$ and $m\in \mathbb Z$ corresponding to the stationary and axisymmetric Killing fields, and decompose general solutions $\psi$ of $(\ref{WAVE})$ into their superradiant and non-superradiant part. For the latter part, one can prove boundedness as in Schwarzschild. For the superradiant part, it turns out that one can explicitly prove *both* boundedness and integrated local energy decay perturbing a Schwarzschild energy identity yielding both positive boundary and space-time terms that do not degenerate. The non-degeneracy of this estimate encodes the fact that the *superradiant part of $\psi$ is not trapped*. We shall return to this insight later in our discussion of the general $|a|<M$ case.
Turning to the issue of proving decay for $|a|\ll M$, the problem of capturing the good properties of trapping was overcome using *frequency-localised* generalisations of the Morawetz multipliers applied in Schwarzschild. There have been three independent approaches [@jnotes; @tattoh; @anblue], which all crucially rely on the additional hidden symmetries of Kerr that are reflected in the existence of a Killing tensor and separability properties of both geodesic flow and the wave equation itself. In our approach [@jnotes; @dr7], the frequency localisation uses directly Carter’s separation of the wave equation [@cartersep2], which introduces, in addition to $\omega$ and $m$ above, a real frequency parameter $\lambda_{m\ell}(a\omega)$ parameterised by an additional parameter $\ell\in \mathbb N_0$ such that, localised to each frequency triple $(\omega, m, \lambda_{m\ell}(a\omega))$, the wave equation $(\ref{WAVE})$ reduces to the following second order o.d.e. $$\label{oderedu}
u''+\omega^2 u =V(a\omega, m, \lambda_{m\ell})u$$ where $V$ is a potential and $'$ denotes differentiation in a rescaled $r^*(r)$ coordinate. The frequencies $\lambda_{m\ell}$ are themselves eigenvalues of an associated elliptic equation whose eigenfunctions (known as oblate spheroidal harmonics) appear in the formula reconstructing $\psi$ from $u$.
Note that in the Schwarzschild ($a=0$) case, the reduction to $(\ref{oderedu})$ corresponds to the classical separation by spherical harmonics, and we have explicitly $\lambda_{m\ell}={\ell(\ell+1)}$ (independent of $\omega$) and $$V_{\rm Schw}(r)=(r-2M)\left(\frac{\lambda_{m\ell}}{r^3}+\frac{2M}{r^4}\right).$$ For all $\lambda_{m\ell}$, the potential $V_{\rm Schw}$ has a unique simple maximum at an $r$-value $r_{\rm max}(\lambda_{m\ell})$ such that $$r_{\rm max}(\lambda_{m\ell})\to 3M$$ as $\ell\to \infty$. One sees thus the relation of this potential to the trapping phenomenon. Indeed, the radial dependence of null geodesics with fixed energy and angular momentum is governed by a potential which coincides with $V$ above in the high frequency rescaled limit.
In the Kerr case, the eigenvalues $\lambda_{m\ell}(a\omega)$ are no longer explicit expressions, and the form of $V$ is considerably more complicated. It was shown, however, in [@jnotes], that for $|a|\ll M$ and for frequencies in the “trapping” regime $$\label{trappingfr}
1\ll \omega^2 \sim \lambda_{m\ell},$$ the potential $V$ retains its “good” Schwarzschild properties. Specifically, the potential $V$ in $(\ref{oderedu})$ can be seen to again have a unique simple maximum in this frequency range, whose $r$-value $r_{\rm max}$ depends on the frequency parameters $$r_{\rm max}=r_{\rm max}(a\omega, m, \lambda_{m\ell}).$$ This allows, separately for each $(\omega, m , \lambda_{m\ell}(a\omega))$, the construction of an analogue of the current $f(r)\partial_r$ vanishing exactly at $r_{\rm max}$, yielding the desired positivity properties. Unlike the Schwarzschild case, however, there does not exist a unique high frequency limit of $r_{\rm max}$, consistent with the fact [@alinhac] that these currents cannot be replaced by a classical vector field defined in physical space. See however [@anblue]. At the level of geodesic flow, this precisely reflects the fact that trapped null geodesics exist for a full range of $r$-values in a neighbourhood of $r=3M$.[^2]
In the remaining frequency regimes, one can in fact simply carry over the previous physical-space Schwarzschild constructions (see our argument in [@jnotes]) to the more general $|a|
\ll M$ case, as, restricted to those ranges, the relevant Schwarzschild estimates *do not* degenerate and thus their positivity properties are manifestly stable to small perturbation. Alternatively, as in the first parts of this series [@dr7], one can construct new currents taylored specifically to these frequency ranges. The latter approach is more flexible (it has the additional benefit of providing [@dr7] an independent second proof of the Schwarzschild case) and will be more useful for the general subextremal case $|a|<M$.
Let us note that our proof of integrated decay in the first part of this series [@dr7] in fact gives a separate proof of the boundedness statement of [@dr6], when the latter is specialised to Kerr. Here, one exploits the fact that when $|a|\ll M$, superradiance is controlled by a small parameter and thus boundary terms of the wrong sign can be absorbed by a small multiple of the red-shift current added to the conserved $\partial_t$ energy. One obtains thus boundedness and integrated local energy decay at the same time, without separating the solution into its superradiant and non-superradiant parts. We shall see, however, that for the general case $|a|<M$, the original insight of [@dr6] will again be fundamental.
In discussing our frequency analysis for both the problems of boundedness and integrated local energy decay, we have suppressed an important point: To define frequencies $\omega\in
\mathbb R$, we are applying the Fourier transform in time. Since solutions are not known *a priori* to be sufficiently integrable in time, however, one must first apply suitable cutoffs in the future, generating error terms which must then be absorbed. For this, some weak *a priori* control of these terms is essential–and here we have used in both [@dr6; @dr7] yet again[^3] the closeness to Schwarzschild. The analogue of this procedure for the general $|a|<M$ case is our appeal to the continuity argument of Section \[continuityintro\]. We defer further discussion of this till then.
### Structure of trapping and its disjointness from superradiance {#highfreqinsec}
In passing to the general subextremal case $|a|<M$, the first fundamental issues that must be addressed are the “high frequency” ones.
The most obvious question is whether the structure of trapping retains its “good” properties which allow in principle for degenerate integrated decay statements. At the same time, since superradiance is no longer governed by a small parameter, one has to understand what is it which finally quantifies its strength, or, in the context of the proof, how does one guarantee the control of boundary terms in space-time energy identities.
Though geodesic flow and various measures of the strength of superradiance have been thoroughly investigated in physics [@carter; @chandrasekhar; @Starobinsky], the properties that turn out to be essential for our argument do not appear to have been considered explicitly in the classical literature. Indeed, it is somewhat of a miracle that the Kerr geometry turns out to be well behaved on both accounts, , specifically:
1. *The structure of trapping is as in Schwarzschild.*
2. *Superradiant frequencies are not trapped.*
The embodiment of properties (a) and (b) we shall need were proven already in our survey paper [@stabi] by analysing the potential $V$ in $(\ref{oderedu})$. Concerning (a), the “trapping” frequency range $(\ref{trappingfr})$, the potential $V$ was shown to have a unique simple maximum $V_{\rm max}$ at a (frequency dependent) $r$-value $r_{\rm max}$, just as in the slowly rotating case $|a|\ll M$. (This shows *a fortiori* that the underlying null geodesic flow near trapping is hyperbolic.) Concerning (b), it was shown that for high *superradiant* frequencies, $$\label{highsupers}
1\ll \lambda + \omega^2,\qquad \omega\left(\omega-\frac{am}{2M(M+\sqrt{M^2+a^2}}\right)<0$$ the maximum of $V$ is always “quantitatively” above the energy level $\omega^2$, in the sense $$\label{quantabove}
V_{\rm max} \ge (1+\epsilon) \omega^2.$$ (This in turn shows *a fortiori* that future directed null geodesics whose tangent vector has nonnegative inner product with $\partial_t$ are never future trapped; they will always cross the event horizon $\mathcal{H}^+$.[^4] *Note that in contrast to the $|a|\ll M$ case, the stronger statement that $\partial_t$ is eventually timelike along any future trapped null geodesic is not true; i.e. it is not true that all future trapped null geodesics must leave the ergoregion.*)
As with the $|a|\ll M$ case, it is not statements about geodesic flow that we appeal to, but rather we use directly the properties of the potential $V$ to construct appropriate energy currents. More specifically, the above properties of the potential are used to construct frequency dependent multiplier currents yielding both positive bulk *and positive boundary terms* for all high frequency ranges. In the superradiant case, property $(\ref{quantabove})$ can be exploited to arrange such that the bulk term is in fact non-degenerate; this can be thought of as the definitive embodiment of (b). Note that these high-frequency multiplier constructions all appeared explicitly in our survey [@stabi]. We will repeat these constructions here with very minor modifications. See the outline in Section \[outlinesection\] below.
It is interesting to note that property (b) above in fact degenerates in the extremal limit $|a|\to M$ in the following sense: At the endpoint of the superradiant frequency range $(\ref{highsupers})$, one loses the $\epsilon$ in $(\ref{quantabove})$. This is an additional (and separate) phenomenon to the degeneration of the red-shift and could have interesting implications for the remaining problem of understanding non-axisymmetric $\psi$ in the extremal case $|a|=M$. See [@AndGlam] and the discussion in [@aretakisKerr].
### Absence of bounded frequency superradiant obstructions {#nolow}
The above still leaves us with the problem of understanding *bounded* (i.e. $|\omega|\lesssim1$) frequencies.
One must first distinguish the near stationary case $|\omega|\ll1$. This frequency range is very sensitive to global aspects of the geometry. It turns out that here an explicit multiplier construction is possible which adapts our construction of the first parts of this series [@dr7]. (Interestingly, the cases of $|a|\le \tilde{a}_0$ and $|a|\ge \tilde{a}_0$ are here handled differently.) These multiplier constructions appear for the first time in the present paper.
Turning now to the remaining bounded frequencies, as explained in our survey [@stabi], whereas in the *non-superradiant* regime, one can explicitly construct multipliers with both nonnegative bulk and boundary terms, for bounded non-superradiant frequencies, adapting the constructions of [@dr7] from the $|a|\ll M$ case, there does not appear to be a straightforward such construction for the superradiant regime, when neither can superradiance be treated as a small parameter, nor can one exploit $(\ref{quantabove})$ together with either $\omega$, $m$ or $\lambda_{m\ell}$ as a large parameter. One can indeed construct currents with a non-negative *bulk* term, but these generate a boundary term of the wrong sign which still must be controlled.
As announced already in [@stabi], to control the remaining term one requires a quantitative extension of Whiting’s celebrated mode stability [@whiting], which in particular excludes the presence not just of growing modes but also resonances on the real axis. This was achieved in the recent [@shlapRot]. Appeal to [@shlapRot] will indeed allow us to control the remaining boundary term. Again, see the outline in Section \[outlinesection\] below.
### Higher order estimates {#HOEhere}
To obtain higher-order integrated local energy decay in the slowly rotating case $|a|\ll M$, it was sufficient to commute $(\ref{WAVE})$ with $\partial_t$ (which is Killing) and also with the red-shift vector field (the latter an argument first applied in [@dr6]), exploiting the fact that the latter, though not Killing, generates positive terms in appropriate energy estimates modulo terms which can be controlled by the $\partial_t$-commutation. To show this fact, one uses in turn that control of a second derivative of $\psi$ in a timelike direction allows control of all second derivatives of the solution via elliptic estimates (in view of equation $(\ref{WAVE})$).
For the general case $|a|<M$, one appeals to yet another fundamental fact about Kerr geometry:
1. *The span of the stationary $\partial_t$ and axisymmetric $\partial_\phi$ Killing fields is timelike outside the horizon for the full range $|a|<M$.*
Thus, commuting with $\partial_t$, $\chi\partial_\phi$ (where $\chi$ is a cutoff function with compact support in $r$) and the red-shift vector field, one can essentially apply the same argument as before.
### Continuity argument {#continuityintro}
We now return to the issue that we have suppressed at the end of Section \[reviewsr\], namely, the question of how can one justify in the first place a frequency analysis based on real frequencies $\omega$ defined via the Fourier transform in time. In the case $|a|\ll M$, closeness to Schwarzschild gave a small parameter that could be exploited here. For the general $|a|<M$ case, however, as explained already in our survey [@stabi], one must exploit a continuity argument in $|a|$.
Note first that to justify the Fourier assumption and thus prove integrated local energy decay, one sees easily that it is sufficient to assume the non-quantitative assumption that the energy through $\Sigma_\tau$ of the projection $\psi_m$ of $\psi$ to each azimuthal frequency is finite. This is the statement that we show by continuity: For each azimuthal frequency number $m\in\mathbb Z$, we define the subset $$\mathcal{A}_m\subset [0,M)=\{|a|: \psi {\rm\ satisfying\ } (\ref{WAVE})
{\rm\ with\ } g_{a,M}\implies {\rm energy\ of\ }\psi_m{\rm\ remains\ finite}\}$$ We will show that $\mathcal{A}_m$ is a non-empty open and closed subset of $[0,M)$, and thus, $\mathcal{A}_m=[0,M)$.
We turn to a brief account of the continuity argument.
The non-emptyness of $\mathcal{A}_m$ follows from the general boundedness result for black hole spacetimes without ergoregions proven in [@jnotes], specialised to the Schwarzschild case $a=0$.
For openness, one shows that if $\mathring{a}\in \mathcal{A}_m$, then $|a-\mathring{a}|<\epsilon$ satisfies $a\in \mathcal{A}_m$ for sufficiently small $\epsilon$. One exploits here $\epsilon$ as a small parameter. The issues associated to openness already appeared in the small $|a|\ll M$ case; see [@dr6] and [@dr7]. The fact that we have fixed the azimuthal mode $m$ makes the argument here technically easier to implement. For this, the fundamental insight is that
1. trapping occurs *outside the ergoregion*.
2. using the energy identity for a vector field of the form $\partial_t+\alpha(r)\partial_\phi$, one can obtain boundedness modulo lower order terms supported only in the ergoregion.
To exploit the above, we first construct from a fixed-$m$ solution $\psi_m$ to $(\ref{WAVE})$ on $g_{a,M}$ and for each $\tau\ge 0$, a solution $\Psi$ of the inhomogeneous wave equation $\Box_g\Psi_m=F_m$ on an interpolating metric $g$ which coincides with $g_{a,M}$ in the region between $\Sigma_0$ and $\Sigma_{\tau-\delta_0}$ and coincides with $g_{\mathring{a},M}$ in the region in the future of $\Sigma_{\tau}$ and to which the integrability properties apply (since $\mathring{a}\in\mathcal{A}_m$). Applying our estimates and using 1. and 2., we may now absorb (for sufficiently small $\epsilon$) the error terms arising from the inhomogeneity to obtain an integrated decay statement for $\psi_m$. We note that the fixed-$m$ currents used for 1. and 2. may find additional applications.
Closedness is easy given the estimates shown and the smooth dependence of the Kerr family on the parameter $a$.
### Non-degenerate boundedness from integrated local energy decay {#boundednessintro}
The frequency analysis on which our proof of integrated local energy decay is based does not directly “see” the energy flux on fixed time hypersurfaces $\Sigma_\tau$, only the energy fluxes on the horizon $\mathcal{H}^+$ and future null infinity $\mathcal{I}^+$. Thus, it remains to show boundedness of the energy (and higher-order energies) through $\Sigma_\tau$.
In the slowly rotating case $|a|\ll M$, it is clear that given integrated local energy decay, boundedness of the energy flux through a spacelike foliation easily follows *a posteriori*[^5] by revisiting the physical space energy identity of a globally timelike vector field which coincides with $T$ where the latter is timelike, noting that, if $|a|$ is sufficiently small $T$ is timelike near trapping.
The above argument again uses in an essential way the disjointness of the ergoregion and the set–associated to trapping–on which integrated local energy decay estimate degenerates. As we have remarked earlier, these sets intersect when $|a|\sim M$–it is only in phase space where superradiance can be understood as disjoint of trapping.
One approach to boundedness could be to try to exploit again property (b) from Section \[highfreqinsec\]. It is technically easier to simply exploit the physical space fact (c) of Section \[HOEhere\], namely that Killing fields $\partial_t$ and $\partial_\phi$ together span a timelike subspace outside the horizon. Specifically, in a small neighbourhood of any $r$-value there exists a combination of $\partial_t$ and $\partial_\phi$ which is timelike and Killing. We use our frequency analysis to partition a solution $\psi$ of the wave equation into finitely many pieces $\tilde\psi_i$, each of which satisfies an analogue of integrated local energy decay degenerating only in a small neighbourhood of some $r_i$. Applying the energy estimates corresponding to a suitable $i$-dependent combination of $\partial_t$ and $\partial_\phi$ to each $\tilde\psi_i$, and summing, one obtains the desired non-degenerate uniform boundedness of the energy flux through $\Sigma_\tau$.
Outline of the paper {#outlinesection}
--------------------
We end this introduction with an outline of the structure of the paper.
In Section \[notats\], we will review the set-up and various notations from the first parts of the series [@dr7], including the ambient manifold, the form of the Kerr family of metrics and useful vector fields, hypersurfaces and formulas. This will allow us to give precise formulations of the main theorems in Section \[statementsections\]. (The reader may wish to refer to this outline again when reading Section \[theLogic\], which will describe the logical flow of the proofs of the various statements.)
Section \[prelimn\] contains various preliminaries, including a review of the propositions from [@dr7] capturing the redshift effect, an estimate for large $r$, Hardy inequalities and finally, various statements concerning the span of the Killing fields $\partial_t$ and $\partial_\phi$.
Our frequency localisation based on Carter’s separation will be reviewed in Section \[cartersupersection\]. The natural setting for this will be the class of sufficiently integrable outgoing functions $\Psi:\mathcal{R}\to \mathbb R$, a useful notion which we shall define in Section \[suffInt\]. The resulting coefficients $u$ and their corresponding radial o.d.e. (cf. $(\ref{oderedu})$ above) are obtained in Section \[separationSubsection\] and the “outgoing” boundary conditions in Section \[outgoingcondsec\].
The next three sections, Sections 6, 7, and 8, concern the study of the o.d.e. $(\ref{oderedu})$ and the proof of uniform estimates in the frequency parameters $\omega$, $m$, and $\Lambda$.
In Section \[Vpropsec\], we will give salient properties of the potential $V$ of $(\ref{oderedu})$ which embody (a) and (b) of Section \[highfreqinsec\]. Versions of the lemmas of Sections \[Vtrsec\] and \[Vsrsec\] together with proofs have in fact already been given in our survey paper [@stabi]; we repeat these here for completeness. The lemma of Section \[Vnewsec\], reflecting the properties of trapping for fixed $m$, is new and will be used in the context of the continuity argument of Section \[continuityargsec\] discussed below.
In Section \[sct\], we shall review our notation for fixed frequency current templates, which, upon selection of the free functions, will be used to obtain multiplier estimates for solutions to $(\ref{oderedu})$.
Section \[freqLocEst\] is the heart of the paper. Here, with the help of well-chosen functions in the current templates of Section \[sct\], we construct suitable currents for all relevant frequency ranges yielding positive bulk terms and thus an estimate for solutions of the radial o.d.e. $(\ref{oderedu})$ uniform in frequency parameters. In the trapping regime, the currents degenerate at $r_{\rm max}$. All these currents have appeared previously in our survey paper [@stabi] with the exception of the near-stationary range of Section \[nearstat\]. The boundary terms can also be made positive, with the exception of a range of bounded frequencies, which give rise to an extra horizon boundary term on the right hand side of the resulting estimate, which must still be absorbed.
In Section \[summation\], we apply the results of the previous section to the coefficients $u$ arising from the setting of Section \[cartersupersection\], summing the resulting frequency localised estimates to obtain control of a non-negative definite space-time integral. We note Section \[whitinghere\], where the extra horizon term (arising from low superradiant frequencies) is bounded by appeal to Proposition \[propShlapRot\], a result of [@shlapRot]. One obtains finally an integrated local decay statement for “future integrable” solutions of the wave equation, and a similar statement for the inhomogeneous equation in Section \[ILEDinhomo\].
Higher order decay estimates are then provided in Section \[higher\], using the structure described in Section \[HOEhere\].
In Section \[continuityargsec\], we implement our new continuity argument discussed in Section \[continuityintro\] above, which will allow us to drop the *a priori* assumption of future integrability, and extend our results to general solutions of the Cauchy problem for $(\ref{WAVE})$. The reduction to fixed azimuthal frequency is accomplished in Section \[reduxfixed\]. The most difficult part of the argument is openness, handled in Section \[opennesssec\], while closedness is considered in Section \[closednesssec\].
Section \[precise\] will state the more precise integrated local energy statement which has actually been obtained in the proof.
Finally, in Section \[boundSuff\], we prove the boundedness statements, following our discussion in Section \[boundednessintro\]. This will conclude the paper.
Acknowledgements
----------------
During the period where this research has been carried out, MD was supported by the European Research Council and the Engineering and Physical Sciences Research Council. IR acknowledges support through NSF grant DMS-1001500 and DMS-1065710. YS acknowledges support through NSF grants DMS-0943787 and DMS-1065710. The authors thank S. Aretakis, D. Civin and G. Moschidis for discussions and comments on the manuscript.
Review of the setup {#notats}
===================
In this section, we review the setup and certain notations from the first parts of the series [@dr7], so that the present paper can be read independently. The reader wishing for a more leisurely exposition of this material should refer back to [@dr7]; he or she familiar with [@dr7] can skip to Section \[statementsections\].
Ambient manifold and coordinate systems
---------------------------------------
The first task is to define an ambient manifold-with-boundary on which the Kerr family in its subextremal range defines a smooth two-parameter family of metrics. The differential structure of the smooth manifold is defined by what we shall call *fixed coordinates*, while the Kerr metric itself will be defined with the help of auxilliary coordinates depending on the parameters. We review this here:
### Fixed coordinates $(y^*,t^*,\theta^*,\phi^*)$
We define first the manifold-with-boundary $$\label{manwb}
\mathcal{R}=\mathbb R^+\times\mathbb R\times \mathbb S^ 2.$$ Fixed coordinates are just the standard $y^*\in\mathbb R^+$, $t^*\in\mathbb R$ and a choice of standard spherical coordinates $(\theta^*,\phi^*)\in\mathbb S^2$. Associated to this ambient differentiable structure are the *event horizon* $\mathcal{H}^+\doteq\partial\mathcal{R}=\{y^*=0\}$, the vector fields $T=\partial_{t^*}$, $\Phi=\partial_{\phi^*}$ and the one-parameter group of transormations $\varphi_\tau$ generated by $T$.
### Kerr-star coordinates $(r,t^*,\theta^*,\phi^*)$ {#kerrstardefsec}
We define a new coordinate system which depends on parameters $|a|<M$.
For each choice $|a|<M$, we first set $r_\pm=M\pm \sqrt{M^2-a^2}$ and then define a new coordinate $r$ which is related smoothly to $y^*$, depends smoothly on the parameters and such that, for fixed parameters, we have $r=r_+(a,M)$ on $\mathcal{H}^+$.[^6] Associated to these coordinates is the vector field $Z^*$, defined to be the smooth extension of the Kerr-star coordinate vector field $\partial_{r}$ to $\mathcal{R}$.
We will sometimes replace $r$ by a rescaled version, $r^*$, defined only in the interior of $\mathcal{R}$, by $$\label{r*def}
\frac{dr^*}{dr}=\frac{r^2+a^2}{\Delta}, \qquad r^*(3M)=0,$$ where $\Delta= (r-r_+)(r-r_-)$. Here we note that $\Delta$ vanishes to first order on $\mathcal{H}^+$, and the coordinate range $r>r_+$ corresponds to the range $r^*>-\infty$.
### Boyer-Lindquist coordinates $(r,t, \theta,\phi)$
We define a final coordinate system, again depending on a choice of fixed parameters $|a|<M$, by further transforming Kerr star coordinates, by defining $$t(t^*,r)= t^* - \bar t(r), \qquad
\phi(\phi^*,r)= \phi^*- \bar \phi(r) \mod 2\pi, \qquad
\theta=\theta^*$$ where $\bar{t}$ is a smooth function (see [@dr7] for details) chosen to satisfy $$\label{tnearh}
\bar t(r)=r^*(r)-r-r^*(9M/4)+9M/4, \qquad {\rm for}\qquad r_+\le r\le 15M/8,$$ $$\label{tnonearh}
\bar t(r)=0 \qquad {\rm for}\qquad r\ge 9M/4,$$ $$\label{forspacel}
\frac{d(r^*-\bar{t})}{dr}>0, \qquad 2-\left(1-\frac{2Mr}\rho^2\right)\frac{d(r^*-\bar{t})}{dr}>0.$$ Associated to these coordinates is the vector field $Z$ defined to be (the extension to ${\rm int}(\mathcal{R})$ of) the Boyer-Lindquist coordinate vector field $\partial_{r}$.[^7]
The Kerr metric and its properties
----------------------------------
Given these coordinate systems, we may now define the Kerr metric as a smooth $2$-parameter family on $\mathcal{R}$.
### Explicit form of the metric
For fixed parameters $|a|<M$, in addition to $\Delta$ above, let us first set $\rho^2=r^2+a^2\cos^2\theta$. The Kerr metric is then defined with respect to Boyer-Lindquist coordinates by $$\begin{aligned}
\label{eleme}
g_{a,M}=
-\frac{\Delta}{\rho^2}\left(dt-a\sin^2\theta d\phi\right)^2
+\frac{\rho^2}{\Delta}dr^2+\rho^2d\theta^2 +\frac{\sin^2\theta}{\rho^2}
\left(a\,dt-(r^2+a^2)d\phi\right)^2.\end{aligned}$$ Though a priori this is only well defined on ${\rm int}(\mathcal{R})$, by transforming the above into regular coordinates (see [@dr7]), one sees that the metric $(\ref{eleme})$ extends uniquely to the boundary so that for each $|a|<M$, indeed $(\mathcal{R},g_{a,M})$ defines a smooth Lorentzian manifold-with-boundary, and such that moreover the metric smoothly depends on the parameters $a$, $M$.[^8] These metrics are Ricci flat (i.e. they satisfy Einstein’s vacuum equations).
### Killing fields {#Killingfieldsec}
We note that the fixed vector fields $T$ and $\Phi$ on $\mathcal{R}$ defined in Section \[kerrstardefsec\] are Killing for $g_{a,M}$ for all parameter values $|a|<M$.
For each given $|a|<M$, the span of $T$ and $\Phi$ yields a timelike subspace of $T_p\mathcal{R}$ for all $p\in
{\rm int}(\mathcal{R})$ (in particular, $T$ is a timelike vector when $\Phi = 0$). The event horizon $\mathcal{H}^+=\partial\mathcal{R}$ is also a *Killing horizon*: the Killing field given by the linear combination $$K=T+ \upomega_+\Phi,$$ where $\upomega_+ \doteq \frac{a}{2Mr_+}$ is the “anuglar velocity” of the event horizon, is null and normal to $\mathcal{H}^+$; thus, $\mathcal{H}^+$ is in particular a null hypersurface. Note that along $\mathcal{H}^+$ we have $$\label{eutuxws}
\nabla_K K=
\kappa \, K,\qquad
\kappa= \frac{r_+-r_-}{2(r_+^2+a^2)}>0.$$ The quantity $\kappa$ is known as the *surface gravity*. The positivity $(\ref{eutuxws})$ is what determines the red-shift property, essential for our estimates (see Section \[Nmult\]). We note that $\kappa$ in fact vanishes in the extremal case $|a|=M$; this gives rise to the Aretakis instability [@aretakisHor].
We recall moreover that the vector $K$ restricted to $\mathcal{H}^+$ coincides with the smooth extension of the coordinate vector field $\partial_{r^*}$ of the $(r^*,t,\theta, \phi)$ coordinate system.
### The photon sphere and trapping parameters {#weritsdef}
It is well known that in the Schwarzschild case $a=0$, all future-trapped null geodesics asymptote to the timelike hypersurface $r=3M$.
In the statement of Theorem 1.1 of [@dr7], we defined $s_\pm(a_0,M)$ such that for all $|a|\le a_0$, then $r_+<3M-s_-(a_0,M)$ and all future trapped null geodesics enter the region $3M-s_-(a_0,M)< r< 3M-s_+(a_0,M)$. We have shown in Section 10.4 of [@stabi] the existence of such parameters again, for the full subextremal range $|a|<M$. We will repeat this proof in Section \[trappingParam\]. We note that in the extremal limit $a_0\to M$, $3M-s_-\to r_+(M,M)$.
Given the above parameters, let $\eta_{[3M-s^-,3M+s^+]}(r)$ denote the indicator function, and let us define, for each $a_0<M$, the function $$\label{degenerationfunc}
\zeta(r)= (1-3M/r)^2(1- \eta_{[3M-s^-,3M+s^+]}(r)).$$ This function will encode physical space degeneration of the “integrated local energy decay” estimate of Theorem \[theResult\]. The presence of the $(1-3M/r)^2$ factor ensures uniformity of the estimate as $a_0\to 0$ so as to retrieve our original Schwarzschild result [@dr3].
Finally, since it is derivatives with respect to the vector field $Z$ which do not degenerate at trapping, but it is the vector field $Z^*$ which extends to the horizon, it will be convenient to define a hybrid vector field that has both good properties. For this let us define, for each $|a|<M$, a cutoff funtion $\chi(r)$ such that $\chi=1$ for $r\ge 3M-s^{-}$ and $\chi=0$ for $r\le (r_++3M-s^-)/2$, We define then a new vector field $\tilde{Z}^*=\chi Z+(1-\chi)Z^*$. This will be the vector field which appears in the statement of Theorem \[theResult\].
### The ergoregion
The region $\mathcal{S} \subset \mathcal{R}$ where $T$ is spacelike is known as the *ergoregion*; more explicitly, it is exactly the subset of $\mathcal{R}$ defined by $$\label{ergoregionS}
\mathcal{S}=\{ \Delta - a^2\sin^2\theta < 0\}.$$ The boundary $\partial\mathcal{S}$ is called the *ergosphere*.
### The $\Sigma_\tau$ hypersurfaces, and the regions $\mathcal{R}_{(0,\tau)}$, $\mathcal{H}^+_{(0,\tau)}$ {#hypersurfaceDef}
We have arranged the definition of Kerr-star coordinates in Section \[kerrstardefsec\] so that the hypersurfaces $t^*=c$ are spacelike (see the conditions $(\ref{forspacel})$) with respect to the metric $g_{a,M}$, for all values of parameters $|a|<M$.
In the region $r\le 15M/8$, we have in fact $$g(\nabla t^*, \nabla t^*)= -1-\frac{2Mr}{\rho^2}.$$
We will define $$\Sigma_\tau=\{t^*=\tau\},$$ $$\mathcal{R}_{(0,\tau)}=\cup_{0\le \tau^*\le\tau} \Sigma_{\tau^*}$$ and $$\mathcal{R}_0=\cup \mathcal{R}_{(0,\tau)}.$$ Note that $\Sigma_0$ is a past Cauchy hypersurface for the regions $\mathcal{R}_{(0,\tau)}$, $\mathcal{R}_0$. Let us also define $$\mathcal{H}^+_{(0,\tau)} = \mathcal{R}_{(0,\tau)}\cap\mathcal{H}^+,
\qquad
\mathcal{H}^+_0 = \mathcal{R}_{0}\cap\mathcal{H}^+.$$
### Angular derivatives and the volume form {#usefulcomps}
For future reference, let us introduce here the notation ${\mbox{$g \mkern-9mu /$\,}}$, ${\mbox{$\nabla \mkern-13mu /$\,}}$ to denote the induced metric and covariant derivative from $g_{a,M}$ (\[eleme\]) on the $\mathbb S^2$ factors of $\mathcal{R}$ in the product $(\ref{manwb})$.
We record finally from [@dr7] some useful properties of the volume form $dV$ of the metric $g_{a,M}$: With respect to Boyer-Lindquist coordinates, we have $$dV= v(r,\theta) \, dt\, dr\, dV_{{\mbox{$g \mkern-9mu /$\,}}} \qquad \text{\rm with\ } v\sim 1$$ whereas using the alternative $r^*$ coordinate, $$dV =v(r^*,\theta) \, dt\, dr^*\, dV_{{\mbox{$g \mkern-9mu /$\,}}} \qquad \text{\rm with\ } v\sim \Delta/r^2.$$ With respect to Kerr-star coordinates, we have $$dV =v(r,\theta^*) \, dt^*\, dr\, dV_{{\mbox{$g \mkern-9mu /$\,}}} \qquad \text{\rm with\ } v\sim 1.$$
Let $\gamma$ denote the standard unit metric on the sphere in $(\theta,\phi)$ coordinates. We have that ${\mbox{$g \mkern-9mu /$\,}}\sim r^2\gamma$, and thus we may replace $dV_{{\mbox{$g \mkern-9mu /$\,}}}$ in the above using $$dV_{{\mbox{$g \mkern-9mu /$\,}}} = v(r,\theta)\, r^2\sin\theta\, d\theta \,d\phi
\qquad \text{\rm with\ } v\sim
1.$$ Finally, we note that $$\label{easyCoArea}
dV \sim d\tau\, dV_{\Sigma_{\tau}}.$$
For $a_0<M$ and $|a|\le a_0$, note that the implicit constants in the above are uniformly bounded, depending only on $a_0$ and $M$.
Multiplier currents and the general energy identity {#multsandcomts}
---------------------------------------------------
We shall repeat our standard notation for vector field multiplier current identities associated to “multiplier” vector fields $V$ which will be applied to $\psi$ as well as to $\Xi\psi$ for various commutation vector fields $\Xi$. See [@dr7] for more details and [@book2] for a systematic discussion. See [@muchT] for an early application of non-trivial energy currents to the problem of decay for the wave equation on Minkowski space.
### Currents
Given a general Lorentzian manifold $(\mathcal{M},g)$, let $\Psi$ be a sufficiently regular function. We define $${\bf T}_{\mu\nu}[\Psi]
\doteq \partial_\mu\Psi\partial_\nu\Psi -\frac12 g_{\mu\nu}g^{\alpha\beta}
\partial_\alpha\Psi \partial_\beta\Psi.$$ Given a sufficiently regular vector field $V_\mu$ and function $w$ on $\mathcal{M}$, we will define the currents $${\bf J}^V_\mu[\Psi] = {\bf T}_{\mu\nu}[\Psi] V^\nu, \qquad
{\bf J}^{V,w}_\mu[\Psi] =
{\bf J}_\mu^V[\Psi]+\frac18w\partial_\mu (\Psi^2)-\frac18(\partial_\mu w)\Psi^2,$$ $${\bf K}^V[\Psi] ={\bf T}_{\mu\nu}[\Psi]\nabla^\mu V^\nu, \qquad
{\bf K}^{V,w}[\Psi] = {\bf K}
^V[\Psi] -\frac18\Box_gw (\Psi^2) +\frac14w \nabla^\alpha\Psi\nabla_\alpha\Psi,$$ $$\mathcal{E}^V[\Psi] = -(\Box_g\Psi)V^\nu \Psi_{,v}, \qquad
\mathcal{E}^{V,w}[\Psi]= \mathcal{E}^V(\Psi)-\frac14w\Psi \Box_g\Psi.$$
Note that even if one is only interested in the study of solutions $\psi$ to the homogeneous (\[WAVE\]), inhomogeneous terms will arise from applying cutoffs to $\psi$ and also from applying commutation vector fields (like vector field $Y$ from Section \[Nmult\] below) which do commute with $\Box_g$.
### The divergence identity
The divergence identity between two homologous spacelike hypersurfaces $S^-$, $S^+$, bounding a region $\mathcal{B}$, with $S^+$ in the future of $S^-$, yields $$\label{ingeneralform}
\int_{S^+}{\bf J}^V_\mu [\Psi]n^\mu_{S^+}+
\int_{\mathcal{B}} ({\bf K}^V[\Psi] + \mathcal{E}^V[\Psi])
=\int_{S^-}{\bf J}^V_\mu [\Psi]n^\mu_{S^-},$$ where $n_{\Sigma_i}$ denotes the future directed timelike unit normal. The induced volume forms are to be understood. A similar identity holds for the ${\bf J}^{V,w}_\mu$ currents, etc.
We shall typically apply $(\ref{ingeneralform})$ for the Kerr metric $g_{a,M}$ in the case where $S^-=\Sigma_0$ and $S^+=\Sigma_\tau \cup \mathcal{H}^+_{(0,\tau)}$ and $\Psi$ is compactly supported in $\mathcal{R}_{(0,\tau)}$ to obtain $$\label{ingeneralform2}
\int_{\Sigma_\tau}{\bf J}^V_\mu [\Psi]n^\mu_{\Sigma_\tau}+
\int_{\mathcal{H}^+_{(0,\tau)}}{\bf J}^V_\mu [\Psi]n^\mu_{\mathcal{H}^+}+
\int_{\mathcal{R}_{(0,\tau)}} ({\bf K}^V[\Psi] + \mathcal{E}^V[\Psi])
=\int_{\Sigma_0}{\bf J}^V_\mu [\Psi]n^\mu_{\Sigma_0}.$$ Let us note that the compactness of the support justifies the absence of an additional boundary term even though $S^\pm$ are not homologous. Since $\mathcal{H}^+$ is null, its induced normal form is coupled to the choice of $n^\mu_{\mathcal{H}^+}$. In writing the above, we shall assume such a choice has been made such that the formula indeed holds.
### Superradiance in Kerr {#supInKerr}
As already mentioned in the introduction, the presence of the ergoregion $\mathcal{S}$ is one of the fundamental difficulties associated with the passage from Schwarzschild to a rotating Kerr spacetime. One particular consequence is that for $a\ne 0$, the conserved $\mathbf{J}^T_{\mu}[\psi]$ energy flux for a solution to (\[WAVE\]) may be negative on the horizon $\mathcal{H}^+$. Hence, applying (\[ingeneralform2\]), the energy on $\Sigma_{\tau}$ can be larger than the energy on $\Sigma_0$; this phenomenon is known as *superradiance*.[^9]
An explicit computation in (\[ingeneralform2\]) shows the $\mathbf{J}^T_{\mu}[\psi]$ energy flux along $\mathcal{H}^+(0,\infty)$ is given by $$\int_{\mathcal{H}^+(0,\infty)}\text{Re}\left(T\psi\overline{\left(T\psi + \upomega_+\Phi\psi\right)}\right),$$ where $\upomega_+$ was defined in Section \[Killingfieldsec\]. In particular, if one formally considers a (complex-valued) solution of the form $$\psi(t^*,r,\theta,\phi^*) = e^{-i\omega t^*}e^{im\phi^*}\psi_0\left(r,\theta\right),$$ then the sign of the $\mathbf{J}^T_{\mu}[\psi]$ flux on the horizon is determined by the sign of $$\omega\left(\omega - \upomega_+m\right).$$ Thus, we say that the parameters $\omega$ and $m$ are superradiant if $$\label{superradiantParam}
\omega\left(\omega - \upomega_+m\right) < 0.$$ Observe that in the case $a \geq 0$, the condition (\[superradiantParam\]) is equivalent to $$\label{superradiantParamPosa}
m\omega \in \left(0,\frac{am^2}{2Mr_+}\right]$$ We will return to a discussion of the significance of this frequency range in Section \[Vsrsec\].
The main theorems {#statementsections}
=================
With the notations of Section \[notats\] we may now give precise statements of the results.
Boundedness and integrated local energy decay {#boundedandiled}
---------------------------------------------
Recall the notations of Section \[notats\], in particular the hypersurfaces $\Sigma_\tau$, the region $\mathcal{R}_0$, the vector fields $T$, $\tilde{Z}^*$ and the degeneration function $\zeta$ defined in $(\ref{degenerationfunc})$. Let $n^\mu_{\Sigma_\tau}$, $n^\mu_{\mathcal{H}^+}$ denote the corresponding normals. The vector field $N$ below can be taken[^10] to be $n_{\Sigma_\tau}$, thought of now as a smooth vectorfield on $\mathcal{R}$.
The main theorem of the present paper is
\[theResult\] Fix $M>0$, $0\le a_0<M$ and $\delta>0$. There exists a constant $C=C(a_0,M,\delta)$ such that for all $|a|\le a_0$, and all sufficiently regular solutions $\psi$ of the wave equation $\Box_{g_{a,M}}\psi = 0$ on $\mathcal{R}_0$, the following estimates hold: $$\label{protasn1b}
\int_{\mathcal{R}_0}\Big(r^{-1}\zeta |{\mbox{$\nabla \mkern-13mu /$\,}}\psi|^2+r^{-1-\delta}\zeta (T\psi)^2+r^{-1-\delta}(\tilde Z^*\psi)^2+ r^{-3-\delta} (\psi-\psi_\infty)^2\Big) \le C
\int_{\Sigma_0} {\bf J}_\mu^N[\psi]n^\mu_{\Sigma_0},$$ $$\label{fluxho...}
\int_{\mathcal{H}^+_0}\left( {\bf J}^N_\mu[\psi]n^\mu_{\mathcal{H}^+}
+(\psi-\psi_{\infty})^2\right)
\le C \int_{\Sigma_0} {\bf J}_\mu^N[\psi]n^\mu_{\Sigma_0},$$ $$\label{fluxNullInf}
\int_{\mathcal{I}^+}{\bf J}^T_\mu[\psi]n^\mu_{\mathcal{I}^+}
\le C \int_{\Sigma_0} {\bf J}_\mu^N[\psi]n^\mu_{\Sigma_0},$$ $$\label{bndts1b}
\int_{\Sigma_\tau} {\bf J}_\mu^N[\psi]
n^\mu_{\Sigma_\tau}\le C \int_{\Sigma_0} {\bf J}_\mu^N[\psi]n^\mu_{\Sigma_0}, \qquad
\forall\tau\ge 0,$$ where $4\pi\psi_\infty^2= \lim_{r'\to\infty}\int_{\Sigma_0\cap\{r=r'\}} r^{-2}\psi^2$.
Estimate $(\ref{protasn1b})$ is an [**integrated local energy decay**]{} statement degenerating at trapping. The full statement obtained in the proof is more precise but cannot be expressed in physical space; see Proposition \[preciseILED\] of Section \[precise\].
Estimate $(\ref{fluxho...})$ is the [**boundedness of the energy flux through the event horizon**]{} $\mathcal{H}^+_0$ (as measured by a local observer), while estimate $(\ref{fluxNullInf})$ is the [**boundedness of the energy flux to null infinity**]{} $\mathcal{I}^+$. (The latter will be explained in Section \[toayplus\].) These two estimates are obtained concurrently with $(\ref{protasn1b})$.
Estimate $(\ref{bndts1b})$ is the statement of [**uniform energy boundedness through the foliation**]{} $\Sigma_\tau$. We note that the proof of this statement, which is obtained a posteriori, requires the more precise version of $(\ref{protasn1b})$ given in Proposition \[preciseILED\]. Note that $$\label{sobol1}
\int_{\Sigma_\tau} {\bf J}_\mu^N[\psi]n^\mu_{\Sigma_\tau}\sim
\| \psi\|^2_{\mathring{H}^1(\Sigma_\tau)} +\|n_{\Sigma_\tau}\psi\|^2_{L^2(\Sigma_\tau)}
\sim \int_{\theta,\phi^*} \int_{r_+}^\infty
( |\partial_{t^*}\psi|^2 + |\partial_{r}\psi|^2 +|{\mbox{$\nabla \mkern-13mu /$\,}}\psi|^2_{{\mbox{$g \mkern-9mu /$\,}}})\, dr\, dV_{{\mbox{$g \mkern-9mu /$\,}}}$$ with respect to coordinates $(t^*, r, \theta, \phi^*)$, where here $f\left(\psi\right) \sim g\left(\psi\right)$ means there exist constants $c$ and $C$ not depending on $\psi$ such that $cg\left(\psi\right) \leq f\left(\psi\right) \leq Cg\left(\psi\right)$. Thus, $(\ref{bndts1b})$ gives uniform geometric $\mathring{H}^1$ bounds on the solution.
The reader familiar with Penrose-diagrammatic notation may find the following useful $$\PandocStartInclude{partiii.pstex_t}\PandocEndInclude{input}{1453}{24}$$
The higher order statement
--------------------------
For various applications, it is essential to have a higher-order analogue of the above. This is given by
\[h.o.s.\] Let $M$, $a_0$, $a$ be as in Theorem \[theResult\]. Then, for all $\delta>0$ and all integers $j\ge 1$, there exists a constant $C=C(a_0, M, \delta,j)$ such that the following inequalities hold for all sufficiently regular solutions $\psi$ to the wave equation $\Box_{g_{a,M}}\psi = 0$ on $\mathcal{R}_0$ $$\begin{aligned}
\nonumber
\label{protasn1}
\int_{\mathcal{R}_0} &
r^{-1-\delta}\zeta \sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\\
\nonumber
&+r^{-1-\delta}\sum_{1\le i_1+i_2+i_3\le j-1}
\left(|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3+1}\psi|^2+|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(Z^*)^{i_3}\psi|^2\right)\\
&\le C\int_{\Sigma_0} \sum_{0\le i \le j-1}
{\bf J}^N_\mu[N^{i}\psi]n^\mu_{\Sigma_0},\end{aligned}$$ $$\label{bndts2}
\int_{\mathcal{H}^+_0} \sum_{0\le i \le j-1}{\bf J}^N_\mu[N^{i}\psi]n^\mu_{\mathcal{H}^+}
\le C\int_{\Sigma_0} {\sum_{0\le i \le j-1}
{\bf J}^N_\mu[N^{i}\psi]n^\mu_{\Sigma_0}},$$ $$\label{fluxNullInfHigher}
\int_{\mathcal{I}^+}\sum_{0\le i \le j-1}{\bf J}^T_\mu[N^{i}\psi]n^\mu_{\mathcal{I}^+}
\le C \int_{\Sigma_0} \sum_{0\le i \le j-1}{\bf J}_\mu^N[N^i\psi]n^\mu_{\Sigma_0},$$ $$\label{bndts1}
\int_{\Sigma_\tau } \sum_{0\le i \le j-1}{\bf J}^N_\mu[N^{i}\psi]n^\mu_{\Sigma_\tau}
\le C\int_{\Sigma_0} {\sum_{0\le i \le j-1}
{\bf J}^N_\mu[N^{i}\psi]n^\mu_{\Sigma_0}}, \qquad
\forall\tau\ge 0.$$
Let us note that by an elliptic estimate, we have $$\label{inasyflatcas}
\int_{\Sigma_\tau } \sum_{0\le i \le j-1}{\bf J}^N_\mu[N^{i}\psi]n^\mu_{\Sigma_\tau}
\sim \sum_{1\le i\le j} \|\psi\|^2_{\mathring{H}^i(\Sigma_\tau)} +\|n_{\Sigma_\tau}\psi\|^2_{\mathring{H}^{i-1}(\Sigma_\tau)}.$$ Thus, as with $(\ref{bndts1b})$ before, we may reexpress statement $(\ref{bndts1})$ of the above theorem as the statement of the uniform boundedness of geometric Sobolev norms. Note that [**uniform pointwise bounds**]{} on $|\psi|$ and its derivatives to *arbitrary* order $|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|$ in $\mathcal{R}_0$ follow as an immediate consequence of the above Theorems in view of the Sobolev inequality applied on each $\Sigma_\tau$.
The above theorems also imply pointwise decay statements and decay for the energy flux through suitable hypersurfaces. We turn to this now.
Corollaries {#corollariessec}
-----------
Let us note first that by a reduction proven as Proposition 4.6.1 of [@dr7], Theorems \[theResult\] and \[h.o.s.\] hold where $\Sigma_0$ is replaced by an arbitrary “admissible” hypersurface $\widetilde{\Sigma}_0$ (see Section 4.4 of [@dr7] for this notion), $\Sigma_\tau$ is replaced by $\widetilde{\Sigma}_\tau\doteq {\varphi_\tau}(\widetilde\Sigma_0)$, $n_{\Sigma_\tau}$ is replaced by $n_{\widetilde{\Sigma}_\tau}$, $\mathcal{R}_0$, $\mathcal{H}^+_0$ are redefined as $D^+(\widetilde\Sigma_0)$, $D^+(\widetilde\Sigma_0)
\cap\mathcal{H}^+$, respectively, and $N$ is kept as is. This notion includes both asymptotically flat hypersurfaces terminating at spatial infinity (a special case of admissible hypersurfaces of the first kind) and asymptotically hyperboloidal hypersurfaces terminating at null infinity (a special case of admissible hypersurface of the second kind). The latter case is depicted below $$\PandocStartInclude{partiii2.pstex_t}\PandocEndInclude{input}{1549}{25}$$ Note, however, that in the latter case, $(\ref{inasyflatcas})$ (with the above substitutions) will never hold. It is for this reason that we prefer to state Theorems \[theResult\] and \[h.o.s.\] in the form given.
As a consequence of this more general statement, the above theorems allow us to apply our “black box” result of [@icmp] (see [@schlue] and [@moschidis] for detailed treatments). We obtain
\[thecorol\] Let $a_0$, $M$, $a$, $\delta$ be as in Theorems \[theResult\]–\[h.o.s.\], and let $R>r_+$. Let $\widetilde{\Sigma}_0$ be an asymptotically hyperboloidal hypersurface terminating at null infinity, and denote $\widetilde{\Sigma}_\tau= \varphi_\tau(\widetilde{\Sigma}_0)$. Then for sufficiently regular solutions of the wave equation, we have the following estimates for the energy flux $$\int_{\widetilde{\Sigma}_\tau} {\bf J}^N_\mu [\psi] n^\mu_{\widetilde{\Sigma}_\tau} \le
C(a_0,M) E\tau^{-2}$$ $$\int_{\widetilde{\Sigma}_\tau\cap \{r\le R\}} {\bf J}^N_\mu [N\psi] n^\mu_{\widetilde{\Sigma}_\tau}
\le
C(a_0, M,\delta, R) E\tau^{-4+2\delta}$$ and the following pointwise estimates $$\sup_{\widetilde{\Sigma}_\tau}
r
|\psi-\psi_\infty|\le C(a_0, M) \sqrt{E}\, \tau^{-1/2},$$ $$\label{pdecay1}
\sup_{\widetilde{\Sigma}_\tau\cap\{r\le R\}} |\psi-\psi_{\infty}|\le C(a_0, M, \delta, R) \sqrt{E} \tau^{-3/2+\delta},$$ $$\label{pdecay2}
\sup_{\widetilde\Sigma_\tau\cap\{r\le R\}}|n_{\widetilde\Sigma}\psi|+
|\nabla_{\widetilde\Sigma}\psi|
\le C(a_0, M, \delta, R)E \tau^{-2+\delta},$$ where in each inequality, $E$ denotes an appropriate higher order weighted energy on $\widetilde{\Sigma}_0$ (or alternatively on an asymptotically flat $\Sigma_0$ in the past of $\widetilde{\Sigma}_0$).
From the point of view of nonlinear applications, the main significance of the powers on the right hand side of $(\ref{pdecay1})$ and $(\ref{pdecay2})$ is that they are integrable in time.
The logic of the proof {#theLogic}
----------------------
Now that we have given precise formulations of the main theorems, we will give a brief account of the logic of the proof, highlighting where each statement is proven.
The reader may wish to refer back to the outline of Section \[outlinesection\]. Recall that Section \[prelimn\] concerns various preliminary propositions, including a reduction (in Section \[WPosed\]) to considering $\psi$ arising from smooth compactly supported data on $\Sigma_0$, whereas Section \[cartersupersection\] defines a class of functions for which Carter’s separation to the radial o.d.e. (\[e3iswsntouu\]) and appropriate boundary conditions can be justified *a priori*. Sections \[Vpropsec\], \[sct\] and \[freqLocEst\], on the other hand, are logically independent of the rest of the paper; they are concerned with the study of classical solutions $u$ to the o.d.e. (\[e3iswsntouu\]) assumed to satisfy appropriate boundary conditions. The culmination is Theorem \[phaseSpaceILED\] which establishes estimates on $u$ independent of the frequency parameters in the potential.
The logic of the proof of Theorem \[theResult\] can be properly thought to commence in Section \[summation\]. We define a class of solutions to (\[WAVE\]) which we call “future-integrable” and which allows us to apply Carter’s separation of Section \[cartersupersection\] to a suitably defined function, with the help of a cutoff. We then apply Theorem \[phaseSpaceILED\] to the resulting $u$. Summing via Plancherel, and using in addition the preliminary propositions of Section \[prelimn\] and the refined mode stability of [@shlapRot], we establish in Proposition \[closedILED\] the integrated energy decay statement (\[protasn1b\]), the horizon energy flux bound (\[fluxho...\]) and the null infinity flux bound (\[fluxNullInf\]) *for this class of future-integrable solutions to (\[WAVE\]).*
In Proposition \[h.o.s.suff\] we will upgrade these to the higher order statements (\[protasn1\]), (\[fluxNullInfHigher\]) and (\[bndts1\]) of Theorem \[h.o.s.\], *again for the class of future-integrable solutions*.
Next, in Proposition \[allIntegrable\] we will use a continuity argument to show that solutions to (\[WAVE\]) arising from smooth compactly supported data (according to the reduction of Section \[WPosed\]) are future-integrable. We thus unconditionally obtain the statements (\[protasn1b\]), (\[fluxho...\]), (\[fluxNullInf\]), (\[protasn1\]), (\[bndts2\]) and (\[fluxNullInfHigher\]).
Finally, in Proposition \[sufficientIntBound\] we unconditionally establish the statements (\[bndts1b\]) and (\[bndts1\]). This will complete the proof of Theorems \[theResult\] and \[h.o.s.\].
Preliminaries {#prelimn}
=============
Well posedness, regularity and smooth dependence {#WPosed}
------------------------------------------------
Let us note that the wave equation $(\ref{WAVE})$ is well posed in $\mathcal{R}_0$ with initial data $(\uppsi, \uppsi')$ defined on $\Sigma_0$ in $H^j_{\rm loc}(\Sigma_0)\times H^{j-1}_{\rm loc}(\Sigma_0)$ (cf. Proposition 4.5.1 of [@dr7]). Moreover, if the initial data are smooth and of compact support on $\Sigma_0$, then $\psi$ will be smooth, and of compact support on all $\Sigma_\tau$ for $\tau\ge 0$.
In the proof of our theorems, by standard density arguments (applied to $\psi-\psi_{\infty}$), we may thus assume that $\psi$ indeed arises from such data and thus is smooth and of compact support for fixed $\Sigma_\tau$ for all $\tau\ge 0$.
Lastly, we observe that the solution $\psi$ to $(\ref{WAVE})$ depends smoothly on $a$, e.g.
\[lemmaSmoothDependence\]Let $\left|a_{\infty}\right| < M$, $\{a_k\}_{k=1}^{\infty}$ satisfy $a_k \to a_{\infty}$, $\Box_{g_{a_k,M}}\psi_k = 0$, $\psi_k|_{\Sigma_0} = \psi_{\infty}|_{\Sigma_0}$ and $n_{\Sigma_0}\psi_k|_{\Sigma_0} = n_{\Sigma_0}\psi_{\infty}|_{\Sigma_0}$. Then, for every $j \geq 1$ and $\tau \geq 0$, $$\lim_{k\to\infty}\int_{\Sigma_{\tau}}\sum_{1\leq i_1 + i_2 + i_3 \leq j}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}\left(\tilde Z^*\right)^{i_3}\psi_k\right|^2 = \int_{\Sigma_{\tau}}\sum_{1\leq i_1 + i_2 + i_3 \leq j}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}\left(\tilde Z^*\right)^{i_3}\psi_{\infty}\right|^2.$$
(We shall appeal to the above lemma at the end of Section \[closednesssec\] in the context of the closedness part of our continuity argument.)
The sign of $a$ {#signOfa}
---------------
For given $a$, $M$, given a solution $\psi$ of $\Box_{g_{M,a}}\psi=0$, then, defining $\tilde\psi(y^*, t^*,\theta^*, \phi^*)=\psi(y^*,t^*,\theta^*,2\pi-\phi^*)$, we have that $\tilde\psi$ satisfies $\Box_{g_{M,-a}}\tilde\psi=0$. Moreover, the estimates of Theorems \[theResult\] and \[h.o.s.\] for $\tilde\psi$ with quantities defined respect to the metric $g_{M,-a}$ are equivalent to the analogous estimates for $\psi$ with respect to the metric $g_{M,a}$. Thus, it suffices to prove our Theorems for $a \geq 0$. This reduction is of no conceptual significance, but it slightly simplifies the notation for discussing the superradiant frequency range, which then can be given by $(\ref{superradiantParamPosa})$. For notational convenience we will indeed use the reduction to $a \geq 0$ in Sections \[Vpropsec\]–\[freqLocEst\] the context of describing the properties of the potential $V$ in various frequency regimes and defining the frequency dependent multiplier currents. The reader can assume that $a\ge 0$ globally in this paper, but it is strictly speaking only necessary for those statements which refer explicitly to frequency-dependent functions in the separation.
Hardy inequalities
------------------
As in the previous parts [@dr7] of this series, at various points we shall refer to Hardy inequalities. In view of our comments concerning the volume form (see Section \[usefulcomps\]), the reader can easily derive these from the one-dimensional inequalities $$\label{hardy1}
\int_{0}^2 x^{-1} |\log x|^{-2} f^2(x) \le C\int_{0}^{2} \left(\frac{df}{dx}\right)^2(x)dx+
C\int_1^2f^2(x) dx,$$ $$\label{hardy2}
\int_1^\infty f^2(x)\le C\int_1^\infty x^2\left(\frac{df}{dx}\right)^2 (x)dx,$$ where the latter holds for functions $f$ of compact support.
Generic constants in inequalities and fixed parameters {#genconstsec}
------------------------------------------------------
Let us recall our conventions from [@dr7] regarding constants depending on the Kerr geometry.
As in the statement of Theorem \[theResult\], all propositions in this paper providing estimates will explicitly refer to two fixed parameters $a_0<M$ delineating the range of Kerr parameters allowed. In the context of inequalities, we shall denote by $B$ potentially large positive constants, whereas we shall denote by $b$ potentially small positive constants, . *This dependence is always to be understood.* We record the resulting algebra of constants: $$b+b=b, \, B+B=B, \, B\cdot B=B, \, B^{-1}=b, \ldots$$ We note that these constants will often blow up $B\to\infty$, $b^{-1}\to\infty$ in the extremal limit $a_0\to M$.
Our constructions will depend on various additional parameters, for instance, the parameters $\omega_{\text{high}}$, $E$, etc., which are free in the statements of Propositions \[odeEst1\], etc., but are chosen by the end of the proof of Theorem \[phaseSpaceILED\], in whose statement they appear as fixed parameters.
When a parameter is required to be “sufficiently large” or “sufficiently small” without further clarification, this always means that there exists a constant such that the parameter can be taken to be an arbitrary value bigger than that constant. If a parameter is required to be “sufficiently large” *given another parameter*, this means that there again exists such a constant depending on $a_0$ and $M$ the other parameter.
Until a parameter has been fixed, e.g. the parameter $\omega_{\text{high}}$, we shall use the notation $B(\omega_{\text{high}})$, etc., in the context of inequalities, to denote constants depending on $\omega_{\text{high}}$ *in addition to $M$ and $a_0$*. For a parameter, say $c$ which is an explicit function of other parameter(s), say $\omega_{\text{high}}$, together with $M$ and $a_0$, we will write $c(\omega_{\text{high}})$. Again, the dependence on $M$ and $a_0$ is to be understood.
The final choices of all initially free parameters used in the present paper will be made to depend only on $M$ and $a_0$. Once such choices are made, $B(\omega_{\text{high}})$ is replaced by $B$, following our conventions.
The red-shift {#Nmult}
-------------
Understanding the red-shift is an essential part of the dynamics. Definitive constructions have been given in Section 7 of [@jnotes]. These depend only on the positivity of the surface gravity $\kappa$, recalled in Section \[Killingfieldsec\].
### The vectorfield $N$
Let us recall briefly from [@dr7] the construction of a vector field $N$ capturing the red-shift effect.
\[specialises..\] Let $|a|\le a_0<M$, $g_{a, M}$ be the Kerr metric and $\mathcal{R}$, etc., be as before. There exist positive constants $b$ and $B$, parameters $r_1(a,M)>r_{\rm red}(a,M)>r_+$ and a $\varphi_\tau$-invariant timelike vector field $N=N(a,M)$ on $\mathcal{R}$, normalised so that $N - K$ is future oriented, traverse to $\mathcal{H}^+$, and null with $g(N,K)=-2$, such that
1. \[fir\] ${\bf K}^N[\Psi] \ge b\, {\bf J}^N_\mu [\Psi] N^\mu$ for $r\le r_{\rm red}$
2. \[item2\] $-{\bf K}^N[\Psi] \le B\, {\bf J}^N_\mu [\Psi] N^\mu$, for $r\ge r_{\rm red}$
3. \[lastitem\] $T= N$ for $r\ge r_1$,
where the currents are defined with respect to $g_{a,M}$.
Note the implicit $a_0$ and $M$ dependence of constants $b$ and $B$ as described in Section \[genconstsec\] above. This proposition would fail in the case $a_0=M$. See [@aretakisKerr; @Sbierski].
### The red-shift estimate
The above leads immediately to the following estimate (see [@dr7])
\[ftrs\] Let $g=g_{a,M}$ for $|a|\le a_0<M$, and let $r_{\rm red}$ be as in the above Proposition. Then the following is true. For all $r_+\le \tilde{r}\le r_{\rm red}$ and $\tilde\delta>0$, there exists a positive constant $B(\tilde{r},\tilde\delta)$, such that for all functions $\Psi$ on $\mathcal{R}_0$, then $$\begin{aligned}
\int_{\mathcal{R}_{(0,\tau)}\cap\{r\le \tilde{r}\}} &
({\bf J}_\mu^N[\Psi]N^\mu +|\log(|r-r_+|)^{-2}||r-r_+|^{-1}\Psi^2)+\int_{\mathcal{H}^+_{(0,\tau)}}
{\bf J}^N_\mu[\Psi]n_{\mathcal{H}^+}^\mu
+\int_{\Sigma_\tau\cap\{r\le \tilde{r}\}}{\bf J}^N_\mu[\Psi] n^\mu \\
&\le B(\tilde{r},\tilde\delta)\int_{\Sigma_0} {\bf J}^N_\mu[\Psi] n^\mu + B(\tilde{r},\tilde\delta)
\int_{\mathcal{R}_{(0,\tau)}\cap \{\tilde{r}\le r\le \tilde{r}+\tilde{\delta}
\}}
({\bf J}_\mu^N[\Psi]N^\mu
+\Psi^2)-\mathcal{E}^N\left[\Psi\right].\end{aligned}$$
Again, recall that the additional dependence of $B$ on $M$ and $a_0$ is now implicit according to our conventions. Note that the proof of this estimate uses the Hardy inequality $(\ref{hardy1})$, so as to include the useful zeroth order term on the left hand side. We note that the same estimate holds with the above zeroth order terms removed from both the right and the left hand sides.
### Red-shift commutation and the vector field $Y$ {#rscsec}
We specialise Theorem 7.2 of [@jnotes] to the Kerr case.
\[commuprop\] Let $g=g_{a,M}$, let $K$ be the vector field of Section \[Killingfieldsec\], let $Y=N-K$, and let $E_1$, $E_2$ be $\varphi_\tau$-invariant vector fields such that $\{K, Y, E_1, E_2\}$ form a local null frame on $\mathcal{H}^+$. Then for all $k\ge 0$ and multi-indices ${\bf m}=(m_1,m_2,m_3,m_4)$, $$\Box_g(Y^k\Psi)=\kappa_k Y^{k+1}\Psi + \sum_{|{\bf m}|\le k+1, m_4\le k} c_{{\bf m}}
E_1^{m_1}E_2^{m_2}L^{m_3}Y^{m_4}\Psi +Y^k(\Box_g\Psi)$$ where $\kappa_k>0$ and the $c_{\bf m}$ are smooth $\varphi_\tau$-invariant functions.
The above proposition, which is another manifestation of the red-shift effect, effectively allows us not only to apply a transversal vector field to the horizon as a multiplier, but also as a commutation vector field. This is fundamental for retrieving higher order statements as in Theorem \[h.o.s.\].
An estimate for large $r$ {#largeR}
-------------------------
We will also need the following estimate.
\[lrp\] Fix $M > 0$ and $a_0 < M$. For each $\delta>0$, there exist positive values $2M < \tilde{R} <R_{\rm large}$, and positive constants $B(\delta)$ such that if $\left|a\right| \leq a_0$, $\psi$ denotes a solution of $(\ref{WAVE})$ and $\psi_{\infty}=0$, then for all $\tau\ge 0$ $$\begin{aligned}
\int_{\mathcal{R}_{(0,\tau)}\cap\{r\ge R_{\rm large}\}}& r^{-1}( r^{-\delta}|\partial_r\psi|^2
+r^{-\delta}|\partial_t\psi|^2+|{\mbox{$\nabla \mkern-13mu /$\,}}\psi|^2_{{\mbox{$g \mkern-9mu /$\,}}}+
r^{-2-\delta}\psi^2)\\
\le& B(\delta)\int_{\Sigma_0} {\bf J}^{N}_\mu[\psi] n^\mu_{\Sigma_0}+
B(\delta)\int_{\Sigma_\tau} {\bf J}^{N}_\mu[\psi] n^\mu_{\Sigma_\tau} + B(\delta)\int_{\mathcal{R}_{(0,\tau)}\cap\{ r \ge \tilde{R}\}}r^{1+\delta}\left|F\right|^2
\\
&+ B(\delta)
\int_{\mathcal{R}_{(0,\tau)}\cap\{\tilde{R} \le r\le R_{\rm large}\}}
( |\partial_r\psi|^2 +|\partial_t\psi|^2+|{\mbox{$\nabla \mkern-13mu /$\,}}\psi|^2_{{\mbox{$g \mkern-9mu /$\,}}}
+\psi^2).\end{aligned}$$
The homogeneous case is treated in [@dr7], and the inhomogeneous case follows by the same argument.
A timelike vector field
-----------------------
We have noted in Section \[Killingfieldsec\] that the span of $T$ and $\Phi$ is a null subspace on the horizon $\mathcal{H}^+$ and a timelike subspace on $\mathcal{R}\setminus\mathcal{H}^+$. These statements are in particular implied by the following Lemma, which will be important later.
\[timelikeVector\] The vector field $$T + \frac{2Mar}{\left(r^2+a^2\right)^2}\Phi$$ is timelike in $\mathcal{R}\setminus\mathcal{H}^+$ and null on $\mathcal{H}^+$.
On the horizon $\mathcal{H}^+$, the vector field reduces to $$T + \frac{a}{2Mr_+}\Phi,$$ which can immediately be seen to be its standard null generator.
Off the horizon, computing $g\left(T + \frac{2Mar}{\left(r^2+a^2\right)^2}\Phi,T + \frac{2Mar}{\left(r^2+a^2\right)^2}\Phi\right)$ in Boyer-Lindquist coordinates reduces the assertion to checking that $$-\Delta + \sin^2\theta\left(a^2 - \frac{4M^2r^2a^2}{(r^2+a^2)^2} - \frac{4M^2r^2a^4\sin^2\theta\Delta}{(r^2+a^2)^4}\right) < 0.$$ It suffices to consider the case where the quantity in parentheses is positive. In this case, it is sufficient to check that $$\label{check}
-\Delta + a^2 - \frac{4M^2r^2a^2}{(r^2+a^2)^2} < 0.$$
Observe that $$(r^2+a^2)^2 = (\Delta + 2Mr)^2 = \Delta^2 + 4Mr\Delta + 4M^2r^2.$$ Multiplying through by $(r^2+a^2)^2$ then reduces (\[check\]) to $$-\Delta^3 - 4Mr\Delta^2 - 4M^2r^2\Delta + a^2\Delta^2 + 4Mra^2\Delta < 0 \Leftrightarrow
-\Delta^2 - (4Mr-a^2)\Delta - 4Mr(Mr-a^2) < 0.$$ Now it suffices to observe the inequalities $r > M>a$.
We also note the following, which actually is a trivial consequence of positivity of surface gravity $(\ref{eutuxws})$.
\[horizonTimelike\]There exists $\epsilon_0 > 0$ such that the vector field $$T + \frac{a}{2Mr_+}\Phi$$ is timelike for $r \in (r_+,r_++\epsilon_0)$.
A computation gives $$g\left(T+\frac{a}{2Mr_+}\Phi,T+\frac{a}{2Mr_+}\Phi\right) = \rho^{-2}\left(-\Delta + \sin^2\theta\left(a^2 - \frac{2a^2r}{r_+} + \frac{a^2\left(r^2+a^2\right)^2}{4M^2r_+^2} - \frac{a^4\sin^2\theta\Delta}{4M^2r_+^2}\right)\right).$$ Consider the function $$F(r) := -\Delta + a^2 - \frac{2a^2r}{r_+} + \frac{a^2\left(r^2+a^2\right)^2}{4M^2r_+^2}.$$ The lemma follows noting that $F(r_+) = 0$ and $$\begin{aligned}
\frac{dF}{dr}(r_+) &= -(r_+-r_-) - \frac{2a^2}{r_+} + \frac{2a^2}{M}
\\ \nonumber &= -2\sqrt{M^2-a^2} - \frac{2a^2}{M + \sqrt{M^2-a^2}} + \frac{2a^2}{M}
\\ \nonumber &= \frac{1}{M(M + \sqrt{M^2-a^2})}\left(-2M(M^2 - a^2) -2\sqrt{M^2-a^2}(M^2-a^2)\right)
\\ \nonumber &< 0.\end{aligned}$$
The sufficiently integrable outgoing class and Carter’s separation {#cartersupersection}
==================================================================
In this section we will define a suitable class of functions $\Psi$ for which one can apply Carter’s separation, and for which moreover, the resulting radial ordinary differential equation for $u$ will satisfy appropriate “outgoing” boundary conditions. We shall define this class in Section \[suffInt\] below, and then review the separation in Section \[karteri\].
The class of sufficiently integrable outgoing functions {#suffInt}
-------------------------------------------------------
We define here a class of functions $\Psi$ for which frequency analysis is well defined. To give some motivation for the class, we give a brief preview of how separation will be applied.
As described in the introduction, the frequency analysis necessary for our proof of integrated local energy decay requires taking a Fourier transform in $t$. However, a priori, solutions $\psi$ to the wave equation $(\ref{WAVE})$ could even grow exponentially in time. We shall at first instance, however, restrict to smooth solutions of the wave equation (arising from compactly supported data) which are *assumed* to be $L^2$ in time in the future of $\Sigma_0$.[^11] We shall first appeal to our estimates with $\Psi = \xi\psi$ where $\psi$ is a solution which is known to be future integrable, and $\xi(\tau)$ is a suitable cutoff so that $\xi=1$ for $\tau \ge 1$ and $\xi=0$ for $\tau\le 0$. See Proposition \[closedILED\]. Note that $\Psi$ satisfies an inhomogeneous equation $$\label{newinh}
\Box_{g_{a,M}}\Psi=F$$ with compactly supported $F$, and $\Psi=0$ to the past of $\Sigma_0$.
In the context of the openness argument, we shall apply estimates to $\Psi = \xi\psi$ with two different choices of $\psi$ and the cutoff $\xi(\tau)$. First, we will take $\psi$ to solve the wave equation (\[WAVE\]), and the cutoff $\xi(\tau)$ will be taken to vanish for $\tau \leq 0$ and $\tau \geq \tau_0$. Next, we will take $\psi$ to be be a solution of $\Box_g\psi=0$ where $g$ is an “interpolating metric” between $g_{a,M}$ and a $g_{\mathring{a},M}$, with interpolation region between hypersurfaces $\Sigma_{\tau-\delta_0}$ and $\Sigma_{\tau}$, and $\xi(\tau)$ will be take to equal $1$ for $\tau \ge 1$ and equal $0$ for $\tau \le 0$. This will make $\Psi$ an $L^2$ in time solution of the inhomogeneous $(\ref{newinh})$, where again $F$ is compactly supported in spacetime and $\Psi=0$ in the past of $\Sigma_0$. See Proposition \[cutoffILED\] and Section \[interpolatingSection\].
In the closedness argument, we shall be able to work with solutions $\psi$ which are *a priori* square integrable in time. Thus, we shall apply our estimates to $\Psi = \xi\psi$ where $\xi(\tau)$ is an appropriate cutoff such that $\xi = 1$ for $\tau \geq 1$ and $\xi = 0$ for $\tau \leq 0$. Then, $\Psi$ will satisfy an inhomogeneous equation (\[newinh\]) with a compactly supported right hand side, and $\Psi = 0$ to the past of $\Sigma_0$.
Finally, in the context of the boundedness argument, $\psi$ has already been proven to be $L^2$ in time (both to the future, and, after suitable extension, to the past), and the argument is applied to $\Psi=\tilde\chi\psi$ where $\tilde\chi(r)$ is a cutoff in $r$ away from the horizon and null infinity. See Section \[boundSuff\].
In all these cases, we note that we apply frequency analysis to $\Psi$ which satisfies (a) $\Psi\left(r\right)$ is square integrable in $t$ for each $r \in [r_+,\infty)$. and (b) $\Psi$ is supported away from the past event horizon $\mathcal{H}^-$ and “past null infinity” $\mathcal{I}^-$ (cf. the Penrose diagram of Section \[boundedandiled\]), with $\Box_{g_{a,M}}\Psi$ vanishing for sufficiently large $r$. It is these properties that motivate the following definitions.
### Sufficiently integrable
We first introduce the most basic integrability assumptions that will allow us to apply the separation and make sense of the radial o.d.e. $(\ref{oderedu})$ discussed already in Section \[reviewsr\].
\[sufficient\] Let $a_0<M$, $|a|\le a_0$ and let $g=g_{a,M}$. We say that a smooth function $\Psi:\mathcal{R}\to \mathbb R$ is “sufficiently integrable” if for every $j \geq 1$ and $A > r_+$, we have $$\sup_{r \in [r_+,A]}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\sum_{0 \leq i_1 + i_2 + i_3 \leq j}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}\left(Z^*\right)^{i_3}\Psi\right|^2\sin\theta\, dt\, d\theta\, d\phi < \infty,$$
$$\sup_{r \in [r_+,A]}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\sum_{0 \leq i_1 + i_2 + i_3 \leq j}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}\left(Z^*\right)^{i_3}\Box_g\Psi\right|^2\sin\theta\, dt\, d\theta\, d\phi < \infty.$$
Observe that each fixed-$r$ integral is unchanged under the change of variables $t \mapsto t^*$ and $\phi \mapsto \phi^*$.
### The “outgoing” condition {#outgoingcondsec}
We next introduce an assumption that will imply that solutions of the radial ODE $(\ref{oderedu})$ have “outgoing” boundary conditions.
\[sufficient2\] Let $a_0<M$ and $|a|\le a_0$. We shall say a smooth function $\Psi$ is “outgoing” if there exists an $\epsilon>0$ such that $\Psi$ vanishes in $\Sigma_\tau \cap \{r\le r_++\epsilon\}$ and $\Sigma_\tau \cap\{r\ge \epsilon^{-1}\}$ for all $\tau \le -\epsilon^{-1}$, and $\Box_{g_{a,M}}\Psi$ vanishes for sufficiently large $r$.
We shall see the significance of each of these assumptions individually in Sections \[karteri\] and \[boundarySection\] below. From Section \[summation\] onward, we shall always work in the class defined by assuming *both* Definitions \[sufficient\] and \[sufficient2\], i.e. $\Psi$ will always be a sufficiently integrable outgoing function.
Of course, one could significantly weaken the assumptions in Definitions \[sufficient\] and \[sufficient2\]; however, this class of functions is very easy to work with, and weakening the assumptions would not simplify the proof of Theorem \[theResult\].
Review of Carter’s separation {#karteri}
-----------------------------
As we have already discussed in our summary of the first parts of the series in Section \[reviewsr\], we shall view Carter’s separation of the wave equation as a convenient geometric framework for frequency-localising energy estimates, closely tied to the Kerr geometry. In the present section, we shall review the relevant formalism from [@dr7].
### The oblate spheroidal harmonics {#oblat}
Let $\nu \in \mathbb{R}$. We begin by recalling the collection $$\{S_{m\ell}(\nu,\cos \theta)e^{im\phi}\}_{m\ell}$$ of the eigenfunctions of the self-adjoint operator $$P(\nu)\, f= -\frac 1{\sin\theta} \frac{\partial}{\partial\theta} \left (\sin\theta \frac{\partial}{\partial\theta}f\right)
-\frac{\partial^2 f}{\partial\phi^2}\frac{1}{\sin^2\theta}
- \nu^2 \cos^2\theta f$$ on $L^2(\sin\theta\, d\theta\, d\phi)$. These form a complete orthonormal basis on $L^2(\sin\theta\, d\theta\, d\phi)$. The eigenfunctions are parametrised by a set of real discrete eigenvalues $\lambda^{(\nu)}_{m\ell}$ $$\label{eigenvals}
P(\nu)\, S_{m\ell}(\nu,\cos \theta)e^{im\phi}=\lambda_{m\ell}(\nu) S_{m\ell}(\nu,\cos \theta)e^{im\phi},$$ which have the property that $$\label{eq:lam}
\lambda_{m\ell}^{(\nu)}+\nu^2\ge |m|(|m|+1),$$ $$\label{lamBound}
\lambda_{m\ell}^{(\nu)} + \nu^2 \geq 2\left|m\nu\right|.\footnote{This follows immediately from the variational characterization of $\lambda_{m\ell}$ and the inequality $\frac{m^2}{\sin^2\theta} + \nu^2\sin^2\theta \geq 2\left|m\nu\right|$.}$$ For $\nu=0$ the [*oblate spheroidal harmonics*]{} $S_{m\ell}(\nu,\cos \theta)e^{im\phi}$ reduce to the standard spherical harmonics $Y_{m\ell}$ with the corresponding eigenvalues $\lambda^{(0)}_{m\ell}=\ell(\ell+1)$.
### The coefficients $\Psi_{m\ell}^{(a\omega)}$ {#separationSubsection}
Given parameters $a$, $M$, let $\Psi$ be “sufficiently integrable” in the sense of Definition \[sufficient\]. We write $$\Psi(t,r,\theta,\phi)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty e^{-i\omega t} \widehat\Psi (\omega,r,\theta,\phi) d\omega,$$ and then, setting $\nu = a\omega$ for each $\omega\in \mathbb R$, further decompose $$\widehat\Psi(\omega, r,\theta,\phi) = \sum_{m\ell} \Psi^{(a\omega)}_{m\ell}(r)
S_{m\ell}(a\omega,\cos \theta)e^{im\phi},$$ to arrive at $$\Psi(t,r,\theta,\phi)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty\sum_{m\ell} e^{-i\omega t} \Psi^{(a\omega)}_{m\ell}(r)S_{m\ell}(a\omega,\cos\theta)e^{im\phi} d\omega.$$ The sufficiently integrable assumption implies that for each $r$, the first and third equality may be interpreted in $L^2_tL^2_{\mathbb{S}^2}$, while the second equality may be interpreted in $L^2_{\omega}L^2_{\mathbb{S}^2}$.
Furthermore, if $\Psi$ satisfies Definition \[sufficient\], then so do $\partial_t\Psi$, $\partial_\phi\Psi$ and, by the well-known properties of the Fourier transform, we have $$\partial_t\Psi(t,r,\theta,\phi)=\frac{-i}{\sqrt{2\pi}}\int_{-\infty}^\infty\sum_{m\ell} \omega e^{-i\omega t} \Psi^{(a\omega)}_{m\ell}(r)S_{m\ell}(a\omega,\cos\theta)e^{im\phi} d\omega.$$ $$\partial_{\phi}\Psi(t,r,\theta,\phi)=\frac{i}{\sqrt{2\pi}}\int_{-\infty}^\infty\sum_{m\ell}m e^{-i\omega t} \Psi^{(a\omega)}_{m\ell}(r)S_{m\ell}(a\omega,\cos\theta)e^{im\phi} d\omega.$$ As above, for each $r$ these equalities may be interpreted in $L^2_tL^2_{\mathbb{S}^2}$.
Let us take the opportunity to observe the following consequences of Plancherel’s formula and the orthonormality of the $S_{ml}(a\omega,\cos\theta)e^{-im\phi}$: $$\begin{aligned}
&\int_0^{2\pi}\int_0^\pi\int_{-\infty}^{\infty}
|\Psi|^2(t,r,\theta,\varphi) \sin\theta\, d\varphi\, d\theta\, dt= \int_{-\infty}^\infty \sum_{m\ell}
|\Psi^{(a\omega)}_{m\ell}(r)|^2\, d\omega,\\
&\int_0^{2\pi}\int_0^\pi\int_{-\infty}^{\infty} \Psi\cdot\overline{\Upsilon} \sin\theta d\varphi\, d\theta\, dt=
\int_{-\infty}^\infty\sum_{m\ell} \Psi^{(a\omega)}_{m\ell}\cdot\bar \Upsilon^{(a\omega)}_{m\ell}
d\omega,\\
&\int_0^{2\pi}\int_0^\pi\int_{-\infty}^{\infty}
|\partial_r\Psi|^2(t,r,\theta,\varphi) \sin\theta\, d\varphi\, d\theta\, dt = \int_{-\infty}^\infty \sum_{m\ell}
\left|\frac{d}{dr}\Psi^{(a\omega)}_{m\ell}(r)\right|^2
d\omega,\\
&\int_0^{2\pi}\int_0^\pi\int_{-\infty}^{\infty}
|\partial_t\Psi|^2(t,r,\theta,\varphi) \sin\theta\, d\varphi\, d\theta\, dt = \int_{-\infty}^\infty \sum_{m\ell}
\omega^2 |\Psi^{(a\omega)}_{m\ell}(r)|^2
d\omega.\end{aligned}$$
Finally, we note that a straightforward integration by parts, Plancherel and the orthonormality of $S_{m\ell}(a\omega, \cos\theta)e^{im\phi}$ imply $$\begin{aligned}
\int_0^{2\pi}\int_0^\pi\int_{-\infty}^{\infty}
&\left[\left|\frac{\partial \Psi}{\partial\theta}\right|^2 + \left|\frac{\partial \Psi}{\partial\phi}\sin^{-1}\theta\right|^2\right](t,r,\theta,\varphi) \sin\theta\, d\varphi\, d\theta\, dt \\ &=\int_{-\infty}^\infty \sum_{m,\ell}
\lambda_{m\ell}^{(a\omega)}|\Psi^{(a\omega)}_{m\ell}(r)|^2
d\omega
-a^2\int_0^{2\pi}\int_0^\pi\int_{-\infty}^{\infty}\cos^2\theta\left|\partial_t\Psi\right|^2 \sin\theta\, d\varphi\, d\theta\, dt.\end{aligned}$$
### The radial ordinary differential equation and the potential $V$ {#theRadialODEPot}
If $\Psi$ is sufficiently integrable in the sense of Definition \[sufficient\], define $$\label{homogdefeq}
F=\Box_g\Psi.$$ for $g=g_{a,M}$ a Kerr metric with $|a|<M$.
The sufficiently integrable assumption implies that we may define the coefficients $\Psi_{m\ell}^{(a\omega)}(r)$, $\left(\rho^2F\right)_{m\ell}^{(a\omega)}(r)$ as above.
Carter’s formal separation [@cartersep2] of the wave operator yields:
Let $a_0<M$, $|a|\le a_0$, $\Psi$ be sufficiently integrable, and let $F$ be defined by $(\ref{homogdefeq})$. Then $$\begin{aligned}
\label{CartersODE}
\Delta \frac{d}{dr} \left (\Delta \frac{d\Psi_{m\ell}^{(a\omega)}}{dr}\right)& + \left (a^2m^2 + (r^2+a^2)^2\omega^2-4Mra\omega m - \Delta (\lambda_{m\ell}+a^2\omega^2) \right) \Psi_{m\ell}^{(a\omega)}=\Delta\,
\left(\rho^2F\right)_{m\ell}^{(a\omega)}.\end{aligned}$$ Note that the sufficiently integrable assumption allows us to interpret this equality for each $r$ in $L^2_{\omega}l^2_{m\ell}$.
Using the definition $(\ref{r*def})$ of $r^*$ and setting $$\label{uDef}
u^{(a\omega)}_{m\ell}(r)=(r^2+a^2)^{1/2}
\Psi^{(a\omega)}_{m\ell} (r),$$ $$\label{hDef}
H^{(a\omega)}_{m\ell}(r)=\frac{\Delta \left(\rho^2F\right)^{(a\omega)}_{m\ell}(r)}{(r^2+a^2)^{3/2}},$$ we obtain $$\label{e3iswsntouu}
\frac{d^2}{(dr^*)^2}u^{(a\omega)}_{m\ell}+(\omega^2 - V^{(a\omega)}_{m\ell }(r))u =
H^{(a\omega)}_{m\ell}$$ where $$\label{defofV}
V^{(a\omega)}_{m \ell}(r)= \frac{4Mram\omega-a^2m^2+\Delta (\lambda_{m\ell}+\omega^2a^2)}{(r^2+a^2)^2}
+\frac{\Delta(3r^2-4Mr+a^2)}{(r^2+a^2)^3}
-\frac{3\Delta^2 r^2}{(r^2+a^2)^4}.$$ In the Schwarzschild case: $$\label{at0freq}
V^{(0\omega)}_{m\ell}(r) = (r-2M)\left(\frac{\ell\left(\ell+1\right)}{r^3}+\frac{2M}{r^4}\right),$$ $$\label{at0freq'}
\left(\frac{dV}{dr^*}\right)^{(0\omega)}_{m\ell}(r)= \frac{r-2M}{r}\left(\frac{2\ell\left(\ell+1\right)(3M-r)}{r^4}+\frac{2M(8M-3r)}{r^5}\right).$$ Again, for each $r$, (\[e3iswsntouu\]) is to be interpreted in $L^2_{\omega}l^2_{m\ell}$.
### Notational conventions
Following well-established convention, in what follows, as in [@dr7], we shall suppress the dependence of $u$, $H$ and $V$ on $a\omega$, $m$, $\ell$ in our notation. We will also use the notation $$\label{primenotation}
'=\frac{d}{dr^*}.$$
Since for fixed $g_{a,M}$, $r$ is a smoothly invertible function of $r^*$, we shall often refer to $r^*$-ranges by their corresponding $r$-ranges (in particular, given an $r$-parameter such as $R$, we shall very often use the notation $R^*=r^*(R)$ without further comment), and we shall express functions appearing in most estimates as functions of $r$. The derivative $'$ always is to denote $(\ref{primenotation})$!
Boundary conditions {#boundarySection}
-------------------
In this section, we shall establish boundary conditions for the radial ODE $(\ref{e3iswsntouu})$. We will require Definitions \[sufficient\] and \[sufficient2\].
\[boundHorizon\] Let $a_0<M$, $|a|\le a_0$, and $\Psi$ be sufficiently integrable and outgoing. Define $u^{(a\omega)}_{m\ell}(r)$ by (\[uDef\]). Then $$\int_{-\infty}^{\infty}\sum_{m\ell}\left|u'(r)+i\left(\omega-\frac {a m}{2Mr_+}\right)u(r)\right|^2\, d\omega$$ is a smooth function of $r$ which vanishes at $r = r_+$.
The assumptions on $\Psi$ imply that $$\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\sum_{0 \leq i_1 + i_2 + i_3 \leq j}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}\left(Z^*\right)^{i_3}\Psi\right|^2\sin\theta\, dt\, d\theta\, d\phi$$ is a smooth function of $r$. Combining this with the fact that $\partial_{r^*} = T + \frac{a}{2Mr_+}\Phi$ on $\mathcal{H}^+$, we conclude that $$\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left|\partial_{r^*}\left((r^2+a^2)^{1/2}\Psi\right) - \left(T + \frac{a}{2Mr_+}\right)\left((r^2+a^2)^{1/2}\Psi\right)\right|^2\sin\theta\, dt\, d\theta\, d\phi = O(r-r_+) \Leftrightarrow$$
$$\int_{-\infty}^{\infty}\sum_{m\ell}\left|u'(r)+i\left(\omega-\frac {a m}{2Mr_+}\right)u(r)\right|^2\, d\omega = O\left(r-r_+\right)\text{ as }r \to r_+.$$
\[boundInfinity\] Let $a_0<M$, $|a|\le a_0$, and let $\Psi$ be sufficiently integrable and outgoing. Define $u^{(a\omega)}_{m\ell}(r)$ by (\[uDef\]). Then, there exists a sequence $\{r_n\}_{n=1}^{\infty}$ such that $r_n \to \infty$ as $n\to \infty$ and $$\lim_{n\to\infty}\left|\left(u^{(a\omega)}_{m\ell}\right)'(r_n)-i\omega u^{(a\omega)}_{m\ell}(r_n)\right| \to 0$$ for almost every $\omega$.
The “$r^p$-estimate” from [@icmp] with $p = 1$, and Definitions \[sufficient\] and \[sufficient2\] immediately imply that for $R$ sufficiently large, $$\int_{-\infty}^{\infty}\int_{r\geq R}\int_0^{2\pi}\int_0^{\pi}\left|\left(\partial_t+\partial_{r^*}\right)\left(\left(r^2+a^2\right)^{1/2}\Psi\right)\right|^2\sin\theta\, dt\, dr\, d\theta\, d\phi < \infty.$$ An application of Plancherel and a standard pigeonhole argument imply that there exists a dyadic subsequence $\{r_n\}_{n=1}^{\infty}$ such that $$\lim_{n\to\infty}\int_{-\infty}^{\infty}\left|\left(u^{(a\omega)}_{m\ell}\right)'(r_n)-i\omega u^{(a\omega)}_{m\ell}(r_n)\right|^2\ d\omega \to 0.$$ Finally, we recall the standard fact that convergence in $L^2$ implies almost everywhere pointwise convergence along a subsequence.
As noted in [@stabi] we may formally write these boundary conditions as $$\begin{aligned}
&u'+i\left(\omega-\frac {a m}{2Mr_+}\right)u=0,\qquad r=r_+,\label{eq:b-}\\
&u'-i\omega u =0,\qquad\qquad\qquad\quad r=\infty\label{eq:b+}.\end{aligned}$$
On the almost everywhere regularity of $u^{(a\omega)}_{m\ell}$
--------------------------------------------------------------
The most natural setting for our frequency analysis is the class of functions of $r$ with values in $L^2_{\omega}l^2_{m\ell}$ referred to already after equality $(\ref{CartersODE})$. However, for convenience, in Sections \[sct\] and \[freqLocEst\], we shall study *classical* solutions $u$ to the o.d.e. (\[e3iswsntouu\]). The following lemma will be used in Section \[summation\] to justify the reduction to classical solutions.
\[aeRegular\] Let $a_0<M$, $|a|\le a_0$, let $\Psi$ be sufficiently integrable and outgoing, and define $u^{(a\omega)}_{m\ell}(r)$ by (\[uDef\]). Then, for almost every $\omega$, for all $m$ and $\ell$, $H$ is smooth and $u$ is a smooth solution to the o.d.e. (\[e3iswsntouu\]) satisfying the boundary conditions (\[eq:b+\]) and (\[eq:b-\]).
Pick any labeling of the eigenvalues $\lambda_{m\ell}$ such that $\lambda_{m\ell}$ is a measurable function $\omega$. Then, using the fact that a countable union of measure zero sets is measure zero, it clearly suffices to prove the lemma for each fixed value of $m$ and $\ell$.
For any $j \geq 1$ and $R > r_+$, the fundamental theorem of calculus implies $$\begin{aligned}
\label{fromFundCalc}
\sum_{0 \leq i \leq j}\int_{-\infty}^{\infty}\sup_{r \in [r_+,R]}\left|\left(Z^*\right)^iu\right|^2\, d\omega \leq& \sum_{0 \leq i \leq j}\int_{-\infty}^{\infty}\sum_{m\ell}\sup_{r \in [r_+,R]}\left|\left(Z^*\right)^iu\right|^2\, d\omega
\\ \nonumber \leq& \sum_{0 \leq i \leq j}\int_{-\infty}^{\infty}\sum_{m\ell}\left|\left(Z^*\right)^iu\right|^2\Big|_{r = r_+}\, d\omega +
\sum_{0 \leq i \leq j+1}\int_{-\infty}^{\infty}\sum_{m\ell}\int_{r_+}^R\left|\left(Z^*\right)^iu\right|^2\, d\omega\, dr.\end{aligned}$$ Next, Plancherel (see the explicit formulas in Section \[separationSubsection\]), Sobolev inequalities on $\mathbb{S}^2$ and the sufficiently integrable assumption imply that (\[fromFundCalc\]) is less than $$\begin{aligned}
B\sum_{0 \leq i \leq j+1}&\int_{-\infty}^{\infty}\int_{r_+}^R\int_{\mathbb{S}^2}\left|\left(Z^*\right)^i\Psi\right|^2\sin\theta\, dt^* dr\, d\theta\, d\phi^* +
B\sum_{0 \leq i \leq j}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left|\left(Z^*\right)^i\Psi\right|^2\Big|_{r = r_+}\sin\theta\, dt^* d\theta\, d\phi^*
\\ \nonumber \le& B\sum_{0 \leq i \leq j+1,\, k=0,1,2}\int_{-\infty}^{\infty}\int_{r_+}^R\int_{\mathbb{S}^2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^k\left(Z^*\right)^i\Psi\right|^2\sin\theta\, dt^*\, dr\, d\theta\, d\phi^*
\\ \nonumber&+
B\sum_{0 \leq i \leq j,\, k=0,1,2}\int_{\mathbb{S}^2}\int_{-\infty}^{\infty}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^k\left(Z^*\right)^i\Psi\right|^2\Big|_{r = r_+}\sin\theta\, dt\, d\theta\, d\phi^* < \infty.\end{aligned}$$
Thus, we conclude that for each $n \in \mathbb{Z}_+$ and $j \geq 0$, $\sum_{0 \leq i \leq j}\sup_{r \in [r_+,r_++n]}\left|\left(Z^*\right)^iu\right|^2$ is an $L^2$ function of $\omega$. Consequently, we may find a set $U^{(j)}_n \subset \mathbb{R}$ such that $\left|\left(U^{(j)}_n\right)^c\right| = 0$ and $\omega \in U^{(j)}_n$ implies that $u^{(a\omega)}_{m\ell}(r)$ is $C^j$ on the interval $(r_+,r_++n)$. Observe that $$\left|\left(\cap_{j,n=1}^{\infty}U^{(j)}_n\right)^c\right| = \left|\cup_{j,n=1}^{\infty}\left(U^{(j)}_n\right)^c\right| \leq \sum_{j,n=1}^{\infty}\left|\left(U^{(j)}_n\right)^c\right| = 0.$$ Thus, we have a set $U \doteq \cap_{j,n=1}^{\infty}U^{(j)}_n$ such that the complement of $U$ has measure $0$, and $\omega \in U$ implies that $u^{(a\omega)}_{m\ell}$ is a smooth function of $r$. Of course, the same procedure may be carried out for $H^{(a\omega)}_{m\ell}$. We conclude that for almost every $\omega$, $u$ and $H$ are smooth functions of $r$, and hence $u$ is a classical solution of the radial o.d.e. (\[e3iswsntouu\]).
Next, we turn to the boundary condition (\[eq:b+\]). For every $\omega$ such that $u$ is a classical solution of the radial o.d.e. (\[e3iswsntouu\]), an asymptotic analysis of the o.d.e. (\[e3iswsntouu\]) as $r^*\to\infty$ implies that we can find constants $A_{out}$ and $A_{in}$ such that $$u^{(a\omega)}_{m\ell} = A_{out}e^{i\omega r^*} + A_{in}e^{-i\omega r^*} + O\left(r^{-1}\right)\text{ as }r^*\to\infty,$$ where $O\left(r^{-1}\right)$ is preserved upon differentiation. Lemma \[boundInfinity\] implies that we must have $A_{in} = 0$, and hence that the boundary condition (\[eq:b+\]) holds.
Similarly, an asymptotic analysis of the o.d.e. (\[e3iswsntouu\]) as $r^*\to-\infty$ implies that we can find constants $C_{out}$ and $C_{in}$ such that $$u^{(a\omega)}_{m\ell} = C_{out}e^{-i\left(\omega -\upomega_+m\right)r^*} + C_{in}e^{i\left(\omega-\upomega_+m\right) r^*} + O\left(\left|r^*\right|^{-1}\right)\text{ as }r^*\to-\infty.$$ Lemma \[boundHorizon\] implies that we must have $C_{in} = 0$, and hence that the boundary condition (\[eq:b-\]) holds.
Properties of the potential $V$ {#Vpropsec}
===============================
In this section, we prove certain fundamental properties of the potential $V$ appearing in $(\ref{e3iswsntouu})$, defined by the expression $(\ref{defofV})$. In particular, we shall prove high-frequency regime properties which will be essential for the coercivity of the currents of Section \[freqLocEst\] in the high frequency ranges. Sections \[decomppotsec\]–\[Vsrsec\] below follow closely Section 11.1 of our survey [@stabi]. Section \[Vnewsec\], relevant for the fixed-$m$ case which will be used in our continuity argument of Section \[continuityargsec\], is new. Finally, we record explicitly in Section \[ASIDEsec\] the relation of the properties of $V$ proven here to properties of geodesic flow on Kerr.
Recall from the outline in Section \[outlinesection\] and the discussion of Section \[theLogic\] that the present section, together with Sections \[sct\] and \[freqLocEst\], can be understood to form an independent logical unit of this paper which culminates in Theorem \[phaseSpaceILED\] giving frequency independent estimates for classical solutions $u$ of the o.d.e. $(\ref{e3iswsntouu})$ satisfying the boundary conditions $(\ref{eq:b-})$ and $(\ref{eq:b+})$. Note that for convenience, this analysis will use the reduction to $a\ge 0$ discussed in Section \[signOfa\]. We shall return to the study of $(\ref{WAVE})$ in Section \[summation\].
Admissible frequencies
----------------------
Recall that the set of eigenvalues $\{\lambda_{m\ell}(a\omega)\}$ defined by $(\ref{eigenvals})$ are not known explicitly in closed form. As is clear from $(\ref{defofV})$, the potential depends on $\lambda_{m\ell}(a\omega)$ only through the quantity $$\label{LAMBDAfreq}
\Lambda=\lambda_{m\ell}(a\omega)+a^2\omega^2,$$ which according to and obeys $$\label{needsanumber}
\Lambda\ge |m|(|m|+1),$$ $$\label{useful}
\Lambda \ge 2\left|am\omega\right|.$$ It turns out that the results of this section depend *only* on the constraints $(\ref{needsanumber})$ and (\[useful\]), not on the precise values of the set $\{\lambda_{m\ell}(a\omega)\}$ In what follows, we may thus consider $\omega\in \mathbb R$, $m\in \mathbb Z$, $\Lambda\in \mathbb R$ to be *independent* parameters[^12] constrained only by $(\ref{needsanumber})$ and $(\ref{useful})$. This motivates
We call a frequency triple $(\omega, m, \Lambda)$ admissible if $\omega\in \mathbb R$, $m\in \mathbb Z$, $\Lambda\in \mathbb R$, where $\Lambda\ge |m|(|m|+1)$ and $\Lambda \geq 2\left|am\omega\right|$.
Decomposition of the potential {#decomppotsec}
------------------------------
Given Kerr parameters $0\le a<M$, and an admissible frequency triple $(\omega, m, \Lambda)$, we may now *define* the potential as $$\label{NOWADEF}
V(\omega,m,\Lambda)=V_0(\omega, m,\Lambda)+V_1,$$ where $$\begin{aligned}
\label{NOWADEF2}
V_0&=&\frac{4Mram\omega -a^2m^2+\Delta\Lambda}{(r^2+a^2)^2},\\
\nonumber
V_1&=&\frac{\Delta(3r^2-4Mr+a^2)}{(r^2+a^2)^3} -\frac{3\Delta^2 r^2}{(r^2+a^2)^4}.\end{aligned}$$ Note that $V_0$ dominates for high frequencies since $V_1$ does not contain any frequency parameters $m$, $\omega$, $\Lambda$. Note also the nonnegativity property: $$V_1=\frac {\Delta}{(r^2+a^2)^4}\left[a^2\Delta+2Mr(r^2-a^2)\right]\ge 0.$$
The critical points of $V_0$ and the structure of trapping {#Vtrsec}
----------------------------------------------------------
To understand the nature of trapping, one must first identify the critical points of $V_0$. This is provided by the following Lemma. (This appeared as Lemma 11.1.1 of [@stabi]; we repeat its statement and proof here.)
\[lem:1\] Let $M>0$, $a_0<M$ and $0\le a\le a_0$. Then for all admissible frequency triples $(\omega, m, \Lambda)$ with $\Lambda>0$, the potential function $V_0$ defined by $(\ref{NOWADEF2})$ as a function $V_0:(r_+,\infty)\to \mathbb R$ is either (a) strictly decreasing, (b) has a unique critical value $r^0_{\rm max}$ which is a global maximum, or (c) has exactly two critical values $r^0_{\rm min}<r^0_{\rm max}$ which are a local minimum and maximum respectively. The value $r^0_{\rm max}$ is bounded independently of the frequency parameters $$r^0_{\rm max}\le B.$$
We have $$\begin{aligned}
\frac {d}{dr} V_0&=4maM\omega \left(\frac 1{(r^2+a^2)^2} - \frac {4r^2}{(r^2+a^2)^3}\right)
+\frac {4ra^2m^2}{(r^2+a^2)^3} + \frac \Lambda{(r^2+a^2)^2} \left(2(r-M) -\frac {4r\Delta}{r^2+a^2}\right)\\
&=\frac 1{(r^2+a^2)^3} \left(4maM\omega(-3r^2+a^2) + 4ra^2m^2-2\Lambda(r^3+a^2r-3Mr^2+Ma^2)\right),\end{aligned}$$ and thus, $$\begin{aligned}
\frac d{dr} \left((r^2+a^2)^3 \frac d{dr} V_0\right)&=&-24Mam\omega r + 4a^2m^2-2\Lambda(3r^2-6Mr+a^2)\\ &=&-6\Lambda\left(r^2-2Mr +4Mr\sigma+\frac{a^2}3-\frac 23 a^2 \frac{m^2}\Lambda\right),\end{aligned}$$ where we have set $$\sigma=\frac {am\omega}\Lambda.$$
It follows that any critical points of the function $(r^2+a^2)^3 \frac d{dr} V_0$ must be roots of the quadratic $$r^2-2Mr(1-2\sigma)+\frac {a^2}3\left(1-\frac {2m^2}{\Lambda}\right)$$ which we may denote as $$r_{1,2}=M(1-2\sigma)\pm\sqrt{M^2(1-2\sigma)^2-\frac {a^2}3\left(1-\frac {2m^2}\Lambda\right)}.$$
Recalling that $r_+>M$, then if $m\omega\geq 0$ (and thus $\sigma\geq 0$), it follows that $Re(r_2)<M$ and thus the only possible critical point on the interval $(r_+,\infty)$ would be $$r_1=M(1-2\sigma)+\sqrt{M^2(1-2\sigma)^2-\frac {a^2}3\left(1-\frac {2m^2}\Lambda\right)}.$$ Noting that since $\Lambda>0$, we have $$\lim_{r\to\infty} (r^2+a^2)^3 \frac d{dr} V_0= -\infty,$$ it follows that $\frac d{dr} V_0$ either (a\*) vanishes nowhere, (b\*) vanishes at a unique point to be denoted $r^0_{\rm max}$, or (c\*) vanishes at two points, denoted $r^0_{\rm min}<r^0_{\rm max}$, where $$\frac{ d^2}{dr^2} V_0(r^0_{\rm min})\ge 0,\qquad
\frac{ d^2}{dr^2} V_0(r^0_{\rm max})\le 0.$$
In case (a\*), it follows that $V_0$ is strictly decreasing (case (a) of the lemma). In case (b\*), it follows that either $r_{\rm max}$ is an inflection point and $V_0$ is again strictly decreasing (corresponding again to case (a) of the statement of the lemma), or $r_{\rm max}$ is a global maximum (case (b) of the statement of the lemma). In case (c\*), it is moreover easy to see that these inequalities are in fact strict, and thus $r^0_{\rm min}$ and $r^0_{\rm max}$ correspond to the unique minumum and maximum of $V_0$ on $(r_+,\infty)$ (corresponding to case (c) of the statement of the lemma).
If $m\omega<0$ (and thus $\sigma<0$), then let us reexpress the root $r_2$ by $$r_2=
M(1-2\sigma)\left(1-\sqrt{1-\frac {a^2(1-\frac {2m^2}\Lambda)}{3M^2(1-2\sigma)^2}}\right).$$ Since $a <M$ and $\sigma<0$, we have $$\frac {a^2(1-\frac {2m^2}\Lambda)}{3M^2(1-2\sigma)^2}<\frac 13.$$ Noting for $0\le x<\frac 13$ the inequality $
\sqrt{1-x}\ge 1-\frac{2x}3$, it follows that $$Re(r_2)<\frac{2M(1-2\sigma)a^2(1-\frac{2m^2}\Lambda)}{9M^2(1-2\sigma)^2}=
\frac {2a^2(1-\frac{2m^2}\Lambda)}{9M(1-2\sigma)}<\frac {2M}9<r_+.$$ This now implies that $r_1$ is the only possible zero of $\frac d{dr} \left [(r^2+a^2)^3 \frac d{dr} V_0\right]$ on the interval $[r_+,\infty)$ and the previous argument applies.
The last statement of the lemma easily follows from observing that for all $\Lambda > 0$ $$(r^2+a^2)^3 \frac d{dr} V_0 = (6\La M -12 Mam\omega)r^2-2\Lambda r^3 + O\left(am\omega,m^2\right)r\text{ as }r\to\infty,$$ and we have $\Lambda \geq \left|m\right|\left(\left|m\right|+1\right)$, $\Lambda\geq
2a|m\omega|$.
The next statement effectively establishes that even if $r^0_{\rm min}$ exists, it can only be ‘trapped’ for the value $\omega=\upomega_+m$. (Again this appeared as Lemma 11.1.2 of [@stabi]. We repeat its statement and proof here.)
\[lem:2\] Let $M>0$, $a_0<M$ and $0\le a \le a_0$. For all admissible frequency triples $(\omega, m, \Lambda)$ we have $$\label{anineq}
\omega^2\ge V(r_+)$$ with equality achieved if and only if $\omega=\upomega_+m$. In particular, in the notations of the previous lemma, this implies that $$\omega^2>V_0(r^0_{\rm min}).$$
We simply compute $$\omega^2-V(r_+)=\omega^2-\frac {4Mr_+am\omega-a^2m^2}{(r_+^2+a^2)^2}=
\frac {(2Mr_+\omega-am)^2}{4M^2r_+^2}.$$
Note that the case of equality in $(\ref{anineq})$ occurs precisely at the threshold of the superradiance condition (\[superradiantParamPosa\]): $$\omega=\upomega_+m=\frac {am}{2Mr_+}.$$
Superradiant frequencies are not trapped {#Vsrsec}
----------------------------------------
We now turn specifically to the superradiant frequencies, which under the assumption $a\ge 0$ are defined by $(\ref{superradiantParamPosa})$. We will show that these are in fact , in the sense that, for such frequencies, the maximum of $V$ is always (quantitatively) above the energy level $\omega^2$.
First, let us show that for a range of frequency parameters including the superradiant regime, $V_0$ can only have a critical point at a maximum, that is the point $r^0_{\rm min}$ is absent. (This was Lemma 11.1.3 of [@stabi] augmented by Remark 11.1.)
\[lem:3\] Let $M>0$, $a_0<M$ and $0\le a\le a_0$. Then for all admissible frequency triples $(\omega, m, \Lambda)$ satisfying in addition $$m\omega\le \frac {am^2}{2Mr_+},$$ we have $$\label{increaseHorizon}
\frac {d}{dr} V(r_+)\ge \frac {d}{dr} V_0(r_+)\ge b\Lambda \ge 0.$$ Recall that Lemma \[lem:1\] showed that if $r^0_{\rm min}$ exists, we either have $r^0_{\rm min}<r^0_{\rm max}$ or $\frac{dV_0}{dr} \leq 0$ on $(r_+,\infty)$. Thus (\[increaseHorizon\]) implies that $r^0_{\rm min}$ does not exist and the potential $V_0$ has its unique critical point at $r^0_{\rm max}$.
Moreover, for all $\alpha>0$ sufficiently small[^13], the same statement holds under the weaker assumption $$\label{withAlpha}
m\omega \leq \frac{am^2}{2Mr_+} + \alpha\Lambda.$$
We begin with the first statement of the lemma. Note $$\begin{aligned}
\frac d{dr} V_0(r_+)&=&\frac{4maM\omega}{(r_+^2+a^2)^3} \left (-3r_+^2+a^2\right)
+\frac {4r_+a^2m^2}{(r_+^2+a^2)^3} + \frac {2(r_+-M)\Lambda}{(r_+^2+a^2)^2} \\ &=&
\frac 1{(r_+^2+a^2)^3}\left(4maM\omega (-3r_+^2+a^2) + 4r_+a^2m^2+2(r_+^2+a^2)(r_+-M)\Lambda\right).\end{aligned}$$ For frequency parameters satisfying $m\omega<0$, the conclusion of the lemma is now obvious, since $-3r_+^2+a^2<0$. Otherwise, using the condition $$0\le m\omega\le \frac {am^2}{2Mr_+}$$ we obtain $$\begin{aligned}
(r_+^2+a^2)^3\frac d{dr} V_0(r_+)&\ge \left (\frac {2a^2m^2}{r_+} (-3r_+^2+a^2) +4r_+a^2m^2+2(r_+^2+a^2)(r_+-M)\Lambda\right)\\ &= \left(\frac {2a^2m^2}{r_+} (-r_+^2+a^2) +2(r_+^2+a^2)(r_+-M)\Lambda\right)\\ &=2(r_+-M)\left(\Lambda (r_+^2+a^2)-2a^2m^2\right)\\&=
4(r_+-M)\left(\Lambda Mr_+-a^2m^2\right).\end{aligned}$$ The inequalities $\Lambda\ge m^2$ and $r_+>M>a$ imply that $\frac d{dr} V_0(r_+)\geq b\Lambda$. We finish the proof of the first statement by recalling that $V=V_0+V_1$ and observing the identity $$\frac {d}{dr} V_1(r_+)=\frac {4Mr_+(r_+-M)(r_+^2-a^2)}{(r_+^2+a^2)^4}>0.$$
It is clear that the final assertion of the lemma concerning the weaker assumption (\[withAlpha\]) follows immediately now from the first.
Recall the superradiant condition $(\ref{superradiantParamPosa})$. The statement that superradiant frequencies are not trapped now follows from the following Lemma (again, cf. Lemma 11.1.4 of [@stabi])
\[lem:4\] Let $M>0$, $a_0<M$ and $0\le a\le a_0$. For all $\alpha\ge0$ sufficiently small, then for all admissible frequency triples $(\omega, m, \Lambda)$ satisfying in addition $$0 < m\omega\le \frac {am^2}{2Mr_+} + \alpha\Lambda,$$ the potential $V_0$ satisfies $$b\Lambda \leq V_0(r^0_{max}) - \omega^2.$$
Again, it suffices to prove the lemma with $\alpha = 0$. Let $\epsilon > 0$ be a fixed sufficiently small constant.
We first consider the case when $m\left(\frac{am}{2Mr_+} - \omega\right) \leq \epsilon\left|m\right|\sqrt{\Lambda}$. In this case we have $$\omega^2 - V_0(r_+) = \left(\omega - \frac{am}{2Mr_+}\right)^2 \leq \epsilon^2\Lambda.$$ Combining this with Lemma \[lem:3\] easily shows $$V_0(r_++\delta)- \omega^2 \geq b\Lambda$$ for some sufficiently small $\delta > 0$ and even smaller $\epsilon$.
Next, we consider the case when $\omega^2 \leq \epsilon \Lambda$. Then we clearly have $$V_0(r) - \omega^2 \geq \frac{\Lambda}{r^2} + O\left(\frac{\Lambda}{r^3}\right) - \epsilon\Lambda\text{ as }r\to\infty.$$ Therefore, if we let $\tilde{r}$ be sufficiently large, and then let $\epsilon$ be sufficiently small, we can arrange for $$V_0(\tilde{r}) - \omega^2 \geq b\Lambda.$$
Finally, we consider the case where $m\left(\frac{am}{2Mr_+}-\omega\right) > \epsilon\left|m\right|\sqrt{\Lambda}$ and $\omega^2 > \epsilon \Lambda$. In this case, $r_0 := \frac{am}{2M\omega}$ will satisfy $r_0 \in \left[r_++\delta,R\right]$ for some $\delta > 0$ and $R < \infty$. Letting $\Delta_{r_0}$ denote $r_0^2 - 2Mr_0 + a^2$, we then compute
$$\begin{aligned}
\omega^2-V_0(r_0)&=&\omega^2-\frac {4Mr_0am\omega-a^2m^2+\Delta_{r_0}\Lambda}{(r_0^2+a^2)^2}\\ &=&
\frac 1{(r_0^2+a^2)^2}\left[(r_0^2+a^2)^2\omega^2-4Mr_0am\omega+a^2m^2-\Delta_{r_0}\Lambda\right]\\ &=&
\frac 1{(r_0^2+a^2)^2}\left[4M^2r_0^2\omega^2-4Mr_0am\omega+a^2m^2+\omega^2\left((r_0^2+a^2)^2-4Mr_0^2\right)-\Delta_{r_0}\Lambda\right]\\&=&\frac {\omega^2(r_0^2-2Mr_0+a^2)(r_0^2+2Mr_0+a^2)-\Delta_{r_0}\Lambda}{(r_0^2+a^2)^2}\\
&=&\frac{ \Delta_{r_0}}{(r_0^2+a^2)^2}\left(\frac {a^2 m^2}{4M^2} \left(1+\frac{2M}{r_0}+\frac{a^2}{r_0^2}\right)-\Lambda\right).\end{aligned}$$
We now recall that $a<M<r_0$ and that $\Lambda\ge \left|m\right|(\left|m\right|+1)$ to conclude that $$V_0\left(r_0\right) -\omega^2 \geq b\frac{\Delta_{r_0}}{\left(r_0^2+a^2\right)^2}\Lambda \geq b\Lambda.$$ In the last inequality we have used that $r_0$ is bounded away from $r_+$ and $\infty$ independently of the frequency parameters.
Trapping for fixed-azimuthal mode solutions {#Vnewsec}
-------------------------------------------
The final result of this section shows in the case of a fixed azimuthal frequency $m$, large $\Lambda$ and $\omega^2 \sim \Lambda$, $r^0_{\rm max}$ occurs outside the ergoregion.
\[aziTrap\] Let $M>0$, $a_0<M$ and $|a|\le a_0$. Recall that we previously defined $\sigma = \frac{am\omega}{\Lambda}$. There exists a small constant $c>0$ such that $|\sigma| \leq c$, $m^2 \leq c\Lambda$ and $c^{-1} \leq \Lambda$ imply that $r^0_{\rm max} > (1+\sqrt{2})M$.
A previous computation showed $$\left(r^2+a^2\right)^3\frac{dV_0}{dr} = 4maM\omega\left(-3r^2+ a^2\right) + 4ra^2m^2 -2\Lambda\left(r^3-3Mr^2 + a^2r + a^2M\right).$$ Since $r^0_{\rm max}$ is the final critical point of $V_0$, we have that $r \geq r^0_{\rm max}$ implies $\frac{dV_0}{dr}\left(r\right) \leq 0$. Hence, the lemma will follow if we can check that $\frac{dV_0}{dr}\left(r = \left(1+\sqrt{2}\right)M\right) > 0$:
$$\Lambda^{-1}\left(r^2+a^2\right)^3\frac{dV_0}{dr}\Big|_{r = \left(1+\sqrt{2}\right)M} =O\left(c\right) - 2\left(M^3\left(1+\sqrt{2}\right)^3 - 3M^3\left(1+\sqrt{2}\right)^2 + a^2M\left(1+\sqrt{2}\right) + a^2M\right).$$ Since we have $$\left(1+\sqrt{2}\right)^2 = 3 + 2\sqrt{2},$$ $$\left(1+\sqrt{2}\right)^3 = 7 + 5\sqrt{2},$$ we obtain $$\begin{aligned}
\Lambda^{-1}\left(r^2+a^2\right)^3\frac{dV_0}{dr}\Big|_{r = \left(1+\sqrt{2}\right)M} &=
O\left(c\right) - 2\left(7M^3 + 5\sqrt{2}M^3 - 9M^3 -6\sqrt{2}M^3 + a^2M + \sqrt{2}a^2M + a^2M\right) \\
&=O\left(c\right) - 2\left(2M\left(a^2-M^2\right) + \sqrt{2}M\left(a^2 - M^2\right)\right).\end{aligned}$$ This is positive for sufficiently small $c>0$.
The importance of the value $r = \left(1+\sqrt{2}\right)M$ comes from the fact that this is the unique location of trapping for axisymmetric solutions to the wave equation on an extreme Kerr background, see [@aretakisKerr].
Note that in the case $a = 0$, one may drop the assumptions $|\sigma| \leq c$ and $\left|m\right|^2 \leq c\Lambda$ and the $O(c)$’s which occur in the proof.
\[ergoTrap\]Of course, the Killing vector field $T$ satisfies $$g\left(T,T\right) = -\left(\frac{r^2 -2Mr + a^2\cos^2\theta}{r^2+a^2\cos^2\theta}\right),$$ which is manifestly negative for $r \geq \left(1+\sqrt{2}\right)M > 2M$.
Aside: relation with null geodesic flow {#ASIDEsec}
---------------------------------------
We note that the potential $V_0$ is intimately related to the potential which appears for the radial dependence of solutions of the geodesic equation, i.e. let $\gamma(s) = \left(t\left(s\right),r\left(s\right),\theta\left(s\right),\phi\left(s\right)\right)$ be a null geodesic.
The conserved quantities associated to stationarity and axisymmetry are $$E \doteq g\left(\dot\gamma,T\right) = -\left(1-\frac{2Mr}{\rho^2}\right)\dot t - \frac{2Mra\sin^2\theta}{\rho^2}\dot\phi,$$ $$L \doteq -g\left(\dot\gamma,\Phi\right) = \frac{2Mra\sin^2\theta}{\rho^2}\dot t - \sin^2\theta\frac{(r^2+a^2)^2 - a^2\sin^2\theta\Delta}{\rho^2}\dot\phi.$$ Carter’s hidden conserved quantity is $$Q \doteq \rho^4\left(\dot\theta\right)^2 + \frac{L^2}{\sin^2\theta} - a^2E^2\cos^2\theta.\footnote{Instead of $Q$ one often finds the Carter constant defined as $K := \rho^4\left(\dot\theta\right)^2 + \frac{(L-aE\sin^2\theta)^2}{\sin^2\theta}$, but $Q$ will relate more naturally to our conventions for the wave equation.}$$
Geodesic motion then reduces to the following system (see [@carter])
$$\rho^2\dot t = a\left(Ea\sin^2\theta-L\right) + \frac{\left(r^2+a^2\right)\left(La-\left(r^2+a^2\right)E\right)}{\Delta},$$
$$\rho^2\dot\phi = \frac{Ea\sin^2\theta-L}{\sin^2\theta} + \frac{a\left(La-(r^2+a^2)E\right)}{\Delta},$$
$$\rho^4\left(\dot\theta\right)^2 = Q + a^2E^2 - 2aEL - \frac{\left(L-aE\sin^2\theta\right)^2}{\sin^2\theta},$$
$$\label{rEqn}
\rho^4\left(\dot r\right)^2 = \left((r^2+a^2)E-aL\right)^2 -\Delta\left(Q + a^2E^2 - 2aEL\right).$$
Note that the right hand side of (\[rEqn\]) be re-arranged to $$\begin{aligned}
\label{arrange}
\left(r^2+a^2\right)^2E^2 - 4MarEL + a^2L^2 - \Delta\left(Q + a^2E^2\right).\end{aligned}$$
Under the correspondence $E \mapsto \omega$, $L \mapsto m$ and $Q \mapsto \lambda_{ml}$, (\[arrange\]) is exactly equal to $(r^2+a^2)^2\left(\omega^2-V_0\right)$. Hence, we can write $\dot r$’s equation as $$\frac{\rho^4}{(r^2+a^2)^2}\left(\dot r\right)^2 = E^2 - V_0\left(E,L,Q,r\right).$$
As a corollary of Lemmas \[lem:1\], \[lem:2\], \[lem:3\], \[lem:4\] and \[aziTrap\], one has that (a) null geodesic flow is hyperbolic in a neighborhood of the set of future trapped null geodesics (b) null geodesics $\gamma$ whose future tangent $\dot\gamma$ has $g_{a,M}(\dot\gamma,T)\ge 0$ are *not* future trapped; they intersect $\mathcal{H}^+$ (c) trapped null geodesics orthogonal to $\partial_\phi$ lie outside of the ergoregion. We shall not however make direct use of any of these facts at the level of geodesic flow.
The separated current templates {#sct}
===============================
Before turning to our estimates we recall the separated current templates of [@dr7] and [@stabi].
The frequency-localised virial currents ${\bf J}^{X,w}$
-------------------------------------------------------
First, we define the frequency-localised analogue of the virial currents ${\bf J}^{X,w}$ where $X$ is in the direction of $\partial_{r^*}$, and $w$ is a suitable function.
Fix Kerr parameters $M>0$ and $|a|<M$ and frequency parameters $\omega\in\mathbb R$, $m\in\mathbb Z$, and $\Lambda\in \mathbb R$. Let $f(r^*)$, $h(r^*)$ and $y(r^*)$ be arbitrary sufficiently regular functions.[^14] With the notation $(\ref{primenotation})$, let us define[^15] the currents $$\begin{aligned}
{\text{Q}}^f[u]&=&
f \left ( |u'|^2 + (\omega^2-V ) |u|^2\right) + f'
{\rm Re}\left(u'\bar u\right)-
\frac 12f'' |u|^2,\\
\text {\fontencoding{LGR}\selectfont \Coppa}^h[u] &=& h {\rm Re} (u'\bar u)-\frac12 h' |u|^2,\\
\text {\fontencoding{LGR}\selectfont \koppa}^y[u] &=& y\left(|u'|^2+(\omega^2-V)|u|^2\right),\end{aligned}$$ associated to the choice of an arbitrary smooth function $u(r^*)$.[^16]
For $u$ satisfying (\[e3iswsntouu\]), we compute: $$\label{eq:Qfor}
(\text{Q}^f[u])'= 2 f' |u'|^2 - f V' |u|^2 + {\rm Re}(2 f \bar{H} u'+f' \bar{H} u)-\frac 12 f''' |u|^2,$$ $$\label{eq:Q1for}
(\text {\fontencoding{LGR}\selectfont \Coppa}^h[u])'=h\left(|u'|^2 +(V-\omega^2)|u|^2\right) -\frac12 h'' |u|^2 +h\, {\rm Re} (u\bar H),$$ $$\label{eq:Q2for}
(\text {\fontencoding{LGR}\selectfont \koppa}^y[u])'= y ' \left(|u'|^2+(\omega^2-V)|u|^2\right) -yV'|u|^2+ 2y\,{\rm Re} (u'\bar H).$$
The virial currents we shall employ will be various combinations of $\text{Q}$, $\text {\fontencoding{LGR}\selectfont \Coppa}$, $\text {\fontencoding{LGR}\selectfont \koppa}$ with suitably selected functions $f$, $h$, $y$. Note that the choice of these functions may depend on $a$, $\omega$, $m$, $\Lambda$, but, again, we temporarily suppress this from the notation.
The frequency-localised conserved energy currents
-------------------------------------------------
As in our survey [@stabi], we shall need, in addition to the above, a frequency-localised analogue of the conserved energy current ${\bf J}^T$. Whereas in [@stabi], we introduced also a frequency-localised version of the red-shift current ${\bf J}^N$, here we shall use in its place a frequency-localised version of the (again conserved) current ${\bf J}^K$.
Again, fix Kerr parameters $M>0$ and $|a|<M$ and frequency parameters $\omega\in\mathbb R$, $m\in\mathbb Z$, and $\Lambda\in \mathbb R$. The “frequency-localised” versions of ${\bf J}^T$ and ${\bf J}^K$ are then defined as follows: $$\begin{aligned}
\text{Q}^T[u]&=&\omega\, {\text Im} (u'\overline {u}),\\
\text{Q}^K[u]&=& \left(\omega - \upomega_+m\right)\, {\text Im} (u'\overline {u}),\end{aligned}$$ where $\upomega_+ = \frac{a}{2Mr_+}$ is the “angular velocity” of the event horizon. For $u$ satisfying $(\ref{e3iswsntouu})$, we have $$\label{eq:Q3for}
\left(\text{Q}^T[u]\right)' = \omega\, {\text Im} (H\overline {u}),$$ $$\label{eq:Q4for}
\left(\text{Q}^K[u]\right)' = \left(\omega - \upomega_+m\right)\, {\text Im} (H\overline {u}).$$
The frequency localised multiplier estimates {#freqLocEst}
============================================
In the present section, using the current templates of Section \[sct\], we will estimate smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with a general smooth right hand side $H$ and which satisfy the boundary conditions (\[eq:b-\]) and (\[eq:b+\]). The point is to obtain estimates which are uniform in the frequency parameters $(\omega, m, \Lambda)$. In view of future applications, we will write the result as an independent theorem. We apply this theorem several times in the present paper (in slightly different contexts) in Sections \[summation\], \[continuityargsec\], \[precise\] and \[boundSuff\]. We remark that the theorem can in principle be applied in future applications independently of the specific setup of Section \[cartersupersection\].
Before stating the theorem, given $|a|\le a_0<M$, set $R_- \doteq r_+ + \frac{1}{2}\left(r_{\rm red}-r_+\right)$ where $r_{\rm red}$ is the constant from Proposition \[specialises..\] and set $R_+ \doteq 2R_{\rm large}$, where $R_{\rm large}$ is the constant from Proposition \[lrp\]. These values will be referred to below. The precise statement of the main result of this section is
\[phaseSpaceILED\] Given $0\le a_0<M$, there exist positive parameters $\omega_{\rm high}$, $\omega_{\rm low}$, $\epsilon_{\rm width}$, $E$ and $R_{\infty}^*$, such that the following is true.
Let $0\le a\le a_0$ and let $(\omega, m, \Lambda)$ be an admissible frequency triple.
Then there exist functions $f$, $h$, $y$, $\hat y$, $\tilde y$, $\chi_1$ and $\chi_2$, and a value $r_{\rm trap}$, depending on the parameters $a_0$, $M$, $a$ and the frequency triple $(\omega, m, \Lambda)$ but satisfying the uniform bounds $$|r_{\rm trap}-r_+|^{-1}+|r_{\rm trap}|+
\left|f\right| + \Delta^{-1}r^2\left|f'\right| + \left|h\right| + \left|y\right| + \left|\tilde y\right| + \left|\hat y\right| + \left|\chi_1\right| + \left|\chi_2\right| \leq B,$$ $$f + y = 1,\ f' = 0,\ h = 0,\ \left|\tilde y\right| \leq B\exp\left(-br^*\right),\ \hat y = 0,\ \chi_1 = 0,\ \chi_2 = 1\text{ for }r^* \geq R^*_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]), we have, $$\begin{aligned}
\label{fromPhaseSpace2}
b&\int_{R^*_-}^{R_+^*}\left[\left|u'\right|^2 + \left(\left(1-r^{-1}r_{\rm trap}\right)^2\left(\omega^2 + \Lambda\right) + 1\right)\left|u\right|^2\right]\, dr^*
\\ \nonumber &\le \int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\, dr^* + 1_{\{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}\}\cap \{\Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2.\end{aligned}$$ The symbol $1_{\{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}\}\cap \{\Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}$ denotes the indicator function for the set $\{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}\}\cap \{\Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}$, and $$\begin{aligned}
H \cdot (f, h, y, \chi) \cdot (u, u') \doteq &-2f\text{Re}\left(u'\overline{H}\right) -f'\text{Re}\left(u\overline{H}\right) + h\text{Re}\left(u\overline{H}\right) -E\chi_2\omega\text{Im}\left(H\overline{u}\right)
\\ \nonumber &\,\,- E\chi_1\left(\omega-\upomega_+m\right)\text{Im}\left(H\overline{u}\right) -2y\text{Re}\left(u'\overline{H}\right)-2\tilde y\text{Re}\left(u'\overline{H}\right) -2\hat y\text{Re}\left(u'\overline{H}\right).\end{aligned}$$
Before discussing the proof of the theorem, we give a few remarks pertaining to the application of Theorem \[freqLocEst\] in Section \[summation\] in the context of $u$ arising from Carter’s separation applied to a solution $\Psi$ of the inhomogeneous wave equation.
\[Traprem\] For frequencies in the trapping regime, $r_{\rm trap}$ will denote the unique trapped value of $r$ associated to the triple $\left(\omega,m,\Lambda\right)$. Otherwise, $r_{\rm trap}$ will be set to $0$. This will capture the degeneration due to trapping.
\[R2\] The specific behaviour of the functions $f$, $h$, $y$, $\hat y$, $\tilde y$, $\chi_1$ and $\chi_2$ in the region $r^* \geq R_{\infty}^*$ will be useful in Section \[summation\] when we sum (\[fromPhaseSpace2\]) to produce a physical space estimate.
\[thisIsPhaseILED\] If we consider the right hand side of the estimate (\[fromPhaseSpace2\]) as “data”, a direct application of Plancherel (see the explicit formulas in Section \[separationSubsection\]) shows that (\[fromPhaseSpace2\]) is the phase space versions of integrated local energy decay.
Let us draw particular attention to the term $1_{\{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}\}\cap \{\Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2$ on the right hand side of the estimate (\[fromPhaseSpace2\]). This term must initially be put on the right hand side of the corresponding integrated energy decay statement (cf. Remark \[thisIsPhaseILED\]). Eventually, this term will be dealt with in Section \[whitinghere\] using the quantitative refinement [@shlapRot] of mode stability.
The proof proper of Theorem \[phaseSpaceILED\] will be given in Section \[putting\]. It will be based on a series of propositions proven in Sections \[largeSuper\]–\[whit\] below, where $(\ref{fromPhaseSpace2})$ is successively obtained for various ranges of admissible frequency triples. These frequency ranges, however, are determined by parameters which must be suitably optimised so as for our constructions to be possible. We begin thus with a discussion of these ranges and an overview of the constructions.
The frequency ranges {#freerange}
--------------------
Let $a_0<M$. Fix a parameter $\alpha$ (depending only on $a_0$, $M$) satisfying the statement of Lemma \[lem:3\]. For each $0\le a\le a_0$, and all $\omega_{\rm high}>0$, $\epsilon_{\rm width}>0$, we define the frequency ranges $\mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm high})$, $\mathcal{G}_{{{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}}(\omega_{\rm high}, \epsilon_{\rm width})$, $\mathcal{G}_{\mbox{$\natural$}}(\omega_{\rm high}, \epsilon_{\rm width})$, $\mathcal{G}_{\mbox{$\sharp$}}(\omega_{\rm high},\epsilon_{\rm width})$, $\mathcal{G}^{\mbox{$\sharp$}}(\omega_{\rm high},\epsilon_{\rm width})$ by
- $\mathcal{G}^{\mbox{$\sharp$}}=\{(\omega, m, \Lambda){\rm\ admissible\ }$ : $\Lambda \ge (\frac{a}{2Mr_+} + \alpha)^{-2}\omega_{\rm high}^2$, $m\omega\in (0,\frac {am^2}{2Mr_+}+\alpha\Lambda]\}$,
- $\mathcal{G}_{\mbox{$\sharp$}}=\{(\omega, m, \Lambda){\rm\ admissible\ }$ : $|\omega|\ge \omega_{\rm high}$, $\Lambda<\epsilon_{\rm width} \omega^2, m\omega\not\in (0,\frac {am^2}{2Mr_+}+\alpha\Lambda]\}$,
- $\mathcal{G}_{\mbox{${{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}$}}=\{(\omega, m, \Lambda){\rm\ admissible\ }$ : $\Lambda \geq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2$, $\epsilon_{\rm width} \Lambda > \omega^2, m\omega\not\in (0,\frac {am^2}{2Mr_+} + \alpha\Lambda]\}$,
- $\mathcal{G}_{\mbox{$\natural$}}=\{(\omega, m, \Lambda){\rm\ admissible\ }$ : $|\omega|\ge \omega_{\rm high}$, $\epsilon_{\rm width} \Lambda \le \omega^2\le \epsilon_{\rm width}^{-1} \Lambda, m\omega\not\in (0,\frac {am^2}{2Mr_+}+\alpha\Lambda]\}$,
- $\mathcal{G}_{\mbox{$\flat$}}=\{(\omega, m, \Lambda){\rm\ admissible\ }$ : $|\omega| < \omega_{\rm high}$, $\Lambda < \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}$.
The parameters $\omega_{\rm high}$ and $\epsilon_{\rm width}$ will be fixed in the course of the proof of Theorem \[freqLocEst\], see Section \[putting\].
We see easily that
\[everythingCovered\] With the above notation, for all $0\le a\le a_0$ if $(\omega, m, \Lambda)$ is admissible, then, for all choices of parameters $\omega_{\rm high}$, $\epsilon_{\rm width}$, $(\omega, m, \Lambda)$ lies in exactly one of the frequency ranges $\mathcal{G}^{\mbox{$\sharp$}}$, $\mathcal{G}_{\mbox{$\sharp$}}$, $\mathcal{G}_{\mbox{${{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}$}}$, $\mathcal{G}_{\mbox{$\natural$}}$, or $\mathcal{G}_{\mbox{$\flat$}}$.
To see this, observe that $$|\omega| \geq \omega_{\rm high} \text{ and }m\omega \in \left(0,\frac{am^2}{2Mr_+}+\alpha\Lambda\right] \Rightarrow
\Lambda \ge \left(\frac{a}{2Mr_+} + \alpha\right)^{-2}\omega_{\rm high}^2.$$
Our constructions of currents will vary according to the frequency range of the triple $(\omega, m,\Lambda)$. We now give an overview of these constructions.
Overview
--------
For each admissible triple $(\omega, m, \Lambda)$, we would like to find a current $\text{Q}$ consisting of various combinations of $\text{Q}^f$, $\text {\fontencoding{LGR}\selectfont \koppa}^y$, $\text {\fontencoding{LGR}\selectfont \Coppa}^h$, $\text{Q}^T$ and $\text{Q}^K$ satisfying the bulk coercivity property $$\label{yaxvw}
\int_{-\infty}^{\infty} \text{Q}'[u] \ge b\int_{R^*_-}^{R^*_+} \left(|u'|^2 +\left(1-r^{-1}r_{\rm trap}\right)^2(\Lambda +\omega^2)|u|^2 +|u|^2\right) - \int_{-\infty}^{\infty}
H \cdot (f, h, y, \chi) \cdot (u, u'),$$ and, ideally, the boundary positivity property $$\label{boundaryOK}
\text{Q}\left(\infty\right) - \text{Q}\left(-\infty\right) \leq 0.$$
The terms Q, $r_{\rm trap}$, $H$, $f$, $h$, $y$ should all be understood to depend on $\omega$, $m$, and $\Lambda$, here omitted for brevity, and the integrals are with respect to $r^*$. One restricts the domain of integration on the first term to $[R_-^*,R_+^*]$ on the right hand side because one expects this virial current not to control things at the horizon and infinity.
The most difficult aspect of establishing (\[yaxvw\]) is the need to understand trapping. In order to do this this we will heavily rely on the analysis of the potential $V_0$ carried out in Section \[Vpropsec\]. For frequencies for which trapping is relevant, $r_{\rm trap}$ will denote the unique value of $r$, associated with the frequency triple, where the estimate must degenerate. For frequencies where trapping is not relevant, $r_{\rm trap} = 0$.
The fundamental obstruction to achieving (\[boundaryOK\]), on the other hand, is superradiance (see Section \[supInKerr\]). For non-superradiant frequencies, i.e. frequencies which satisfy $\omega\left(\omega - \upomega_+m\right) \geq 0$, one may easily[^17] control these fluxes via a sufficiently large multiple of the conserved Q$^T$ current: $$\label{microEnergyEst}
\int_{-\infty}^{\infty}\text{Im}\left(H\overline{u}\right) = \int_{-\infty}^{\infty}\left(\text{Q}^T\right)' = \text{Q}^T(\infty) - \text{Q}^T(-\infty) = \omega^2\left|u(\infty)\right|^2 + \omega\left(\omega-\upomega_+m\right)\left|u(-\infty)\right|^2.$$ However, for superradiant frequencies, where $\omega\left(\omega-\upomega_+m\right) < 0$, no conserved current gives a coercive estimate for the boundary terms and it is thus no longer clear how to arrange for (\[boundaryOK\]). As it turns out, see Section \[largeSuper\] below, one of the miracles of the Kerr geometry is that trapping and superradiance are disjoint; exploiting this, one may indeed establish (\[boundaryOK\]) for sufficiently large frequencies with the help of (\[yaxvw\]) and a large positive parameter. Unfortunately, for bounded superradiant frequencies, one does not have a large parameter at hand. We will not be able to carry out such a scheme, and we will not in fact establish (\[boundaryOK\]); see Section \[theseAreBounded\].
We now turn to a more detailed discussion of the difficulties in each frequency range. The reader may wish to refer to this when reading Sections \[largeSuper\]–\[whit\] below.
### The $\mathcal{G}^{\mbox{$\sharp$}}$ range {#largeSuper000}
This is the large frequency superradiant regime. Lemma \[lem:4\] shows that these frequencies are not trapped. Thus, it is not difficult to establish (\[yaxvw\]) via the combination of a $\text{Q}^f$ and $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ current with a monotonically increasing $f$ which switches signs at the unique maximum of the potential and a positive function $h$ which peaks near the maximum of the potential.
As for the boundary terms, despite the lack of a coercive conserved current, we will appeal to the aforementioned miracle that superradiant frequencies are not trapped to find a large parameter which will still allow us to achieve (\[boundaryOK\]). Briefly put, Lemma \[lem:4\] shows that we have a quantitatively large “classically forbidden region”, and from this one expects to derive an estimate for $u$ near $r_{\max}$ which comes with a large parameter.
### The $\mathcal{G}_{\mbox{$\sharp$}}$ range {#lanosudiscu}
This is a non-superradiant regime where the time frequency $\omega$ is large and dominates the other parameters. It is easy to see that a $\text {\fontencoding{LGR}\selectfont \koppa}^y$ current with an appropriate choice of $y$ will establish (\[yaxvw\]).
Of course, the boundary terms may be easily controlled with (\[microEnergyEst\]).
### The $\mathcal{G}_{\mbox{${{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}$}}$ range {#boundaryStillaProblem}
This is a non-superradiant regime where the angular frequency $\Lambda$ is large and dominates the other parameters. One may easily show that the conclusions of Lemma \[lem:4\] still hold, and, as in Section \[largeSuper\], it is not difficult to establish (\[yaxvw\]).
Turning to the boundary terms, note that $\omega\left(\omega-\upomega_+m\right)$ and $\left(\omega-\upomega_+m\right)^2$ are not necessarily comparable in this regime. Thus, even though the flux $\text{Q}^T[u]|_{r=\infty}$ may be controlled with (\[microEnergyEst\]), the estimate (\[microEnergyEst\]) does not provide sufficient control of the flux $\text{Q}^K[u]|_{r=r_+}$. Fortunately, we may apply the same argument as in the Section \[largeSuper\] to control the horizon flux.
### The $\mathcal{G}_{\mbox{$\natural$}}$ range {#spartaheretoo}
This is a non-superradiant regime where the angular frequency $\Lambda$ and the time frequency $\omega$ are large and comparable. and hence the only frequency range where $r_{\rm trap}$ will be non-zero. The estimate (\[yaxvw\]) is achieved via a Q$^f$ current with a monotonically increasing function $f$ which switches sign at the unique maximum of the potential. The construction the function $f$ will heavily depend on the critical point analysis of $V_0$ carried out in Section \[Vtrsec\].
The estimate (\[boundaryOK\]) is easily achieved via (\[microEnergyEst\]).
### The $\mathcal{G}_{\mbox{$\flat$}}$ range {#theseAreBounded}
This is a bounded frequency regime. It turns out to be useful to further split this frequency regime into the following four sub-regimes.
1. $\left|\omega\right| \leq \omega_{\rm low}$, $0\le a < \tilde a_0$ and $m \neq 0$.
2. $\left|\omega\right| \leq \omega_{\rm low}$ and $m =0$.
3. $\left|\omega\right| \leq \omega_{\rm low}$, $m \neq 0$ and $a \geq \tilde a_0$.
4. $|\omega|\ge \omega_{\rm low}$.
Here $\omega_{\rm low}$ is the small parameter mentioned in Theorem \[freqLocEst\] and $\tilde a_0$ is a small parameter to be fixed in the course of the proof.
For the estimate (\[yaxvw\]) we will exploit two types of estimates. If $|\omega|\ge \omega_{\rm low}$ or $\left|\omega\right| \leq \omega_{\rm low}$, $m \neq 0$ and $a \geq \tilde a_0$, then we will either have $\omega^2 \sim 1$ or $\left(\omega - \upomega_+m\right)^2 \sim 1$. In this case we will employ $\text {\fontencoding{LGR}\selectfont \koppa}^y$ currents with exponential multipliers $y \doteq \exp\left(\int \upsilon\right)$ and appropriate functions $\upsilon$. If $\left|\omega\right| \leq \omega_{\rm low}$ and $\omega_{\rm low}$ is sufficiently small, then in regions with $1 \lesssim V$ we will have $1 \lesssim V - \omega^2$. We will apply $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ currents to exploit this positivity of $V - \omega^2$.
As in Section \[largeSuper\], the fundamental difficulty is a lack of control of the boundary terms for superradiant frequencies. It turns out that when $\omega^2 \ll 1$, i.e. $\left|\omega\right| \leq \omega_{\rm low}$ for $\omega_{\rm low}$ sufficiently small, then $\omega$ arises naturally as a small parameter and we will again be able to achieve (\[boundaryOK\]). However, for bounded frequencies with $|\omega| \geq \omega_{\rm low}$ there is no large or small parameter to exploit. Instead, for this frequency range we will only be able to establish the weaker $$\label{asthenes}
\text{Q}(\infty) - \text{Q}(-\infty) \leq B\left|u\left(-\infty\right)\right|^2.$$ This is the origin of the term $1_{\{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}\}\cap \{\Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2$ on the right hand side of the estimate (\[fromPhaseSpace2\]).
We now turn to the detailed constructions of the currents for each frequency regime.
The $\mathcal{G}^{\mbox{$\sharp$}}$ range {#largeSuper}
-----------------------------------------
As discussed in Section \[largeSuper000\], this defines a large frequency superradiant regime, and by the results of Section \[Vsrsec\], frequencies in this regime can be viewed as non-trapped.
Once we have made our final choice of the parameter $\omega_{\rm high}$, then for $(\omega, m, \Lambda)\in \mathcal{G}^{\mbox{$\sharp$}}(\omega_{\rm high})$, we will set the functions $y$, $\hat y$ and $\tilde y$ together with the parameter $r_{\rm trap}$ from the statement of Theorem \[phaseSpaceILED\] to be $0$. The desired coercivity in this range and remaining functions $f$, $h$, $\chi_1$ and $\chi_2$ are given by the following:
\[odeEst1\] Let $a_0<M$. Then, for all $E\ge 2$, for all $\omega_{\rm high}$ sufficiently big depending on $E$, for all $R_{\infty}$ sufficiently big, for all $0\le a\le a_0$, $(\omega, m, \Lambda)\in \mathcal{G}^{\mbox{$\sharp$}}(\omega_{\rm high})$, there exist functions $f$, $h$, $\chi_1$ and $\chi_2$ satisfying the uniform bounds $$\left|f\right| + \Delta^{-1}r^2\left|f'\right| + \left|h\right| + \left|\chi_1\right| + \left|\chi_2\right| \leq B\left(\omega_{\rm high}\right),$$ $$f = 1,\ h = 0,\ \chi_1 = 0\text{ and }\chi_2 = 1\text{ for }r\geq R_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]) we have the estimate $$\begin{aligned}
\label{sharpEst}
b&\int_{R^*_-}^{R^*_+}\left (|u'|^2+(\omega^2+\La) |u|^2\right)
\\ &\leq \int_{-\infty}^\infty \left (-2f \,{\text{Re}} (u'\overline{H}) - (f'+h)\, {\text{Re}} (u\overline{H})+E\chi_2\omega {\text{Im}} (H\overline u) + E\chi_1\left(\omega-\upomega_+m\right){\text{Im}}(H\overline u)\right ).\end{aligned}$$
As $\mathcal{G}^{\mbox{$\sharp$}}$ is a superradiant regime with $\Lambda > 0$, the conclusions of both Lemma \[lem:3\] and Lemma \[lem:4\] apply. In particular, the potential $V_0$ has a unique $r^0_{\rm max}$ which is a maximum, and satisfies $$\label{reallyNotTrapped}
V_0(r^0_{\rm max}) - \omega^2 \geq c\Lambda,$$ for some positive constant $c$ depending only on $a_0$ and $M$.
We shall first need to establish the following lemma, which shows that the full potential $V$ behaves similarly in the range $\mathcal{G}^{\mbox{$\sharp$}}\left(\omega_{\rm high}\right)$ for sufficiently large $\omega_{\rm high}$.
\[theOtherLemma\]There exists a $\delta > 0$ depending only on $a_0$ and $M$ such that for sufficiently large $\omega_{\rm high}$ and $(\omega,m,\Lambda) \in \mathcal{G}^{\mbox{$\sharp$}}\left(\omega_{\rm high}\right)$, then $V$ has a unique critical point $r_{\rm max}$ and satisfies $$V(r)-\omega^2\ge b\Lambda,\qquad \forall r\in (r_{\max}-\de,r_{\max}+\de),$$ $$\label{eq:Vdeg}
-(r-r_{\max}) \frac d{dr} V(r)\ge b \Lambda \frac {(r-r_{\max})^2}{r^{4}}, \qquad \forall r\in [r_+,\infty),$$ $$\left|r_{\rm max} - r^0_{max}\right| \leq B\Lambda^{-1}.$$
Let us first refine our estimates on $V_0$ for frequencies $(\omega, m, \Lambda)\in
\mathcal{G}^{\mbox{$\sharp$}}\left(\omega_{\rm high}\right)$.
Using the fact that $\left|\frac{dV_0}{dr}\right| \leq B\Lambda$, (\[reallyNotTrapped\]) implies that we may find a $\delta_1 > 0$ depending only on $a_0$ and $M$ such that $$V_0(r)-\omega^2\ge \frac{c}{2} \Lambda,\qquad \forall\, r\in [r^0_{max}-\delta_1,r^0_{max}+\delta_1].$$
Lemma \[lem:1\] implies that $r^0_{\rm max}$ is bounded from above independently of the frequency parameters: $$r^0_{\rm max} \le B.$$ Furthermore, Lemma \[lem:3\] and the bound $\left|\frac{d^2V_0}{dr^2}\right| \leq B\Lambda$ implies $r^0_{\rm max}$ is also bounded away from $r_+$ independently of the frequency parameters: $$r^0_{\rm max} - r_+ \ge b.$$
Lemma \[lem:3\] also implies that the full potential $V=V_0+V_1$ satisfies $$\label{increaseOnHorizon}
\frac {d}{dr} V(r_+)\ge \frac {d}{dr} V_0(r_+)\ge \hat c \Lambda$$ for a positive constant $\hat c$ depending only on $a_0$ and $M$.
Recall now that the proof of Lemma \[lem:1\] showed that the function $$\frac {d}{dr} \left ((r^2+a^2)^3\frac {d}{dr} V_0(r)\right)$$ is either non-positive on $[r_+,\infty)$ or there exists a unique point $r_+ \leq r_1 < r^0_{\rm max}$ such that $\frac {d}{dr} \left ((r^2+a^2)^3\frac {d}{dr} V_0(r)\right)$ is positive on $[r_+,r_1)$ and negative on $(r_1,\infty)$.
We first consider the case where the point $r_1$ exists. Then, $$\frac {d}{dr} V_0(r)\ge \hat c \frac{(r_+^2+a^2)^3}{(r_1^2+a^2)^3}\Lambda\qquad \forall\, r\in [r_+,r_1].$$ Next, recall from the proof of Lemma \[lem:1\] that $$\frac{d}{dr}\left((r^2+a^2)^3\frac{dV_0}{dr}\right) = -6\Lambda\left(r^2-2Mr +4Mr\sigma+\frac{a^2}3-\frac 23 a^2 \frac{m^2}\Lambda\right),$$ and furthermore, by definition, $\frac{d}{dr}\left((r^2+a^2)^3\frac{dV_0}{dr}\right)$ is negative on $(r_1,\infty)$. Thus, we can choose a value $r'_1\in (r_1,r_{max}^0)$ such that $$\label{aProp1}
\frac {d}{dr} V_0(r)\ge \frac {\hat c}{2}\frac{(r_+^2+a^2)^3}{(r_1^2+a^2)^3}\Lambda,\qquad \forall r\in [r_+,r_1'],$$ and $$\label{aProp2}
\frac {d}{dr} \left ((r^2+a^2)^3\frac {d}{dr} V_0(r)\right)\le -\tilde c \Lambda r^2,\qquad \forall r\in [r_1',\infty),$$ for a positive constant $\tilde c$ independent of the frequency parameters.
In the case where $r_1$ does not exists, the same argument *mutatis mutandis* will produce a value $r_1'$ with the properties (\[aProp1\]) and (\[aProp2\]).
Now, we simply observe that the potential $V_1$ satisfies the bounds $$\left|V_1\right|\le B r^{-3},\qquad
\left|\frac {d}{dr} V_1(r)\right|\le B r^{-4},\qquad \left|\frac {d}{dr} \left ((r^2+a^2)^3\frac {d}{dr} V_1(r)\right)\right|\le B r.$$
For $\omega_{\rm high}$ sufficiently large (and hence large $\Lambda$), it immediately follows that, for $(\omega, m, \Lambda)\in
\mathcal{G}^{\mbox{$\sharp$}}\left(\omega_{\rm high}\right)$, the full potential $V=V_0+V_1$ cannot have any critical points on $[r_+,r_1']$ and has a unique maximum $r_{\max}\in [r_1',\infty)$ which satisfies $\left|r_{\max} - r^0_{max}\right| \leq B\Lambda^{-1}$.
The proof concludes by applying the fact that $\left|\frac{dV}{dr}\right| \leq B\Lambda$.
We now proceed to the construction of a suitable current for the regime $\mathcal{G}^{\mbox{$\sharp$}}$. The current will be of the form: $$\text{Q}=\text{Q}^f + \text {\fontencoding{LGR}\selectfont \Coppa}^h - E\chi_1\text{Q}^K - E\chi_2\text{Q}^T,$$ for appropriate functions $f$, $h$, $\chi_1$ and $\chi_2$ and large constant $E$.
It is simpler to describe this procedure in three stages.
[**Stage 1.**]{} We first apply current $\text{Q}^f$ where $f$ is a function chosen such that $$\label{choices1}
f=-1 {\rm\ at\ } r=r_+, \qquad f=0 {\rm\ at\ } r=r_{\max},\qquad
f=1 {\rm\ when\ } r^* \geq R^*_{\infty},$$ $$\label{choices2}
f'(r^*) > 0 {\rm\ for\ all\ } r \leq R_1, \qquad f'(r^*) \geq 0 {\rm\ for\ all\ } r>r_+, \qquad \left|f\right| + \Delta^{-1}r^2\left|f'\right| \leq B.$$
Application of $(\ref{eq:Qfor})$ yields then $$\begin{aligned}
\label{metastra}
\nonumber
\int_{-\infty}^\infty \left (2f'|u'|^2-fV' |u|^2 -\frac 12 f{'''}|u|^2\right )&=
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} + \left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty}
\\ &\qquad- \int_{-\infty}^\infty \left (2 f \,{\text{Re}} (u'\overline{H}) + f'\, {\text{Re}} (u\overline{H})\right).\end{aligned}$$ Let us moreover require that $f$ above has been chosen so that in addition to $(\ref{choices1})$, $(\ref{choices2})$, the following coercivity property holds $$\label{choices3}
-fV'-\frac 12 f'''\ge \La \frac {\Delta (r-r_{\max})^2}{r^7}, {\rm\ for\ all\ } r>r_+.$$ Since $f$ vanishes at $r=r_{\max}$ and $V'$ obeys the property , we can easily arrange such that in addition to $(\ref{choices1})$, $(\ref{choices2})$ and $(\ref{choices3})$, we have $$\label{choices4}
fV'\ge b \La \frac {\Delta (r-r_{\max})^2}{r^7}.$$ It remains to impose $$\label{choices5}
f'''(r)<0 {\rm\ in\ a\ small\ neighbourhood\ of\ }r_{\max}, \qquad
|f'''(r)|\le B \Delta r^{-5}.$$
Note that the reader may easily construct a function $f$ satisfying the conditions (\[choices1\]), (\[choices2\]), (\[choices3\]), (\[choices4\]) and (\[choices5\]). With the above choice of $f$, the left hand side of $(\ref{metastra})$ is now non-negative, but still degenerate at $r=r_{\max}$. As discussed in Section \[Vsrsec\], the bound $V(r_{\max})-\omega^2\ge b\La$ indicates this regime is non-trapped and thus the degeneracy may be removed with the help of the current $\text {\fontencoding{LGR}\selectfont \Coppa}^h$. The more serious problem is a lack of control of the boundary terms on the right hand side, due to the superradiant condition. However, as we shall see below, we will be able to overcome this by exploiting the largeness of the potential in the region $(r_{\max}-\de,r_{\max}+\de)$.
[**Stage 2.**]{} We now add a $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ current with a function $h \doteq A\tilde h$ such that $$\label{choices6}
h \geq 0,\qquad \left|\tilde h\right| \leq B,$$ $$\label{choices7}
\text{supp}\left(h\right) \subset [r_{\max}-\de,r_{\max} + \de],\qquad \tilde h = 1\text{ for } r \in [r_{\max}-\de/2,r_{\max}+\de/2]$$ and $A$ is a constant to be determined.
We obtain $$\begin{aligned}
\label{metatastra2}
\nonumber
\int_{-\infty}^\infty &\left ((2f'+Ah) |u'|^2+\left (A\tilde h(V-\omega^2)-fV'\right) |u|^2 -\frac 12 (f{'''}+Ah'')|u|^2\right )\\&=
\left(|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} + \left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty}
- \int_{-\infty}^\infty \left(2 f \,{\text{Re}} (u'\overline{H}) + (f'+h)\, {\text{Re}} (u\overline{H})\right).\end{aligned}$$ Note that as long as $A \leq \tilde\epsilon\omega_{\rm high}^2$ for a sufficiently small constant $\tilde \epsilon$ only depending on $a_0$ and $M$, the integrand of the left hand side of $(\ref{metatastra2})$ will be positive. Moreover, this integrand has the property that it satisfies $$\ge bA (|u'|^2 + \Lambda |u|^2),\quad \forall\, r\in [r_{\max}-\frac{\delta}2,r_{\max}+\frac\delta 2].$$
[**Stage 3.**]{} We now let $\chi_1(r)$ be a smooth function such that $$\label{somerequirements}
\chi_1 = 1\text{ for } r\in [r_+,r_{\max}-\frac{\de}{2}],\qquad \chi_1 = 0\text{ for }r \in [r_{\max}+\frac{\delta}{2},\infty),\qquad \left|\chi_1\right| \leq B.$$ Since $E \geq 2$, we have $$\begin{aligned}
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} &\leq E\int_{-\infty}^{\infty}(\chi_1\text{Q}^K)' \\&=
E\int_{r_{\max}-\frac{\de}{2}}^{r_{\max}+\frac{\de}{2}}\chi_1'\left(\omega-\upomega_+m\right)\text{Im}\left(u'\overline{u}\right) + E\int_{-\infty}^{\infty}\chi_1\left(\omega-\upomega_+m\right)\text{Im}\left(H\overline{u}\right).\end{aligned}$$ Now, we require that $\omega_{\rm high}$ be sufficiently large so as to satisfy $E\delta^{-1} \ll (1/2)\tilde\epsilon\omega_{\rm high}^2$, and then we set $A \doteq (1/2)\tilde\epsilon\omega_{\rm high}^2$. This choice of $A$ will both maintain the coercivity of the left hand side of (\[metatastra2\]) and yield $$\begin{aligned}
E\left|\int_{r_{\max}-\frac{\de}{2}}^{r_{\max}+\frac{\de}{2}}\chi_1'\left(\omega-\upomega_+m\right)\text{Im}\left(u'\overline{u}\right)\right| &\leq& E\delta^{-1}\int_{r_{\max}-\frac{\de}{2}}^{r_{\max}+\frac{\de}{2}}\left(\left|u'\right|^2 + \left(\omega^2+m^2\right)\left|u\right|^2\right) \\
&\ll& A\int_{r_{\max}-\frac{\de}{2}}^{r_{\max}+\frac{\de}{2}}\left(\left|u'\right|^2 + \Lambda\left|u\right|^2\right).\end{aligned}$$ We can, of course, carry out an analogous construction with a cutoff $\chi_2$, satisfying $$\label{somerequirements2}
\chi_2 = 1\text{ for }r \in [r_{\max} + \frac{\de}{2},\infty),\qquad \chi_2 = 0\text{ for }r\in [r_+,r_{\max}-\frac{\de}{2}],\qquad \left|\chi_2\right| \leq B,$$ and the current $\text{Q}^T$. Then, adding the currents $-E\chi_1\text{Q}^K-E\chi_2\text{Q}^T$ will give us the necessary control of the boundary terms.
Observing that the left hand side of the resulting estimate is coercive (with weights which degenerate however as $r^* \to \pm\infty$), restricting the domain of integration of the left hand side then yields (\[sharpEst\]).
The $\mathcal{G}_{\mbox{$\sharp$}}$ range {#sub:sharp}
-----------------------------------------
As discussed in Section \[lanosudiscu\], $\mathcal{G}_{\mbox{$\sharp$}}$ defines a large frequency regime (whose definition still depends on parameters $\omega_{\rm high}$ and $\epsilon_{\rm width}$, yet to be fixed) where time frequencies will dominate angular frequencies. The regime is manifestly non-superradiant, and, for suitable choice of parameters, non-trapped.
Once we have made our final choice of parameters $\omega_{\rm high}$ and $\epsilon_{\rm width}$, then for $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\sharp$}}(\omega_{\rm high},\epsilon_{\rm width})$, we will set the functions $f$, $h$, $\hat y$, $\tilde y$ and $\chi_1$ appearing in Theorem \[phaseSpaceILED\] together with the parameter $r_{\rm trap}$ to be $0$. The remaining function $y$ and the desired coercivity property are given by
\[odeEst2\]Let $a_0<M$. Then, for all $\omega_{\rm high}$, $\epsilon_{\rm width}^{-1}$, $R_{\infty}$ sufficiently big, for all $E\ge 2$, $0\le a\le a_0$, $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\sharp$}}(\omega_{\rm high},\epsilon_{\rm width})$, there exists a function $y$ satisfying the uniform bounds $$\left|y\right| \leq B,$$ $$\label{someyStuff}
y = 1\text{ for }r^* \geq R^*_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]) we have the estimate $$\begin{aligned}
b\int_{R^*_-}^{R^*_+}\left( |u'|^2+{(\omega^2+\La)} |u|^2\right)\le
\int_{-\infty}^\infty \left (-2 y \,{\text{Re}} (u'\overline{H}) +E\omega {\text{Im}} (H \overline u)\right ).\end{aligned}$$
The construction of our currents will exploit the fact that the range $\mathcal{G}_{\mbox{$\sharp$}}$ defines a large frequency regime in which $\Lambda \ll \omega^2 $ (and thus also $m^2 \ll \omega^2$). To handle the boundary terms, we will use that this regime is moreover manifestly non-superradiant, and thus addition of a sufficiently large multiple of the $\text{Q}^T$ current provides positive terms at $r=r_+$ and $r=\infty$.
We turn to the details. First of all, it is easy to see that the admissibility inequalities $\Lambda \geq 2a|m\omega|$ and $\Lambda \geq \left|m\right|(\left|m\right|+1)$ imply that there exists a constant $R_{\rm dec}^* \geq 2R^*_+$ only depending on $a_0$ and $M$ such that $$\label{decrease}
V' < 0\text{ for }r^* \geq R_{\rm dec}^*.$$
Define a current given by the following expression: $$\text{Q}=\text {\fontencoding{LGR}\selectfont \koppa}^y - E \text{Q}^T.$$ We require that $$\label{choices8}
\left|y'\right| \leq B,\qquad y' \geq 0\text{ for }r \in [r_+,\infty),\qquad y' > 0\text{ for }r^* \in
[R^*_-,R^*_1],$$ $$\label{choices9}
\frac{1}{2} \leq y \leq 1\text{ for }r^* \in (-\infty,R^*_+],\qquad y(-\infty) = 1/2,\qquad y = 1\text{ for }r^* \geq R^*_{\rm dec}.$$ Such a $y$ is trivial to construct.
We obtain from (\[eq:Q2for\]) and (\[eq:Q3for\]) the identity $$\begin{aligned}
\label{whatweobtainh}
\nonumber
&\int_{-\infty}^\infty \left (y' |u'|^2+\left (y'(\omega^2-V)-yV'\right) |u|^2 \right )\\&\qquad
\nonumber
-\left (\frac{1}{2}|u'|^2+\left(\frac{1}{2}(\omega-\upomega_+m)^2-E\omega(\omega-\upomega_+m)\right) |u|^2\right)_{r=r_+} - \left (|u'|^2+(1-E)\omega^2 |u|^2\right)_{r=\infty}
\\ &= \int_{-\infty}^\infty \left (-2 y \,{\text{Re}} (u'\overline{H})+E\omega {\text{Im}} (\overline H u) \right).\end{aligned}$$ Next, we observe the bound $$\label{someBoundS}
|V|\le B\left(\epsilon^{-1}\Lambda + \epsilon\omega^2 + \omega_{\rm high}^{-2})\omega^2\right),\qquad |V'|\le \frac{B\Delta}{r^5}
\left(\epsilon^{-1}\Lambda + \epsilon\omega^2 + \omega_{\rm high}^{-2})\omega^2\right)$$ where $\epsilon > 0$ is arbitrary.
Now, we fix a sufficiently small $\epsilon > 0$, require that $\epsilon_{\rm width}$ is sufficiently small depending on $\epsilon$, and combine the inequality $\omega^2 > \epsilon_{\rm width}^{-1}\Lambda$ with the inequalities (\[someBoundS\]) and (\[decrease\]). We conclude the integrand on the left hand side of $(\ref{whatweobtainh})$ is non-negative and bounds from above the expression $$b \int_{R_-^*}^{R_+^*}\left(|u'|^2+(\omega^2+\La) |u|^2\right).$$ The boundary terms are non-negative due to the boundary conditions (\[eq:b-\]) and (\[eq:b+\]), the non-superradiance condition and the requirement $E \geq 2$. Requiring that $R^*_{\infty} > R^*_{\rm dec}$ ensures that (\[someyStuff\]) is satisfied.
The $\mathcal{G}_{\mbox{${{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}$}}$ range {#angularDominated}
----------------------------------------------------------------------------
As described in Section \[boundaryStillaProblem\], this is again a large frequency regime (whose definition still depends on parameters $\omega_{\rm high}$ and $\epsilon_{\rm width}$ yet to be fixed), but where angular frequencies will now dominate time frequencies. The regime is again manifestly non-superradiant, and, for suitable parameters, non-trapped, but as we shall see, we will have to handle the horizon boundary term as in the superradiant regime.
Once we have made our final choice of the parameters $\omega_{\rm high}$ and $\epsilon_{\rm width}$, then for $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{${{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}$}}(\omega_{\rm high},
\epsilon_{\rm width})$, we set the functions $y$, $\tilde y$ and $\hat y$ together with the parameter $r_{\rm trap}$ to be $0$. The remaining functions $f$, $h$ and $\chi_1$ and the desired coercivity properties are given by
\[odeEst3\] Let $a_0<M$. Then, for all $\omega_{\rm high}$, $R_{\infty}$ and $\epsilon_{\rm width}^{-1}$ sufficiently large, for all $E\ge 2$, $0\le a\le a_0$, $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\sharp$}}(\omega_{\rm high},
\epsilon_{\rm width})$, there exist functions $f$, $h$ and $\chi_1$ satisfying the uniform bounds $$\left|f\right| + \Delta^{-1}r^2\left|f'\right| + \left|h\right| + \left|\chi_1\right| + \left|\chi_2\right| \leq B\left(\omega_{\rm high},\epsilon_{\rm width}\right),$$ $$f = 1,\ h = 0,\ \chi_1 = 0\text{ for }r^* \geq R^*_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]), we have the estimate $$\begin{aligned}
b\int_{R^*_-}^{R^*_+}&\left (|u'|^2+(\omega^2+\La) |u|^2\right)
\\ &\leq \int_{-\infty}^\infty \left (-2f \,{\text{Re}} (u'\overline{H}) - (f'+h)\, {\text{Re}} (u\overline{H})+E\omega {\text{Im}} (H\overline u) + \chi_1\left(\omega-\upomega_+m\right){\text{Im}}(H\overline u)\right ).\end{aligned}$$
For the construction of our currents, we again shall exploit that $\mathcal{G}_{\mbox{${{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}$}}$ defines a large frequency regime, where now, however, $\omega^2\ll \Lambda$. Since this is a non-superradiant regime, the boundary term of $r^* = \infty$ may be controlled with the Q$^T$ current; however, we shall handle the boundary term at the horizon as we did for the regime $\mathcal{G}^{\mbox{$\sharp$}}$. As we explained in Section \[boundaryStillaProblem\] this is necessary because the boundary term at the horizon is proportional to $\left(\omega - \upomega_+m\right)^2\left|u\left(-\infty\right)\right|^2$, and the Q$^T$ current would only give an estimate for $\omega\left(\omega-\upomega_+m\right)\left|u\left(-\infty\right)\right|^2$. In the frequency regime under consideration these are *not* necessarily comparable.
Turning to the proof, we begin by arguing that $\epsilon_{\rm width}$ sufficiently small implies $m\omega \leq 0$. Suppose $m\omega > 0$. Then we have $$\label{mBigOmega}
m\omega \geq \frac{am^2}{2Mr_+} + \alpha\Lambda \geq \alpha\epsilon_{\rm width}^{-1}\omega^2 \Rightarrow \left|m\right| \geq \alpha\epsilon_{\rm width}^{-1}\left|\omega\right|.$$ On the other hand, $$\label{omegaBigm}
m\omega \geq \frac{am^2}{2Mr_+} + \alpha\Lambda \Rightarrow \left|\omega\right| \geq \alpha\Lambda\left|m\right|^{-1} \geq \alpha\left|m\right|.$$ Combining (\[mBigOmega\]) and (\[omegaBigm\]) implies $$\left|\omega\right| \geq \alpha^2\epsilon_{\rm width}^{-1}\left|\omega\right|.$$ This is a contradiction if we take $\epsilon_{\rm width} < \alpha^2$. Thus, we indeed have $m\omega < 0$.
From the above inequality, it follows that Lemma \[lem:3\] applies, and we may thus conclude that the potential $V_0$ is increasing at $r_+$, and hence has only one critical point at $r=r^0_{\rm max}$ where it attains a maximum. As in the proof of Proposition \[odeEst1\] concerning the regime $\mathcal{G}^{\mbox{$\sharp$}}$, we again infer that, for $\omega_{\rm width}$ sufficiently large, in the regime $\mathcal{G}_{\mbox{${{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}$}}\left(\omega_{\rm high},\epsilon_{\rm width}\right)$, the potential $V$ has a unique non-degenerate critical point at $r_{\max}$, where it attains a maximum, and that $r_{\max}$ is uniformly bounded away from $r_+$ and is uniformly bounded from above. Similarly, we also obtain the existence of an interval $(r_{\max}-\de,r_{\max}+\de)$, where $\delta$ is independent of frequency parameters, such that $V$ satisfies the two relations $$V(r)-\omega^2\ge b\Lambda,\qquad \forall r\in (r_{\rm max}-\delta,r_{\rm max}+\delta)$$ and $$(r-r_{\rm max}) \frac d{dr} V(r)\ge b \Lambda \frac {(r-r_{\rm max})^2}{r^4}, \qquad r>r_+.$$
We may now follow the construction given in Proposition \[odeEst1\] for the range $\mathcal{G}^{\mbox{$\sharp$}}$. We define first a current $\text{Q}=\text{Q}^f+\text {\fontencoding{LGR}\selectfont \Coppa}^h$ with the same choice of functions $f$ and $h = A\tilde h$ as for $\mathcal{G}^{\mbox{$\sharp$}}$. This gives the inequality $$\begin{aligned}
\label{someEstimateWithA}
\int_{-\infty}^\infty &\left ((2f'+Ah) |u'|^2+\left (Ah(V-\omega^2)-fV'\right) |u|^2 -\frac 12 (f{'''}+Ah'')|u|^2\right )\\&\nonumber=
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} + \left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty}
\\ \nonumber &\qquad- \int_{-\infty}^\infty \left (2 f \,{\text{Re}} (u'\overline{H}) + (f'+Ah)\, {\text{Re}} (u\overline{H})\right),\end{aligned}$$ where the integrand on the left hand side is positive definite. As in the $\mathcal{G}^{\mbox{$\sharp$}}$ regime, we may gain a large parameter in the region $(r_{\max}-\de,r_{\max}+\de)$ by observing that there exists a small constant $\tilde\delta$ only depending on $a_0$ and $M$ so that, as long as $A \leq \tilde\delta\epsilon_{\rm width}^{-1}\omega_{\rm high}^2$, the left-hand side of (\[someEstimateWithA\]) will give a coercive estimate. We fix such an $A$. Finally, using a $\chi_1\left(\omega-\upomega_+m\right)\text{Q}^K$ current we may handle the boundary term at the horizon, *mutatis mutandis*, as we did for the $\mathcal{G}^{\mbox{$\sharp$}}$ regime. We obtain $$\begin{aligned}
b\int_{-\infty}^\infty &\left ((2f'+Ah) |u'|^2+\left (Ah(V-\omega^2)-fV'\right) |u|^2 -\frac 12 (f{'''}+Ah'')|u|^2\right )\\&\leq \left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty}
- \int_{-\infty}^\infty \left (2 f \,{\text{Re}} (u'\overline{H}) + (f'+Ah)\, {\text{Re}} (u\overline{H}) + \chi_1\left(\omega-\upomega_+m\right)\text{Im}\left(H\overline u\right)\right).\end{aligned}$$ Finally, for any $E \geq 2$, the boundary term at infinity is controlled easily with a Q$^T$ current: $$\begin{aligned}
b\int_{-\infty}^\infty &\left ((2f'+Ah) |u'|^2+\left (Ah(V-\omega^2)-fV'\right) |u|^2 -\frac 12 (f{'''}+Ah'')|u|^2\right )\\&\leq- \int_{-\infty}^\infty \left (2 f \,{\text{Re}} (u'\overline{H}) + (f'+Ah)\, {\text{Re}} (u\overline{H}) + \chi_1\left(\omega-\upomega_+m\right)\text{Im}\left(H\overline u\right) + E\omega\text{Im}\left(H\overline u\right)\right).\end{aligned}$$ Restricting the domain of integration of the left hand side of our estimate then finishes the proof.
The $\mathcal{G}_{\mbox{$\natural$}}$ range {#thisissparta}
-------------------------------------------
This range is manifestly non-superradiant. By the results of Section \[Vsrsec\], it will follow that, after suitable such choices of $\omega_{\rm high}$ and $\epsilon_{\rm width}$, this will be the only range which can contain trapping phenomena; thus, it is only in this range for which we will define a non-zero parameter $r_{\rm trap}$.
After the final choices of parameters $\omega_{\rm high}$ and $\epsilon_{\rm width}$ have been made, then for $(\omega, m, \Lambda)\in\mathcal{G}_{\mbox{$\natural$}}(\omega_{\rm high}, \epsilon_{\rm width})$, we set the functions $h$, $\tilde y$, $\hat y$ and $\chi_1$ appearing in the statement of Theorem \[phaseSpaceILED\] to be identically $0$. The remaining functions $f$ and $\hat{y}$, the parameter $r_{\rm trap}$, and the desired coercivity properties are given by the following:
\[odeEst4\] Let $a_0<M$. Then, for all $\epsilon_{\rm width} > 0$, for all $\omega_{\rm high}$, $R_{\infty}$ and $E$ sufficiently big depending on $\epsilon_{\rm width}$, and for all $0\le a\le a_0$, $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\natural$}}(\omega_{\rm high},
\epsilon_{\rm width})$, there exist functions $f$ and $\hat y$ and a value $r_{\rm trap}$ satisfying the uniform bounds $$r_{\rm trap} =0 \qquad {\rm\ or\ } \qquad 0<b< r_{\rm trap}-r_+<B,$$ $$\left|f\right| + \Delta^{-1}r^2\left|f'\right| + \left|y\right| \leq B\left(\epsilon_{\rm width}\right),$$ $$f = 1,\ \hat y = 0\text{ for }r^* \geq R^*_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]), we have the estimate $$\begin{aligned}
\label{thisisspartaEst}
b\left(\epsilon_{\rm width}\right)\int_{R^*_-}^{R^*_+}&\left (|u'|^2+ \left((\omega^2+\La)\left(1-r^{-1}r_{\rm trap}\right)^2+1\right)|u|^2\right)\\ \nonumber &\le \int_{-\infty}^\infty \left (-2 f \,{\text{Re}} (u'\overline{H}) -f'{\text{Re}} (u\overline H)+E\omega {\text{Im}} (H\overline u)\right) + \int_{-\infty}^{\infty} 2\hat y {\text{Re}} (u'\overline H).\end{aligned}$$
As noted above, this frequency range, where $\omega^2$ is comparable to $\Lambda$, contains the trapping phenomena, but is non-superradiant. For frequencies $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\natural$}}$, Lemma \[lem:1\] implies that the potential $V_0$ may have at most two critical points. Furthermore, either a maximum $r^0_{\rm max}$ exists or $V_0$ is non-increasing on $[r_+,\infty)$; if the maximum exists, then there may also exist a minimum $r^0_{\rm min}$ which will satisfy $r^0_{\rm min} < r^0_{\rm max}$.
In analogy to Lemma \[theOtherLemma\] we first must show that for $\omega_{\rm high}$ sufficiently large, the full potential $V$ enjoys similar properties.
For $\epsilon_{\rm width}$ as above, for all $\omega_{\rm high}$ sufficiently large depending on $\epsilon_{\rm width}$ and for $(\omega,m,\Lambda) \in \mathcal{G}_{\mbox{$\natural$}}(\omega_{\rm high},\epsilon_{\rm width})$, there exists an $r_3 \in (r_+,\infty]$ depending on the frequency triple but bounded away from $r_+$, $$r_3 - r_+ \geq b(\epsilon_{\rm width}),$$ such that for $ r\in [r_+,r_3]$ $$V(r)\le \omega^2-b\left(\epsilon_{\rm width}\right) \La.$$ Furthermore, in the case when $r_3 < \infty$, then in fact $r_3 \leq B\left(\omega_{\rm high},\epsilon_{\rm width}\right)$, $r^0_{max}$ exists and the potential $V$ has a unique non-degenerate maximum $r_{\max} \in [r_3,\infty)$, $\left|r_{\rm max}-r_{\rm max}^0\right|\leq B\left(\epsilon_{\rm width}\right)\Lambda^{-1}$ and $\frac {d^2}{dr^2} V(r_{\max})<-b\left(\epsilon_{\rm width}\right) \La$.
Since $m\omega\not\in(0,\frac {am^2}{2Mr_+}+\alpha \La]$ and $\epsilon_{\rm width}\La \leq \omega^2 \leq \epsilon_{\rm width}^{-1}\La$ , we find $$\omega^2-V(r_+)=\omega^2-V_0(r_+)\ge c \Lambda,$$ where $c = c\left(\epsilon_{\rm width}\right)$ only depends on the value of $\epsilon_{\rm width}$. We define $r_0 \in (r_+,\infty]$ to be the largest value with the property that for all $r\in [r_+,r_0)$ $$V_0(r)\le V_0(r_+)+\frac {c}2\La.$$ If $r_0$ is finite then we must have a maximum $r_{\rm max}^0$. Furthermore, $\frac{dV_0}{dr}(r_0) \geq 0$; hence, Lemma \[lem:1\] implies that if $r_{\rm min}^0$ exists, then $$r_{\rm min}^0<r_0\le r_{\rm max}^0.$$ Moreover, Lemma \[lem:1\] implies that $r_{\rm max}^0$ is bounded from above by a constant only depending on $\epsilon_{\rm width}$. On the other hand, since $\left|\frac {d}{dr} V_0(r)\right|\le B\left(\epsilon_{\rm width}\right) \La r^{-3}$, the value $r_0-r_+$ and thus $r_{max}^0-r_+$ is bounded from below by a constant only depending on $\epsilon_{\rm width}$.
We continue to consider the case where $r_0 < \infty$. Recall from the proof of Lemma \[lem:1\] that either $\frac{d}{dr}\left((r^2+a^2)^3\frac{dV_0}{dr}\right)$ is negative on $[r_+,\infty)$ or there exists a unique value $r_1 \in [r_+,r_{\rm max}^0)$ such that $\frac{d}{dr}\left((r^2+a^2)^3\frac{dV_0}{dr}\right)$ is positive on $[r_+,r_1)$ and negative on $(r_1,\infty)$. Moreover, since for frequency triples in $\mathcal{G}_{\mbox{$\natural$}}$, the parameter $\sigma=am\omega/\Lambda$ is bounded by a constant only depending on $\epsilon_{\rm width}$, the value of $r_1$ is uniformly bounded from above by a constant only depending on $\epsilon_{\rm width}$. We first consider the case where the point $r_1$ exists and further split the analysis into two sub-cases based on the value of $V_0(r_1)$.
If $V_0(r_1)\le V_0(r_+)+\frac {3c}4\La$, then, in view of the fact that $V_0$ has a unique maximum at $r^0_{\rm max}$, we have that $$V_0(r)\le V_0(r_+)+\frac {3c}4\La,\qquad \forall r\in [r_+,\max(r_0,r_1)].$$ Moreover, using that $\left|\frac{dV_0}{dr}\right| \le B\left(\epsilon_{\rm width}\right)\La r^{-3}$ and that $\Lambda^{-1} \frac {d}{dr} \left ((r^2+a^2)^3 \frac d{dr} V_0(r)\right)$ is a quadratic polynomial, with coefficients bounded by $\epsilon_{\rm width}$, vanishing at the unique point $r_1$ on the interval $[r_+,\infty)$, we can find a small constant $\de = \de\left(\epsilon_{\rm width}\right) >0$ only depending on $\epsilon_{\rm width}$ such that $$V_0(r)\le V_0(r_+)+\frac {3c}5\La,\qquad \forall r\in [r_+,\max(r_0,r_1)+\de]$$ and $$\frac {d}{dr} \left ((r^2+a^2)^3 \frac d{dr} V_0(r)\right)<-c_1\left(\epsilon_{\rm width}\right)\La r^2,\qquad\forall r\in [\max(r_0,r_1)+\de,\infty),$$ where the positive constant $c_1$ only depends on $\epsilon_{\rm width}$.
Now we consider the case where $V_0(r_1)\ge V_0(r_+)+\frac {3c}4\La$. Then, once again using the bound $\left|\frac{dV_0}{dr}\right| \leq B\left(\epsilon_{\rm width}\right)\La$, we conclude that $r_1-r_0$ is bounded from below by a small positive constant just depending on $\epsilon_{\rm width}$. Furthermore, since $\frac d{dr} V_0(r_0)\ge 0$, we can find a value $r_0'\in [r_0,r_1]$ such that $$V_0(r)\le V_0(r_+)+\frac {3c}4\La,\qquad \forall r\in [r_+,r_0']$$ and $$\frac d{dr} V_0(r)\ge c_2 \La,\qquad \forall r\in [r_0',r_1],$$ where $c_2 = c_2\left(\epsilon_{\rm width}\right)$ is a positive constant which only depends on $\epsilon_{\rm width}$.
Moreover, after slightly changing $c_2$, the last property can be easily extended to a slightly larger interval $$\frac d{dr} V_0(r)\ge c_2 \La,\qquad \forall r\in [r_0',r_1+\de],$$ so that $\delta$ only depends on the constant $\epsilon_{\rm width}$. $$\frac {d}{dr} \left ((r^2+a^2)^3 \frac d{dr} V_0(r)\right)<-c_3\La r^2,\qquad\forall r\in [r_1+\de,\infty),$$ for a positive constant $c_3 = c_3\left(\epsilon_{\rm width}\right)$ which only depends on $\epsilon_{\rm width}$.
If $r_1$ does not exists, the above arguments *mutatis mutandis* will produce a value $r_0'$ only depending on the value $\epsilon_{\rm width}$ such that $$V_0(r)\le V_0(r_+)+\frac {3c}4\La,\qquad \forall r\in [r_+,r_0'],$$ $$\frac {d}{dr} \left ((r^2+a^2)^3 \frac d{dr} V_0(r)\right)<-c_4\left(\epsilon_{\rm width}\right)\La r^2,\qquad\forall r\in [r_0',\infty),$$ for a positive constant $c_4 = c_4\left(\epsilon_{\rm width}\right)$ only depending on $\epsilon_{\rm width}$.
Finally, in both cases $r_0 < \infty$ and $r_0 = \infty$ we may therefore claim the existence of a value $r_3$ (possibly infinite), bounded away from $r_+$ by a constant only depending on $\epsilon_{\rm width}$, such that $$V_0(r)\le V_0(r_+)+\frac {3c}4\La\le \omega^2-\frac c4 \La,\qquad \forall r\in [r_+,r_3]$$ and, such that for any $r\in [r_3,\infty)$, either $$\frac d{dr} V_0(r)\ge b\left(\epsilon_{\rm width}\right) \La$$ or $$\frac {d}{dr} \left ((r^2+a^2)^3 \frac d{dr} V_0(r)\right)<-b\left(\epsilon_{\rm width}\right)\La r^2.$$ We note that if $r_3$ is finite, then it is bounded from above by a constant only depending on $\epsilon_{\rm width}$.
Now, just as we argued in the frequency range $\mathcal{G}^{\mbox{$\sharp$}}$, adding the bounded potential $V_1$, and requiring that $\omega_{\rm high}$ is sufficiently large finishes the proof.
Before constructing our current, it will be useful to recall that, as observed in Section \[sub:sharp\], the inequalities $\Lambda \geq \left|m\right|\left(\left|m\right| + 1\right)$ and $\Lambda \geq 2a\left|m\omega\right|$ imply that there exists a constant $R_{\rm dec}^* \geq 2R^*_+$ only depending on $a_0$ and $M$ such that $$\label{decrease2}
V' < 0\text{ for }r^* \geq R_{\rm dec}^*.$$
We now construct our current, first under the assumption that $r^*_3 < R^*_{\rm dec}$. Given $E$ sufficiently large depending on $\epsilon_{\rm width}$, we shall use a combination $$\text{Q}=\text{Q}^f-\text {\fontencoding{LGR}\selectfont \koppa}^y-E \text{Q}^T$$ of the currents $\text{Q}^f, \text {\fontencoding{LGR}\selectfont \koppa}^y$ and $Q^T$ where $f$, $y$ are chosen as described below.
The current $\text{Q}^f$ is applied with a function $f$ such that $$\label{choices10}
\left|f\right| + \Delta^{-1}r^2\left|f'\right| \leq B\left(\epsilon_{\rm width}\right),\qquad f(r_+) = 0,\qquad f' > 0\text{ for }r \in [r_3,R_{\infty}],$$ $$\label{choices11}
f\text{ switches from negative to positive at }r = r_{\max},\qquad f = 1\text{ for }r^* \geq R^*_{\rm dec},$$ $$\label{choices12}
-fV' - \frac 12 f'''(r) > b(\epsilon_{\rm width})\La \frac {\Delta (r-r_{\max})^2}{r^7},\qquad \forall r\in [r_3,\infty).$$ In view of the properties of $V$ proven above, such a function can easily be constructed.
The second current will be $\text {\fontencoding{LGR}\selectfont \koppa}^{\hat y}$, with $$\label{choices13}
\hat y = 0\text{ for }r \geq r_3,\qquad \hat y' > 0\text{ for }r \leq r_3,\qquad \left|\hat y\right| + \left|\hat y'\right| \leq B\left(\epsilon_{\rm width}\right).$$ Such a $\hat{y}$ is now trivial to construct.
Finally, we subtract the multiple $E\text{Q}^T$ of the current $\text{Q}^T$. We obtain: $$\begin{aligned}
\int_{-\infty}^{r_3} &\left(\hat y'\left (|u'|^2+(\omega^2-V)|u|^2\right)-\hat yV'|u|^2\right)+
\int_{-\infty}^\infty \left (2f' |u'|^2-(fV'+\frac 12 f''') |u|^2 \right )\\+
&\left(-f|u'|^2+(\frac{1}{2}E-f)\omega^2 |u|^2\right)_{r=\infty}+\left(\frac{1}{2}E\omega(\omega-\upomega_+m)-2\hat y(\omega-\upomega_+m)^2\right) |u|^2|_{r=r_+}
\\ & =- \int_{-\infty}^\infty \left (2 f \,{\text{Re}} (u'\overline{H})+f' {\text{Re}} (u\overline{H})-E\omega {\text{Im}} (H\overline{u}) \right)+ \int_{-\infty}^{r_3}
2\hat y \,{\text{Re}} (u'\overline{H}) .\end{aligned}$$ By the described properties of the potential $V$, the expression $-(fV'+\frac 12 f''')$ is positive on the interval $[r_3,\infty)$. On the interval $(r_+,r_3]$, we need to choose a function $\hat y$ so that in addition to (\[choices13\]) we have $$\label{somanychoices}
\hat y' (\omega^2-V)-\hat y V'-(fV'+\frac 12 f''')\ge 0.$$ Since for these values of $r$ $$(\omega^2-V)\ge b\left(\epsilon_{\rm width}\right) \La,\qquad |V'|\le B\left(\epsilon_{\rm width}\right)\Lambda \frac {\Delta}{r^2}, \qquad |f|+|f'''|\le B\left(\epsilon_{\rm width}\right)\frac {\Delta}{r^2},$$ it suffices to fulfill the inequality $$\label{aSplendidIneq}
\frac {d}{dr} \hat y\ge -\hat y C + C,$$ provided that $C$ is sufficiently large only depending on $\epsilon_{\rm width}$. The function $$\hat y= 1-e^{C(r_3-r)}$$ satisfies all the above criteria. Note that the constant $C$ only depends on $\epsilon_{\rm width}$. Finally, for all $E$ such that $C \ll E$, the non-superradiant condition $m\omega\not\in (0,m\upomega_+]$ and the boundary condition $u'=i\omega u$ at $r=\infty$ ensure that both boundary terms at $r=r_+$ and $r=\infty$ are positive. After restricting the domain of integration of the left hand side of our estimate, we have obtained $(\ref{thisisspartaEst})$, defining $$r_{\rm trap} = r_{\rm max}.$$
In the case $\infty\ge r_3 \geq R^*_{\rm dec}$ we construct our current as follows. As above we will have $$\label{aNiceCurrent}
\text{Q}=\text{Q}^f+\text {\fontencoding{LGR}\selectfont \koppa}^{\hat y}-E \text{Q}^T.$$ We define $$\hat y = 1-e^{\hat C\left(R_{\rm dec} + 2 - r\right)}\text{ for }r \leq R_{\rm dec} + 2,$$ $$\hat y = 0\text{ for }r\geq R_{\rm dec} + 2.$$ Note we shall satisfy (\[aSplendidIneq\]) with $C$ replaced by $\hat C$. Thus, arguing just as in the case when $r_3 < R_{\rm dec}$, for a sufficiently large $\hat C$ we will have $$\int_{-\infty}^{R_{\rm dec} + 1} \left(\hat y'\left (|u'|^2+(\omega^2-V)|u|^2\right)-\hat yV'|u|^2\right) \geq b\left(\epsilon_{\rm width}\right)\int_{R^*_-}^{R_{\rm dec} + 1}\left(\left|u'\right|^2 + \omega^2\left|u\right|^2\right).$$ Next, we let $f$ be any smooth function such that $$\label{thechoicesneverend}
f' \geq 0,\qquad f = 0\text{ for }r\in [r_+,R_{\rm dec}], \qquad f = 1\text{ for }[R_{\rm dec}+1,\infty),\qquad \left|f\right| +|f'| + \left|f'''\right| \leq B.$$ Such an $f$ is trivial to construct.
Requiring $\omega_{\rm high}$ to be sufficiently large depending on $\epsilon_{\rm width}$, we shall have $$\int_{R_{\rm dec}}^{\infty} \left(\hat y'\left (|u'|^2+(\omega^2-V)|u|^2\right)-\hat yV'|u|^2\right)+
\int_{R_{\rm dec}}^\infty \left (2f' |u'|^2-(fV'+\frac 12 f''') |u|^2 \right )$$ $$\geq \int_{R_{\rm dec}^*}^{R_{\rm dec}^*+1}\left(b\omega_{\rm high}^2 - \frac 12f'''\right)\left|u\right|^2 \geq 0.$$ Thus, the bulk term of the estimate corresponding to $\text{Q}$ is positive. Just as in the case $r_3 < R_{\rm dec}$, requiring that $E$ is large enough depending on $\epsilon_{\rm width}$ will guarantee that the boundary terms are controlled. Finally, we require that $R^*_{\infty} \geq R_{\rm dec}^* + 1$. This gives again $(\ref{thisisspartaEst})$ defining $r_{\rm trap}=0$.
The $\mathcal{G}_{\mbox{$\flat$}}$ range {#whit}
----------------------------------------
This range again depends on $\omega_{\rm high}$, and $\epsilon_{\rm width}$. As opposed to the Propositions concerning the other ranges which restrict the choices of one or both these parameters, in the range $\mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm high}, \epsilon_{\rm width})$, estimates can be obtained for $\omega_{\rm high}>0$ and $\epsilon_{\rm width}>0$, but the relevant constants will degenerate as $\omega_{\rm high}\to\infty$, $\epsilon_{\rm width}\to 0$.
We shall split the frequency range $\mathcal{G}_{\mbox{$\flat$}}$ into four subcases, considering each separately. We will see the above degeneration in the last of the cases. We note that our decomposition will not however distinguish between superradiant and non-superradiant frequencies. It should be clear to the reader how the constructions could be simplified if restricted to the non-superradiant case.
The split will rely on the introduction of a further small parameter $\tilde a_0$. This parameter is for now free–we choose it in Section \[putting\].
### The subrange $\left|\omega\right| \leq \omega_{\rm low}$, $0\le a < \tilde a_0$ and $m \neq 0$
Given the final choice of parameters, $\omega_{\rm high}$, $\epsilon_{\rm width}$ and $\omega_{\rm low}$, then for $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm high}, \epsilon_{\rm width})$ such that $|\omega|\le \omega_{\rm low}$ and $a < \tilde a_0$, we will set the functions $f$ and $\tilde y$ together with the parameter $r_{\rm trap}$ to be $0$. The remaining functions $y$, $\hat y$, $h$, $\chi_1$, $\chi_2$ and the desired coercivity properties are given by the following
\[odeEst5\] Let $a_0<M$. Then, for all $\omega_{\rm high}>0$, $\epsilon_{\rm width}>0$, for all $\omega_{\rm low} > 0$, $\tilde{a}_0 > 0$ sufficiently small depending on $\omega_{\rm high}$ and $\epsilon_{\rm width}$, for all $R_{\infty}$ sufficiently large, for all $E \geq 2$, $0\le a\le a_0$, and for all $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm high}, \epsilon_{\rm width})$ such that $\left|\omega\right| \leq \omega_{\rm low}$ and $0\le a < \tilde a_0$, there exist functions $y$, $\hat y$, $\chi_1$, $\chi_2$ and $h$, satisfying the uniform bounds $$\left|y\right| + \left|\hat y\right| + \left|h\right| + \left|\chi_2\right| \leq B,$$ $$\chi_2 = 1,\ \chi_1 = 0,\\ y = 1,\ \hat y = 0,\ h = 0\text{ for }r^* \geq R^*_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]), we have the estimate $$\begin{aligned}
b\int_{R^*_-}^{R^*_+}\left(\left|u'\right|^2 + \left|u\right|^2\right) &\leq
\int_{-\infty}^\infty \left (2(y+\hat y) \,{\text{Re}} (u'\overline{H})+ h\,{\text{Re}}(u\overline H)+E\omega \chi_1{\text{Im}} (H\overline u)+\chi_2\left(\omega-\upomega_+m\right) {\text{Im}} (H\overline u)\right).\end{aligned}$$
The construction of our current is inspired by the treatment of similar frequency regimes in [@aretakisKerr] and [@holz-smul].
The following three properties are easily verified:
1. For every $-\infty < \alpha < \beta < \infty$, if we require $\tilde a$ and $\omega_{\rm low}$ sufficiently small, both depending on $\alpha$ and $\beta$, then $r \in [\alpha,\beta] \Rightarrow V - \omega^2 > 0$.
2. For sufficiently large $r^*$, independent of the frequency parameters, we have $V' < 0$.
3. For sufficiently small $\tilde a_0$ and sufficiently negative $r^*$, independent of the frequency parameters, we have $V' > 0$.
Let’s introduce the set of relevant constants.
1. Requiring that $\tilde a_0$ is sufficiently small, let $R^*_1 < R^*_-$ be a fixed negative constant chosen so that $r^* \leq R^*_1$ implies that $V' > 0$ and $\left(r^*\left(V - V|_{r=r_+}\right)\right)' > 0$.
2. Let $R^*_2 > R_+^*$ be a fixed positive constant chosen so that $r^* \geq R^*_2$ implies $V' < 0$ and $\left(r^*V\right)' < 0$.
3. Let $\epsilon > 0$ be a sufficiently small positive constant to be fixed later.
4. Let $p = p\left(\epsilon\right) > 0$ be a sufficiently small positive constant depending on $\epsilon$.
We now construct our current Q in a step by step fashion. Choose a function $h$ satisfying $$\label{ccchoices}
h=1 {\rm\ for\ } r^* \in [R^*_1,R^*_2], \qquad h=0 {\rm\ for\ } r^* \in (-\infty,e^{p^{-1}}R^*_1],\qquad h \geq 0,$$ $$\label{ccchoices2}
h=0 {\rm\ for\ } r^* \in [e^{p^{-1}}R^*_2,\infty), \quad \left|h''\right| \leq \frac{Bp}{\left|r^*\right|^2} {\rm\ when\ } r^* \in [e^{p^{-1}}R_1^*,R^*_1] \cup [R_2^*,e^{p^{-1}}R_2^*].$$ Note that one may easily construct a function $h$ satisfying (\[ccchoices\]) and (\[ccchoices2\]).
We then apply a $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ current: $$\begin{aligned}
\label{axisym1}
\int_{-\infty}^\infty &\left (h |u'|^2+\left(h(V-\omega^2)-\frac 12 h''\right)|u|^2\right ) = - \int_{-\infty}^\infty h\, {\text{Re}} (u\overline{H}).\end{aligned}$$ The integrand of the left hand side of the estimate (\[axisym1\]) will cease to be non-negative for $r^* \in [e^{p^{-1}}R_1^*,R^*_1] \cup [R_2^*,e^{p^{-1}}R_2^*]$. We will produce a non-negative integrand by adding in $\text {\fontencoding{LGR}\selectfont \koppa}^y$ and $\text {\fontencoding{LGR}\selectfont \koppa}^{\hat y}$ currents.
Define a function $y$ by $$\label{ccchoices3}
y = 0{\rm\ for\ } r^* \in (-\infty,R^*_2-1), \qquad y = \frac{r^*-R_2^*+1}{2}{\rm\ for\ }r^* \in [R^*_2-1,R^*_2),$$ $$\label{ccchoices4}
y = \epsilon\left(\frac{1}{r^*V} - \frac{1}{R_2^*V|_{r^* = R^*_2}}\right) + \frac{1}{2}{\rm\ for\ }r^* \in [R^*_2,e^{p^{-1}}R^*_2],$$ $$\label{cccoices5}
y = \epsilon\left(\frac{1}{e^{p^{-1}}R_2^*V|_{r^* = e^{p^{-1}}R_2^*}} - \frac{1}{R_2^*V|_{r^* = R_2^*}}\right) + \frac{1}{2}{\rm\ for\ }r^* \in [e^{p^{-1}}R^*_2,\infty).$$ Note that we have chosen $R^*_2$ so that we will have $y' \geq 0$. Of course, we also have $y \geq 0$.
Now we add in a $\text {\fontencoding{LGR}\selectfont \koppa}^y$ current to (\[axisym1\]) and obtain $$\begin{aligned}
\label{axisym2}
\int_{-\infty}^\infty &\left (\left(h + y'\right)|u'|^2+\left(y'\omega^2 + h(V-\omega^2) - (yV)' - \frac{1}{2}h''\right)|u|^2\right )=
\\ \nonumber &y(\infty)\left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty}- \int_{-\infty}^\infty \left (2y \,{\text{Re}} (u'\overline{H}) + h\, {\text{Re}} (u\overline{H})\right).\end{aligned}$$
We will now show that if we require $\omega_{\rm low}$ and $\tilde a_0$ to be sufficiently small depending on appropriate choices of $\epsilon$ and $p$, the integrand of the left hand side of (\[axisym2\]) is non-negative in the region $r^* \in [R^*_+,\infty)$. Since $h$, $y' \geq 0$ it suffices to show that the term $h(V-\omega^2) - (yV)' - \frac{1}{2}h''$ is non-negative.
The function $y$ vanishes and $h = 1$ in the region $r^* \in [R^*_+,R^*_2-1)$, and thus we have $$r^* \in [R^*_+,R^*_2-1) \Rightarrow h(V-\omega^2) - (yV)' - \frac{1}{2}h'' = V - \omega^2.$$ If $\omega_{\rm low}$ and $\tilde a_0$ are sufficiently small, then $V - \omega^2$ will be positive in this region.
Next, we have $$r^* \in [R^*_2-1,R^*_2) \Rightarrow h(V-\omega^2) - (yV)' - \frac{1}{2}h'' = V - \omega^2 - \frac{1}{2}V - yV'.$$ Recall that we chose $R^*_2$ so that $V' < 0$ in this region. Since $y > 0$, we then get $$r^* \in [R^*_2-1,R^*_2) \Rightarrow V - \omega^2 - \frac{1}{2}V - yV' \geq \frac{1}{2}V - \omega^2.$$ Now, it is clear that if $\omega_{\rm low}$ and $\tilde a_0$ are sufficiently small, then $\frac{1}{2}V - \omega^2$ will be positive in this region.
Next we consider the region $r^* \in [R^*_2,e^{p^{-1}}R^*_2)$. As usual, we start by noting that if we require $\omega_{\rm low}$ and $\tilde a_0$ to be sufficiently small depending on $p$, then $V - \omega^2 > 0$ in the region $r^* \in [R^*_2,e^{p^{-1}}R^*_2)$. Hence, we will have $$\label{someInequality}
r^* \in [R^*_2,e^{p^{-1}}R^*_2) \Rightarrow h(V - \omega^2) - (yV)' - \frac{1}{2}h'' \geq \frac{\epsilon - Bp}{(r^*)^2} - \left(\frac{1}{2} - \frac{\epsilon}{R^*_2V|_{r^* = R^*_2}}\right)V'.$$ Again, we recall that $V' < 0$ for $r^* > R^*_5$. Furthermore, as long as we require $\epsilon$ to be sufficiently small, we will have $\frac{1}{2} - \frac{\epsilon}{R^*_2V|_{r^* = R^*_2}} > 0$. Finally, we may choose $p$ small enough depending on $\epsilon$ so that the first term on the right hand side of (\[someInequality\]) is also positive.
In the region $r^* \in [e^{p^{-1}}R^*_2,\infty)$ we have that $h = 0$ and $y$ is constant. Since $V' < 0$ in this region, we have $$h(V - \omega^2) - (yV)' - \frac{1}{2}h'' = -y(\infty)V' > 0.$$
Thus as long as $\epsilon > 0$ is sufficiently small, $p$ is sufficiently small depending on $\epsilon$ and $\omega_{\rm low}$ and $\tilde a_0$ are sufficiently small depending on $p$, the integrand of the left hand side of (\[axisym2\]) is non-negative for $r^* \geq R^*_2$; however, it is still not non-negative for $r^* < R^*_1$. To remedy this we will employ a $\text {\fontencoding{LGR}\selectfont \koppa}^{\hat y}$ current with a function $\hat y$ whose properties as $r^* \to -\infty$ will mimic the properties of $y$ as $r^* \to \infty$. The key point which allows us to carry out an analogous construction is that $V' > 0$ for $r^*$ sufficiently close to $-\infty$.
We define $$\label{ccchoices6}
\hat y = 0{\rm\ for\ } r^* \in (R^*_1+1,\infty), \qquad \hat y = \frac{r^*-R_1^*-1}{2}{\rm\ for\ }r^* \in [R^*_1,R^*_1+1),$$ $$\label{ccchoices7}
\hat y = \epsilon\left(\frac{1}{r^*\tilde V} - \frac{1}{R_1^*\tilde V|_{r^* = R^*_1}}\right) - \frac{1}{2}{\rm\ for\ }r^* \in [e^{p^{-1}}R^*_1,R^*_1),$$ $$\label{ccchoices8}
\hat y = \epsilon\left(\frac{1}{e^{p^{-1}}R_1^*\tilde V|_{r^* = e^{p^{-1}}R_1^*}} - \frac{1}{R_1^*\tilde V|_{r^* = R_1^*}}\right) - \frac{1}{2}{\rm\ for\ }r^* \in (-\infty,e^{p^{-1}}R^*_1].$$ Here $\tilde V \doteq V - V|_{r=r_+}$. Note that we have chosen $R^*_1$ so that $\hat y' \geq 0$. Of course, we also have $\hat y \leq 0$.
Now we add a $\text {\fontencoding{LGR}\selectfont \koppa}^{\hat y}$ current to (\[axisym2\]). We obtain $$\begin{aligned}
\label{axisym3}
\int_{-\infty}^\infty &\left (\left(h + y' + \hat y '\right)|u'|^2+\left(y'\omega^2 + \hat y'\left(\omega-\upomega_+m\right)^2 + h(V-\omega^2) - (yV)' - (\hat y \tilde V)'- \frac{1}{2}h''\right)|u|^2\right )=
\\ \nonumber &y(\infty)\left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty} + \left|\hat y(-\infty)\right|\left(|u'|^2 + (\omega-\upomega_+m)^2|u|^2\right)_{r = r_+} - \int_{-\infty}^\infty \left ((y+\hat y) \,{\text{Re}} (u'\overline{H}) + h\, {\text{Re}} (u\overline{H})\right).\end{aligned}$$ Now, keeping in mind that $V' > 0$ for sufficiently negative $r^*$ and repeating the argument, *mutatis mutandis*, which showed that $r^* \geq R^*_2 \Rightarrow h(V-\omega^2) - (yV)' - \frac{1}{2}h'' \geq 0$ we obtain that $$r^* \leq R^*_1 \Rightarrow h(V-\omega^2) - (\hat y\tilde V)' - \frac{1}{2}h'' \geq 0.$$ We conclude that the integrand of the left hand side of (\[axisym3\]) is non-negative and is greater than $$b\int_{R^*_-}^{R^*_+}\left(|u'|^2 + |u|^2\right).$$ We may now fix the constants $\epsilon$ and $p$.
It remains to absorb the boundary terms on the right hand side of (\[axisym3\]). We start with the term at $r = \infty$. Let $\chi_2$ be a function which is identically $1$ for $r^* \geq R_+^*$ and identically $0$ for $r^* \leq R^*_-$. Requiring that $E \geq 2$, we obtain $$\begin{aligned}
\label{getThatBoundaryTerm00}
y(\infty)\left(|u'|^2 + \omega^2|u|^2\right)_{r = \infty} &\leq Ey(\infty)\int_{-\infty}^{\infty}\left(\chi_2\text{Q}^T\right)'
\\ \nonumber &\leq B(\omega_{\rm high},\epsilon_{\rm width})\omega_{\rm low}\int_{R^*_-}^{R^*_+}\left(|u'|^2 + |u|^2\right) + Ey(\infty)\omega\int_{-\infty}^{\infty}\chi_2\text{Im}(H\overline u).\end{aligned}$$
Taking $\omega_{\rm low}$ sufficiently small, we may add this in to our previous estimate and obtain $$\begin{aligned}
\label{axisym4}
b\int_{R^*_-}^{R^*_+}&\left(|u'|^2 + |u|^2\right) \leq
\\ \nonumber &B\left(|u'|^2 + (\omega-\upomega_+m)^2|u|^2\right)_{r = r_+} - \int_{-\infty}^\infty \left ((y+\hat y) \,{\text{Re}} (u'\overline{H}) + h\, {\text{Re}} (u\overline{H})+ Ey(\infty)\omega\chi_2\text{Im}(H\overline{u})\right).\end{aligned}$$
Let $\chi_1$ be a function which is identically $1$ for $r^* \in (-\infty,R^*_1)$ and identically $0$ for $r^* \geq R^*_2$. We obtain $$\begin{aligned}
\label{getThatBoundaryTerm}
B\left(|u'|^2 + (\omega-\upomega_+m)^2|u|^2\right)_{r = r_+} &= B\int_{-\infty}^{\infty}\left(\chi_1\text{Q}^K\right)'
\\ \nonumber &\leq B(\omega_{\rm high},\epsilon_{\rm width})\left(\omega_{\rm low} + \tilde a_0\right)\int_{R^*_-}^{R^*_+}\left(|u'|^2 + |u|^2\right) + B\left(\omega-\upomega_+m\right)\int_{-\infty}^{\infty}\chi_1\text{Im}(H\overline u).\end{aligned}$$ Thus, it is clear that if require that $\omega_{\rm low}$ and $\tilde a_0$ are sufficiently small, depending on $\omega_{\rm high}$ and $\epsilon_{\rm width}$, we may multiply $\chi_1$ by a bounded constant, add in $\chi_1\text{Q}^K$ to our current and obtain $$\begin{aligned}
\label{axisym40}
b\int_{R^*_-}^{R^*_+}&\left(|u'|^2 + |u|^2\right) \leq - \int_{-\infty}^\infty \left ((y+\hat y) \,{\text{Re}} (u'\overline{H}) + h\, {\text{Re}} (u\overline{H})+ Ey(\infty)\omega\chi_2\text{Im}(H\overline{u}) + (\omega-\upomega_+m)\chi_1\text{Im}(H\overline u)\right).\end{aligned}$$
Finally, we may rescale all of the multipliers so that $y(\infty) = 1$. We obtain $$\begin{aligned}
\label{axisym5}
b\int_{R^*_-}^{R^*_+} &\left (|u'|^2+|u|^2\right) \leq - \int_{-\infty}^\infty \left ((y+\hat y) \,{\text{Re}} (u'\overline{H}) + h\, {\text{Re}} (u\overline{H})+ E\chi_2\text{Im}(H\overline{u})+ (\omega-\upomega_+m)\chi_1\text{Im}(H\overline u)\right).\end{aligned}$$
Of course, $R^*_{\infty}$ is simply required to be larger than $e^{p^{-1}}R^*_2$.
The above proof does not use the assumption $m \neq 0$. We only include $m \neq 0$ in the definition of the frequency range so that the set of frequencies covered by Proposition \[odeEst5\] is disjoint from the set of frequencies covered by Proposition \[odeEst5b\].
### The subrange $\left|\omega\right| \leq \omega_{\rm low}$ and $m =0$
Given the final choice of parameters, $\omega_{\rm high}$, $\epsilon_{\rm width}$ and $\omega_{\rm low}$, then for $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm hig}, \epsilon_{\rm width})$ such that $|\omega|\le \omega_{\rm low}$ and $m=0$, we will set the functions $f$, $\tilde y$ and $\chi_1$ together with the parameter $r_{\rm trap}$ to be $0$. The remaining functions $y$, $\hat y$, $h$ and the desired coercivity properties are given by the following
\[odeEst5b\] Let $a_0<M$. Then, for all $\omega_{\rm high}>0$, $\epsilon_{\rm width}>0$, for all $\omega_{\rm low} > 0$ sufficiently small depending on $\omega_{\rm high}$ and $\epsilon_{\rm width}$, for all $R_{\infty}$ sufficiently large, for all $E \geq 2$, $0\le a\le a_0$, and for all $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm high}, \epsilon_{\rm width})$ such that $\left|\omega\right| \leq \omega_{\rm low}$ and $m =0$, there exists functions $y$, $\hat y$ and $h$, satisfying the uniform bounds $$\left|y\right| + \left|\hat y\right| + \left|h\right| \leq B,$$ $$y = 1,\ h = 0,\ \hat y= 0\text{ for }r^* \geq R^*_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]), we have the estimate $$\begin{aligned}
b\int_{R^*_-}^{R^*_+}\left(\left|u'\right|^2 + \left|u\right|^2\right) &\leq-\int_{-\infty}^\infty \left (2(y+\hat y) \,{\text{Re}} (u'\overline{H})+ h\,{\text{Re}}(u\overline H)+E\omega {\text{Im}} (H\overline u)\right).\end{aligned}$$
Observe that the properties of the potential $V$ used in the proof of Proposition \[odeEst5\] also hold here:
1. For every $-\infty < \alpha < \beta < \infty$, if we require $\omega_{\rm low}$ sufficiently small, both depending on $\alpha$ and $\beta$, then $r \in [\alpha,\beta] \Rightarrow V - \omega^2 > 0$.
2. For sufficiently large $r^*$, independent of the frequency parameters, we have $V' < 0$.
3. For sufficiently negative $r^*$, independent of the frequency parameters, we have $V' > 0$.
Using these observations, one may repeat, *mutatis mutandis*, the current construction from the proof of Proposition \[odeEst5\]. In fact, the situation is strictly better here; since this proposition concerns a non-superradiant regime, we may set $\chi_1 = 0$. One obtains $$\begin{aligned}
\label{axisym6}
b\int_{R^*_-}^{R^*_+} &\left (|u'|^2+|u|^2\right) \leq - \int_{-\infty}^\infty \left (2(y+\hat y) \,{\text{Re}} (u'\overline{H}) + h\, {\text{Re}} (u\overline{H})+ E\text{Im}(H\overline{u})\right).\end{aligned}$$
### The subrange $\left|\omega\right| \leq \omega_{\rm low}$, $m \neq 0$ and $a \geq \tilde a_0$ (the near stationary subcase) {#nearstat}
Although these frequencies are near-stationary, we will here be able to effectively exploit the non-vanishing of $a$ and the bound $\left|m\right|\ge 1$.
For $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\flat$}}$, $m\ne0$, and $a\ge \tilde a_0$, we set the functions $f$ and $\hat y$ together with the parameter $r_{\rm trap}$ to $0$. The remaining functions $\tilde{y}$, $y$, $h$, $\chi_1$ and $\chi_2$ and the desired coercivity properties are given by the following:
\[odeEst6\]Let $a_0<M$. Then, for all $\omega_{\rm high}>0$, $\epsilon_{\rm width}>0$, $E\geq 2$, for all $\omega_{\rm low} > 0$ sufficiently small depending on $\tilde a_0$ and $E$, for all $R_{\infty}$ sufficiently large depending on $\tilde a_0$, $0\le a\le a_0$, and for all $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm high}, \epsilon_{\rm width})$ such that $\left|\omega\right| \leq \omega_{\rm low}$, $m\neq 0$ and $a \geq \tilde a_0$, there exist functions $\tilde y$, $y$, $h$, $\chi_1$ and $\chi_2$, satisfying the uniform bounds $$\left|\tilde y\right| + \left|y\right| + \left|h\right| + \left|\chi_1\right| + \left|\chi_2\right| \leq B\left(\tilde a_0\right),$$ $$\left|\tilde y\right| \leq B\exp\left(-br\right),\ y = 1,\ h = 0\text{ for }r^* \geq R^*_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]), we have the estimate $$\begin{aligned}
\label{someEstimate}
b\left(\tilde a_0\right)&\int_{R_-^*}^{R_+^*} \left (|u'|^2+|u|^2 \right ) \\ &\nonumber\le
\int_{-\infty}^\infty \left (-2\tilde y \,{\text{Re}} (u'\overline{H}) - E\chi_2\omega\,{\text{Im}}(H \overline u) - 2\chi_1\left(\omega-\upomega_+m\right)\,{\text{Im}}( H \overline u)- h\,{\text{Re}}(u\overline H)- 2y\,{\text{Re}}(u'\overline H)\right).\end{aligned}$$
Our current will be of the form $$\text{Q} = \text {\fontencoding{LGR}\selectfont \koppa}^{\tilde y} + \text {\fontencoding{LGR}\selectfont \Coppa}^{h} + \text {\fontencoding{LGR}\selectfont \koppa}^{y} - \chi_1\text{Q}^K - E\chi_2\text{Q}^T,$$ for suitable functions $\tilde{y}$, $h$, $y$, $\chi_1$ and $\chi_2$.
As we did for the frequency range $\mathcal{G}^{\mbox{$\sharp$}}$, for the purposes of exposition we shall construct the current in a step by step fashion. The first important observation is that the assumptions $m \neq 0$ and $a \geq \tilde a_0$ imply that $\omega_0^2 := \left(\omega - \upomega_+m\right)^2 \geq b\left(\tilde a_0\right)$ as long as $\omega_{\rm low} \ll \tilde a_0$. The second important observation is that $V = \frac{\Lambda}{r^2} + O(r^{-3})\text{ as }r\to\infty$ and, since $m \neq 0$, $\Lambda \geq 2$. This implies that for any $1 \ll \alpha \ll \beta < \infty$, then requiring that $\omega_{\rm low}$ is small enough, depending on $\alpha$ and $\beta$, we have $$\label{potPos}
r \in [\alpha,\beta]\Rightarrow V - \omega^2 \geq \frac{b}{r^2}$$ in this frequency range. We shall exploit this positivity via the use of a $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ current.
Let us now introduce the set of relevant constants. Let $p > 0$, $R_1 < R_2 < R_3 < e^{p^{-1}}R_3$, $C > 0$ and $c > 0$ be constants such that
1. $C = C(\tilde a_0)$ is sufficiently large.
2. $c$ is sufficiently small.
3. $R_1$ is sufficiently large.
4. $c\omega_0^{-2}R_1\exp\left(BCR_1\right) \ll R_2$.
5. $R_2 \ll R_3$.
6. $p \ll R_3^{-3}$.
7. $\omega_{\rm low}^2 \ll cE^{-1}\omega_0^{2}\exp\left(-p^{-1}\right)\exp\left(-BCR_2\right)R_3^{-4}$ and, requiring $\omega_{\rm low}$ sufficiently small, $r \in [R_1,e^{p^{-1}}R_3]\Rightarrow V - \omega^2 \geq br^{-2}$.
We write $$\omega^2 - V =: \omega_0^2 - \tilde{V}$$ where $\tilde{V}(r_+) = 0$. Let $\upsilon(r)$ be a positive function such that $$\label{evenmorechoices}
\upsilon = \Delta\text{ near }r_+,\qquad \upsilon = 1\text{ when }r^* \geq R^*_{\infty},\qquad \left|\upsilon\right| \leq B.$$ Then we define $$\tilde y(r^*) := -\exp\left(-C\int_{-\infty}^{r^*}\upsilon dr^*\right),$$ and consider the current $\text {\fontencoding{LGR}\selectfont \koppa}^{\tilde y}$. Note that $\tilde y\left(-\infty\right) = -1$ and $\tilde y\left(\infty\right) = 0$. We obtain $$\begin{aligned}
\label{smallOmegaFirst}
\int_{-\infty}^\infty &\left (\tilde y' |u'|^2+\left (\tilde y'\omega_0^2 - \left(\tilde y\tilde{V}\right)'\right) |u|^2 \right )=
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} - \int_{-\infty}^\infty \left (2\tilde y \,{\text{Re}} (u'\overline{H})\right).\end{aligned}$$
Note that unlike every other microlocal current we have considered, this $\tilde y$ cannot be taken independent of the frequency parameters when $r^* \geq R^*_{\infty}$ (since this is the only regime which employs a $\text {\fontencoding{LGR}\selectfont \koppa}^{\tilde y}$ current where the seed function $\tilde y$ is negative for large $r$). Nevertheless, the exponential decay of $\tilde y$ as $r \to \infty$ will allow us to handle this when we re-sum (see Section \[summation\]).
We now turn to the $\left(\tilde y\tilde{V}\right)'|u|^2$ term on the left hand side of (\[smallOmegaFirst\]) which threatens to destroy our estimate: $$\begin{aligned}
\left|\int_{-\infty}^{\infty}\left(\tilde y\tilde{V}\right)'\left|u\right|^2\right| = 2\left|\int_{-\infty}^{\infty}2\tilde y\tilde{V}\text{Re}\left(u'\overline u\right)\right| \leq \epsilon \int_{-\infty}^{\infty}\tilde y'\left|u'\right|^2 + B\epsilon^{-1}\int_{-\infty}^{\infty}\tilde y'\omega_0^2\frac{\tilde y^2\tilde{V}^2}{\left(\tilde y'\right)^2\omega_0^2}\left|u\right|^2 =\end{aligned}$$ $$\begin{aligned}
\epsilon \int_{-\infty}^{\infty}\tilde y'\left|u'\right|^2 + B\epsilon^{-1}\int_{-\infty}^{\infty}\tilde y'\omega_0^2\frac{\tilde{V}^2}{C^2\upsilon^2\omega_0^2}\left|u\right|^2 \leq \epsilon \int_{-\infty}^{\infty}\tilde y'\left|u'\right|^2 + B\epsilon^{-1}C^{-2}\omega_0^{-2}\int_{-\infty}^{\infty}\tilde y'\omega_0^2\left|u\right|^2.\end{aligned}$$ Hence, taking $\epsilon$ sufficiently small and then $C = C(\tilde a_0)$ sufficiently large gives us the estimate $$\begin{aligned}
\label{smallOmegaSecond}
b\int_{-\infty}^\infty &\left(\tilde y' |u'|^2+\tilde y'\omega_0^2 |u|^2 \right)\leq
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} - \int_{-\infty}^\infty 2\tilde y \,{\text{Re}} (u'\overline{H}).\end{aligned}$$
As in the frequency regime $\mathcal{G}^{\mbox{$\sharp$}}$, we need to find a large parameter in order to handle the boundary term $\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+}$. We employ a $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ current where $$\label{evenmorechoices2}
h = 0\text{ for }r \in [r_+,R_1],\qquad h'' = c\tilde y'\omega_0^2\text{ for }r \in [R_1,R_2],\qquad h'' = 0\text{ for }r \in [R_2,R_3],$$ $$\label{evenmorechoices3}
\left|h'\right| \leq \frac{BR_3p}{r}\text{ for }r \in [R_3,e^{p^{-1}}R_3],\qquad \left|h''\right| \leq \frac{BR_3p}{r^2}\text{ for }r\in[R_3,e^{p^{-1}}R_3],$$ $$\label{evenmorechoices4}
h = 0\text{ for }r \in [e^{p^{-1}}R_3,\infty).$$ Note that one may easily construct an $h$ satisfying (\[evenmorechoices2\]), (\[evenmorechoices3\]) and (\[evenmorechoices4\]).
In order to help orient the reader for the estimates below, let us briefly describe the rationale behind the construction of $h$. First of all, the $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ current gives a good estimate when $h$ is positive, $h\left(V-\omega^2\right)$ is positive and if the error terms from the $-\frac{1}{2}h''$ term can be controlled. Since $V-\omega^2$ is only positive for large enough $r$, we set $h$ to be $0$ for $r \leq R_1$. In order for $h$ to become non-zero, it is necessary for $-\frac{1}{2}h'' < 0$. Thus, the definition of $h$ on $[R_1,R_2]$ is motivated by the desire to increase $h$ as fast as possible while still being able to absorb the error term $-\frac{1}{2}h''$ with the estimate (\[smallOmegaSecond\]). This successfully produces a positive $h$, but we still need to find a large parameter. In the region $[R_2,R_3]$ we achieve this by setting $h'' = 0$, and then letting $h$ grow linearly. Since we have taken $\omega^2$ small enough so that $V-\omega^2$ is positive on $[R_2,R_3]$, by taking $R_3$ very large we can arrange for $h$ to be as large as we wish. The crucial estimates for absorbtion of the boundary term $\left|u(-\infty)\right|^2$ are $$r \in [R_2,R_3] \Rightarrow bc\omega_0^2\exp\left(-BCR_1\right)\left(r - R_1\right) \leq h \leq BR_3,$$ $$r \in [R_2,R_3] \Rightarrow h^{-1} \leq B\left(V - \omega^2\right)h,$$ see the estimates (\[smallOmegaHorizon\]), (\[smallOmegaHorizonError\]) and (\[smallOmegaSixth\]). Now that we have succeeded in finding a large parameter to absorb the boundary term, we need to take $h$ back down to $0$. Keeping in mind that $pR_3 \ll 1$, the choice of $h$ on $[R_3,e^{p^{-1}}R_3]$ is motivated by the desire to take $h$ down to $0$ in a such a way that the error term $-\frac{1}{2}h''$ is as small as possible. See estimates (\[smallOmegaIntermediate\]), (\[yProperty\]) and (\[smallOmegaFifth\]) for the details of how these error terms are dealt with.
We now turn to the specifics. Applying the current $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ gives $$\begin{aligned}
\label{smallOmegaThird}
\int_{-\infty}^\infty &\left (\left(b\tilde y'+h\right) |u'|^2+\left(b\tilde y'\omega_0^2 + h\left(V-\omega^2\right) - \frac{1}{2}h''\right)|u|^2 \right ) \\ &\nonumber \leq
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} - \int_{-\infty}^\infty \left (2\tilde y \,{\text{Re}} (u'\overline{H}) + h\,{\text{Re}}(u\overline H)\right).\end{aligned}$$ Recall that we explicitly required that $\omega_{\rm low}$ be sufficiently small so that in particular $V-\omega^2$ is positive on $[R_1,e^{p^{-1}}R_3]$. Given this, the only negative terms on the left hand side of this estimate come from the $-\frac{1}{2}h''$ term on the intervals $[R_1,R_2]$ and $[R_3,e^{p^{-1}}R_3]$. By construction of $h$ and the assumption that $c\ll 1$, the negative terms on $[R_1,R_2]$ can be controlled by the $by'\omega_0^2\left|u\right|^2$ term. Therefore, we have $$\begin{aligned}
\label{smallOmegaFourth}
b\int_{-\infty}^\infty &\left (\left(\tilde y'+h\right) |u'|^2+\left(\tilde y'\omega_0^2 + h\left(V-\omega^2\right)\right)|u|^2 \right ) \\ &\nonumber\le
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} + BpR_3\int_{R_3}^{e^{p^{-1}}R_3}\left|u\right|^2r^{-2} -\int_{-\infty}^\infty \left (2\tilde y \,{\text{Re}} (u'\overline{H}) + h\,{\text{Re}}(u\overline H)\right).\end{aligned}$$
The left hand side now is sufficiently strong to absorb the boundary term on the right hand side in a similar fashion as in the $\mathcal{G}^{\mbox{$\sharp$}}$ regime, i.e. by an application of the current $\chi Q^K$ for a suitable cutoff $\chi$. However, we still need to address the term $BpR_3\int_{R_3}^{e^{p^{-1}}R_3}\left|u\right|^2r^{-2}$. For this we use a $\text {\fontencoding{LGR}\selectfont \koppa}^y$ current with a function $y$ which is determined by $$\label{evenmorechoices5}
y = 0\text{ for }r \in [r_+,R_2],\qquad y' = \frac{h}{2}\text{ for }r \in [R_2,R_3],$$ $$\label{evenmorechoices6}
y = \frac{r-R_3}{R_3^2} + \frac{1}{2}\int_{R_2}^{R_3}h\text{ for }r \in [R_3,e^{p^{-1}}R_3],$$ $$\label{evenmorechoices7}
y = \frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \frac{1}{2}\int_{R_2}^{R_3}h\text{ for }r \in [e^{p^{-1}}R_3,\infty).$$
We obtain $$\begin{aligned}
\label{smallOmegaIntermediate}
\int_{R_2}^{\infty}&\left(y'\left|u'\right|^2 + \left(y'\omega^2 - \left(yV\right)'\right)\left|u\right|^2\right) \\ \nonumber &=
\left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \frac{1}{2}\int_{R_2}^{R_3}h\right)\left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty} - \int_{R_2}^{\infty}2y\,{\text{Re}}(u'\overline H).\end{aligned}$$ Observe that $r \in [R_3,e^{p^{-1}}R_3]$ implies $$\begin{aligned}
\label{yProperty}
-\frac{d}{dr}\left(yV\right) &= -\left(R_3^{-2}\left(\frac{\Lambda}{r^2} + O\left(r^{-3}\right)\right) + \left(\frac{r-R_3}{R_3^2}+\frac{1}{2}\int_{R_2}^{R_3}h\right)\left(\frac{-2\Lambda}{r^3} + O\left(r^{-4}\right)\right)\right)\\
\nonumber & \geq R_3^{-2}\frac{\Lambda}{r^2} + R_3^{-1}O\left(r^{-3}\right) + \left(\frac{1}{2}\int_{R_2}^{R_3}h\right)\frac{2\Lambda}{r^3} \geq bR_3^{-2}r^{-2}.\end{aligned}$$ We have used that $$\int_{R_2}^{R_3}h \geq bc\omega_0^2\exp\left(-BCR_1\right)\left(R_3^2 - R_2^2 - R_1\right),$$ and that $R_3$ has been chosen to dominate $R_2 + c\omega_0^{-2}\exp\left(BCR_1\right)$. Of course, $\frac{d}{dr^*}\left(yV\right) = \left(1 + O(r^{-1})\right)\frac{d}{dr}\left(yV\right)$. We conclude that $$-\left(yV\right)' \geq bR_3^{-2}r^{-2}.$$
Next, keeping in mind that $p \ll R_3^{-3}$ and that $R_1$ sufficiently large implies that $V' < 0$ for $r \in [R_1,\infty)$, we may add (\[smallOmegaIntermediate\]) to (\[smallOmegaFourth\]) to obtain $$\begin{aligned}
\label{smallOmegaFifth}
b\int_{-\infty}^\infty &\left (\left(\tilde y'+h\right) |u'|^2+\left(b\tilde y'\omega_0^2 + h\left(V-\omega^2\right)\right)|u|^2 \right )\\ &\nonumber\leq
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} + \left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \int_{R_2}^{R_3}h\right)\left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty}
\\ &\nonumber\qquad - \int_{-\infty}^\infty \left (2\tilde y \,{\text{Re}} (u'\overline{H}) + h\,{\text{Re}}(u\overline H)+ 2y\,{\text{Re}}(u'\overline H)\right).\end{aligned}$$
Lastly, it remains to absorb the boundary terms on the right hand side. We start with the horizon term. Let $\chi_1$ be a smooth function such that $$\label{evenmorechoices8}
\chi_1 = 1\text{ for }r \in [r_+,R_2],\qquad \chi_1 = 0\text{ for }r \in [R_3,\infty),$$ $$\label{evenmorechoices9}
\left|\chi_1'\right| \leq B\left(R_3-R_2\right)^{-1},\qquad \left|\chi_1\right| \leq B.$$ We have $$\begin{aligned}
\label{smallOmegaHorizon}
\left (|u'|^2+(\omega-\upomega_+m)^2 |u|^2\right)_{r=r_+} &= 2\int_{-\infty}^{\infty}\left(\chi_1\text{Q}^K\right)' \\ \nonumber &= \left(\omega-\upomega_+m\right)\int_{R_2}^{R_3}\chi_1'\text{Im}\left(u'\overline u\right) + \left(\omega-\upomega_+m\right)\int_{-\infty}^{\infty}\chi_1\text{Im}\left(H\overline u\right).\end{aligned}$$
Now, recall that $r \in [R_2,R_3]$ implies that $h \geq bc\omega_0^2\exp\left(-BCR_1\right)\left(r-R_1\right)$ which in turn implies $h\left(V-\omega^2\right) \geq bc\omega_0^2\exp\left(-BCR_1\right)r^{-1} - BR_1R_2^{-1}r^{-1} - \omega_{\rm low}^2R_3 \geq bc\omega_0^2\exp\left(-BCR_1\right)r^{-1}$. Thus, keeping in mind that $R_1R_2^{-1} \ll h\left(V-\omega^2\right)$ on $[R_2,R_3]$, we conclude that $$r \in [R_2,R_3]\Rightarrow h^{-1} \leq Bh\left(V-\omega^2\right).$$ Thus, $$\begin{aligned}
\label{smallOmegaHorizonError}
&\left|\int_{R_2}^{R_3}\chi_1'\text{Im}\left(u'\overline u\right)\right| \leq B\left(R_3-R_2\right)^{-1}\int_{R_2}^{R_3}\left|u'\right|\left|u\right| \\ \nonumber &\qquad\leq B\left(R_3-R_2\right)^{-1}\int_{R_2}^{R_3}\left(h\left|u'\right|^2 + h^{-1}\left|u\right|^2\right) \leq B\left(R_3-R_2\right)^{-1}\int_{R_2}^{R_3}\left(h\left|u'\right|^2 + h\left(V-\omega^2\right)\left|u\right|^2\right).\end{aligned}$$
Hence, we may combine (\[smallOmegaHorizon\]), (\[smallOmegaFifth\]) and (\[smallOmegaHorizonError\]) to obtain $$\begin{aligned}
\label{smallOmegaSixth}
b\int_{-\infty}^\infty &\left (\left(\tilde y'+h\right) |u'|^2+\left(\tilde y'\omega_0^2 + h\left(V-\omega^2\right)\right)|u|^2 \right) \\ &\nonumber
\le \left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \int_{R_2}^{R_3}h\right)\left (|u'|^2+\omega^2 |u|^2\right)_{r=\infty}
\\ &\qquad -\nonumber \int_{-\infty}^\infty \left (2\tilde y \,{\text{Re}} (u'\overline{H}) + 2\chi_1\left(\omega-\upomega_+m\right)\,{\text{Im}}(H\overline u)+ h\,{\text{Re}}(u\overline H)+ 2y\,{\text{Re}}(u'\overline H)\right).\end{aligned}$$
Now we shall handle the boundary term at $\infty$. Let $\chi_2$ be a smooth function such that $$\label{evenmorechoices10}
\chi_2 = 1\text{ for }r \in [R_2,\infty),\qquad \chi_2 = 0 \text{ for }r \in [r_+,R_1],\qquad \left|\chi_2\right| \leq B.$$ Then, we have $$\begin{aligned}
\label{smallOmegaInfinity}
&\frac{E}{2}\left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \int_{R_2}^{R_3}h\right)\left [|u'|^2+\omega^2 |u|^2\right]_{r=\infty}= E\left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \int_{R_2}^{R_3}h\right)\int_{-\infty}^{\infty}\left(\chi_2\text{Q}^T\right)' \\ \nonumber &= E\left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \int_{R_2}^{R_3}h\right)\omega\int_{R_1}^{R_2}\chi_2'\text{Im}\left(u'\overline u\right)+ E\left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \int_{R_2}^{R_3}h\right)\omega\int_{-\infty}^{\infty}\chi_2\text{Im}\left(H\overline u\right).\end{aligned}$$
We have $$\begin{aligned}
\label{smallOmegaHorizonError2}
&E\left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \int_{R_2}^{R_3}h\right)\left|\omega\int_{R_1}^{R_2}\chi_2'\text{Im}\left(u'\overline u\right)\right| \\ \nonumber &\qquad\le BE\omega_0^{-2}\exp\left(BCR_2\right)\left(\frac{e^{p^{-1}}R_3 - R_3}{R_3^2} + \int_{R_2}^{R_3}h\right)\omega_{\rm low}\int_{R_1}^{R_2}\left(\tilde y'\left|u'\right|^2 + \omega_0^2\tilde y'\left|u\right|^2\right).\end{aligned}$$
Thus, using that $E \geq 2$, we may combine (\[smallOmegaInfinity\]) and (\[smallOmegaSixth\]) and obtain $$\begin{aligned}
\label{smallOmegaSeventh}
b&\int_{-\infty}^\infty \left (\left(\tilde y'+h\right) |u'|^2+\left(\tilde y'\omega_0^2 + h\left(V-\omega^2\right)\right)|u|^2 \right ) \\ &\nonumber\le
\int_{-\infty}^\infty \left (-2\tilde y \,{\text{Re}} (u'\overline{H}) - Ey(\infty)\chi_2\omega\,{\text{Im}}(H \overline u) - 2\chi_1\left(\omega-\upomega_+m\right)\,{\text{Im}}(H\overline u)- h\,{\text{Re}}(u\overline H)- 2y\,{\text{Re}}(u'\overline H)\right).\end{aligned}$$ At this point, it is clear that we may rescale the functions $\tilde{y}$, $h$ and $y$ by an $\tilde a_0$ dependent constant so that $y$ is identically $1$ for $r^* \geq R^*_{\infty}$.
### The subrange $|\omega|\ge \omega_{\rm low}$ (the non-stationary subcase) {#nonStat}
We turn finally to our last frequency range. It is only this range which gives rise to the term $1_{\{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}\}\cap \{\Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u(-\infty)\right|^2$ on the right hand side of $(\ref{fromPhaseSpace2})$ in the statement of Theorem \[phaseSpaceILED\].
Let $(\omega, m,\Lambda)\in \mathcal{G}_{\mbox{$\flat$}}$ where $|\omega|\ge \omega_{\rm low}$. When the final choices of $\omega_{\rm high}$ $\epsilon_{\rm width}$, $\omega_{\rm low}$ have been made, we will set the functions $f$, $h$, $\tilde y$, $\hat y$ and $\chi_1$ together with the trapping parameter $r_{\rm trap}$ to be $0$. The remaining function $y$ and desired coercivity property is given by
\[odeEst7\]Let $a_0<M$. Then, for all $\omega_{\rm high}>0$, $\epsilon_{\rm width}>0$, $\omega_{\rm low} > 0$, $E\geq 2$, for all $R_{\infty}$ sufficiently large depending on $\omega_{\rm high}$, $\omega_{\rm low}$, $0\le a\le a_0$, and for all $(\omega, m, \Lambda)\in \mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm high}, \epsilon_{\rm width})$ such that $|\omega|\ge \omega_{\rm low}$, there exists a function $y$ satisfying the uniform bounds $$\left|y\right| \leq B,$$ $$y = 1\text{ for }r^* \geq R^*_{\infty},$$ such that, for all smooth solutions $u$ to the radial o.d.e. (\[e3iswsntouu\]) with right hand side $H$, satisfying moreover the boundary conditions (\[eq:b-\]) and (\[eq:b+\]), we have $$\begin{aligned}
b\left(\omega_{\rm low}, \omega_{\rm high}, \epsilon_{\rm width}\right)\int_{R_-^*}^{R_+^*} &\left(|u'|^2+|u|^2 \right)
\\ &\le B\left (\left|\omega\left(\omega-\upomega_+m\right)\right||u|^2\right)_{r=r_+}- \int_{-\infty}^\infty \left (2 y \,{\text{Re}} (u'\overline{H})-E\omega {\text{Im}} (\overline H u) \right).\end{aligned}$$
As in the previous section we will treat the superradiant and non-superradiant frequencies concurrently. However, as previously discussed, it is only for the sake of the superradiant frequencies for which we include the first term on the right hand side of the estimates of the proposition.
Our current will be of the form $$\text{Q} = \text {\fontencoding{LGR}\selectfont \koppa}^y - E\text{Q}^T$$ for $$y(r^*) := \exp\left(-C\int_{r^*}^{\infty}\chi_{R_{\infty}^*}r^{-2}dr\right),$$ where $C = C(\omega_{\rm low},\omega_{\rm high},\epsilon_{\rm width})$ is a sufficiently large constant. The function $\chi_{R_{\infty}^*}$ is a smooth function which is identically $1$ on $[r_+,R_{\infty}-1)$ and identically $0$ on $[R_{\infty},\infty)$. Note that $y|_{r^* \geq R^*_{\infty}} = 1$ and $y\left(-\infty\right) = 0$. Applying the current $\text{Q}$ gives $$\begin{aligned}
\label{nonStatFirst}
\int_{-\infty}^{\infty} &\left (y' |u'|^2+\left (y'\omega^2-\left(yV\right)'\right) |u|^2 \right) - \left (|u'|^2+(1-E)\omega^2 |u|^2\right)_{r=\infty}
\\ &=\left(\omega(\omega-\upomega_+m) |u|^2\right)_{r=r_+}- \int_{-\infty}^\infty \left (2 y \,{\text{Re}} (u'\overline{H})+E\omega {\text{Im}} (u\overline H) \right).\end{aligned}$$
Let $R_{\infty}$ be sufficiently large and $R_6$ be chosen such that $1 \ll R_6 \ll R_{\infty}-1$. Then, let $\chi_2$ be a smooth function such that $$\label{evenevenmorechoices}
\chi_2 = 1\text{ for }r\in [r_+,R_6],\qquad \chi_2 = 0\text{ for }R \in [R_{\infty}-1,\infty),$$ $$\label{evenevenmorechoices2}
\left|\chi_2'\right| \leq B\left(R_{\infty}-R_6\right)^{-1},\qquad \left|\chi_2\right| \leq B.$$ Then set $V_{\leq} := \chi_2V$ and $V_{\geq} = \left(1-\chi_2\right)V$. Of course, we have $V = V_{\leq} + V_{\geq}$. For $V_{\leq}$ we have $$\begin{aligned}
\label{absorbV1}
\left|\int_{-\infty}^{\infty}\left(yV_{\leq}\right)'\left|u\right|^2\right| &= 2\left|\int_{-\infty}^{\infty}yV_{\leq}\text{Re}\left(u'\overline u\right)\right| \\
\nonumber
&\leq \epsilon \int_{-\infty}^{\infty}y'\left|u'\right|^2 + B\epsilon^{-1}\int_{-\infty}^{\infty}y'\omega^2\frac{y^2V_{\leq}^2}{\left(y'\right)^2\omega^2}\left|u\right|^2\\ \nonumber &=\epsilon \int_{-\infty}^{\infty}y'\left|u'\right|^2 + B\left(\omega_{\rm high},\epsilon_{\rm width}\right)\epsilon^{-1}\int_{-\infty}^{\infty}y'\omega^2\frac{V_{\leq}^2}{C^2r^{-4}\omega^2}\left|u\right|^2\\
&\nonumber \leq \epsilon \int_{-\infty}^{\infty}y'\left|u'\right|^2 + B\left(\omega_{\rm high},\epsilon_{\rm width}\right)\epsilon^{-1}C^{-2}\omega^{-2}\int_{-\infty}^{\infty}y'\omega^2\left|u\right|^2,\end{aligned}$$ while for $V_{\geq}$ we have $$\begin{aligned}
\label{absorbV2}
-\int_{-\infty}^{\infty}\left(yV_{\geq}\right)'\left|u\right|^2 &= \int_{-\infty}^{\infty}\left(-y'V_{\geq} - yV_{\geq}'\right)\left|u\right|^2 \\ \nonumber
&\ge \int_{-\infty}^{\infty}\left(-B\left(\omega_{\rm high},\epsilon_{\rm width}\right)R_6^{-2}\omega^{-2}\left(y'\omega^2\right)\right.\\
\nonumber &\qquad\qquad\left.
-B\left(\omega_{\rm high},\epsilon_{\rm width}\right)\left(R_{\infty}-R_6\right)^{-1}1_{\text{supp}\left(\chi_2'\right)}yV + by\left(1-\chi_2\right)r^{-3}\right)\left|u\right|^2 \\ \nonumber & \ge-B\left(\omega_{\rm high},\epsilon_{\rm width}\right)\text{max}\left(R_{\infty}^{-2}\omega^{-2},\left(R_{\infty}-R_6\right)^{-1}\right)\int_{-\infty}^{\infty}\left(y'\omega^2\right)\left|u\right|^2.\end{aligned}$$
It is now clear that choosing $C$, $R_{\infty}$ and $R_{\infty} - R_6$ sufficiently large depending on $\omega_{\rm low}$ or $\omega_{\rm low}$ and $\omega_{\rm high}$ and $\epsilon_{\rm width}$ and combining (\[nonStatFirst\]), (\[absorbV1\]) and (\[absorbV2\]) will imply the proposition.
Putting everything together {#putting}
---------------------------
In this section we will combine the propositions of the above sections to prove Theorem \[phaseSpaceILED\].
First of all, keeping Lemma \[everythingCovered\] in mind, we observe that for any choice of $\omega_{\rm high}$ and $\epsilon_{\rm width}$, every admissible frequency triple $\left(\omega,m,\Lambda\right)$ lies in exactly one of the frequency ranges: $\mathcal{G}_{\mbox{$\flat$}}(\omega_{\rm high}, \epsilon_{\rm width})$, $\mathcal{G}_{{{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}}(\omega_{\rm high}, \epsilon_{\rm width})$, $\mathcal{G}_{\mbox{$\natural$}}(\omega_{\rm high}, \epsilon_{\rm width})$, $\mathcal{G}_{\mbox{$\sharp$}}(\omega_{\rm high}, \epsilon_{\rm width})$, $\mathcal{G}^{\mbox{$\sharp$}}(\omega_{\rm high}, \epsilon_{\rm width})$. Thus, it only remains to choose the constants $\epsilon_{\rm width}$, $E$, $\omega_{\rm high}$, $\omega_{\rm low}$, $\tilde a_0$ and $R_{\infty}$ in the correct order so that it is possible to apply simultaneously all of the above propositions.
The first constant we fix is a sufficiently small $\epsilon_{\rm width}$, consistent with the requirements of Propositions \[odeEst2\] and \[odeEst3\]. Then, depending on the choice of $\epsilon_{\rm width}$, for all large enough $\omega_{\rm high}$, $R_{\infty}$ and $E$ we may apply Propositions \[odeEst2\], \[odeEst3\] and, in addition, Proposition \[odeEst4\], corresponding to the frequency regimes $\mathcal{G}_{\mbox{$\sharp$}}$, $\mathcal{G}_{{{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}}$ and $\mathcal{G}_{\mbox{$\natural$}}$. Now we fix the choice of $E$ consistent with the above requirement. Then, depending on this choice of $E$, for all large enough $\omega_{\rm high}$ and $R_{\infty}$ we may apply, in addition to the above Propositions, also Proposition \[odeEst1\], corresponding to the frequency regime $\mathcal{G}^{\mbox{$\sharp$}}$. Finally, we fix the constant $\omega_{\rm high}$ consistent also with this requirement.
Since $\epsilon_{\rm width}$ and $\omega_{\rm high}$ are both fixed, the frequency ranges $\mathcal{G}_{\mbox{$\sharp$}}$, $\mathcal{G}_{{{{\mbox{$\flat \mkern-12mu {}^{\_}$}}}}}$, $\mathcal{G}_{\mbox{$\natural$}}$, $\mathcal{G}^{\mbox{$\sharp$}}$, $\mathcal{G}_{\mbox{$\flat$}}$ are now determined.
We still must determine the four subranges of $\mathcal{G}_{\mbox{$\flat$}}$ which depend on additional parameters $\omega_{\rm low}$ and $\tilde a_0$, and make our final choice of $R_{\infty}$.
We choose first $\tilde a_0$ and $\omega_{\rm low}$ sufficiently small so that for $R_\infty$ sufficiently large, we can apply Propositions \[odeEst5\] and \[odeEst5b\]. We then fix our choice of $\tilde a_0$. Then chose sufficiently small $\omega_{\rm low}$ depending on $\tilde a_0$, and note that for sufficiently large $R_{\infty}$ depending on $\tilde a_0$ we may apply (in addition to all previous Propositions) also Proposition \[odeEst6\]. Finally, choose $R_{\infty}$ so that we may apply, in addition to all the previous Propositions, also Proposition \[odeEst7\].
With these choices, all frequency ranges are determined so as to indeed simultaneously satisfy the assumptions of Propositions \[odeEst1\]–\[odeEst7\]. We now for each frequency range define the functions $f$, $y$, $r_{\rm trap}$, etc., as given in the corresponding Proposition or else set them to $0$ (as explained before each statement). The statement of Propositions \[odeEst1\]–\[odeEst7\] then give Theorem \[phaseSpaceILED\] for frequencies $(\omega, m, \Lambda)$ in the corresponding range. Since these ranges cover all admissible frequencies, the proof is complete.
Trapping parameters {#trappingParam}
-------------------
We finally define the trapping parameters $s_{\pm}$ which appear in the definition of the degeneration function $\zeta$ (see (\[degenerationfunc\])) which in turn appears in the statement of Theorem \[theResult\].
Let $0\le a_0 < M$ and let $\omega_{\rm high} = \omega_{\rm high}(a_0,M)$ and $\epsilon_{\rm width} = \epsilon_{\rm width}(a_0,M)$ be the parameters from Theorem \[phaseSpaceILED\]. We define the trapping parameters $s_{\pm}$ by $$\label{sminus}
s_-\left(a_0,M\right) \doteq 3M - \inf_{0\le a \leq a_0,(\omega,m,\Lambda) \in \mathcal{G}_{\mbox{$\natural$}}, r_{\rm trap}\ne 0}r_{\rm trap}\left(\omega,m,\Lambda\right) - \varepsilon\left(a_0\right),$$ $$\label{splus}
s_+\left(a_0,M\right) \doteq \sup_{0\le a \leq a_0,(\omega,m,\Lambda) \in \mathcal{G}_{\mbox{$\natural$}}r_{\rm trap}\ne 0}r_{\rm trap}\left(\omega,m,\Lambda\right) - 3M + \varepsilon\left(a_0\right).$$ where $\varepsilon\left(a_0\right)$ is a fixed choice of continuous function such that $\varepsilon(0)=0$ and $\varepsilon\left(a_0\right) > 0$ for $a_0>0$, and such that $s_\pm$ satisfy for all $0\le a\le a_0$ the relations $$r_+(a,M)<3M-s_-(a_0,M)
<3M + s_+\left(a_0,M\right) < \infty.$$
The proof of Proposition \[odeEst4\] shows that $\varepsilon(a_0)$ can be chosen ensuring that $s_\pm$, $\varepsilon$ enjoy the properties claimed in the above definition.
Let us observe that we then necessarily have $$\lim_{a_0\to 0}s_{\pm}\left(a_0,M\right) = 0, \qquad
\lim_{a_0\to M} s_-(a_0,M) =2M=3M-r_+(M,M)$$ From the latter, it follows that we must also have $\lim_{a_0\to M}\varepsilon\left(a_0\right) = 0$.
Recall the definition of the physical space degeneration function $\zeta$ (see (\[degenerationfunc\])) which in particular required the definition of the points $s_{\pm}$. It follows from our definition of $r_{\rm trap}$ in Theorem \[phaseSpaceILED\] that we now have for all admissible $(\omega, m, \Lambda)$ then $r_{\rm trap}=0$ or $$3M-s_-+\varepsilon(a_0) \le r_{\rm trap} \le 3M+s_+-\varepsilon(a_0)$$ It follows that for all admissible $(\omega, m, \Lambda)$, we have the uniform bound $$\label{insteadofthis}
b\zeta \le (1-r^{-1}r_{\rm trap})^2.$$ In particular, the statement of Theorem \[phaseSpaceILED\] holds with $(1-r^{-1}r_{\rm trap})^2$ replaced by $\zeta$. It is this weaker statement that we will in fact apply in the following section.
Summing and integrated local energy decay for future-integrable solutions {#summation}
=========================================================================
In this section, we will combine the estimates of Sections \[Nmult\], \[largeR\] and the o.d.e. analysis of Section \[freqLocEst\] to prove integrated local energy decay for solutions of the wave equation sufficiently integrable towards the future. We begin by defining this class and stating the main proposition.
Future-integrable solutions of the wave equation {#futIntSec}
------------------------------------------------
Let $a_0<M$, $|a| \le a_0$ and let $\psi$ be as in the reduction of Section \[WPosed\], i.e., a solution of the wave equation $(\ref{WAVE})$ on $\mathcal{R}_0$ arising from smooth compactly supported data at $\Sigma_0$. Let $\xi\left(\tau\right)$ be smooth function which is $0$ in the past of $\Sigma_0$ and identically $1$ in the future of $\Sigma_1$. Then we define $$\psi_{\text{\Rightscissors}} \doteq \xi\psi.$$ We have $$\label{inhogo}
\Box_g\psi_{\text{\Rightscissors}} = F \doteq 2\nabla^{\mu}\xi\nabla_{\mu}\psi + \left(\Box_g\xi\right)\psi.$$
Let $|a|<M$ and let $\psi$ be a solution of $(\ref{WAVE})$ as in the reduction of Section \[WPosed\]. We shall say that $\psi$ is future-integrable if $\psi_{\text{\Rightscissors}}$ satisfies Definition \[sufficient\].
Note that $\psi_{\text{\Rightscissors}}$ by its construction will then automatically satisfy Definitions \[sufficient2\].
Recall the degeneration function $\zeta$ defined by $(\ref{degenerationfunc})$ in Section \[weritsdef\], and \[easyCoArea\]. The main result of this section is
\[closedILED\]Let $a_0<M$, $|a|\le a_0$, and let $\psi$ be a future integrable solution of $(\ref{WAVE})$. Then, for every $\delta > 0$ $$\int_{\mathcal{H}^+_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\mathcal{H}^+}+ \int_{\mathcal{I}^+}\mathbf{J}^T_{\mu}[\psi]n^{\mu}_{\mathcal{I}^+}+
\int_0^{\infty}\int_{\Sigma_{\tau}}\left(\left|\tilde{Z}^*\psi\right|^2r^{-1-\delta} + r^{-3-\delta}\left|\psi\right|^2 +
\zeta \left|T\psi\right|^2r^{-1-\delta} + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\psi\right|^2r^{-1}\right)\, d\tau$$ $$\leq
B\left(\delta\right)\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.$$
The proof of this proposition will be carried out in Sections \[cutoffSec\]–\[whitinghere\] below. In view of the reduction of Section \[signOfa\], we may assume in this proof that $a\ge 0$, in order to appeal to the the results of Section \[freqLocEst\] as stated.
Finite in time energy estimate {#cutoffSec}
------------------------------
Defining $\psi_{\text{\Rightscissors}}$ as above, by Section \[karteri\], we may apply Carter’s separation to the inhomogeneous equation $(\ref{inhogo})$ to define the function $u^{(a\omega)}_{m\ell}$. Lemma \[aeRegular\] implies that for almost every $\omega$, then for all $m$, $\ell$. the function $H^{(a\omega)}_{m\ell}$ is smooth and $u^{(a\omega)}_{m\ell}$ is a smooth solution to the radial o.d.e. (\[e3iswsntouu\]) satisfying the boundary conditions (\[eq:b+\]) and (\[eq:b-\]). For each such $\omega$ we may apply the estimates of Section \[freqLocEst\] to the admissible triples $(\omega, m, \Lambda= \Lambda_{m\ell}(a\omega))$ and conclude that Theorem \[phaseSpaceILED\] holds.
Thus, integrating $(\ref{fromPhaseSpace2})$ in $\omega$, summing in $m$ and $\ell$ and applying (\[insteadofthis\]) yields the estimate $$\begin{aligned}
\label{fromPhaseSpace}
b&\int_{-\infty}^{\infty}\sum_{m\ell}\int_{R_-^*}^{R_+^*}\left(\left|u'\right|^2 + \left(\zeta\left(\omega^2 + \Lambda_{m\ell}\right) + 1\right)\left|u\right|^2\right)\, dr^*\, d\omega
\\ \nonumber &\le \int_{-\infty}^{\infty}\sum_{m\ell}\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\,dr^*\, d\omega + \int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega.\end{aligned}$$
An application of Plancherel to $(\ref{fromPhaseSpace})$ (see the explicit formulas in Section \[separationSubsection\] and the discussion of the volume form in Section \[usefulcomps\]) yields $$\begin{aligned}
\label{conseqSecFreqLocEst}
&b\int_0^{\infty}\int_{\Sigma_{\tau} \cap [R_-,R_+]}\left(\left|\partial_{r^*}\psi_{\text{\Rightscissors}}\right|^2 + \left|\psi_{\text{\Rightscissors}}\right|^2 +
\zeta \left|T\psi_{\text{\Rightscissors}}\right|^2 + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\psi_{\text{\Rightscissors}}\right|^2\right)\, d\tau \\ \nonumber &\,\,\le \int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega+ \int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega.\end{aligned}$$
Recall that the last term on the right hand side of both of these estimates arises from Section \[nonStat\]. We further remark that this term would be controlled by the physical space quantity $\int_{\mathcal{H}^+(0,\infty)}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\mathcal{H}^+}$, if we had control for the latter–in general, we do not, however! In Section \[whitinghere\] we shall exploit the localisation of the integral in $\omega$ to control this using the quantitative mode stability result [@shlapRot].
The first thing we observe is that $\partial\psi_{\text{\Rightscissors}}$ only differ from $\partial\psi$ when $\tau \in [0,1]$. However, in this region, the energy can simply be controlled by a finite in time energy inequality and a Hardy inequality.
For the $\left|\psi_{\text{\Rightscissors}}\right|^2$ term, we observe $$\begin{aligned}
\int_0^1\int_{\Sigma_{\tau}\cap [R_-, R_+]}\left|\psi\right|^2 &\leq B\int_0^1\int_{\Sigma_{\tau}}\frac{\left|\psi\right|^2}{r^2} \leq B\int_0^1\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}},\end{aligned}$$ where we have used a Hardy inequality and a finite in time energy estimate.
We conclude $$\begin{aligned}
\label{addFiniteEnergyEst}
&b\int_0^{\infty}\int_{\Sigma_{\tau}\cap \left[R_-,R_+\right]}\left(\left|\partial_{r^*}\psi\right|^2 + \left|\psi\right|^2 +
\zeta \left|T\psi\right|^2 + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\psi\right|^2\right)\, d\tau \\ \nonumber &\le \int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega
\\ \nonumber &\qquad+\int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
Adding in the red-shift {#addInTheRedShift}
-----------------------
Next, we add a small constant (depending only on $M$) times the red-shift estimate of Proposition \[ftrs\] to (\[addFiniteEnergyEst\]). Recalling that $R_- = r_+ + \frac{1}{2}\left(r_{\rm red} - r_+\right)$, we thus obtain $$\begin{aligned}
\label{redshiftAdded}
&b\int_0^{\infty}\int_{\Sigma_{\tau}\cap \left[r_+,R_+\right]}\left[\left|\tilde{Z}^*\psi_{\text{\Rightscissors}}\right|^2 + \left|\psi_{\text{\Rightscissors}}\right|^2 +
\zeta \left|T\psi_{\text{\Rightscissors}}\right|^2 + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\psi_{\text{\Rightscissors}}\right|^2\right]\, d\tau +b\int_{\mathcal{H}^+(0,\infty)}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\mathcal{H}^+} \\ \nonumber &\le \int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega
\\ \nonumber &\qquad + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega.\end{aligned}$$
Adding in the large $r$ current
-------------------------------
Next, we would like to add in a small constant times the large $r$ estimate of Section \[largeR\]. However, we must be careful because that estimate produces an error proportional to $\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}$, and we do not yet have a uniform energy boundedness statement.
We surmount this difficulty as follows. Since $\psi$ is future-integrable, we know that $\int_{\Sigma_{\tau} \cap [r_+,R_+]}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}$, as a function of $\tau$, is in $L^1_{\tau}[0,\infty)$. A pigeonhole argument implies that we may find a constant $C\left(\psi\right)$ and a dyadic sequence $\{\tau_n\}_{n=1}^{\infty}$ with $\lim_{n\to\infty}\tau_n = \infty$ such that $$\label{dyadicEnergyDecay}
\int_{\Sigma_{\tau_n}\cap[r_+,R_+]}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau_n}} \leq \frac{C}{\tau_n}.$$ Note that $R_+ = 2R_{\rm large} \geq 4M$, so that $T$ is timelike in the region $r \geq R_+$. Then, a $\mathbf{J}_{\mu}^T$ energy estimate implies $$\begin{aligned}
\label{conseqDyadicEnergyDecay}
\int_{\Sigma_{\tau_n}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma} &\leq \int_{\Sigma_{\tau_n}\cap [r_+,R_+]}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma} + B\int_{\Sigma_{\tau_n}\cap [R_+,\infty)}\mathbf{J}^T_{\mu}[\psi]n^{\mu}_{\Sigma}
\\ \nonumber &\leq B\frac{C}{\tau_n} + B\int_{\mathcal{H}^+(0,\tau_n)}\mathbf{J}^N_{\mu}[\psi]n^{\mu} + B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}.\end{aligned}$$ Now, combine Proposition \[lrp\] with (\[conseqDyadicEnergyDecay\]), multiply the result by a sufficiently small constant, and then add the result to (\[redshiftAdded\]). In particular, the horizon term on the right hand side of (\[conseqDyadicEnergyDecay\]) will be multiplied by a small constant, and hence can be absorbed into the left hand side of (\[redshiftAdded\]). We obtain $$\begin{aligned}
\label{addedLargeR}
b\int_0^{\tau_n}\int_{\Sigma_{\tau}}&\left[|\tilde{Z}^*\psi|^2r^{-1-\delta} + \left|\psi\right|^2r^{-3-\delta} +
\zeta \left|T\psi\right|^2r^{-1-\delta} + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\psi\right|^2r^{-1}\right]\, d\tau \\ \nonumber \le&
\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega + B\left(\delta\right)\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} \\ \nonumber &+ \int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega + \frac{C}{\tau_n}.\end{aligned}$$ Taking $n \to \infty$ gives $$\begin{aligned}
\label{addedLargeR2}
\int_0^{\infty}\int_{\Sigma_{\tau}}&\left[\left|\tilde{Z}^*\psi\right|^2r^{-1-\delta} + \left|\psi\right|^2r^{-3-\delta} +
\zeta \left|T\psi\right|^2r^{-1-\delta} + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\psi\right|^2r^{-1}\right]\, d\tau \\ \nonumber \le& \int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega
\\\nonumber&+
B\left(\delta\right)\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega.\end{aligned}$$
Boundedness of the energy flux to $\mathcal{I}^+$ {#toayplus}
-------------------------------------------------
The estimates of the previous section give in addition an estimate for the energy flux to null infinity $$\label{tonullinfinity}
\int_{\mathcal{I}^+}\mathbf{J}^T_{\mu}[\psi]n^{\mu}_{\mathcal{I}^+},$$ which is well-defined by an easy limiting operation (see [@dr6]) which we omit here.
To bound $(\ref{tonullinfinity})$, we only need the easily proven property of its definition, that applying $\mathbf{J}^T$ energy estimates outside the ergoregion, $(\ref{tonullinfinity})$ can be seen to satisfy $$\label{toBoundNullInf}
\int_{\mathcal{I}^+}\mathbf{J}^T_{\mu}[\psi]n^{\mu}_{\mathcal{I}^+} \leq \limsup_{n\to\infty}\int_{\Sigma_{s_n}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s_n}},$$ where $\{s_n\}$ is any sequence with $\lim_{n\to\infty}s_n = \infty$.
Now, taking the limit as $n\to\infty$ in the inequality (\[conseqDyadicEnergyDecay\]) and then applying (\[redshiftAdded\]) gives $$\begin{aligned}
\label{almostNullInfBound}
\limsup_{n\to\infty}\int_{\Sigma_{\tau_n}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma} &\leq B\int_{\mathcal{H}^+(0,\infty)}\mathbf{J}^N_{\mu}[\psi]n^{\mu} + B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}
\\ \nonumber &\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$ Then (\[toBoundNullInf\]) implies $$\begin{aligned}
\label{boundNullInf}
\int_{\mathcal{I}^+}\mathbf{J}^T_{\mu}[\psi]n^{\mu}_{\mathcal{I}^+} \leq
\limsup_{n\to\infty}\int_{\Sigma_{\tau_n}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma} \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
An alternative approach for controlling $\int_{\mathcal{I}^+}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\mathcal{I}^+}$ can be found in the proof of the inequality (\[boundedEnergyInfinity\]) where a cut-off $\mathbf{J}^T$ energy estimate is applied and the errors are absorbed with the integrated energy decay statement. Lastly, we note that yet another approach would be to first show that (up to a normalisation constant) $$\int_{\mathcal{I}^+}\mathbf{J}^T_{\mu}[\psi]n^{\mu}_{\mathcal{I}^+} = \int_{-\infty}^{\infty}\sum_{m\ell}\omega^2\left|u\left(\infty\right)\right|^2,$$ and then use the fact that the estimates of Section \[freqLocEst\] give a bound for the right hand side of this equality.
Error terms associated to the cutoff
------------------------------------
In this subsection we will study closely the terms $\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega$ on the right hand side of (\[addedLargeR2\]). Recall that when $r \geq R_{\infty}$, we have arranged for our multipliers to be independent of the frequency parameters or decay exponentially in $r$ (see Remark \[R2\]). In particular, we may split the error terms associated to the cutoff into: $$\begin{aligned}
\label{errorsFromCutoff}
&\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega\\
\nonumber&\qquad = \int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}\chi_{R_{\infty}}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega -\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}2\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(u'\overline H\right)\right)\, d\omega
\\ \nonumber &\qquad\qquad+\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}E\omega\left(1-\chi_{R_{\infty}}\right)\text{Im}\left(H\overline u\right)\right)\, d\omega
\\ \nonumber &\qquad\qquad+2\int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left(\int_{-\infty}^{\infty}\left(1-\chi_{R_{\infty}}\right)\tilde y\text{Re}\left(u'\overline H\right)\right)\, d\omega.\end{aligned}$$ Here $\chi_{R_{\infty}}$ is a cutoff which is identically $1$ on $[r_+,R_{\infty}]$ and $0$ on $[R_{\infty}+1,\infty)$.
### The bounded $r$ error terms {#boundedRError}
The error terms in the region $[r_+,R_{\infty}]$ are the easiest to deal with. We simply observe that an application of Plancherel (see Sections \[separationSubsection\] and \[usefulcomps\]), finite in time energy estimates and Hardy inequalities $$\begin{aligned}
&\left|\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}\chi_{R_{\infty}}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega\right|
\\ \nonumber &\qquad\leq B\int_{-\infty}^{\infty}\sum_{m\ell}\int_{-\infty}^{R_{\infty}+1}\left|H\right|\left(\left|u'\right| + \left|u\right|\right)
\\ \nonumber &\qquad\leq B\epsilon^{-1}\int_0^{\infty}\int_{\Sigma_{\tau}\cap [r_+,R_{\infty}+1)}\left|F\right|^2 + \epsilon\int_0^{\infty}\int_{\Sigma_{\tau}\cap [r_+,R_{\infty}+1)}\left(\left|\partial_{r^*}\psi\right|^2 + \left|\psi\right|^2\right)
\\ \nonumber &\qquad\leq B\epsilon^{-1}\int_0^{2}\int_{\Sigma_{\tau}\cap [r_+,\infty)}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} + \epsilon\int_0^{\infty}\int_{\Sigma_{\tau}\cap [r_+,R_{\infty}+1)}\left(\left|\partial_{r^*}\psi\right|^2 + \left|\psi\right|^2\right)
\\ \nonumber &\qquad\leq B\epsilon^{-1}\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \epsilon\int_0^{\infty}\int_{\Sigma_{\tau}\cap [r_+,R_{\infty}+1)}\left(\left|\partial_{r^*}\psi\right|^2 + \left|\psi\right|^2\right).\end{aligned}$$
### Large $r$: $\text {\fontencoding{LGR}\selectfont \koppa}^y$ error terms {#largeRkoppa}
For error terms supported in the $r \in [R_{\infty},\infty)$ region we must be careful that lower order terms come with appropriate $r$-weights so that either a Hardy inequality can be applied or they can be absorbed into the left hand side of (\[addedLargeR2\]). First of all, an application of Plancherel (see Section \[separationSubsection\]) gives $$\begin{aligned}
&\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(u'\overline H\right)\right)\, d\omega
\\ \nonumber\qquad = &\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\partial_{r^*}\left(\left(r^2+a^2\right)^{1/2}\psi_{\text{\Rightscissors}}\right)\overline{\Delta\left(r^2+a^2\right)^{-1/2}F}\right)\ \sin\theta\, dt\, dr^*\, d\theta\, d\phi.\end{aligned}$$ Recall that $$F = \left(r^2+a^2\right)^{-1}\rho^2\left(2\nabla^{\mu}\xi\nabla_{\mu}\psi + \left(\Box_g\xi\right)\psi\right).$$ For sufficiently large $r$, $\xi$ is just a function of $t$. Hence, $$F = \left(r^2+a^2\right)^{-1}\rho^2\left(2g^{tt}\dot\xi\partial_t\psi + 2g^{t\phi}\dot\xi\partial_{\phi}\psi + g^{tt}\ddot\xi\psi\right).$$ Thus, (suppressing the $\sin\theta\, dt\, dr^*\, d\theta\, d\phi$) $$\begin{aligned}
&\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\partial_{r^*}\left(\left(r^2+a^2\right)^{1/2}\psi_{\text{\Rightscissors}}\right)\overline{\Delta\left(r^2+a^2\right)^{-1/2}F}\right)
\\& =\nonumber \int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\partial_{r^*}\left(\left(r^2+a^2\right)^{1/2}\psi_{\text{\Rightscissors}}\right)\overline{\Delta\left(r^2+a^2\right)^{-3/2}\rho^2\left(2g^{tt}\dot\xi\partial_t\psi + 2g^{t\phi}\dot\xi\partial_{\phi}\psi\right)}\right)
\\ \nonumber &\qquad+\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\partial_{r^*}\left(\left(r^2+a^2\right)^{1/2}\psi_{\text{\Rightscissors}}\right)\overline{\Delta\left(r^2+a^2\right)^{-3/2}\rho^2g^{tt}\ddot\xi\psi}\right).\end{aligned}$$
First we consider the term with $g^{tt}\dot\xi\partial_t\psi$: $$\begin{aligned}
\label{gttTerm}
&\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\partial_{r^*}\left(\left(r^2+a^2\right)^{1/2}\psi_{\text{\Rightscissors}}\right)\overline{\Delta\left(r^2+a^2\right)^{-3/2}\rho^2\left(2g^{tt}\dot\xi\partial_t\psi\right)}\right) \right|
\\ \nonumber \le&B\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\left(\partial_{r^*}\psi_{\text{\Rightscissors}}\right)\overline{\Delta\left(r^2+a^2\right)^{-1}\rho^2\left(2g^{tt}\dot\xi\partial_t\psi\right)}\right)\right|
\\ \nonumber &+B\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\frac{r}{(r^2+a^2)^{1/2}}\text{Re}\left(\left(\psi_{\text{\Rightscissors}}\right)\overline{\Delta\left(r^2+a^2\right)^{-3/2}\rho^2\left(2g^{tt}\dot\xi\partial_t\psi\right)}\right)\right|
\\ \nonumber\leq & B\int_0^1\int_{\Sigma_{\tau}\cap [R_{\infty},\infty)}\left(\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} + \frac{\left|\psi\right|^2}{r^2}\right)
\nonumber\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}.\end{aligned}$$
Keeping in mind that $g^{t\phi} = O\left(r^{-3}\right)$, the term with $g^{t\phi}\dot\xi\partial_{\phi}\psi$ can be treated like (\[gttTerm\]). Finally, recalling that $\partial_{r^*}\xi = 0$ for $r \geq R_{\infty}$, we have $$\begin{aligned}
\label{lowerOrder}
&\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\partial_{r^*}\left(\left(r^2+a^2\right)^{1/2}\psi_{\text{\Rightscissors}}\right)\overline{\Delta\left(r^2+a^2\right)^{-3/2}\rho^2g^{tt}\ddot\xi\psi}\right)\right|
\\ \nonumber &= \left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\left(1-\chi_{R_{\infty}}\right)\Delta\left(r^2+a^2\right)^{-2}\rho^2g^{tt}\ddot\xi\xi\text{Re}\left(\partial_{r^*}\left(\left(r^2+a^2\right)^{1/2}\psi\right)\overline{\left(r^2+a^2\right)^{1/2}\psi}\right)\right|
\\ \nonumber &=
\frac{1}{2}\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\partial_{r^*}\left(\left(1-\chi_{R_{\infty}}\right)\Delta\left(r^2+a^2\right)^{-2}\rho^2g^{tt}\ddot\xi\xi\right)\left(r^2+a^2\right)\left|\psi\right|^2\right|
\\ \nonumber &\leq B\int_0^1\int_{\Sigma_{\tau}\cap [R_{\infty},\infty)}\frac{\left|\psi\right|^2}{r^2}
\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
Combining everything implies $$\left|\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}2\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(u'\overline H\right)\right)\, d\omega\right| \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.$$
### Large $r$: $\text{Q}^T$ error terms {#largeRQT}
An application of Plancherel (see Section \[separationSubsection\]) gives $$\begin{aligned}
&\left|\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}\omega\left(1-\chi_{R_{\infty}}\right)\text{Im}\left(H\overline u\right)\right)\, d\omega\right|\\ \nonumber &=\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\Delta\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\partial_t\psi_{\text{\Rightscissors}}\overline{F}\right)\ \sin\theta\, dt\, dr^*\, d\theta\, d\phi\right|.\end{aligned}$$ We have $$F = \left(r^2+a^2\right)^{-1}\rho^2\left(2g^{tt}\dot\xi\partial_t\psi +
2g^{t\phi}\dot\xi\partial_{\phi}\psi + g^{tt}\ddot\xi\psi\right),\qquad
\partial_t\psi_{\text{\Rightscissors}} = \dot\xi\psi + \xi\partial_t\psi.$$ To ease the notation, let us introduce $$W\left(r,\theta\right) := \Delta\left(r^2+a^2\right)^{-1}\left(1-\chi_{R_{\infty}}\right)\rho^2.$$ Then, suppressing the $\sin\theta\, dt\, dr^*\, d\theta\, d\phi$, we have $$\begin{aligned}
\left|\int_0^{\infty}\right.&\left.\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}\Delta\left(1-\chi_{R_{\infty}}\right)\text{Re}\left(\partial_t\psi_{\text{\Rightscissors}}\overline{F}\right)\right| \\
\nonumber
\le &2\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}W\left(r,\theta\right)g^{t\phi}\dot\xi\text{Re}\left(\left(\dot\xi\psi + \xi\partial_t\psi\right)\overline{\partial_{\phi}\psi}\right)\right|
\\ \nonumber &+\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}W\left(r,\theta\right)\text{Re}\left(\left(\dot\xi\psi + \xi\partial_t\psi\right)\overline{\left(2g^{tt}\dot\xi\partial_t\psi + g^{tt}\ddot\xi\psi\right)}\right)\right|.\end{aligned}$$ Several of the above terms are easy to deal with: $$\begin{aligned}
&\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}W\left(r,\theta\right)g^{t\phi}\dot\xi\text{Re}\left(\left(\dot\xi\psi\right)\overline{\partial_{\phi}\psi}\right)\right| = \left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}W\left(r,\theta\right)g^{t\phi}\left(\dot\xi\right)^2\partial_{\phi}\left|\psi\right|^2\right| = 0.\end{aligned}$$ $$\begin{aligned}
&\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}W\left(r,\theta\right)g^{t\phi}\dot\xi\xi\text{Re}\left(\left(\partial_t\psi\right)\overline{\partial_{\phi}\psi}\right)\right| \leq \int_0^1\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} \leq \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$ $$\begin{aligned}
2\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}W\left(r,\theta\right)g^{tt}\xi\dot\xi\left|\partial_t\psi\right|^2\right| \leq \int_0^1\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} \leq \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$ Combining everything yields $$\begin{aligned}
\label{partialProgAbsorbQT}
&\left|\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}\omega\left(1-\chi_{R_{\infty}}\right)\text{Im}\left(H\overline u\right)\right)\, d\omega\right|
\\ \nonumber &\leq\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}W\left(r,\theta\right)\left(\text{Re}\left(\dot\xi\psi\overline{\left(2g^{tt}\dot\xi\partial_t\psi + g^{tt}\ddot\xi\psi\right)}\right) + \text{Re}\left(\xi\partial_t\psi\overline{g^{tt}\ddot\xi\psi}\right)\right)\right| + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
We now focus on the first term on the right hand side: $$\begin{aligned}
&\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}g^{tt}W\left(r,\theta\right)\left(2\left(\dot\xi\right)^2\text{Re}\left(\psi\overline{\partial_t\psi}\right) + \dot\xi\ddot\xi\left|\psi\right|^2 + \xi\ddot\xi\text{Re}\left(\partial_t\psi\overline{\psi}\right) \right)\right|
\\ \nonumber &=\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}g^{tt}W\left(r,\theta\right)\left(2\left(\dot\xi\right)^2\text{Re}\left(\psi\overline{\partial_t\psi}\right) - \left(\dot\xi\right)^2\text{Re}\left(\psi\overline{\partial_t\psi}\right) + \xi\ddot\xi\text{Re}\left(\partial_t\psi\overline{\psi}\right) \right)\right|
\\ \nonumber &=\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}g^{tt}W\left(r,\theta\right)\left(\left(\dot\xi\right)^2\text{Re}\left(\psi\overline{\partial_t\psi}\right)+ \xi\ddot\xi\text{Re}\left(\partial_t\psi\overline{\psi}\right) \right)\right|
\\ \nonumber &=\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}g^{tt}W\left(r,\theta\right)\left(\left(\dot\xi\right)^2\text{Re}\left(\psi\overline{\partial_t\psi}\right)- \left(\dot\xi\right)^2\text{Re}\left(\partial_t\psi\overline{\psi}\right)-\xi\dot\xi\text{Re}\left(\partial_t^2\psi\overline{\psi}\right) - \xi\dot\xi\left|\partial_t\psi\right|^2 \right)\right|
\\ \nonumber
&=\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}g^{tt}W\left(r,\theta\right)\left(\xi\dot\xi\text{Re}\left(\partial_t^2\psi\overline{\psi}\right) + \xi\dot\xi\left|\partial_t\psi\right|^2 \right)\right|
\\ \nonumber
&\leq \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} +
\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}g^{tt}W\left(r,\theta\right)\left(\xi\dot\xi\text{Re}\left(\partial_t^2\psi\overline{\psi}\right)\right)\right|.\end{aligned}$$ Instead of additional integration by parts on this last term, we use that $\psi$ solves the wave equation, which we write out here for reference: $$\begin{aligned}
\label{waveEqn}
&g^{tt}\partial_t^2\psi = \frac{4Mar}{\rho^2\Delta}\partial_{t,\phi}^2\psi - \frac{\Delta-a^2\sin^2\theta}{\Delta\rho^2\sin^2\theta}\partial_{\phi}^2\psi -
\frac{r^2+a^2}{\Delta\rho^2}\partial_{r^*}\left(\left(r^2+a^2\right)\partial_{r^*}\psi\right) - \frac{1}{\rho^{2}\sin\theta}\partial_{\theta}\left(\sin\theta\partial_{\theta}\psi\right).\end{aligned}$$ Substituting the right hand side of (\[waveEqn\]) for $g^{tt}\partial_t^2\psi$ , carrying out a straightforward integration by parts, and applying a finite in time energy inequality shows $$\left|\int_0^{\infty}\int_{-\infty}^{\infty}\int_{\mathbb{S}^2}g^{tt}W\left(r,\theta\right)\left(\xi\dot\xi\text{Re}\left(\partial_t^2\psi\overline{\psi}\right)\right)\right| \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.$$
Thus, we have shown $$\left|\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}\omega\left(1-\chi_{R_{\infty}}\right)\text{Im}\left(H\overline u\right)\right)\, d\omega\right| \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.$$
### Large $r$: $\text {\fontencoding{LGR}\selectfont \koppa}^{\tilde y}$ error terms {#largeRkoppa2}
Since $\left|\tilde{y}\right| \leq \exp\left(-br^*\right)\text{ as }r^*\to \infty$, we may estimate the term $\int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left(\int_{-\infty}^{\infty}\left(1-\chi_{R_{\infty}}\right)\tilde y\text{Re}\left(u'\overline H\right)\right)\, d\omega$ exactly as in Section \[boundedRError\]. We obtain $$\begin{aligned}
&\left|\int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left(\int_{-\infty}^{\infty}\left(1-\chi_{R_{\infty}}\right)\tilde y\text{Re}\left(u'\overline H\right)\right)\, d\omega\right| \\
&\qquad \leq B\epsilon^{-1}\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \epsilon\int_0^{\infty}\int_{\Sigma_{\tau}\cap [R_{\infty},\infty)}e^{-br^*}\left(\left|\partial_{r^*}\psi\right|^2 + \left|\psi\right|^2\right).\end{aligned}$$
### Absorbing the error terms
Combining the results of Sections \[boundedRError\], \[largeRkoppa\], \[largeRQT\] and \[largeRkoppa2\] gives $$\begin{aligned}
\label{cutoffErrorHandled}
&\int_{-\infty}^{\infty}\sum_{m\ell}\left(\int_{-\infty}^{\infty}H \cdot (f, h, y, \chi) \cdot (u, u')\right)\, d\omega
\\ \nonumber &\qquad\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} +\epsilon\int_0^{\infty}\int_{\Sigma_{\tau}}r^{-1-\delta}\left(\left|\partial_{r^*}\psi\right|^2 + r^{-2}\left|\psi\right|^2 + \zeta\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$ Taking $\epsilon$ sufficiently small and combining (\[cutoffErrorHandled\]) with (\[addedLargeR2\]), (\[redshiftAdded\]) and (\[boundNullInf\]) implies $$\begin{aligned}
\label{cutoffsDealtWith}
&b\int_{\mathcal{H}^+(0,\infty)}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\mathcal{H}^+} + b\int_{\mathcal{I}^+_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\mathcal{I}^+}\\
\nonumber &\qquad+b\int_0^{\infty}\int_{\Sigma_{\tau}}\left(\left|\partial_{r^*}\psi\right|^2r^{-1-\delta} + \left|\psi\right|^2r^{-3-\delta} +
\zeta \left|T\psi\right|^2r^{-1-\delta} + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\psi\right|^2r^{-1}\right)\, d\tau \\
\nonumber &\le B\left(\delta\right)\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega.\end{aligned}$$
The non-stationary bounded frequency horizon term {#whitinghere}
-------------------------------------------------
Finally, we come to the term $\int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega$. Since we do not have a small parameter, we cannot hope to absorb this error term into the left hand side of (\[cutoffsDealtWith\]); however, this term has already been dealt with in the context of the quantitative mode stability work [@shlapRot]:
\[propShlapRot\]Let $\psi$ be a future-integrable solution to (\[WAVE\]). Define $u$ by (\[uDef\]) with $\Psi = \psi_{\text{\Rightscissors}}$. Then $$\int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.$$
This follows immediately from Theorem 1.9 of [@shlapRot] if we replace $\Sigma_0$ with a hyperboloidal hypersurface $\tilde\Sigma_0$. For any $1 \ll R$ one can easily find a hyperboloidal hypersurface $\tilde\Sigma_0$ which agrees with $\Sigma_0$ on $\{r \leq R\}$ and which lies to the future of $\Sigma_0$. If we choose $R$ large enough so that $T$ is timelike in the region $\{r \geq R\}$, then a $\mathbf{J}^T_{\mu}$ energy estimate will immediately imply $$\begin{aligned}
\int_{\omega_{\rm low} \leq \left|\omega\right| \leq \omega_{\rm high}}\sum_{\{m\ell : \Lambda \leq \epsilon_{\rm width}^{-1}\omega_{\rm high}^2\}}\left|u\left(-\infty\right)\right|^2\, d\omega &\leq B\int_{\tilde\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\tilde \Sigma_0}\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
We observe that the appeal to [@shlapRot] is not necessary in the case of $a \ll M$ or if $\psi$ is only supported on sufficiently high azimuthal frequencies.
Combining (\[cutoffsDealtWith\]) with Proposition \[propShlapRot\] finishes the proof of Proposition \[closedILED\].
An inhomogeneous estimate {#ILEDinhomo}
-------------------------
In Sections \[higher\] and \[continuityargsec\] we will need to consider future-integrable solutions $\Psi$ to the *inhomogeneous* wave equation $\Box_{g_{a,M}}\Psi = F$.
Let us first generalise the definition of future-integrability to apply to general smooth $\Psi$.
With cutoff $\xi\left(\tau\right)$ as in Section \[futIntSec\], we say that $\Psi:\mathcal{R}_0\to\mathbb R$ is future-integrable if $\xi\Psi$ satisfies Definitions \[sufficient\] and \[sufficient2\].
\[closedILEDinhomo\]Let $\Psi$ be a future integrable solution of the inhomogeneous wave equation $\Box_{g_{a,M}}\Psi = F$. Then, for every $\delta > 0$ and $\epsilon > 0$, $$\begin{aligned}
\label{inhomo1}
&\int_{\mathcal{H}^+(0,\infty)}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\mathcal{H}^+}+\int_{\mathcal{I}^+}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\mathcal{I}^+}\\
\nonumber&\qquad+
\int_0^{\infty}\int_{\Sigma_{\tau}}\left(\left|\tilde{Z}^*\Psi\right|^2r^{-1-\delta} + \left|\Psi\right|^2r^{-3-\delta} +
\zeta \left|T\Psi\right|^2r^{-1-\delta} + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\Psi\right|^2r^{-1}\right)\, d\tau
\\ \nonumber &\leq
B\left(\delta\right)\left(\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\Sigma_0} + \int_{\Sigma_0}\left|\Psi\right|^2 + \int_0^{\infty}\int_{\Sigma_{\tau}}\left[\epsilon^{-1}r^{1+\delta}\left|F\right|^2 + \epsilon\left(1-\zeta\right)\left(\left|T\Psi\right|^2+\left|\Phi\Psi\right|^2\right)\right]\right).\end{aligned}$$ If $F$ is supported in the region $\{r \leq R\}$, then one may drop the $\int_{\Sigma_0}\left|\Psi\right|^2$ term: $$\begin{aligned}
\label{inhomo2}
{\rm L.H.S.\ of\ }(\ref{inhomo1})\,\, \le B\left(\delta,R\right)\left(\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\Sigma_0}+ \int_0^{\infty}\int_{\Sigma_{\tau}}\left[\epsilon^{-1} r^{1+\delta}\left|F\right|^2 + \epsilon\left(1-\zeta\right)\left(\left|T\Psi\right|^2+\left|\Phi\Psi\right|^2\right)\right]
\right).\end{aligned}$$ If $F$ is supported in the region $\{r \geq 3M + s_+\}$, then one may drop the $\left(1-\zeta\right)\left(\left|T\Psi\right|^2+\left|\Phi\Psi\right|^2\right)$ term: $$\begin{aligned}
\label{inhomo3}
{\rm L.H.S.\ of\ }(\ref{inhomo1})\,\, \le B\left(\delta \right)\left(\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\Sigma_0}+ \int_0^{\infty}\int_{\Sigma_{\tau}}r^{1+\delta}\left|F\right|^2
+\int_{\Sigma_0}|\Psi|^2\right).\end{aligned}$$ If $F$ is supported in the region $\{R\ge r \geq 3M + s_+\}$, then one may drop both: $$\begin{aligned}
\label{inhomo4}
{\rm L.H.S.\ of\ }(\ref{inhomo1})\,\, \le B\left(\delta,R\right)\left(\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\Sigma_0}+ \int_0^{\infty}\int_{\Sigma_{\tau}}r^{1+\delta}\left|F\right|^2
\right).\end{aligned}$$
Repeating the proof of Proposition \[closedILED\] *mutatis mutandis* yields $$\begin{aligned}
\int_{\mathcal{H}^+(0,\infty)}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\mathcal{H}^+}&+\int_{\mathcal{I}^+}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\mathcal{I}^+}
\\
&+
\int_0^{\infty}\int_{\Sigma_{\tau}}\left(\left|\tilde{Z}^*\Psi\right|^2r^{-1-\delta} + \left|\Psi\right|^2r^{-3-\delta} +
\zeta \left|T\Psi\right|^2r^{-1-\delta} + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\Psi\right|^2r^{-1}\right)\, d\tau \\
\le&
B\left(\delta\right)\left(\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\Sigma_0} + \int_0^{\infty}\int_{\Sigma_{\tau}}\left[\epsilon^{-1}r^{1+\delta}\left|F\right|^2 + \epsilon\left(1-\zeta\right)\left(\left|T\Psi\right|^2+\left|\Phi\Psi\right|^2\right)\right]\right)
\\ \nonumber &+ \int_0^1\int_{\Sigma_s}\mathbf{J}^N_{\mu}[\Psi_{\text{\Rightscissors}}]n^{\mu}_{\Sigma_s}\, ds.\end{aligned}$$ We cannot absorb the $\epsilon\left(1-\zeta\right)\left(\left|T\Psi\right|^2+\left|\Phi\Psi\right|^2\right)$ term into the left hand side because of the degeneration due to trapping. The final term on the right hand side arises due to the fact that $\Psi_{\text{\Rightscissors}}$ and $\Psi$ differ when $\dot\xi \neq 0$; since there are no weights in $r$, we cannot hope to absorb this term into the left hand side. However, an application of the fundamental theorem of calculus and Hardy inequalities easily allows us to finish the proof of (\[inhomo1\]).
In the case where $F$ is compactly supported in the region $\{r \leq R\}$, the proof of (\[closedILED\]) yields $$\begin{aligned}
\int_{\mathcal{H}^+(0,\infty)}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\mathcal{H}^+}
&+\int_{\mathcal{I}^+}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\mathcal{I}^+}\\
&\qquad+
\int_0^{\infty}\int_{\Sigma_{\tau}}\left(\left|\tilde{Z}^*\Psi\right|^2r^{-1-\delta} + \left|\Psi\right|^2r^{-3-\delta} +
\zeta \left|T\Psi\right|^2r^{-1-\delta} + \zeta\left|{\mbox{$\nabla \mkern-13mu /$\,}}\Psi\right|^2r^{-1}\right)\, d\tau \\
\le&
B\left(\delta\right)\left(\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\Sigma_0} + \int_0^{\infty}\int_{\Sigma_{\tau}}\left[\epsilon^{-1}r^{1+\delta}\left|F\right|^2 + \epsilon\left(1-\zeta\right)\left(\left|T\Psi\right|^2+\left|\Phi\Psi\right|^2\right)\right]\right)
\\ \nonumber &+ \int_0^1\int_{\Sigma_s\cap [r_+,R]}\mathbf{J}^N_{\mu}[\Psi_{\text{\Rightscissors}}]n^{\mu}_{\Sigma_s}\, ds.\end{aligned}$$ The proof of (\[inhomo2\]) follows from an application of Hardy inequalities and a finite in time energy estimate to the last term on the right hand side.
The proof of (\[inhomo3\]) and $(\ref{inhomo4})$ follow from the same reasoning as above *mutatis mutandis*.
After one has proved Theorem \[theResult\] it is possible to revisit the inhomogeneous problem and prove a sharper version of this proposition; however, we shall refrain from a systematic treatment of the inhomogeneous problem.
The higher order statement for future-integrable solutions {#higher}
==========================================================
Section \[summation\] has established the integrated decay statement (\[protasn1b\]) for the class of future-integrable solutions to the wave equation $(\ref{WAVE})$. Retaining this restriction, we will in this section upgrade this statement to the higher order (\[protasn1\]).
\[h.o.s.suff\] Let $M>0$, $0\le a_0<M$. Let $|a|\le a_0$ and let $\psi$ be a future integrable solution of $(\ref{WAVE})$ on $\mathcal{R}_0$. Then, for all $\delta>0$ and all integers $j\ge 1$, the following bound holds $$\begin{aligned}
\nonumber
&\int_{\mathcal{H}^+(0,\infty)}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 + \int_{\mathcal{I}^+}\sum_{1 \leq i \leq j-1}\mathbf{J}^N_{\mu}[N^i\psi]n^{\mu}_{\mathcal{I}^+} \\
\nonumber &\qquad +\int_{\mathcal{R}_0}
r^{-1-\delta}\zeta\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\\
\nonumber
&\qquad\qquad +r^{-1-\delta}\sum_{1\le i_1+i_2+i_3\le j-1}
\left(|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3+1}\psi|^2+|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(Z^*)^{i_3}\psi|^2\right)\\
\le& B\left(\de,j\right)\int_{\Sigma_0} \sum_{0\le i \le j-1}
\mathbf{J}^N_{\mu}[N^i\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
Elliptic estimates
------------------
Before turning to the proof of Proposition \[h.o.s.suff\], we will require a few standard elliptic estimates for solutions of the wave equation $(\ref{WAVE})$.
Throughout this section, $M>0$, $0\le a_0<M$, $|a|\le a_0$ will be fixed, and $\chi$ will be a cutoff which is identically $1$ on $[r_+,R_1]$ and identically $0$ on $[R_1+1,\infty)$ for a sufficiently large constant $R_1$ whose $r$-value will in particular lie outside the ergoregion $\mathcal{S}$, and $Y$ will be the red-shift commutation vector field from Section \[Nmult\].
Lastly, for the reader’s benefit we recall the following pointwise relation which follows immediately from the algebraic properties of the energy-momentum tensor: $$\mathbf{J}^N_{\mu}[\Psi]n^{\mu}_{\Sigma_{\tau}} \geq b\left(\left(T\Psi\right)^2 +
(\tilde{Z}^*\Psi)^2+ \left|{\mbox{$\nabla \mkern-13mu /$\,}}\Psi\right|^2\right).$$
All the lemmas below refer to solutions $\psi$ of the wave equation $(\ref{WAVE})$ as in the reduction of Section \[WPosed\].
\[Nelliptic\] For $\psi$ as above, we have $$\begin{aligned}
&\int_{\Sigma_{\tau}}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 \leq B\int_{\Sigma_{\tau}}\left(\mathbf{J}^N_{\mu}[N\psi] + \mathbf{J}^N_{\mu}[\psi]\right)n^{\mu}_{\Sigma_{\tau}}.\end{aligned}$$
This is standard: Let $\hat{\Sigma}_{\tau}$ be an extension of $\Sigma_{\tau}$ from $r \in [r_+,\infty)$ to $r \in [r_+-\epsilon,\infty)$. By a standard extension lemma, one may extend $\psi$ to a function $\hat\psi$ on $\hat{\Sigma}_{\tau}$ in such a way that $\left\vert\left\vert \Delta_{\hat{\Sigma}_{\tau}}\hat\psi\right\vert\right\vert_{L^2\left(\hat{\Sigma}_{\tau}\right)} \leq B\left\vert\left\vert \Delta_{\Sigma_{\tau}}\psi\right\vert\right\vert_{L^2\left(\Sigma_{\tau}\right)}$. The lemma then follows from a local elliptic estimate.
\[ellipticEstimates\] For $\psi$ as above, we have $$\begin{aligned}
&\int_{\Sigma_{\tau}}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 \le B\int_{\Sigma_{\tau}}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}} +\mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_{\tau}}+\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{\tau}} +\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$
This is standard: One uses elliptic estimates on spheres near the horizon and an elliptic estimate on $\Sigma_{\tau} \cap \{r \geq r_0\}$ away from the horizon. The key point is that $T$ and $\Phi$ span a timelike direction away from the horizon, and $Y$, $T$ and $\Phi$ span a timelike direction near the horizon.
\[horizonSphere\] For $\psi$ as above, we have $$\begin{aligned}
&\int_{\mathcal{H}^+(0,\infty)}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2
\\ \nonumber &\qquad \le B\int_{\mathcal{H}^+(0,\infty)}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\mathcal{H}^+} +\mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\mathcal{H}^+}+\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\mathcal{H}^+} +\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\mathcal{H}^+}\right).\end{aligned}$$
This follows from elliptic estimates on spheres.
One can, of course, localise Lemma \[ellipticEstimates\]:
\[ellipticEstimatesLocal\] For $\psi$ as above, then for any $R < \infty$, we have $$\begin{aligned}
&\int_{\Sigma_{\tau}\cap [r_+,R]}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2
\\ \nonumber &\qquad \le B\int_{\Sigma_{\tau}\cap [r_+,R+1]}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}} +\mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_{\tau}}+\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{\tau}} +\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$
The next four lemmas give control of the solution without including a $Y$-commuted energy on the right hand side.
\[ellipticOffHorizon\] For $\psi$ as above, then for any $r_0 > r_+$, $$\begin{aligned}
&\int_{\Sigma_{\tau} \cap \{r \geq r_0\}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{\tau}} \leq
B(r_0)\int_{\Sigma_{\tau}}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}} + \mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_{\tau}} +\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$
This follows from an elliptic estimate away from the horizon using the fact that the span of $T$ and $\Phi$ is timelike. The straightforward proof is omitted.
\[ellipticOffHorizonLocal\] For $\psi$ as above, then for any $r_+ < r_0 < r_1 < \infty$, $\delta > 0$, $$\begin{aligned}
&\int_{\Sigma_{\tau} \cap \{[r_0,r_1]\}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{\tau}} \leq
B(r_0,r_1,\delta)\int_{\Sigma_{\tau}\cap [r_0-\delta,r_1+\delta]}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}} + \mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_{\tau}} +\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$
This straightforward proof is omitted.
\[ellipticOffHorizonDecay\] For $\psi$ as above, then for any, $r_+ < r_0 < \infty$ and $\delta_1,\delta_2 > 0$, $$\begin{aligned}
&\int_{\Sigma_{\tau}\cap [r_0,\infty)}\sum_{1\le i_1+i_2+i_3\le 2}
r^{-1-\delta_1}|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2
\\ \nonumber &\qquad\le B(r_0,\delta_2)\int_{\Sigma_{\tau}\cap [r_0-\delta_2,\infty)}r^{-1-\delta_1}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}} +\mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_{\tau}} +\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$
This straightforward proof is omitted.
\[ellipticOffHorizonLargeDecay\] For $\psi$ as above, then for any $2M + 1 \leq r_0 < \infty$ and $\delta_1,\delta_2 > 0$,$$\begin{aligned}
&\int_{\Sigma_{\tau} \cap \{[r_0,\infty)\}}r^{-1-\delta_1}\mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_{\tau}}\leq B(\delta_2)\int_{\Sigma_{\tau}\cap [r_0-\delta_2,\infty)}\left(r^{-1-\delta_1}\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}} +r^{-1-\delta_1}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$
It suffices to remark that the region $[2M+1,\infty)$ lies outside the ergoregion (see (\[ergoregionS\])), and apply elliptic estimates as before.
The following lemma will be used in conjunction with red-shift estimate of Proposition \[ftrs\] and the commutation formula for $Y$ given in Proposition \[commuprop\].
\[thetaNearHorizon\] For $\psi$ as above, then for all $\epsilon > 0$, we may find a $r_0 > r_+$ depending on $\epsilon$ such that $$\begin{aligned}
&\int_{\Sigma_{\tau}\cap\{r \leq r_0\}}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^2\psi\right|^2 \leq
B\int_{\Sigma_{\tau}\cap [r_+,r_0)}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}}+ \mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_{\tau}} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right) + \epsilon\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{\tau}}.\end{aligned}$$
Since $Y$ is null on $\mathcal{H}^+$, on $\mathcal{H}^+$ there is no $Y^2$ term in the wave equation. In particular, the second derivative terms in the wave equation which contain a $Y$ derivative may be controlled by $\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}}$ and $\mathbf{J}^N_{\mu}[\Phi\psi]n^{\mu}_{\Sigma_{\tau}}$. Given these observations, the lemma easily follows from elliptic estimates on spheres.
We will also need some integrated in time estimates:
\[iledEllipticT\]Let $\psi$ be as above, and let $R < \infty$. Then $$\begin{aligned}
&\int_0^{\infty}\int_{\Sigma_{\tau}\cap [3M+1,R)}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} \leq
B\int_0^{\infty}\int_{\Sigma_{\tau}\cap [3M,R+1)}\left(\left|T\psi\right|^2 + \left|\partial_{r^*}\psi\right|^2+\left|\psi\right|^2\right) + B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
This is standard: One writes the wave equation as $$\begin{aligned}
&g^{tt}\partial_t^2\psi - \frac{4Mar}{\rho^2\Delta}\partial_{t,\phi}^2\psi = \frac{\Delta-a^2\sin^2\theta}{\Delta\rho^2\sin^2\theta}\partial_{\phi}^2\psi -
\frac{r^2+a^2}{\Delta\rho^2}\partial_{r^*}\left(\left(r^2+a^2\right)\partial_{r^*}\psi\right) - \frac{1}{\rho^{2}\sin\theta}\partial_{\theta}\left(\sin\theta\partial_{\theta}\psi\right),\end{aligned}$$ multiplies by a cutoff to $r \in [3M,R+1)$, multiplies by $\psi$, integrates by parts, remembers the comments concerning the volume form in Section \[usefulcomps\], and applies Hardy inequalities when appropriate.
\[iledEllipticTPhi\] For $\psi$ as above, then for any $r_0 > r_+$, $R < \infty$ and $\delta > 0$, $$\begin{aligned}
&\int_0^{\infty}\int_{\Sigma_{\tau}\cap [r_0+\delta,R-\delta]}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}
\\ \nonumber &\qquad \le B(r_0, \delta)\int_0^{\infty}\int_{\Sigma_{\tau}\cap [r_0,R])}\left(\left|T\psi\right|^2 +\left|\partial_{r^*}\psi\right|^2+\left|\Phi\psi\right|^2 + \left|\psi\right|^2\right) + B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
This is proven in the same fashion as Lemma \[iledEllipticT\].
Proof of Proposition \[h.o.s.suff\]
-----------------------------------
Now we will prove Proposition \[h.o.s.suff\].
Let $a_0$, $M$, $a$ and $\psi$ be as in the statement of the proposition. Let us be given moreover $\delta>0$. We will consider the case $j=2$. The case of $j \geq 3$ follows by induction in a straightforward fashion.
First, we commute the wave equation with $T$ and obtain $\Box_g\left(T\psi\right) = 0$. Since $T\psi$ is future-integrable, the integrated energy decay statement (\[protasn1b\]) holds for $T\psi$.
Next, we commute with $\chi\Phi$, where $\chi$ is a cutoff which is identically $1$ on $[r_+,R_1]$ and identically $0$ on $[R_1+1,\infty)$. We obtain $\Box_g\left(\chi\Phi\psi\right) = \left(\Box_g\chi\right)\Phi\psi + 2\nabla^{\mu}\chi\nabla_{\mu}\Phi\psi$. Now, Lemma \[ellipticOffHorizonLargeDecay\] implies $$\begin{aligned}
\label{errorPhi}
\int_0^{\infty}\int_{\Sigma_{\tau}}\left|\Box_g\left(\chi\Phi\psi\right)\right|^2 &\leq B\int_0^{\infty}\int_{\Sigma_{\tau}\cap [R_1,\infty)}r^{-1-\delta}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right)
\\ \nonumber &\leq B\int_{\Sigma_0}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_0} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\right).\end{aligned}$$
In the last inequality, we used that the integrated energy decay statement holds for $T\psi$. Now, statement (\[inhomo4\]) of Proposition \[closedILEDinhomo\] implies that the integrated energy decay statement holds for $\chi\Phi\psi$ as long as we add $B\int_{\Sigma_0}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_0} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\right)$ to the right hand side of the inequality.
Finally, we turn to commutation with $Y$. We recall Proposition \[commuprop\] which implies $$\label{goodSign}
\Box_g(Y\Psi)=\kappa_1 Y^2\Psi + \sum_{|{\bf m}|\le 2, m_4\le 1} c_{{\bf m}}
E_1^{m_1}E_2^{m_2}L^{m_3}Y^{m_4}\Psi$$ where $\kappa_1>0$ is proportional to the surface gravity. Next, for any $\tilde r \leq r_{\rm red}$, we apply the energy estimate associated to the red-shift vector field $N$, in between the hypersurfaces $\Sigma_0$ and $\Sigma_{\tau}$: $$\begin{aligned}
\label{redImplyHigher2}
&\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{\tau}} + \int_{\mathcal{H}^+(0,\tau)}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\mathcal{H}^+} + \int_{0}^{\tau}\int_{\Sigma_s\cap\{r \leq \tilde r\}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_s}\, ds \\
\nonumber&\qquad \le B\int_{0}^{\tau}\int_{\Sigma_s}\left(1_{r \in \left[\tilde r,\tilde r+\delta\right]}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{s}} + \mathcal{E}^N[Y\psi]\right)\, ds + \int_{\Sigma_{s_1}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{s_1}}.\end{aligned}$$
For any $\epsilon > 0$, we may choose $\tilde r$ close enough to $r_+$, $\delta$ small enough so that $\tilde r + 2\delta < 3M - s^-$ and use (\[goodSign\]), Lemma \[ellipticOffHorizonLocal\], Lemma \[thetaNearHorizon\], Lemma \[ellipticEstimatesLocal\] and the fact that $N = K+Y$, to show that $$\begin{aligned}
\label{errorControl2}
&\int_{\Sigma_s}\mathcal{E}^N[Y\psi]
\\ \nonumber &\,\,\leq \epsilon\int_{\Sigma_s\cap\{r\leq \tilde r\}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_s} + B\epsilon^{-1}\int_{\Sigma_s\cap [r_+,\tilde r+2\delta]}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_s} + \mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_s}+\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}\right)
\\ \nonumber &\,\,\leq \epsilon\int_{\Sigma_s\cap\{r\leq \tilde r\}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_s}+B\epsilon^{-1}\int_{\Sigma_0}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_0}+\mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_0} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\right).\end{aligned}$$
Combining (\[errorControl2\]) and (\[redImplyHigher2\]) implies $$\begin{aligned}
&\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{\tau}} + \int_{\mathcal{H}^+(0,\tau)}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\mathcal{H}^+}+\int_0^{\tau}\int_{\Sigma_s\cap\{r\leq \tilde r\}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_s}\, ds \\ \nonumber &\qquad \le B\int_{\Sigma_0}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_0}+\mathbf{J}^N_{\mu}[\chi\Phi\psi]n^{\mu}_{\Sigma_0} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\right).\end{aligned}$$
Now, the proof concludes with applications of Lemmas \[horizonSphere\], \[ellipticEstimates\], \[ellipticEstimatesLocal\], \[iledEllipticTPhi\], \[ellipticOffHorizonDecay\], \[Nelliptic\], \[iledEllipticT\] and (for the null infinity $\mathcal{I}^+$ term) straightforward $\mathbf{J}^T$ energy estimates in a large $r$ region.
The continuity argument {#continuityargsec}
=======================
In this section, we will prove
\[allIntegrable\] Let $M>0$ and $|a|<M$. All solutions $\psi$ to the wave equation (\[WAVE\]) on $\mathcal{R}_0$ as in the reduction of Section \[WPosed\] (i.e. arising from smooth, compactly supported initial data on $\Sigma_0$) are future-integrable.
The reduction to fixed azimuthal frequency {#reduxfixed}
------------------------------------------
We begin with the following easy but important Lemma.
\[azimuthal\]It suffices to prove Proposition \[allIntegrable\] for solutions $\psi$ to $(\ref{WAVE})$ assumed moreover to be supported on an arbitrary but fixed azimuthal frequency $m$.
Let $\psi$ be a solution to the wave equation arising from smooth, compactly supported initial data, and suppose we have established Proposition \[allIntegrable\] for solutions supported on any fixed azimuthal frequency. We may expand $\psi$ into its azimuthal modes: $\psi = \sum_{m \in \mathbb{Z}}\psi_m$. Since each $\psi_m$ is future-integrable, it follows by Proposition \[closedILED\] that the integrated energy decay statements (\[protasn1b\]) and (\[protasn1\]) hold for $\psi_m$. Orthogonality immediately implies that (\[protasn1b\]) and (\[protasn1\]) hold for $\psi$. Finally, we simply observe that the fundamental theorem of calculus implies that $$\begin{aligned}
&\sup_{r \in [r_+,A]}\int_0^{\infty}\int_{\mathbb{S}^2}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\sin\theta\, dt\, d\theta\, d\phi \\
& \,\,\,\leq
B\left(\int_{\mathcal{H}^+(0,\infty)}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 +\int_0^{\infty}\int_{\Sigma_s\cap [r_+,A]}\sum_{1\le i_1+i_2+i_3\le j+1}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\right).\end{aligned}$$
Lemma \[azimuthal\] thus implies that Proposition \[allIntegrable\] follows from the following proposition:
\[allIntegrablem\] Let $M>0$, $|a|<M$ and $m\in \mathbb Z$. Let $\psi$ be a solution to the wave equation as in the reduction of Section \[WPosed\] such that moreover, $\psi$ is supported only on the azimuthal frequency $m$. Then $\psi$ is sufficiently integrable.
The following Lemma will be very useful for the proof of Proposition \[allIntegrablem\].
\[cutoffILED\] Let $M$, $a$, $m$, and $\psi$ be as in the statement of Proposition \[allIntegrablem\]. Then, for every $\tau \geq 0$ and $\delta > 0$, $$\begin{aligned}
\int_{\mathcal{H}^+(0,\tau)}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\mathcal{H}^+} + &\int_0^{\tau}\int_{\Sigma_s}\Big(r^{-1}(1-\eta_{[\left(1+\sqrt{2}\right)M,3M+s^+]})(1-3M/r)^2 \big(|{\mbox{$\nabla \mkern-13mu /$\,}}\psi|^2+r^{-\delta}\left|T\psi\right|^2\big)\\
&\qquad\qquad+r^{-1-\delta}\left|\tilde Z^*\psi\right|^2+ r^{-3-\delta}\left|\psi\right|^2\Big)\\
\nonumber
&\leq B(\de,m)\left(\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$
One modifies the cutoff $\xi$ from Section \[cutoffSec\]; now let $\xi$ be identically $1$ in between $\Sigma_1$ and $\Sigma_{\tau-1}$ and identically $0$ to the past of $\Sigma_0$ and the future of $\Sigma_{\tau}$. Then, one may easily check that the arguments of Section \[summation\] will imply the lemma. Note that we can write $\eta_{[\left(1+\sqrt{2}\right)M,3M+s^+]}$ instead of $\eta_{[3M-s^-,3M+s^+]}$ because Lemma \[aziTrap\] tells us that for fixed $m$ and large $\Lambda$ the trapped set is contained in $[\left(1+\sqrt{2}\right)M,\infty)$.
Let us emphasise that we do assume that $\psi$ is future-integrable. This is why we must have the term $\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}$ on the right hand side.
As observed in Remark \[ergoTrap\], we see that (for a fixed azimuthal frequency) trapping and the ergoregion are non-overlapping! This will be extremely useful in what follows.
We will also need higher order versions of Lemma \[cutoffILED\].
\[cutoffILEDhigher\] Let $M$, $a$, $m$, and $\psi$ be as in the statement of Proposition \[allIntegrablem\]. Then, for every $\tau \geq 0$, $j \geq 1$ and $\delta > 0$, $$\begin{aligned}
&\int_{\mathcal{H}^+(0,\tau)}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 \\
&+\int_0^{\tau}\int_{\Sigma_s}
r^{-1-\delta}(1-\eta_{[\left(1+\sqrt{2}\right)M,3M+s^+]})(1-3M/r)^2\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\\
\nonumber
&\qquad\qquad\qquad+r^{-1-\delta}\sum_{1\le i_1+i_2+i_3\le j-1}
\left(|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3+1}\psi|^2+|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(Z^*)^{i_3}\psi|^2\right)\\
&\le B(\de,j,m)\left(\int_{\Sigma_0} \sum_{0\le i\le j-1}\mathbf{J}^N_{\mu}[N^i\psi]n^{\mu}_{\Sigma_0}+\int_{\Sigma_\tau} \sum_{0\le i \le j-1}
\mathbf{J}^N_{\mu}[N^i\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$
This follows from repeating the arguments of Section \[higher\].
We have the following easy corollary.
\[boundEquivSuff\] Let $M$, $a$, $m$, and $\psi$ be as in the statement of Proposition \[allIntegrablem\]. Then, $\psi$ is future-integrable if $$\label{aHigherBound}
\sup_{\tau \geq 0}\int_{\Sigma_\tau} \sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 < \infty\quad\forall\ j\geq 1.$$
As in the proof of Lemma \[azimuthal\] we need only observe that $$\begin{aligned}
&\sup_{r \in [r_+,A]}\int_0^{\infty}\int_{\mathbb{S}^2}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\sin\theta\, dt\, d\theta\, d\phi\\
& \leq
B(j)\left(\int_{\mathcal{H}^+(0,\infty)}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 +\int_0^{\infty}\int_{\Sigma_s\cap [r_+,A]}\sum_{1\le i_1+i_2+i_3\le j+1}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\right).\end{aligned}$$
The proof of Proposition \[allIntegrablem\] will be a continuity argument in the rotation parameter $a$ of the black hole. That is, fix $M > 0$, and define for each $m\in\mathbb Z$, the set $$\mathcal{A}_m := \{\left|a\right| \in [0,M) : \text{ the statement~(\ref{aHigherBound}) holds for } g = g_{a,M}\}.$$ We shall prove that $\mathcal{A}_m = [0,M)$ by showing that it is non-empty, open and closed. Proposition \[allIntegrablem\] then follows by Corollary \[boundEquivSuff\].
We note first
For all $m\in \mathbb Z$, the set $\mathcal{A}_m$ is non-empty.
When $a = 0$, it is well known that (\[aHigherBound\]) holds (even without the restriction to a fixed azimuthal frequency). One may find the (relatively short) argument in the lecture notes [@jnotes]. Thus $0\in \mathcal{A}_m$.
We now turn to openness.
Openness {#opennesssec}
--------
In this section, we will prove
\[open\] For all $m\in\mathbb Z$, the set $\mathcal{A}_m$ is open. That is, suppose $\mathring{a} \in \mathcal{A}_m$. Then there exists $\epsilon > 0$ such that $\left|a-\mathring{a}\right| < \epsilon$ implies $a \in \mathcal{A}_m$.
The proof proper will be given in Section \[interpolatingSection\] below. We begin with some preliminaries.
### Gaining derivatives
We start with a definition.
\[aVectorField\] Let $|a|<M$ and let $\epsilon_0 > 0$ be from Lemma \[horizonTimelike\]. Let $\alpha(r)$ be a function such that $V := T + \alpha(r)\Phi$ is a smooth vector field timelike in $\mathcal{R}$ which satisfies $$\begin{aligned}
V& = T + \frac{a}{2Mr_+}\Phi, \text{ when } r \in [r_+,r_+ + \epsilon_0/2],\\
V& = T + \frac{2Mar}{\left(r^2+a^2\right)^2}\Phi, \text{ when }r \in \left[r_++\epsilon_0,\frac{M\left(7+\sqrt{2}\right)}{4}\right],\\
V& = T, \text{ when } r \geq \frac{M\left(3+\sqrt{2}\right)}{2}.\end{aligned}$$
\[VkillTrap\] Note that $2M < \frac{M\left(3+\sqrt{2}\right)}{2} < M\left(1+\sqrt{2}\right)$. In particular, $V$ is Killing in the region where trapping occurs in Lemmas \[cutoffILED\] and \[cutoffILEDhigher\].
The following Lemma can be thought of as a derivative gaining converse to Lemma \[cutoffILED\].
\[prepLemm\] Let $|a|\le a_0<M$, let $m\in\mathbb Z$, and let $\psi$ be a solution the wave equation (\[WAVE\]) as in the reduction of Section \[WPosed\] which is furthermore supported on the fixed azimuthal frequency $m$. Then $$\begin{aligned}
\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}
&\leq B(m)\left(\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\Phi\psi\right|^2\, ds + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\right)\quad \forall\ \tau \geq 0\\
&\leq B(m)\left(\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\psi\right|^2\, ds + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\right)\quad \forall\ \tau \geq 0.\end{aligned}$$
We apply the energy identity associated to the vector field $V$: $$\begin{aligned}
\label{vEstimate}
\int_{\Sigma_{\tau}}\mathbf{J}^V_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} \leq B\int_0^{\tau}\int_{\Sigma_s}\left|\mathbf{K}^V[\psi]\right|\, ds + \int_{\Sigma_0}\mathbf{J}^V_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$ In view of the fact that $T$ and $\Phi$ are Killing vector fields, we have $$\mathbf{K}^V[\psi] = \mathbf{K}^{(\alpha\Phi)}[\psi] = 2\mathbf{T}(\nabla\alpha,\Phi)[\psi] = 2\frac{\Delta}{\rho^2}\frac{d\alpha}{dr}\mathbf{T}(Z,\Phi)[\psi] = 2\frac{\Delta}{\rho^2}\frac{d\alpha}{dr}\text{Re}(Z\psi\overline{\Phi\psi}).$$ Recall that $\frac{d\alpha}{dr}$ is supported away from the horizon, so that $Z$ is a regular vector field when the expression above is non-zero. We may conclude that $$\begin{aligned}
\label{spacetimeControl}
\left|\mathbf{K}^V[\psi]\right| \leq B1_{\text{supp}(\frac{d\alpha}{dr})}\left(\epsilon|\partial_r\psi|^2 + \epsilon^{-1}|\Phi\psi|^2\right).\end{aligned}$$ Lemma \[cutoffILED\] implies $$\begin{aligned}
\label{coupleILED}
\int_0^{\tau}\int_{\Sigma_s}1_{\text{supp}(\frac{d\alpha}{dr})}|\partial_r\psi|^2\, ds \leq B(m)\left(\int_{\Sigma_{\tau}}\mathbf{J}_{\mu}^N[\psi]n^{\mu}_{\Sigma_{\tau}} + \int_{\Sigma_0}\mathbf{J}_{\mu}^N[\psi]n^{\mu}_{\Sigma_0}\right).\end{aligned}$$ Combining $B(m)\epsilon$ times estimate (\[coupleILED\]) with estimates (\[vEstimate\]) and (\[spacetimeControl\]) implies $$\begin{aligned}
\label{almostThere}
\int_{\Sigma_{\tau}}\mathbf{J}^V_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} \leq B(m)\left(\epsilon^{-1}\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\Phi\psi\right|^2\, ds + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{1}} + \epsilon\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right).\end{aligned}$$ In order to finish the Lemma we apply the standard red-shift argument (see the lecture notes [@jnotes]). Set $$A := \sup_{0 \leq s\leq \tau}\epsilon^{-1}\int_0^s\int_{\Sigma_{s'} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\Phi\psi\right|^2\, ds' + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \epsilon\int_{\Sigma_s}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}.$$ For every $0 \leq s_1 < s_2 \leq \tau$ and $\tilde r$ sufficiently close to $r_+$, the red-shift estimate (\[ftrs\]) implies $$\begin{aligned}
\label{redImply}
&\int_{\Sigma_{s_2}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s_2}} + \int_{s_1}^{s_2}\int_{\Sigma_s\cap\{r \leq \tilde r\}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}\, ds \\ \nonumber&\qquad \le
B(m)\int_{s_1}^{s_2}\int_{\Sigma_s\cap\{\tilde r \leq r \leq \tilde r + 1\}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s}}\, ds + \int_{\Sigma_{s_1}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s_1}}.\end{aligned}$$ Now, we observe that in the region $\{\tilde r \leq r\}$ the quantities $\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s}}$ and $\mathbf{J}^V_{\mu}[\psi]n^{\mu}_{\Sigma_s}$ are comparable. Thus, adding $\int_{s_1}^{s_2}\int_{\Sigma_s}\mathbf{J}^V_{\mu}[\psi]n^{\mu}_{\Sigma_s}\, ds$ to both sides of (\[redImply\]) implies $$\begin{aligned}
\label{redImply2}
\int_{\Sigma_{s_2}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s_2}} + b(m)\int_{s_1}^{s_2}\int_{\Sigma_s}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}\, ds \leq B(m)\int_{s_1}^{s_2}\int_{\Sigma_s}\mathbf{J}^V_{\mu}[\psi]n^{\mu}_{\Sigma_{s}}\, ds + \int_{\Sigma_{s_1}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s_1}}.\end{aligned}$$ Now, estimate (\[almostThere\]) (with $\tau$ on the left hand side replaced by $s$) implies $$\begin{aligned}
\label{redImply3}
\int_{\Sigma_{s_2}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s_2}} + b(m)\int_{s_1}^{s_2}\int_{\Sigma_s}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}\, ds \leq B(m)A(s_2-s_1) + \int_{\Sigma_{s_1}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{s_1}}.\end{aligned}$$ Let $$f(s) := \int_{\Sigma_s}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}.$$ We may rewrite equation (\[redImply3\]) as $$f(s_2) + b\int_{s_1}^{s_2}f(s)\, ds \leq B(m)A(s_2-s_1) + f(s_1)\qquad\text{ for every }0 \leq s_1 < s_2 \leq \tau.$$ An easy argument shows that this implies $$f(s) \leq B(m)\left(A + f(0)\right).$$ Writing this out gives $$\begin{aligned}
&\sup_{0\leq s\leq \tau}\int_{\Sigma_s}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s} \leq B(m)\left(\epsilon^{-1}\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\Phi\psi\right|^2\, ds + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{0}} + \epsilon\sup_{0\leq s\leq \tau}\int_{\Sigma_s}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}\right).\end{aligned}$$ We conclude the proof by taking $\epsilon$ sufficiently small.
Observe that the proof does exploit the fact that the ergoregion and trapping are disjoint; indeed, even without the restriction to fixed $m$, we could have proved the first line of the proposition, with a constant $B$ not depending on $m$, simply by exploiting the fact that the $\partial_r$ derivative does not degenerate in the integrated local energy decay. Rather, the point is that for fixed $m$, the presence of the ergoregion is only a low-frequency obstruction to boundedness.
Note that the proof crucially uses that we can upgrade a degenerate energy boundedness statement to a non-degenerate energy boundedness statement without a full integrated local energy decay.
Next, we play Lemmas \[prepLemm\] and Lemma \[cutoffILEDhigher\] off each other. We end up being able to gain an *arbitrary* number of derivatives.
\[higherPrepLemm\] Let $|a|\le a_0<M$, let $m\in\mathbb Z$ and let $\psi$ be a solution the wave equation (\[WAVE\]) as in the reduction of Section \[WPosed\], which is furthermore supported on the fixed azimuthal frequency $m$. Then, for every $j \geq 1$, $$\begin{aligned}
&\int_{\Sigma_{\tau}}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2
\\ \nonumber &\le B(j,m)\left(\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\psi\right|^2\, ds + \int_{\Sigma_0}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\right)\quad \forall\ \tau \geq 0.\end{aligned}$$
We first consider the case $j= 2$. We begin by commuting the wave equation with $T$ and applying Lemma \[prepLemm\]. We obtain $$\begin{aligned}
\label{tCommute}
&\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{\tau}} \leq B(m)\left(\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|T\psi\right|^2\, ds + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{1}}\right).\end{aligned}$$ Now commute the wave equation with the red-shift commutation vector field $Y$. On the horizon $\mathcal{H}^+$ we will have $$\begin{aligned}
\label{yCommute}
\Box_g\left(Y\psi\right) = \kappa_1 Y^2\psi + \sum_{i+j+k \in [0,2], k\leq 1}c_{ijk}T^i\partial_{\theta}^jY^k\psi,\end{aligned}$$ where $\kappa_1 > 0$ is proportional to the surface gravity of $\mathcal{H}^+$.
Next, we apply Proposition \[ftrs\], the energy estimate associated to the red-shift multiplier $N$, to $\Psi=Y\psi$. For every $1 \leq s_1 < s_2 \leq \tau$, we obtain $$\begin{aligned}
\label{redImplyHigher}
&\int_{\Sigma_{s_2}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{s_2}} + \int_{s_1}^{s_2}\int_{\Sigma_s\cap\{r \leq \tilde r\}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_s}\, ds \\
\nonumber&\qquad\le B\int_{s_1}^{s_2}\int_{\Sigma_s}\left(1_{r \in \left[\tilde r,\tilde r+1\right]}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{s}} + \mathcal{E}^N[Y\psi]\right)\, ds + \int_{\Sigma_{s_1}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{s_1}}.\end{aligned}$$ For any $\epsilon > 0$ we may choose $\tilde r$ close enough to $r_+$, and use (\[yCommute\]), Lemma \[ellipticOffHorizon\], Lemma \[thetaNearHorizon\] and the fact that $N = K+Y$, to show that $$\begin{aligned}
\label{errorControl}
\int_{\Sigma_s}\mathcal{E}^N[Y\psi] \leq \epsilon\int_{\Sigma_s\cap\{r\leq \tilde r\}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_s} + B\epsilon^{-1}\int_{\Sigma_s}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_s} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}\right).\end{aligned}$$ Adding $\int_{s_1}^{s_2}\int_{\Sigma_s}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_s} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}\right)\, ds$ to both sides, using Lemma \[ellipticOffHorizon\] and using (\[errorControl\]) implies $$\begin{aligned}
&\int_{\Sigma_{s_2}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{s_2}} + b\int_{s_1}^{s_2}\int_{\Sigma_s}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_s}\, ds \\
\nonumber&\qquad \le B\int_{s_1}^{s_2}\int_{\Sigma_s}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{s}} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}\right)\, ds + \int_{\Sigma_{s_1}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{s_1}}.\end{aligned}$$ Now we use (\[tCommute\]), Lemma \[prepLemm\] and the same argument which occurs at the end of the proof of Lemma \[prepLemm\] to conclude $$\begin{aligned}
\label{redConsequence}
\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{\tau}} \leq &B(m)\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left(\left|T\psi\right|^2 + \left|\psi\right|^2\right)\, ds \\ \nonumber &+B(m)\int_{\Sigma_0}\left(\mathbf{J}^N_{\mu}[T\psi]n^{\mu}_{\Sigma_{1}} + \mathbf{J}^N_{\mu}[Y\psi]n^{\mu}_{\Sigma_{1}} + \mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{1}}\right).\end{aligned}$$ Next, Lemma \[ellipticEstimates\] allows us to combine (\[redConsequence\]) and (\[tCommute\]) to get $$\begin{aligned}
\label{veryClose}
&\int_{\Sigma_{\tau}}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2
\\ \nonumber &\le B(m)\left(\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left(\left|T\psi\right|^2 + \left|\psi\right|^2\right)\, ds + B(m)\int_{\Sigma_0}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\right).\end{aligned}$$ It remains to remove the spacetime integral of $\left|T\psi\right|^2$ from the right hand side; however, we observe the following immediate consequence of Lemmas \[cutoffILED\] and \[prepLemm\]: $$\begin{aligned}
\label{derivativeGain}
\int_0^{\tau}\int_{\Sigma_s\cap\left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|T\psi\right|^2 &\leq B(m)\left(\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\right)\\
\nonumber &\leq B(m)\left(\int_0^{\tau}\int_{\Sigma_s\cap\left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\psi\right|^2 + \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\right).\end{aligned}$$ Combining (\[veryClose\]) and (\[derivativeGain\]) completes the proof for $j = 2$.[^18]
The case for general $j \geq 2$ follows by induction in a straightforward fashion.
The following straightforward corollary will be useful in what follows.
\[combineLemmas\] Let $|a|\le a_0<M$, let $m\in\mathbb Z$, and let $\psi$ be a solution the wave equation (\[WAVE\]) as in the reduction of Section \[WPosed\] which is furthermore supported on the fixed azimuthal frequency $m$. Then, for all $\delta>0$, $j \geq 1$, $$\begin{aligned}
&\sup_{\tau' \leq \tau}\int_{\Sigma_{\tau'}}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 \\
&+\int_0^{\tau}\int_{\Sigma_{\tau}}\Big(\sum_{1\le i_1+i_2+i_3\le j-1}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 + \sum_{1\le i_1+i_2+i_3\le j-1}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3+1}\psi|^2
\\ &\nonumber \qquad\qquad\qquad+r^{-2-\delta}\left|\psi\right|^2 + 1_{\left[r_+,\left(1+\sqrt{2}\right)M\right]}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\Big)\cdot r^{-1-\delta} \\
&\le B(\de,j,m)\left(\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\psi\right|^2\, ds + \int_{\Sigma_0}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\right)\end{aligned}$$
This is a simple combination of Lemmas \[higherPrepLemm\] and \[cutoffILEDhigher\].
### An interpolating metric and the proof of Proposition \[open\] {#interpolatingSection}
We now turn to the proof of Proposition \[open\].
Recall that we have fixed $M>0$. Let us suppose $m\in \mathbb Z$ is fixed and $\mathring{a}\in \mathcal{A}_m$. Let us choose $a_0$ such that $|\mathring{a}|<a_0<M$. We will find an $\epsilon>0$ with $|\mathring{a}|+\epsilon \le a_0$ such that $$\label{1stcondonep}
|a-\mathring{a}|<\epsilon$$ implies $a\in \mathcal{A}_m$.
Let $a$ satisfy $(\ref{1stcondonep})$, for $\epsilon$ to be determined, and $\psi$ be a solution of the wave equation $(\ref{WAVE})$ on $g_{M,a}$, as in the reduction of Section \[WPosed\], such that moreover, $\psi$ is supported on the fixed azimuthal frequency $m$.
Recall that Corollary \[boundEquivSuff\] implies that solutions $\widetilde\psi$ to $\Box_{g_{\mathring{a},M}}\widetilde\psi = 0$ which are supported on a fixed azimuthal frequency are known to be future-integrable. In order to exploit this “black box” knowledge about $g_{\mathring{a},M}$ it is useful to introduce a metric $\tilde{g}_{\tau}$ which interpolates between $g_{\mathring{a},M}$ and $g_{a,M}$. Fortunately, we will *not* need to fine tune $\tilde{g}_{\tau}$.
\[interpolatingMetric\]Pick $\tau \geq 1$. Recalling that the hypersurfaces $\Sigma_s$ are independent of $a$, let $\chi_{\tau}$ be a cut-off which is $0$ in the past of $\Sigma_{\tau-\delta_0}$ and identically one in the future of $\Sigma_{\tau}$ for some sufficiently small $\delta_0 > 0$. We define the interpolating metric $\tilde{g}_{\tau}$ by $$\tilde{g}_{\tau} \doteq \chi_{\tau}g_{\mathring{a},M} + \left(1-\chi_{\tau}\right)g_{a,M}.$$
If $\epsilon$ in $(\ref{1stcondonep})$ is assumed sufficiently small, then $\tilde{g}_{\tau}$ defines a Lorentzian metric on $\mathcal{R}$.
\[stillKilling\]Note that it is easy to see that $\Phi$ is a Killing vector field for the metric $\tilde{g}_{\tau}$, and that for all $\tau\ge 0$, $\Sigma_\tau$ is a past Cauchy hypersurface for $\mathcal{R}_0\setminus \mathcal{R}_{(0,\tau)}$ with respect to $\tilde{g}_\tau$.
Corresponding to our interpolating metric, we will need an “interpolating” solution to the wave equation.
\[interpolatingSolution\] Let $\psi$ be our solution to $\Box_{g_{a,M}}\psi = 0$ defined above. We define the interpolating solution $\tilde{\psi}_{\tau}$ by solving $\Box_{\tilde{g}_{\tau}}\tilde{\psi}_{\tau} = 0$ with the same initial data as $\psi$ on $\Sigma_0$.
Of course, $\tilde{\psi}_{\tau}$ will exactly equal $\psi$ in the past of $\Sigma_{\tau-\delta_0}$, and in the future of $\Sigma_{\tau}$, $\tilde{\psi}_{\tau}$ is a solution to $\Box_{g_{\mathring{a},M}}\tilde{\psi}_{\tau} = 0$. Furthermore, since $\Phi$ is a Killing vector field for $\tilde{g}_{\tau}$, the interpolating solution $\tilde{\psi}_{\tau}$ will be supported on the same azimuthal frequency $m$ as the original solution $\psi$. Hence, by the assumption $\mathring{a}\in\mathcal{A}_m$ and Corollary \[boundEquivSuff\], it follows that $\tilde{\psi}_{\tau}$ is future-integrable with respect to $\mathring{a}$.
We write $$\begin{aligned}
\label{errorCont}
\Box_{g_{\mathring{a},M}}\tilde{\psi}_{\tau} = \left(\Box_{g_{\mathring{a},M}} - \Box_{\tilde{g}_{\tau}}\right)\tilde{\psi}_{\tau},\end{aligned}$$ and observe $$\begin{aligned}
r^{1+\delta}\left|\left(\Box_{g_{\mathring{a},M}} - \Box_{\tilde{g}_{\tau}}\right)\tilde{\psi}_{\tau}\right|^2 \leq B\left(\delta_0^{-1}\right)\left|a-\mathring{a}\right|^2r^{-2}\sum_{1 \leq i_1+i_2+i_3\leq 2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\tilde{\psi}_{\tau}\right|^2.\end{aligned}$$ In this statement, and in what follows, metric defined quantities (such as ${\mbox{$\nabla \mkern-13mu /$\,}}$ and $\mathbf{J}^N_{\mu}[\psi]n^{\mu}$) will refer to $g_{\mathring{a},M}$. Now we apply the $g_{\mathring{a},M}$ integrated local energy estimate to $\tilde{\psi}$.
Keeping in mind that (\[errorCont\]) is supported in the past of $\Sigma_{\tau}$, Proposition \[closedILEDinhomo\] implies $$\begin{aligned}
\label{ILEDinterpolate}
&\int_0^{\tau-\delta_0}\int_{\Sigma_s\cap \left\{r \leq M\left(1+\sqrt{2}\right)
\right\}}\left(\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s} + \left|\psi\right|^2\right)\, ds \\
\nonumber&\qquad\leq B(\delta_0,m)\left|a-\mathring{a}\right|\int_0^{\tau}\int_{\Sigma_s}\sum_{1 \leq i_1+i_2+i_3\leq 2}r^{-2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\tilde{\psi}_{\tau}\right|^2\, ds
\\ \nonumber &\qquad +B\left(\delta_0,m\right)\left|a-\mathring{a}\right|\int_{\tau}^{\infty}\int_{\Sigma_s\cap [r_+,\left(1+\sqrt{2}\right)M]}\left[\left|T\tilde\psi_{\tau}\right|^2 + \left|\tilde\psi_{\tau}\right|^2\right]+ B(m)\int_{\Sigma_0}\left[\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \left|\psi\right|^2\right].\end{aligned}$$ For $\delta_0$ sufficiently small (and then fixing the value of $\delta_0$), finite in time energy estimates (and an easy domain of dependence argument) imply $$\begin{aligned}
\label{moveDelta}
\int_0^{\tau}\int_{\Sigma_s\cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left(\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s} + \left|\psi\right|^2\right)\, ds \leq
B\int_0^{\tau-\delta_0}\int_{\Sigma_s\cap \left\{r \leq M\left(1+\sqrt{2}\right)\right\}}\left(\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s} + \left|\psi\right|^2\right)\, ds.\end{aligned}$$ Since $\tilde{\psi}_{\tau}$ is equal to $\psi$ along $\Sigma_{\tau-\delta_0}$, finite in time energy estimates for $\Box_{\tilde{g}_{\tau}}$ imply $$\begin{aligned}
\label{compareDelta}
\int_{\tau-\delta_0}^{\tau}\int_{\Sigma_s}\sum_{1 \leq i_1+i_2+i_3\leq 2}r^{-2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\tilde{\psi}_{\tau}\right|^2\, ds \leq B\int_{\Sigma_{\tau-\delta_0}}\sum_{1 \leq i_1+i_2+i_3\leq 2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi\right|^2.\end{aligned}$$ Similarly, $$\begin{aligned}
\label{compareDelta2}
\int_{\tau-\delta_0}^{\tau}\int_{\Sigma_s}\sum_{1 \leq i_1+i_2+i_3\leq 2}r^{-2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi\right|^2\, ds \leq B\int_{\Sigma_{\tau-\delta_0}}\sum_{1 \leq i_1+i_2+i_3\leq 2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi\right|^2.\end{aligned}$$ Finally, Proposition \[h.o.s.suff\], the fact that $\tilde\psi_{\tau}$ is future integrable, and finite in time energy inequalities imply $$\begin{aligned}
\label{compareDelta3}
\int_{\tau}^{\infty}\int_{\Sigma_s\cap [r_+,\left(1+\sqrt{2}\right)M]}\left[\left|T\tilde\psi_{\tau}\right|^2 + \left|\tilde\psi_{\tau}\right|^2\right] \leq B\int_{\Sigma_{\tau-\delta_0}}\sum_{1 \leq i_1+i_2+i_3\leq 2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi\right|^2.\end{aligned}$$ Combining (\[ILEDinterpolate\]), (\[moveDelta\]), (\[compareDelta\]), (\[compareDelta2\]) and (\[compareDelta3\]) gives $$\begin{aligned}
\label{continuityEst}
&\int_0^{\tau}\int_{\Sigma_s\cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}
\right\}}\left(\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s} + \left|\psi\right|^2\right)\, ds \\ \nonumber\le & B(m)\left|a-\mathring{a}\right|\left(\int_0^{\tau}\int_{\Sigma_s}\sum_{1 \leq i_1+i_2+i_3\leq 2}r^{-2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi\right|^2\, ds + \int_{\Sigma_{\tau-\delta_0}}\sum_{1 \leq i_1+i_2+i_3\leq 2}\left|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\tilde{\psi}_{\tau}\right|^2\right)
\\ \nonumber &\qquad+B(m)\int_{\Sigma_0}\left[\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + \left|\psi\right|^2\right].\end{aligned}$$ Now combine Corollary \[combineLemmas\] and (\[continuityEst\]): $$\begin{aligned}
&\sup_{\tau' \leq \tau}\int_{\Sigma_{\tau'}}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\\
&+\int_0^{\tau}\int_{\Sigma_{\tau}}\Big(\sum_{1\le i_1+i_2+i_3\le j-1}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 + \sum_{1\le i_1+i_2+i_3\le j-1}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3+1}\psi|^2
\\ &\nonumber\qquad\qquad\qquad + r^{-2-\delta}\left|\psi\right|^2 + 1_{\left[r_+,\left(1+\sqrt{2}\right)M\right]}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\Big)r^{-1-\delta} \\
\le &B(\de,j,m)\left(\int_0^{\tau}\int_{\Sigma_{s} \cap \left\{r \leq \frac{M\left(3+\sqrt{2}\right)}{2}\right\}}\left|\psi\right|^2\, ds + \int_{\Sigma_0}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\right) \\
\le & B(\de,j,m)\left|a-\mathring{a}\right|\left(\int_0^{\tau}\int_{\Sigma_s}r^{-2}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\, ds + \int_{\Sigma_{\tau-\delta_0}}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2\right)
\\ \nonumber &\qquad+B(\de,j,m)\int_{\Sigma_0}\sum_{0\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2.\end{aligned}$$ As long as $j \geq 3$, we may take $\epsilon$ in $(\ref{1stcondonep})$ sufficiently small, absorb the $\left|a-\mathring{a}\right|$ term on the left hand side and conclude $$\sup_{\tau' \leq \tau}\int_{\Sigma_{\tau'}}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 \leq B(j,m)\int_{\Sigma_0}\sum_{0\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 < \infty\quad \forall\ j \geq 3.$$ Lastly, we observe that the final inequality clearly remains true if we define ${\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}$ with respect to $g_{a,M}$ instead of $g_{\mathring{a},M}$.
Closedness {#closednesssec}
----------
To finish the proof, it remains to show
\[closed\]The set $\mathcal{A}_m$ is closed in $[0,M)$. That is, suppose we have a sequence $\{a_k\}_{k=1}^{\infty}$ with $a_k \in \mathcal{A}_m$ and $a_k \to a \in (-M,M)$. Then $a \in \mathcal{A}_m$.
Let $\psi$ be a solution the wave equation $\Box_{g_{a,M}}\psi = 0$ arising from smooth, compactly supported initial data which is furthermore supported on a fixed azimuthal frequency $m$.
We may choose $a_0<M$ such that $|a|<a_0$ and without loss of generality, we may assume that $|a_k|\le a_0$ for all $k$.
We define a sequence of functions $\psi_k$ by solving $\Box_{g_{a_k,M}}\psi_k = 0$ with the same initial data as $\psi$. Using the future-integrability of $\psi_k$, for every $j \geq 1$ we will have $$\begin{aligned}
\label{fromFutureInt}
&\int_0^{\infty}\int_{\Sigma_s}
r^{-1-\delta}(1-\eta_{[\left(1+\sqrt{2}\right)M,3M+s^+]})(1-3M/r)^2\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi_k|^2\\
\nonumber
&\qquad\qquad+r^{-1-\delta}\sum_{1\le i_1+i_2+i_3\le j-1}
\left(|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3+1}\psi_k|^2+|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(Z^*)^{i_3}\psi_k|^2\right)\\
\nonumber& \, \, \le B(\de,j,m)\int_{\Sigma_0} \sum_{0\le i\le j-1}\mathbf{J}^N_{\mu}[N^i\psi_k]n^{\mu}_{\Sigma_0}.\end{aligned}$$ Now, using the fact that the region $\{r \leq M\left(1+\sqrt{2}\right)\}$ contains the ergoregion $\mathcal{S}$, combining (\[fromFutureInt\]) and an $N$-based energy estimate yields $$\begin{aligned}
\label{futIntEnergyBound}
\sup_{\tau \geq 0}\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi_k]n^{\mu}_{\Sigma_{\tau}} &\leq B(m)\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi_k]n^{\mu}_{\Sigma_0} + B(m)\int_0^{\infty}\int_{\Sigma_s\cap\{r \leq M\left(1+\sqrt{2}\right)\}}\mathbf{J}^N_{\mu}[\psi_k]n^{\mu}_{\Sigma_s}
\\ \nonumber &\leq B(m)\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi_k]n^{\mu}_{\Sigma_0}.\end{aligned}$$ It remains to upgrade (\[futIntEnergyBound\]) to its higher order version in the (by now) standard fashion. First we commute with $T$ and obtain $$\begin{aligned}
\label{futIntEnergyBoundT}
\sup_{\tau \geq 0}\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[T\psi_k]n^{\mu}_{\Sigma_{\tau}} &\leq B(m)\int_{\Sigma_0}\mathbf{J}^N_{\mu}[T\psi_k]n^{\mu}_{\Sigma_0}.\end{aligned}$$ Next, we commute with $Y$ and follow the same argument as in the proof of Lemma \[higherPrepLemm\]. We obtain $$\begin{aligned}
\label{futIntEnergyBoundY}
&\sup_{\tau \geq 0}\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[Y\psi_k]n^{\mu}_{\Sigma_{\tau}} \leq B(m)\int_{\Sigma_0}\left(\mathbf{J}^N_{\mu}[\psi_k]n^{\mu}_{\Sigma_0} + \mathbf{J}^N_{\mu}[T\psi_k]n^{\mu}_{\Sigma_0} + \mathbf{J}^N_{\mu}[Y\psi_k]n^{\mu}_{\Sigma_0}\right).\end{aligned}$$ Just as in the proof of Lemma \[higherPrepLemm\], elliptic estimates imply $$\begin{aligned}
&\sup_{\tau \geq 0}\int_{\Sigma_{\tau}}\sum_{1\le i_1+i_2+i_3\le 2}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi_k|^2\leq B(m)\int_{\Sigma_0}\left(\mathbf{J}^N_{\mu}[\psi_k]n^{\mu}_{\Sigma_0} + \mathbf{J}^N_{\mu}[T\psi_k]n^{\mu}_{\Sigma_0} + \mathbf{J}^N_{\mu}[Y\psi_k]n^{\mu}_{\Sigma_0}\right).\end{aligned}$$ Finally, an easy induction argument will imply $$\begin{aligned}
&\sup_{\tau \geq 0}\int_{\Sigma_{\tau}}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi_k|^2 \leq B(j,m)\int_{\Sigma_0}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi_k|^2.\end{aligned}$$ Then, we conclude the proof by observing $$\begin{aligned}
\int_{\Sigma_{\tau}}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2 &= \lim_{k\to\infty}\int_{\Sigma_{\tau}}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi_k|^2
\\ \nonumber &\leq B(j,m)\int_{\Sigma_0}\sum_{1\le i_1+i_2+i_3\le j}
|{\mbox{$\nabla \mkern-13mu /$\,}}^{i_1}T^{i_2}(\tilde Z^*)^{i_3}\psi|^2.\end{aligned}$$ The first equality uses the well-posedness of the wave equation, the smooth dependence of $g_{a,M}$ on $a$ (see Lemma \[lemmaSmoothDependence\]) and the fact that $\psi_k$ and $\psi$ have the same initial data along $\Sigma_0$.
The precise integrated local energy decay statement {#precise}
===================================================
In this section we give will a more precise form of the integrated local energy decay statement. So as to produce a purely physical space estimate, we employed in the proof of Proposition \[closedILED\] a physical space cutoff $\zeta$ (see (\[degenerationfunc\])) in the integrated energy decay statement (\[protasn1b\]). It is clear from the statement of Theorem \[phaseSpaceILED\] in Section \[freqLocEst\] that this throws away information (cf. the discussion in Section \[trappingParam\]).
In order to succinctly state the microlocally precise form of integrated local energy decay, we introduce the following notation: For a sufficiently integrable function $\Psi$ on $\mathcal{R}$, we define $$\label{Ptrapnot}
\mathcal{P}_{\rm trap}\left[\Psi\right] \doteq \frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty\sum_{m\ell} \left|\zeta-(1-\zeta)r^{-1}r_{\rm trap}\right|e^{-i\omega t} \Psi^{(a\omega)}_{m\ell}(r)S_{m\ell}(a\omega,\cos\theta)e^{im\phi} d\omega,$$ where $r_{\rm trap}=r_{\rm trap}(\omega, m, \Lambda_{m\ell})$ is defined in Theorem \[phaseSpaceILED\]. Then, we have
\[preciseILED\] Let $0\le a_0<M$, $0\le a\le a_0$, and let $\psi$ be a solution of (\[WAVE\]) on $\mathcal{R}_0$ as in the reduction of Section \[closedILED\]. Then, $$\begin{aligned}
\label{notCrudeILED1}
b\int_0^{\infty}\int_{\Sigma_{\tau}}\left(\left|\tilde Z^*\psi\right|^2r^{-1-\delta} + \left|\psi\right|^2r^{-3-\delta} +
\left|T\mathcal{P}_{\rm trap}[\xi\psi]\right|^2r^{-1-\delta} + \left|{\mbox{$\nabla \mkern-13mu /$\,}}\mathcal{P}_{\rm trap}[\xi\psi]\right|^2r^{-1}\right)\, dr^*\, d\omega &\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$ where $r_{\rm trap}=r_{\rm trap}(\omega, m, \Lambda_{m\ell})$ is defined in Theorem \[phaseSpaceILED\] and $\xi$ is the cutoff from Section \[summation\].
One revisits the proof of Proposition \[closedILED\] and simply retains the nonnegative term on the left hand side of $(\ref{notCrudeILED1})$ instead of applying the physical space $\zeta$ and the inequality $(\ref{insteadofthis})$ .
Energy boundedness {#boundSuff}
==================
In this section, we establish the uniform boundedness of the energy flux through $\Sigma_\tau$ for solutions $\psi$ to the wave equation (\[WAVE\]):
\[sufficientIntBound\] Let $0\le a_0<M$, $|a|\le a_0$ and let $\psi$ be a solution of the wave equation (\[WAVE\]) on $\mathcal{R}_0$ as in the reduction of Section \[WPosed\]. Then $$\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}\quad \forall\ \tau \geq 0.$$
First, recall that the arguments of Sections \[summation\] and \[continuityargsec\] have shown that $\psi$ is future-integrable and satisfies the integrated decay statements (\[protasn1b\]) and (\[protasn1\]).
Let $\delta > 0$ be a fixed small parameter, $A_0$ be sufficiently close to $r_+$ and $A_1$ be sufficiently large. The proof proceeds in three steps where the cases $r \in [A_0+\delta,A_1-\delta]$, $r \in [r_+,A_0+\delta]$ and $r \in [A_1-\delta,\infty)$ are each dealt with. As one expects, the first region is the most difficult.
Boundedness of $\int_{\Sigma_{\tau}\cap [A_0+\delta,A_1-\delta]}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}$
----------------------------------------------------------------------------------------------------------------
It turns out to be convenient to extend the solution to the entire domain of outer communication $\mathcal{R}$ from $\mathcal{R}_0$.
### Extending the solution
The trace of $\psi$ and $N\psi$ along the hypersurface $\Sigma_0$ only suffice to determine $\psi$ in the future of $\Sigma_0$. However, an easy domain of dependence argument and finite in time energy estimates allow one to extend $\psi$ to the $\mathcal{R}$ in such a way as to guarantee $$\label{aRemark}
\int_{\hat{\Sigma}_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}} \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.$$ Here $\hat{\Sigma}_0$ denotes the image of $\Sigma_0$ under the Boyer-Lindquist coordinate defined map: $t\mapsto -t$.
### Integrated local energy decay for the extended solution
Since the Boyer-Lindquist defined map $t\mapsto -t$ and $a\mapsto -a$ is an isometry, Proposition \[closedILED\] remains true if one goes to the past instead of the future, i.e. if we replace all integrals $\int_0^{\infty}$ with $\int_{-\infty}^0$, and replace $\Sigma_{\tau}$ with $\hat{\Sigma}_{\tau}$. Keeping (\[aRemark\]) in mind, we conclude $$\begin{aligned}
\label{ILEDextended}
&\int_{-\infty}^{\infty}\int_{\Sigma_{\tau}\cap [A_0,A_1]}\left(\left|\partial_{r^*}\psi\right|^2 + r^{-2}\left|\psi\right|^2 + \zeta\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}\right)r^{-1-\delta}\, d\tau \leq B\left(A_0,A_1\right)\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
Unfortunately, this version of integrated local energy decay is too crude for our purposes, and we shall need to appeal to the version (\[notCrudeILED1\]) of integrated local energy decay.
Let $\chi_{[A_0,A_1]}$ be a bump function which is identically $1$ when $r \in [A_0+\delta,A_1-\delta]$ and $0$ when $r \not\in [A_0,A_1]$. We define $$\tilde\psi \doteq \chi_{[A_0,A_1]}\psi.$$ We will have $$\Box_{g_{a,M}}\tilde\psi = \tilde F \doteq \Box_g\chi_{[A_0,A_1]}\psi + 2\nabla^{\mu}\chi_{[A_0,A_1]}\nabla_{\mu}\psi.$$ Observe that $\tilde F$ has compact support in $r$, and $\left|\tilde F\right|^2 \leq B\left(\left|\psi\right|^2 + \left|\partial_{r^*}\psi\right|^2\right)$. In particular, $\tilde\psi$ is sufficiently integrable in the sense of Definition \[sufficient\] and outgoing in the sense of Definition \[sufficient2\]. We also have $$\begin{aligned}
\label{tildeFEst}
\int_{-\infty}^{\infty}\int_{\Sigma_{\tau}}\left|\tilde F\right|^2 \leq B\int_{-\infty}^{\infty}\int_{\Sigma_{\tau}\cap \{[A_0,A_0+\delta]\cup[A_1-\delta,A_1]\}}\left(\left|\psi\right|^2 + \left|\partial_{r^*}\psi\right|^2\right) \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
Now we apply Carter’s separation as defined in Section \[cartersupersection\] and obtain $$\tilde{u}'' + \left(\omega^2 - V\right)\tilde{u} = \tilde{H}.$$ Note that the compact $r$ support in $[A_0,A_1]$ of $\tilde{\psi}$ is inherited by $\tilde{u}$. We apply now Theorem \[phaseSpaceILED\]. In view of the support of $\tilde u$, it follows that the term $\left|\tilde u\left(-\infty\right)\right|^2$ vanishes. Furthermore, the right hand sides of all the frequency localised multiplier estimates $(\ref{fromPhaseSpace2})$ are $O\left(\tilde{H}\right)$, and hence are supported in $[A_0,A_0+\delta]\cup[A_1-\delta,A_1]$. Consequently, we can apply the (now trivial) arguments of Section \[summation\] to conclude the inhomogeneous version of (\[notCrudeILED1\]) $$\begin{aligned}
\label{notCrudeILED}
b\int_{-\infty}^{\infty}\int_{\Sigma_{\tau}}\left(\left|\partial_{r^*}\tilde\psi\right|^2 + \left|\tilde\psi\right|^2 + \mathbf{J}^N_{\mu}[\mathcal{P}_{\rm trap}\tilde\psi]n^{\mu}_{\Sigma_{\tau}}\right)&\leq \int_{-\infty}^{\infty}\int_{\Sigma_{\tau}\cap \{[A_0,A_0+\delta]\cup[A_1-\delta,A_1]\}}\left(\left|\psi\right|^2 + \left|\partial\psi\right|^2\right)
\\ \nonumber &\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0},\end{aligned}$$ where we recall that $\mathcal{P}_{\rm trap}$ is defined by (\[Ptrapnot\]).
### A decomposition
In order to work around the presence of $\mathcal{P}_{\rm trap}$ in (\[notCrudeILED\]), it will be useful to decompose $\tilde\psi$ is pieces, each of which experience trapping near a specific value of $r$. Recalling the definition of $r_{\rm trap}$ from Theorem \[phaseSpaceILED\] we make the following definition.
\[decomp\]Let $\epsilon > 0$ be a sufficiently small parameter to fixed later. We define $$\mathcal{C}_0 := \left\{\left(\omega,m,\Lambda\right) : r_{trap}=0
\right\},$$ $$\mathcal{C}_i:= \left\{\left(\omega,m,\Lambda\right) : r_{trap} \in \left[3M - s^- + \left(i-1\right)\epsilon,3M-s^- +i\epsilon\right)\right\}\forall \quad i = 1,\ldots,\lceil \epsilon^{-1}\left(s^++s^-\right)\rceil.$$
Observe that each value of $\left(\omega,m,\Lambda\right)$ lies in exactly one of the $\mathcal{C}_i$.
We define $\tilde{\psi}_i$ by a phase space multiplication of $\tilde{\psi}$ by $1_{\mathcal{C}_i}$, the indicator function of $\mathcal{C}_i$: $$\tilde{\psi}_i \doteq \frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty\sum_{m\ell} e^{-i\omega t} 1_{\mathcal{C}_i}\tilde{\psi}^{(a\omega)}_{m\ell}(r)S_{m\ell}(a\omega,\cos\theta)e^{im\phi} d\omega.$$
Note that it immediately follows from Plancherel (see Section \[separationSubsection\]) that each $\tilde{\psi}_i$ is sufficiently integrable, and we have $\Box_{g_{a,M}}\tilde{\psi}_i = \tilde{F}_i$ where $\tilde{F}_i$ is defined in the same fashion as $\tilde{\psi}_i$.
It will be useful to observe the following.
\[energyVanish\]For each $i = 0,\ldots,\lceil \epsilon^{-1}\left(s^++s^-\right)\rceil$ there exists a constant $C_i$ and a dyadic sequence $\{\tau^{(i)}_n\}_{n=1}^{\infty}$ such that $\tau^{(i)}_n \to -\infty$ as $n\to\infty$ and $$\int_{\Sigma_{\tau^{(i)}_n}}\mathbf{J}^N_{\mu}\left[\tilde\psi_i\right]n^{\mu}_{\Sigma_{\tau^{(i)}_n}} \leq \frac{C_i}{\tau^{(i)}_n}.$$
Since each $\tilde{\psi}_i$ is sufficiently integrable and compactly supported in $r$, we have $$\int_{-\infty}^{\infty}\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}\left[\tilde\psi_i\right]n^{\mu}_{\Sigma_{\tau}} < \infty.$$ The proof then concludes with a standard pigeonhole argument.
### Boundedness
Finally, we will establish boundedness of the energy of $\tilde\psi$.
\[boundedEnergyInterior\] Under the assumptions of Proposition \[sufficientIntBound\] and with the above notation we have $$\int_{\Sigma_{\tau}}\mathbf{J}^N_{\mu}\left[\tilde\psi\right]n^{\mu}_{\Sigma_{\tau}} \leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}\left[\psi\right]n^{\mu}_{\Sigma_0}\quad \forall\ \tau \in (-\infty,\infty).$$
Since $\tilde\psi = \sum_{i=0}^{\left\lceil \epsilon^{-1}\left(s_++s_-\right)\right\rceil}\tilde{\psi}_i$, it suffices to prove the proposition with $\tilde\psi$ replaced by $\tilde{\psi}_i$.
In Proposition \[timelikeVector\] we showed that the vector field $T + \frac{2Mar}{\left(r^2+a^2\right)^2}\Phi$ is timelike in $\mathcal{R}\setminus \mathcal{H}^+$. Given this and taking $\epsilon$ from Definition \[decomp\] sufficiently small (and then fixing $\epsilon$), it is easy to construct a $\varphi_\tau$-invariant timelike vector field $V_i$ on $\mathcal{R}$ which is Killing in the region $$r \in\left[3M - s^- + \left(i-1\right)\epsilon,3M-s^- + i\epsilon\right).$$ Now we apply the energy identity associated to $V_i$ in between the hypersurfaces $\Sigma_{\tau}$ and $\Sigma_{\tau^{(i)}_n}$. Since $\mathcal{P}_{trap}\tilde\psi_i=\tilde\psi_i$ , we obtain $$\begin{aligned}
\int_{\Sigma_{\tau}}\mathbf{J}^{V_i}_{\mu}\left[\tilde\psi_i\right]n^{\mu}_{\Sigma_{\tau}} &\leq B\int_{\tau^{(i)}_n}^{\tau}\int_{\Sigma_{s}\cap [3M-s^- + \left(i-1\right)\epsilon,3M-s^- + i\epsilon]^c}\mathbf{J}^N_{\mu}[\tilde\psi_i]n^{\mu}_{\Sigma_s} + \int_{\Sigma_{\tau^{(i)}_n}}\mathbf{J}^{V_i}_{\mu}\left[\tilde\psi_i\right]n^{\mu}_{\Sigma_{\tau}}
\\ \nonumber &\leq B\int_{-\infty}^{\infty}\int_{\Sigma_{s}\cap [3M-s^- + \left(i-1\right)\epsilon,3M-s^- + i\epsilon]^c}\mathbf{J}^N_{\mu}[\mathcal{P}_{\rm trap}\tilde\psi_i]n^{\mu}_{\Sigma_s} + \frac{BC_i}{\tau_n^{(i)}}
\\ \nonumber &\leq B\int_{-\infty}^{\infty}\int_{\Sigma_{s}}\mathbf{J}^N_{\mu}[\mathcal{P}_{trap}\tilde\psi]n^{\mu}_{\Sigma_s} + \frac{BC_i}{\tau_n^{(i)}}
\\ \nonumber &\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}\left[\psi\right]n^{\mu}_{\Sigma_0} + \frac{BC_i}{\tau_n^{(i)}},\end{aligned}$$ where we have used $(\ref{notCrudeILED})$ as well as Plancherel. It remains to take $n \to \infty$ and to observe (the trivial fact) that, in view of the support of $\tilde{\psi}_i$ and the $\phi_\tau$-invariance of $V_i$ we have $\mathbf{J}^{V_i}_{\mu}\left[\tilde\psi_i\right]n^{\mu}_{\Sigma_{\tau}} \sim \mathbf{J}^{N}_{\mu}\left[\tilde\psi_i\right]n^{\mu}_{\Sigma_{\tau}}$.
Boundedness of $\int_{\Sigma_{\tau}\cap [r_+,A_0+\delta]}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}$ and $\int_{\Sigma_{\tau}\cap [A_1-\delta,\infty)}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_{\tau}}$
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The following is a trivial consequence of the red-shift estimate (Proposition \[ftrs\]) and Proposition \[closedILED\]: $$\begin{aligned}
\label{boundedEnergyHorizon}
\int_{\Sigma_{\tau}\cap [r_+,A_0+\delta)}\mathbf{J}^N_{\mu}\left[\psi\right]n^{\mu}_{\Sigma_{\tau}} &\leq \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + B\int_0^{\tau}\int_{\Sigma_s\cap [A_0+\delta,A_0+2\delta]}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}
\\ \nonumber &\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
Similarly, we may consider the energy estimate associated to $\chi_{A_1-\delta}T$ where $\chi_{A_1-\delta}$ is a cutoff which is identically $1$ on $[A_1-\delta,\infty)$ and identically $0$ on $[r_+,A_1-2\delta]$. We obtain $$\begin{aligned}
\label{boundedEnergyInfinity}
\int_{\Sigma_{\tau}\cap [A_1-\delta,\infty)}\mathbf{J}^T_{\mu}\left[\psi\right]n^{\mu}_{\Sigma_{\tau}} &\leq \int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0} + B\int_0^{\tau}\int_{\Sigma_s\cap [A_1-2\delta,A_1-\delta]}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_s}
\\ \nonumber &\leq B\int_{\Sigma_0}\mathbf{J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}.\end{aligned}$$
Putting everything together and the higher order statement
----------------------------------------------------------
Combining Proposition \[boundedEnergyInterior\], (\[boundedEnergyHorizon\]) and (\[boundedEnergyInfinity\]) concludes the proof of Proposition \[sufficientIntBound\]. In view of Section \[theLogic\], this completes the proof of Theorem \[theResult\].
For Theorem \[h.o.s.\], we are left only with proving the higher order version of Proposition \[sufficientIntBound\]:
With the notation of Proposition \[sufficientIntBound\], for every $j \geq 1$ $$\int_{\Sigma_\tau } \sum_{0\le i \le j-1}{\bf J}^N_\mu[N^{i}\psi]n^\mu_{\Sigma_\tau}
\le B(j)\int_{\Sigma_0} {\sum_{0\le i \le j-1}
{\bf J}^N_\mu[N^{i}\psi]n^\mu_{\Sigma_0}}, \qquad
\forall\tau\ge 0.$$
We will be brief, since we have already seen multiple times how to upgrade lower order statements to higher order ones. As usual, we will only consider the case $j=2$ as the general case will follow by an easy induction argument.
First we commute $(\ref{WAVE})$ with $T$ and apply Proposition \[sufficientIntBound\]. We obtain $$\int_{\Sigma_\tau }{\bf J}^N_\mu[T\psi]n^\mu_{\Sigma_\tau}
\le B(j)\int_{\Sigma_0} {
{\bf J}^N_\mu[T\psi]n^\mu_{\Sigma_0}}, \qquad
\forall\tau\ge 0.$$
Next, letting $\chi$ be a cutoff which vanishes for large $r$, we commute with $\chi\Phi$. Using the integrated energy decay to the handle resulting error terms, we obtain $$\int_{\Sigma_\tau }{\bf J}^N_\mu[\chi\Phi\psi]n^\mu_{\Sigma_\tau}
\le B(j)\int_{\Sigma_0}\left( {
{\bf J}^N_\mu[N\psi]n^\mu_{\Sigma_0} + {\bf J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}}\right), \qquad
\forall\tau\ge 0.$$
Finally, we commute with the red-shift commutation vector field $Y$ and apply the argument from the proofs of Lemma \[higherPrepLemm\] and Proposition \[h.o.s.suff\] to establish $$\int_{\Sigma_\tau }{\bf J}^N_\mu[Y\psi]n^\mu_{\Sigma_\tau}
\le B(j)\int_{\Sigma_0}\left( {
{\bf J}^N_\mu[N\psi]n^\mu_{\Sigma_0} + {\bf J}^N_{\mu}[\psi]n^{\mu}_{\Sigma_0}}\right), \qquad
\forall\tau\ge 0.$$ The proof concludes via standard elliptic estimates (see the proofs of Lemma \[higherPrepLemm\] and Proposition \[h.o.s.suff\]).
In view of Section \[theLogic\], this obtains the remaining statement $(\ref{bndts1})$ of Theorem \[h.o.s.\]. The proof of both main theorems is thus complete.
[99]{} S. Alexakis, A. Ionescu and S. Klainerman *Uniqueness of smooth stationary black holes in vacuum: small perturbations of the Kerr spaces*, Comm. Math. Phys. [**299**]{} (2010), no. 1, 89–127.
S. Alinhac *Energy multipliers for perturbations of Schwarzschild metric* Comm. Math. Phys. [**288**]{} (2009), no. 1, 199–224.
L. Andersson and P. Blue *Hidden symmetries and decay for the wave equation on the Kerr spacetime*, arXiv:0908.2265.
L. Andersson and P. Blue *Uniform energy bound and asymptotics for the Maxwell field on a slowly rotating Kerr black hole exterior*, arXiv:13102664
N. Andersson and K. Glampedakis *A superradiance resonance cavity outside rapidly rotating black holes* Phys. Rev. Lett. [**84**]{} (2000), 4537–4540.
S. Aretakis *Decay of axisymmetric solutions of the wave equation on extreme Kerr backgrounds*, Journal of Functional Analysis, [**263**]{} (2012), no. 9, 2770–2831.
S. Aretakis *Horizon Instabilities of Extremal Black Holes*, to appear in ATMP, arXiv:1206.6598.
S. Aretakis *Stability and instability of extreme Reissner-Nordström black hole spacetimes for linear scalar perturbations I*, Comm. Math. Phys. [**307**]{} (2011), no. 1, 17–63.
S. Aretakis *Stability and instability of extreme Reissner-Nordström black hole spacetimes for linear scalar perturbations II*, Ann. Henri Poincaré [**12**]{} (2011), no. 8, 1491–1538.
A. Bachelot and A. Motet-Bachelot *Les résonances d’un trou noir de Schwarzschild* Ann. Inst. H. Poincaré Phys. Théor. [**59**]{} (1993), no. 1, 368.
P. Blue and A. Soffer *Semilinear wave equations on the Schwarzschild manifold. I. Local decay estimates*, Adv. Differential Equations [**8**]{} (2003), no. 5, 595–614.
P. Blue and A. Soffer *Errata for “Global existence and scattering for the nonlinear Schrodinger equation on Schwarzschild manifolds”, “Semilinear wave equations on the Schwarzschild manifold I: Local Decay Estimates”, and “The wave equation on the Schwarzschild metric II: Local Decay for the spin 2 Regge Wheeler equation”*, gr-qc/0608073, 6 pages.
P. Blue and J. Sterbenz *Uniform decay of local energy and the semi-linear wave equation on Schwarzschild space* Comm. Math. Phys. [**268**]{} (2006), no. 2, 481–504.
J. F. Bony and D. Häfner *Decay and non-decay of the local energy for the wave equation in the de Sitter-Schwarzschild metric* Comm. Math. Phys. [**282**]{} (2008), no. 3, 697–719.
B. Carter *Black hole equilibrium states*, in Black Holes (Les Houches Lectures), edited B. S. DeWitt and C. DeWitt (Gordon and Breach, New York, 1972).
B. Carter *Hamilton-Jacobi and Schrödinger separable solutions of Einstein’s equations* Comm. Math. Phys. [**10**]{} (1968), 280–310.
S. Chandrasekhar *The mathematical theory of black holes*, Oxford University Press, 1983.
D. Civin *Stability of subextremal Kerr–Newman exterior spacetimes for linear scalar perturbations*, preprint, 2014
D. Christodoulou *The action principle and partial differential equations*, Ann. Math. Studies No. 146, 1999
D. Christodoulou and S. Klainerman *The global nonlinear stability of the Minkowski space*, Princeton University Press, 1993
M. Dafermos *The interior of charged black holes and the problem of uniqueness in general relativity* Comm. Pure Appl. Math. [**58**]{} (2005), 0445–0504.
M. Dafermos and I. Rodnianski *A proof of Price’s law for the collapse of a self-gravitating scalar field*, Invent. Math. [**162**]{} (2005), 381–457.
M. Dafermos and I. Rodnianski *The redshift effect and radiation decay on black hole spacetimes* Comm. Pure Appl. Math. [**52**]{} (2009), 859–919.
M. Dafermos and I. Rodnianski *The wave equation on Schwarzschild-de Sitter spacetimes*, arXiv:0709.2766v1 \[gr-qc\].
M. Dafermos and I. Rodnianski *A note on energy currents and decay for the wave equation on a Schwarzschild background*, arXiv:0710.0171v1 \[math.AP\].
M. Dafermos and I. Rodnianski *A proof of the uniform boundedness of solutions to the wave equation on slowly rotating Kerr backgrounds*, Invent. Math. [**185**]{} (2011) , no. 3, 467–559.
M. Dafermos and I. Rodnianski *Lectures on black holes and linear waves*, Evolution equations, Clay Mathematics Proceedings, Vol. 17. Amer. Math. Soc., Providence, RI, 2013, pp. 97–205, arXiv:0811.0354 \[gr-qc\].
M. Dafermos and I. Rodnianski *A new physical-space approach to decay for the wave equation with applications to black hole spacetimes*, in XVIth International Congress on Mathematical Physics, P. Exner (ed.), World Scientific, London, 2009, arXiv:0910.4957v1 \[math.AP\].
M. Dafermos and I. Rodnianski *The black hole stability problem for linear scalar perturbations*, in Proceedings of the Twelfth Marcel Grossmann Meeting on General Relativity, T. Damour et al (ed.), World Scientific, Singapore, 2011, pp. 132–189, arXiv:1010.5137
M. Dafermos and I. Rodnianski *Decay for solutions of the wave equation on Kerr exterior spacetimes [I–II]{}: The cases $|a|\ll M$ or axisymmetry*, arXiv:1010.5132
M. Dafermos, G. Holzegel and I. Rodnianski *A scattering theory construction of dynamical black hole spacetimes*, arXiv:1306.5534
S. Dyatlov *Quasi-normal modes and exponential energy decay for the for the Kerr–de Sitter black hole* Comm. Math. Physics [**306**]{} (2011), 119–163.
S. Dyatlov *Exponential energy decay for Kerr–de Sitter black holes beyond event horizons*, Mathematical Research Letters [**18**]{} (2011), 1023–1035
S. Dyatlov *Asymptotics of linear waves and resonances with applications to black holes*, arXiv:1305.1723
F. Finster, N. Kamran, J. Smoller, S-T. Yau *Decay of solutions of the wave equation in Kerr geometry* Comm. Math. Phys. [**264**]{} (2006), 465–503.
F. Finster, N. Kamran, J. Smoller, S-T. Yau *Erratum: Decay of solutions of the wave equation in Kerr geometry* Comm. Math. Phys., online first.
A. Franzen, *The wave equation on black hole interiors*, Ph.D. Thesis, 2014
O. Gannot *Quasinormal modes for Schwarzschild-AdS black holes: exponential convergence to the real axis*, arXiv:1212:1907
S. W. Hawking and G. F. R. Ellis *The large scale structure of space-time* Cambridge Monographs on Mathematical Physics, No. 1. Cambridge University Press, London-New York, 1973.
G. Holzegel, *Ultimately Schwarzschildean spacetimes and the black hole stability problem*, arXiv:1010.3216.
G. Holzegel and J. Smulevici *Decay properties of Klein-Gordon fields on Kerr-AdS spacetimes* Comm. Pure Appl. Math. [**66**]{} (2013), no. 11, 1751–1802
G. Holzegel and J. Smulevici *Quasimodes and a Lower Bound on the Uniform Energy Decay Rate for Kerr-AdS Spacetimes*, arXiv:1303.5944
B. Kay and R. Wald *Linear stability of Schwarzschild under perturbations which are nonvanishing on the bifurcation $2$-sphere* Classical Quantum Gravity [**4**]{} (1987), no. 4, 893–898.
S. Klainerman *Uniform decay estimates and the Lorentz invariance of the classical wave equation* Comm. Pure Appl. Math. [**38**]{} (1985), 321–332
I. Laba and A. Soffer *Global existence and scattering for the nonlinear Schrödinger equation on Schwarzschild manifolds* Helv. Phys. Acta [**72**]{} (1999), no. 4, 272–294.
P. Laul, J. Metcalfe, S. Tikare and M. Tohaneanu *Localized energy estimates on Myers–Perry space-times*, arXiv:1401.0465
J. Lucietti and H. S. Reall *Gravitational instability of an extreme Kerr black hole* Phys. Rev. D [**86**]{} (2012), 104030
J. Luk *Improved decay for solutions to the linear wave equation on a Schwarzschild black hole*, Ann. Henri Poincaré [**11**]{} (2010), no. 5, 805–880.
J. Luk *A Vector Field Method Approach to Improved Decay for Solutions to the Wave Equation on a Slowly Rotating Kerr Black Hole*, Anal. PDE [**5**]{} (2012), no. 3, 553–625.
J. Luk *The null condition and global existence for nonlinear wave equations on slowly rotating Kerr spacetimes*, JEMS [**15**]{} (2013), no. 5, 1629–1700.
R. Melrose, A. Sá Barreto, A. Vasy *Asymptotics of solutions of the wave equation on de Sitter-Schwarzschild space*, arXiv:0811.2229.
C. S. Morawetz *Time decay for the nonlinear Klein-Gordon equations* Proc. Roy. Soc. Ser. A [**206**]{} (1968), 291–296.
G. Moschidis, forthcoming.
K. Murata, H. S. Reall and N. Tanahashi *What happens at the horizon(s) of an extreme black hole?* Class. Quantum Grav. [**30**]{} (2013) 235007
J. Ralston *Solutions of the wave equation with localized energy* Comm. Pure Appl. Math. [**22**]{} (1969), 807–823.
J. Sbierski *Characterisation of the Energy of Gaussian Beams on Lorentzian Manifolds - with Applications to Black Hole Spacetimes*, arXiv:1311.2477
V. Schlue *Decay of linear waves on higher dimensional Schwarzschild black holes*, Analysis & PDE [**6**]{}, (2013), 515–600
V. Schlue *Global results for linear waves on expanding Kerr and Schwarzschild de Sitter cosmologies*, arXiv:1207.6055v2
Y. Shlapentokh-Rothman *Quantitative Mode Stability for the Wave Equation on the Kerr Spacetime*, arXiv:1302.6902, to appear in Ann. Henri Poincaré.
Y. Shlapentokh-Rothman *Exponentially growing finite energy solutions for the Klein-Gordon equation on sub-extremal Kerr spacetimes*, arXiv:1302.3448, to appear in Commun. Math. Phys.
A. Starobinsky *Amplification of waves during reflection from a black hole* Soviet Physics JETP [**37**]{} (1973), 28–32.
D. Tataru *Local decay of waves on asymptotically flat stationary space-times*, Amer. J. Math. [**135**]{} (2013), no. 2, 361–401.
D. Tataru and M. Tohaneanu *Local energy estimate on Kerr black hole backgrounds*, IMRN (2011), no. 2, 248–292.
A. Vasy *Microlocal analysis of asymptotically hyperbolic and Kerr-de Sitter spaces, with an appendix by Semyon Dyatlov* Invent. Math. [**194**]{} (2013), 381–513.
B. Whiting *Mode stability of the Kerr black hole* J. Math. Phys. [**30**]{} (1989), 1301.
J. Wunsch and M. Zworski *Resolvent estimates for normally hyperbolic trapped sets* Ann. Inst. Henri Poincaré (A), [**12**]{} (2011), 1349–1385.
S. Yang *Global solutions to nonlinear wave equations in time dependent inhomogeneous media*, Arch. Ration. Mech. Anal. [**209**]{} (2013), no. 2, 683–728.
S. Yang *Global stability of solutions to nonlinear wave equations*, arXiv:1205.4216.
S. Yang *On the quasilinear wave equations in time dependent inhomogeneous media*, arXiv:1312.7246.
[^1]: Let us remark briefly that besides these extensions to $\Lambda\ne 0$, there are a host of other related problems one can also consider, including higher dimensional black holes [@schlue; @lauletal], other hyperbolic equations like Klein–Gordon (for which it is now proven [@shlapRot2] that there are *exponentially growing* solutions for all $|a|\ne 0$), Maxwell [@anblue2], linearised gravity and the nonlinear Einstein vacuum equations themselves (see [@kostakis2; @scattering]). We refer the reader to the many additional references in the first part of this series [@dr7], our survey [@stabi] and our lecture notes [@jnotes].
[^2]: A posteriori, the good structure of trapping in phase space for $|a|\ll M$ can be understood more conceptually, using the structural stability properties of normal hyperbolicity, provided the latter condition is checked for Schwarzschild; see [@WunschZworski]. Note however that these stability properties depend on strong regularity assumptions on the metric, whereas our original boundedness theorem [@dr6] only requires closeness to Schwarzschild in $C^1$. Thus one expects the domain of validity of [@dr6] to be strictly bigger than the class of spacetimes where decay results of the type of our main Theorem hold.
[^3]: Let us note that given the boundedness result of [@dr6], then one need not appeal again to closeness to Schwarzschild in the argument for integrated local energy decay; see our original proof in the lecture notes [@jnotes]. We have used it again in [@dr7] so as for [@dr7] to retrieve independently our previous boundedness result.
[^4]: In the borderline case when the time frequency $\omega$ vanishes, this is intimately related to the fact that there are no trapped null geodesics orthogonal to $\partial_t$. This latter observation turns out to be important in the study of black hole uniqueness (see [@alexakis]).
[^5]: Of course, in our original proof [@dr7], we proved those two statements together as we used the boundedness in our version of the continuity argument. In the new continuity argument presented here, this is not necessary.
[^6]: The precise relation to fixed coordinates as defined in [@dr7] is as follows: Let $\mathcal{P}=\{(x_1,x_2):0\le |x_1|<x_2\}$ denote the parameter space of all admissible subextremal $(a,M)$. We chose a smooth map $r:\mathcal{P}\times(0,\infty) \to (x_2+\sqrt{x_2^2-x_1^2}, \infty)$ such that $r|_{\{(x_1,x_2)\}\times (0,\infty)}$ is a diffeomorphism $(0,\infty)\to (x_2+\sqrt{x_2^2-x_1^2}, \infty)$ which moreover restricts to the identity map restricted to $\{(x_1,x_2)\}\times(3x_2,\infty)$. Note that with this definition, then for $r\ge3M$, $r(y^*)$ is independent of $a$.
[^7]: Recall that this vector field is significant as it will define the directional derivative that does not degenerate in the integrated decay estimate due to trapping. Note that in Boyer–Lindquist coordinates the fixed vector fields $T$ and $\Phi$ correspond to the coordinate vector fields $\partial_t$ and $\partial_\phi$.
[^8]: The latter can be understood in the sense that $$g:\mathcal{P}\times \mathcal{R} \to T^*\mathcal{R}\otimes T^*\mathcal{R}$$ is a smooth map.
[^9]: In this context, it is in fact more appropriate to refer to the energy flux to null infinity $\mathcal{I}^+$.
[^10]: We can alternatively take $N$ to be the vector field of Proposition \[specialises..\]; this is the vector field we shall use in the proof. For the statement of Theorem \[theResult\], the only important feature of $N$ is that it is $\phi_\tau$-invariant, strictly timelike and asymptotic to $T$ for large $r$. Whereas we could have used everywhere $n_{\Sigma_\tau}$ in the statement, we prefer to keep the distinct roles of $n_{\Sigma_\tau}$ and $N$ as this will be important when we replace $\Sigma_\tau$ with hyperboloidal hypersurfaces $\widetilde{\Sigma}_\tau$ in Section \[corollariessec\].
[^11]: Recall that in [@dr7] the Fourier transform was only applied to Schwartz functions in $t$. The added flexibility gained by working with square integrable functions in $t$ will be crucial for the continuity argument (see Section \[continuityargsec\]).
[^12]: In fact, taking $m$ to be integer-valued is of no significance in this analysis, but we will continue to write $m\in\mathbb Z$ to avoid confusion.
[^13]: Recall our conventions from Section \[genconstsec\] on the meaning of this term. This smallness constraint indeed degenerates as $a_0\to M$.
[^14]: In general, $f$ will be bounded and $C^2$, $h$ will be bounded, $C^1$ and piecewise $C^2$ and $y$ will be bounded, $C^0$ and piecewise $C^1$.
[^15]: For better or for worse, we follow here the notation we instituted in the first parts of this series [@dr7]. As this notation proved somewhat unpopular, we suggest that readers who dislike archaic Greek simply substitute ${\text Q}^y$, ${\text Q}^h$ for both $\text {\fontencoding{LGR}\selectfont \koppa}^y$ and $\text {\fontencoding{LGR}\selectfont \Coppa}^h$, as we shall consistently use functions named $f$, $h$ and $y$, according to whether we mean ${\text Q}^f$, $\text {\fontencoding{LGR}\selectfont \Coppa}^h$ or $\text {\fontencoding{LGR}\selectfont \koppa}^y$. Note that in our survey [@stabi], we used the notation $Q^f_0={\text Q}^f$, $Q^h_1=\text {\fontencoding{LGR}\selectfont \Coppa}^h$, $Q^y_2=\text {\fontencoding{LGR}\selectfont \koppa}^y$.
[^16]: Recall that $\text{Q}^f$ is itself the combination $\text {\fontencoding{LGR}\selectfont \Coppa}^{h}[u]+
\text {\fontencoding{LGR}\selectfont \koppa}^{y}[u]$, with $y=f$ and $h=f'$, but sufficiently important to deserve its own name!
[^17]: For the moment we are suppressing the fact that this estimate may be insufficiently strong if $0 \leq \omega\left(\omega-\upomega_+m\right) \ll \left(\omega-\upomega_+m\right)^2$. See Section \[boundaryStillaProblem\].
[^18]: Observe that this final trick relies on the fact that trapping and the ergoregion are disjoint in physical space when the azimuthal frequency is fixed.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We provide an elementary proof of a result by V.P. Fonf and C. Zanco on point-finite coverings of separable Hilbert spaces. Indeed, by using a variation of the famous argument introduced by J. Lindenstrauss and R.R. Phelps [@LP] to prove that the unit ball of a reflexive infinite-dimensional Banach space has uncountably many extreme points, we prove the following result.
[ *Let $X$ be an infinite-dimensional Hilbert space satisfying $\mathrm{dens}(X)<2^{\aleph_0}$, then $X$ does not admit point-finite coverings by open or closed balls, each of positive radius.* ]{}
In the second part of the paper, we follow the argument introduced by V.P. Fonf, M. Levin, and C. Zanco in [@FonfLevZan14] to prove that the previous result holds also in infinite-dimensional Banach spaces that are both uniformly rotund and uniformly smooth.
address: 'Dipartimento di Matematica per le Scienze economiche, finanziarie ed attuariali, Università Cattolica del Sacro Cuore, 20123 Milano,Italy'
author:
- Carlo Alberto De Bernardi
title: 'A note on point-finite coverings by balls'
---
introduction
============
A family of subsets of a real normed space $X$ is called a [*covering*]{} if the union of all its members coincides with $X$. A covering of $X$ is [*point-finite*]{} if each point of $X$ is contained in at most finitely many members of the covering.
The problem concerning existence of point-finite coverings of infinite-dimensional normed spaces by balls was considered for the first time in the paper [@Klee1] in which V. Klee asked the following question.
\[Klee’sProblem\] Let $\Gamma$ be a cardinal such that $|\Gamma|\geq\aleph_0$, does $\ell_1(\Gamma)$ (respectively $\ell_p(\Gamma)$ for $1<p<\infty$) admit a locally finite (respectively point-finite) covering by closed balls or open balls, each of positive radius?
The question above was motivated by the results, contained in the paper itself, implying existence of a covering of $\ell_1(\Gamma)$ by pairwise disjoint closed balls of radius $1$, whenever $\Gamma$ is a suitable uncountable set. In [@FZ06], V.P. Fonf and C. Zanco generalized Corson’s theorem (see Theorem \[corson\] below) by proving that [*if a Banach space $X$ contains an infinite-dimensional closed subspace non-containing $c_0$ then $X$ does not admit any locally finite covering by bounded closed convex bodies*]{}. This completely solved the problem concerning locally finite coverings by balls of $\ell_1(\Gamma)$.
More recently, V.P. Fonf and C. Zanco [@FZHilbert] proved that [*the infinite-dimensional separable Hilbert space does not admit point-finite coverings by closed balls of positive radius*]{}. Then V.P. Fonf, M. Levin and C. Zanco [@FonfLevZan14] extended the result above to separable spaces that are both uniformly smooth and uniformly rotund. However, Klee’s problem about point-finite coverings by balls of $\ell_p(\Gamma)$ spaces ($1<p<\infty$) remained open in the non-separable case.
The proof of the result by V.P. Fonf and C. Zanco, contained in [@FZHilbert], is based on the following ingredients:
1. [@FZHilbert Proposition 2.1], a result excluding existence of certain point-finite families of slices of the unit ball in separable Banach space;
2. [@FZHilbert Theorem 3.1], a characterization of separable isomorphically polyhedral Banach spaces via existence of point-finite countable coverings by slices of the unit sphere;
3. the fact that the intersection among two distinct spheres in any Hilbert space lies in some hyperplane. Indeed, this is a 3-dimensional characterization of inner product spaces [@Ami (15.17)];
4. the fact that no infinite-dimensional dual (and in particular reflexive) Banach space is polyhedral [@Lindenstrauss].
The aim of the present paper is to provide a direct and quite elementary proof of the main result contained in [@FZHilbert] and to present an improvement of the result contained in [@FonfLevZan14], concerning point-finite coverings by balls of Banach spaces that are both uniformly smooth and uniformly rotund. Let us start by describing the result contained in Section \[section-point-finite\]. Our Proposition \[uniformboundedness\] is a restatement of [@FZHilbert Proposition 2.1], the elementary alternative proof presented in our paper is an immediate application of the uniform boundedness principle and it works also in the non-separable case. Theorem \[teo:FollowingLindenstraussPhelps\], excludes existence of certain point-finite families of open or closed slices of the unit ball in reflexive Banach spaces, and it is a variation of the famous argument introduced by J. Lindenstrauss and R.R. Phelps [@LP] to prove that the unit ball of a reflexive infinite-dimensional Banach space has uncountably many extreme points. Theorem \[teo:FollowingLindenstraussPhelps\], combined with (iii), allows us to obtain the following slight improvement of [@FZHilbert Corollary 3.3].
*Let $X$ be an infinite-dimensional Hilbert space.*
1. If the density character of $X$ satisfies $\mathrm{dens}(X)<2^{\aleph_0}$ then it does not admit point-finite coverings by open or closed balls, each of positive radius.
2. $X$ does not admit point-finite coverings by open balls.
Finally, in Section \[section:appendix\], we observe that, following the argument introduced in [@FonfLevZan14], it is possible to extend this latter result to Banach spaces that are both uniformly smooth and uniformly rotund. The new ingredients in our proof are Lemma \[lemma:interiorconvexhull\], that allows us to deal with open and closed balls at the same time, and an easy separable reduction argument used in Theorem \[teo:unifrotundsmooth\]. In particular, our results solve in negative Klee’s problem for point-finite coverings by open balls of $\ell_p(\Gamma)$ spaces ($1<p<\infty$).
Point-finite coverings by slices and balls in Hilbert spaces {#section-point-finite}
============================================================
Throughout the paper, we consider only nontrivial real normed spaces. If $X$ is a normed space then $X^*$ is its dual Banach space. We denote by $B_X$, $U_X$, and $S_X$ the closed unit ball, the open unit ball, and the unit sphere of $X$, respectively. We denote by $U(x,{\varepsilon})$ the open ball with radius ${\varepsilon}> 0$ and center $x$. We denote by $B(x,{\varepsilon})$ the closed ball with radius ${\varepsilon}\geq 0$ and center $x$; in the case ${\varepsilon}=0$, $B(x,{\varepsilon})$ is the [*degenerate ball*]{} containing only the point $x$. In general, by a [*ball*]{} in $X$ we mean a closed ball of non-negative radius or an open ball of positive radius in $X$. For $x,y\in X$, $[x,y]$ denotes the closed segment in $X$ with endpoints $x$ and $y$, and $(x,y)=[x,y]\setminus\{x,y\}$ is the corresponding “open” segment. A set $B\subset X$ will be called a [*body*]{} if it is closed, convex and has nonempty interior. A body is called [*rotund*]{} if its boundary does not contain nontrivial segments. Other notation is standard, and various topological notions refer to the norm topology of $X$, if not specified otherwise.
Let $\mathcal{F}$ be a family of nonempty sets in a normed space $X$. By $\bigcup \mathcal F$ we mean the union of all members of $\mathcal F$. A point $x\in X$ is a [*regular point*]{} for $\mathcal F$ if it has a neighbourhood that meets at most finitely many members of $\mathcal F$. Points that are not regular are called [*singular*]{}. Notice that the set of singular points is a closed set.
\[D:point-locally-finite\] The family $\mathcal F$ is called:
1. [*point-finite*]{} if each $x\in X$ is contained in at most finitely many members of $\mathcal F$;
2. [*locally finite*]{} if each $x\in X$ is a regular point for $\mathcal F$.
A [*minimal covering*]{} is a covering whose no proper subfamily is a covering. A standard application of Zorn’s lemma shows that [*every point-finite covering contains a minimal subcovering.*]{}
In the sequel, we say that $\mathcal F$ is [*a family of open or closed balls*]{} of $X$ if each element of $\mathcal F$ is an open ball (of positive radius) or a closed ball of non-negative radius (i.e., if not differently stated, we admit that $\mathcal F$ contains also degenerate balls).
Let us recall the following famous theorem by H.H. Corson [@Cor61].
\[corson\] Let $\mathcal F$ be a covering of a reflexive infinite-dimensional Banach space by bounded convex sets. Then $\mathcal F$ is not locally finite.
In what follows, we shall use several times the following fact that immediately follows by [@DESOVESTAR Lemma 2.2]. Let us recall that, if $T$ is a topological vector space, $\mathrm{dens}(T)$ denotes its density character (i.e., the smallest cardinality of a dense subset of $T$).
\[density\] Let $T$ be a topological space and let ${\mathcal{B}}$ be a point-finite family of subsets of $T$. Let us denote ${\mathcal{B}}':=\{B\in{\mathcal{B}};\, {\mathrm{int}\:}B\neq\emptyset\}$, then $|{\mathcal{B}}'|\leq\mathrm{dens}(T)$.
The following proposition is a restatement of [@FZHilbert Proposition 2.1]. The elementary alternative proof presented here below is an immediate consequence of the uniform boundedness principle and it works also in the non-separable case.
\[uniformboundedness\] Let $X$ be a Banach space. Let $D\subset X^*$ be an unbounded set. For each $f\in D$, define $S_f=\{x\in X;\, f(x)\geq 1\}$. Then there exist $x\in S_X$ and an infinite set $N\subset D$ such that $x\in {\mathrm{int}\:}S_f$, whenever $f\in N$.
Suppose on the contrary that, for every $x\in S_X$, the set $$N_x:=\{f\in D;\, x\in {\mathrm{int}\:}S_f\}$$ is finite. Fix $x\in S_X$ and observe that, since $N_x$ is finite, the set $x(D)\subset{\mathbb{R}}$ is upper-bounded. By the Banach-Steinhaus uniform boundedness principle, we get a contradiction.
The following theorem is the core of the results of this section and it is a variation of [@LP Theorem 1.1], in which J. Lindenstrauss and R.R. Phelps proved that the unit ball of a reflexive infinite-dimensional Banach space has uncountably many extreme points.
\[teo:FollowingLindenstraussPhelps\] Let $X$ be an infinite-dimensional reflexive Banach space and $\{f_n\}\subset X^*\setminus U_{X^*}$. For each $n\in{\mathbb{N}}$, let $S_n$ be one of the following two sets $$\{x\in X;\, f_n(x)\geq
1\},\ \ \ \ \ \{x\in X;\, f_n(x)>1\}.$$ Let us denote $\mathcal S=\{S_n\}_{n\in{\mathbb{N}}}$ and suppose that $S_X\subset\bigcup \mathcal S$. Then $\mathcal S$ is not point-finite.
Suppose on the contrary that, for every $x\in S_X$, the set $$N_x:=\{n\in{\mathbb{N}};\, x\in S_n\}$$ is finite. By Proposition \[uniformboundedness\], we can assume that $\{f_n\}$ is bounded in $X^*$. For every $n\in{\mathbb{N}}$, let $U_n=f_n^{-1}\bigl{((-\infty,1)\bigr)}$ and put $U=\bigcap_n U_n$. Then $U$ is a convex set and $0\in{\mathrm{int}\:}U$ (since $\{f_n\}$ is bounded in $X^*$). Moreover, $S_X\cap U=\emptyset$ and hence $U\subset U_X$.
We claim that $U$ is open. To see this, let $x\in U\setminus\{0\}$ and suppose on the contrary that $\sup_n f_n(x)=1$. Then, since $f_n(x)<1$ for each $n\in{\mathbb{N}}$ and since $\|x\|<1$, $N_{x/\|x\|}$ is an infinite set. This contradiction proves our claim.
Now, for every $n\in{\mathbb{N}}$, put $F_n=\{x\in\overline U;\,
f_n(x)=p_U(x)\}$ (where $p_U$ denotes the Minkowski gauge of the set $U$) and observe that $F_n$ is closed convex and hence $w$-closed.
Fix $x\in \partial U$ and observe that, for each $n\in{\mathbb{N}}$, $f_n(x)\leq1$; since $x\not\in U$, there exists $\overline n\in{\mathbb{N}}$ such that $f_{\overline n}(x)=1$. Hence $\overline U=\bigcup F_n$. Since $\overline U$ is $w$-compact, by the Baire category theorem, we can suppose without any loss of generality that $F_1$ has nonempty interior in $(\overline U,w)$. So, there exist $x_0\in F_1\cap U$ and $W$, a neighbourhood of the origin in the $w$-topology, such that $(x_0+W)\cap\overline U\subset F_1$. Since $X$ is infinite-dimensional, there exists $y_0\in[x_0+(W\cap\ker
f_1)]\cap\partial U\subset F_1$. Then $$1=p_U(y_0)=f_1(y_0)=f_1(x_0)=p_U(x_0).$$ A contradiction, since $x_0\in U$.
The following observation is an easy consequence of the fact that the intersection among two distinct spheres in any Hilbert space lies in some hyperplane (see [@Ami (15.17)]).
\[obs:slice\] Let $X$ be a Hilbert space and let $B$ be a closed (open, respectively) ball intersecting the unit sphere $S_X$. Then there exists a closed (open, respectively) slice $S$ of $B_X$ such that $S_X\cap B$ coincide with $S_X\cap S$.
We are now ready to prove the main result of this section.
\[teo:separableHilbert\] The following assertions hold true.
1. Let ${\mathcal{B}}$ be a covering of a separable infinite-dimensional Hilbert space by closed or open balls. Suppose that ${\mathcal{B}}$ is point-finite, then $|{\mathcal{B}}|= 2^{\aleph_0}$.
2. If we suppose that $\Gamma$ is an infinite set such that $|\Gamma|<2^{\aleph_0}$, $\ell_2(\Gamma)$ does not admit a point-finite covering by open or closed balls, each of positive radius.
3. Let ${\mathcal{B}}$ be a covering of an infinite-dimensional Hilbert space by open balls. Then ${\mathcal{B}}$ is not point-finite.
Let us observe that (ii) follows easily by (i), indeed assume on the contrary that ${\mathcal{B}}$ is a point-finite cover of $\ell_2(\Gamma)$ by open or closed balls, each of positive radius. Since the density character of $\ell_2(\Gamma)$ is $|\Gamma|$, by Fact \[density\], we have $|{\mathcal{B}}|<2^{\aleph_0}$. Let us consider $Y=\ell_2\subset\ell_2(\Gamma)$ and observe that $${\mathcal{B}}':=\{B\cap Y; B\in{\mathcal{B}},\, B\cap Y\neq\emptyset\}$$ is a cover of a separable infinite-dimensional Hilbert space by open or closed balls such that $|{\mathcal{B}}'|<2^{\aleph_0}$. By (i), we get a contradiction.
Similarly, (i) implies (iii). Indeed, if ${\mathcal{B}}$ is a cover of an infinite-dimensional Hilbert space $X$ by open balls and we consider $Y=\ell_2\subset X$, we have that $${\mathcal{B}}':=\{B\cap Y; B\in{\mathcal{B}},\, B\cap Y\neq\emptyset\}$$ is a cover of a separable infinite-dimensional Hilbert space by open balls. By Fact \[density\], ${\mathcal{B}}'$ is countable. By (i), ${\mathcal{B}}'$ (and hence ${\mathcal{B}}$) is not point-finite.
It remains to prove (i). Let ${\mathcal{B}}$ be a point-finite cover of a separable infinite-dimensional Hilbert space $X$ by open or closed balls. Since $|X|=2^{\aleph_0}$ and ${\mathcal{B}}$ is point-finite, we clearly have $|{\mathcal{B}}|\leq 2^{\aleph_0}$. Now, suppose on the contrary that $|{\mathcal{B}}|< 2^{\aleph_0}$. Since the origin of $X$ is contained in finitely many members of ${\mathcal{B}}$, if we denote $${\mathcal{B}}':=\{B\in {\mathcal{B}};\, 0\not\in B\},$$ there exists $R_0>0$ such that, for each $r\geq R_0$, $r S_X$ is contained in $\bigcup {\mathcal{B}}'$.
Let us consider the set $A\subset[R_0,\infty)$ defined by $$A:=\{r\geq R_0;\, \exists B\in{\mathcal{B}}'\hbox{ such that }B\subset r S_X\};$$ that is, $A$ is the set of all $r\geq R_0$ such that $r S_X$ contains a degenerate ball $B\in{\mathcal{B}}'$. It is clear that $|A|\leq|{\mathcal{B}}'|=|{\mathcal{B}}|< 2^{\aleph_0}$ and hence there exists $\rho\in [R_0,\infty)\setminus A$. By the separability of the space, it is clear that the family $${\mathcal{B}}'':=\{B\in {\mathcal{B}}';\, B\cap \rho S_X\neq\emptyset\},$$ is countable (indeed, each element in ${\mathcal{B}}''$ has nonempty interior). Moreover, $\rho S_X$ is contained in $\bigcup {\mathcal{B}}''$. By Observation \[obs:slice\], there exists a countable point-finite family $\mathcal S$ of closed or open slices of $\rho B_X$ which covers $\rho S_X$ and such that $0\not\in\overline S$, whenever $S\in\mathcal S$. By Theorem \[teo:FollowingLindenstraussPhelps\], we get a contradiction.
Point-finite coverings by balls of Banach spaces that are both uniformly rotund and uniformly smooth {#section:appendix}
====================================================================================================
The aim of this section is to show that, following the argument introduced in [@FonfLevZan14], it is possible to extend Theorem \[teo:separableHilbert\] to Banach spaces that are both uniformly rotund and uniformly smooth. The next two results coincide with [@FonfLevZan14 Proposition 2.3] and [[@FonfLevZan14 Fact 2.4]]{}, respectively. Observe that, if we use Proposition \[uniformboundedness\] instead of [@FZHilbert Proposition 2.1], both the proofs presented in [@FonfLevZan14] work also in the non-separable case and even if we consider families of open or closed balls.
\[prop:locally-point-finite\]
Let ${{\mathcal{B}}}=\{B_n\}_{n\in{\mathbb{N}}}$ be a countable family of open or closed balls in a uniformly smooth Banach space $X$. Let us denote by $R_n$ the radius of $B_n$ ($n\in{\mathbb{N}}$) and suppose that $R_n\to \infty$. If ${\mathcal{B}}$ is not locally finite, then it is not point-finite.
\[F2\] Let ${{\mathcal{B}}}=\{B_n\}_{n\in{\mathbb{N}}}$ be a countable collection of open or closed balls in a uniformly rotund Banach space $X$. Let us denote by $R_n$ the radius of $B_n$ ($n\in{\mathbb{N}}$). Let $b>0$ and $x_0\in X$. Suppose that, for each $n\in{\mathbb{N}}$, $R_n > b$ and $x_0 \notin {\rm int}B_n$. If $$F_n = {\overline{\rm conv}}\bigl( B_n \setminus U(x_0,b) \bigr)$$ and ${\rm dist}(x_0,F_n) \to 0$ then $R_n \to \infty$.
The next lemma coincides with [@FonfLevZan14 Lemma 2.5]. Observe that in their statement it is not necessary to require that the members of ${\mathcal{F}}$ are closed.
\[lemma1\] Let $X$ be a reflexive Banach space. Let $x_0 \in X$, $a>b>c>0$ and ${\mathcal{F}}$ a collection of convex subsets of $X$ contained in $B(x_0,a)\setminus U(x_0,c)$ such that ${\mathcal{F}}$ covers $B(x_0,a)\setminus U(x_0,b)$. Then ${\mathcal{F}}$ is not locally finite in $X$.
The next lemma coincides with [@FonfLevZan14 Lemma 2.6]. Observe that it holds also in the case $X'$ is a closed infinite-dimensional subspace. Moreover, in their statement it is not necessary to require that the members of ${\mathcal{F}}$ are closed. Indeed, it is sufficient in its proof to use Fact \[F2\] and Lemma \[lemma1\] instead of [@FonfLevZan14 Fact 2.4] and [@FonfLevZan14 Lemma 2.5], respectively.
\[lemma2\] Let $X$ be both uniformly rotund and uniformly smooth. Consider a closed infinite-dimensional subspace $X' \subset X$ and let $x_0 \in X',\ a >0$. Assume that ${{\mathcal{B}}}=\{B_n \}_{n=1}^\infty$ is a countable point-finite collection of open or closed balls and ${{\mathcal{F}}}=\{F_n \}_{n=1}^\infty$ is a countable collection of convex sets such that: ${\mathcal{F}}$ covers $B(x_0,a)\cap X'$, $F_n \subset B_n\cap B(x_0,a)$ and $x_0 \notin {\rm int} B_n$, whenever $n\in{\mathbb{N}}$. Then there is a point $y \in B(x_0,a) \cap X', \ y \neq x_0$, that is a singular point for ${\mathcal{F}}$.
\[lemma:interiorconvexhull\] Let $A$ be an open convex subset of an infinite-dimensional Banach space $X$. Let $A_1,\ldots,A_n$ be nonempty convex sets in $X$ such that, for each $i=1,\ldots,n$ and $x\in\partial A_i$, there exists a hyperplane $\Gamma$ supporting $\overline{A_i}$ at $x$ such that $\Gamma\cap \overline{A_i}=\{x\}$. Define $D=A \setminus(A_1 \cup \dots \cup A_n )$, then $D\subset{\mathrm{int}\:}\bigl({\overline{\rm conv}}(D)\bigr)$.
Suppose that $x\in D$, let us prove that $x\in{\mathrm{int}\:}\bigl({\overline{\rm conv}}(D)\bigr)$. If $x\in{\mathrm{int}\:}D$ there is nothing to prove. Suppose that $x\in\partial D$, without any loss of generality, we can suppose that there exists $1\leq m\leq n$ such that:
1. $x\in\partial A_1\cap\ldots\cap\partial A_m$;
2. $x\not\in \overline A_{k}$, whenever $m<k\leq n$.
For each $i=1,\ldots,m$, let $\Gamma_i$ be a hyperplane supporting $\overline{A_i}$ at $x$ such that $\Gamma_i\cap \overline{A_i}=\{x\}$. Since $X$ is infinite-dimensional, $\Gamma=\Gamma_1\cap\ldots\cap\Gamma_m$ is an infinite-dimensional affine subset of $X$. Since $A$ is open, there exists ${\varepsilon}>0$ such that $B(x,2{\varepsilon})\subset A\setminus(\bigcup_{m<k\leq n}A_k)$. Let $v_j\in \Gamma\cap B(x,{\varepsilon})$ ($j=1,2$) be such that $x\in(v_1,v_2)$ and let $0<\delta<{\varepsilon}$ be such that $B(v_j,\delta)\cap \overline{A_i}=\emptyset$ ($j=1,2$, $i=1,\ldots,m$). Then clearly $B(x,\delta)\subset{\overline{\rm conv}}(D)$ and the proof is concluded.
Using the previous lemma we obtain the following easy variation of [@FonfLevZan14 Lemma 2.7]. For the sake of completeness, we include a proof.
\[lemma3\] Let $X$ be both uniformly rotund and uniformly smooth. Let ${{\mathcal{B}}}=\{ B_n \}_{n\in{\mathbb{N}}}$ be a countable point-finite family of open or closed balls in $X$. Let $Y\subset X$ be a separable infinite-dimensional closed subspace of $X$ and suppose that $B'_n=B_n\cap Y\neq\emptyset$ ($n\in{\mathbb{N}}$) and that $\{B'_n\}_{n\in{\mathbb{N}}}$ is a covering of $Y$. Put $B^\#_1 =B'_1$ and, for each $n\in{\mathbb{N}}$, define $$\textstyle B^\#_{n+1}=\begin{cases} {\overline{\rm conv}}((B'_{n+1} \setminus(B'_1 \cup \dots \cup B'_n )) &\ \text{if}\ B'_{n+1}\ \text{is a closed set in $Y$};\\
\mathrm{int}_Y\bigl({\overline{\rm conv}}((B'_{n+1} \setminus(B'_1 \cup \dots \cup B'_n ))\bigr) &\ \text{if}\ B'_{n+1}\ \text{is an open set in $Y$}.
\end{cases}$$ Then ${{\mathcal{B}}}^\#= \{ B_n^\# \}_{n=1}^\infty$ is a point-finite covering of $Y$. Moreover, for every $n\in{\mathbb{N}}$, we have that $B_n^\# \subset B'_n\subset B_n$ and any $x_0 \in \bigcup_{n\in{\mathbb{N}}}\mathrm{int}_Y\, B'_n $ is a regular point for ${{\mathcal{B}}}^\#$.
Observe that, since $X$ is uniformly rotund, for each $n\in{\mathbb{N}}$, one of the following conditions hold:
1. $\overline{B'_n}^Y$ is a rotund body in $Y$;
2. $B'_n$ is a singleton.
In any case, for each $n\in{\mathbb{N}}$ and $x\in\partial_Y B'_n$, there exists a hyperplane $\Gamma$ in $Y$ supporting $\overline{B'_n}^Y$ at $x$ such that $\Gamma\cap \overline{B'_n}^Y=\{x\}$. Applying Lemma \[lemma:interiorconvexhull\], we have that $B'_{n+1}\setminus (B'_n\cup\ldots\cup B'_1)\subset B_{n+1}^\#$, whenever $n\in{\mathbb{N}}$. Hence, ${{\mathcal{B}}}^\#$ is a covering of $Y$.
For the latter part we proceed as in the proof of [@FonfLevZan14 Lemma 2.7]. For $n\in{\mathbb{N}}$, let us denote by $R_{n}$ the radius of the ball $B_{n}$. Assume on the contrary that, for some ${\widetilde}n\in{\mathbb{N}}$, $x_0 \in \mathrm{int}_Y\, B'_{{\widetilde}n}$ is a singular point for ${{\mathcal{B}}}^\#$. Then there exists a subsequence of the integers $\{n_i \}_{i\in{\mathbb{N}}}$ such that, for each $i\in{\mathbb{N}}$: (a) $n_i > {\widetilde}n$; (b) $ x_0 \notin B'_{n_i}$; (c) for every $j \geq i$, $B(x_0, 1/i)$ intersects the set $B^\#_{n_j}$.
Note that $B^\# _{n_i} \subset {\overline{\rm conv}}(B'_{n_i} \setminus B'_{{\widetilde}n})$, whenever $i\in{\mathbb{N}}$. Then $B(x_0, 1/i)$ intersects $ {\overline{\rm conv}}(B_{n_j} \setminus B_{{\widetilde}n})$, whenever $i\in{\mathbb{N}}$ and $j\geq i$. Let $b > 0$ be such that $B(x_0,b) \subset B_{{\widetilde}n}$, then it holds $${\overline{\rm conv}}\bigl(B_{n_j} \setminus B(x_0,b) \bigr)\supset
{\overline{\rm conv}}(B_{n_j} \setminus B_{{\widetilde}n}).$$ Since, for each $i\in{\mathbb{N}}$, $x_0 \notin B_{n_i}$, from Fact \[F2\], we get that $ R_{n_i} {\longrightarrow}\infty $. By Proposition \[prop:locally-point-finite\], this contradicts the assumption that ${\mathcal{B}}$ is point-finite.
\[teo:unifrotundsmooth\] Let $X$ be an infinite-dimensional Banach space. Suppose that $X$ is both uniformly rotund and uniformly smooth. Then the following assertions hold true.
1. If $\mathrm{dens}(X)<2^{\aleph_0}$ then $X$ does not admit a point-finite covering by open or closed balls, each of positive radius.
2. If ${\mathcal{B}}$ is a cover of $X$ by open balls then ${\mathcal{B}}$ is not point-finite.
\(i) Suppose on the contrary that ${{\mathcal{B}}}$ is a point-finite covering of $X$ by open or closed balls, each of positive radius. By Fact \[density\], we have $|{\mathcal{B}}|<2^{\aleph_0}$.
Let $Y$ be a separable infinite-dimensional closed subspace of $X$. Let us denote $${\mathcal{B}}'=\{B\cap Y;\, B\in{\mathcal{B}},\ B\cap Y\neq\emptyset \}.$$ Clearly ${\mathcal{B}}'$ is a point-finite covering of $Y$ and passing to a subcovering we can suppose that ${\mathcal{B}}'$ is a minimal covering of $Y$. If we denote $${\mathcal{C}}'=\{C\in{\mathcal{B}}';\, \mathrm{int}_Y C\neq\emptyset\},\ \ \ {\mathcal{D}}'=\{D\in{\mathcal{B}}';\, |D|=1\},$$ it is clear that ${\mathcal{B}}'={\mathcal{C}}'\cup {\mathcal{D}}'$ and that ${\mathcal{C}}'$ is countable (since $Y$ is separable). Hence, $\bigcup {\mathcal{C}}'$ is a Borel subset of $Y$. By the fact that ${\mathcal{B}}'$ is minimal we have that $\bigcup {\mathcal{D}}'=Y\setminus \bigcup {\mathcal{C}}'$. Hence, $\bigcup {\mathcal{D}}'$ is a Borel subset of a Polish space such that $|\bigcup {\mathcal{D}}'|<2^{\aleph_0}$. By [@Srivastava Theorem 3.2.7], $|{\mathcal{D}}'|=|\bigcup {\mathcal{D}}'|\leq{\aleph_0}$, and hence ${\mathcal{B}}'$ is countable. Let ${\mathcal{B}}'=\{B'_n\}_{n\in{\mathbb{N}}}$ and suppose that, for each $n\in{\mathbb{N}}$, $B'_n=B_n\cap Y$ for some $B_n\in{\mathcal{B}}$. Now, we proceed as in the proof of [@FonfLevZan14 Theorem 1.5]. Consider the covering ${{\mathcal{B}}}^\#$ of $Y$ from Lemma \[lemma3\] and let $S\subset Y$ be the set of the points that are singular for ${{\mathcal{B}}}^\#$. By Theorem \[corson\], we have $S \neq \emptyset $ and, by Lemma \[lemma3\], we have $S\subset \cup_n \partial_Y B'_n$. Since $S$ is closed in $Y$, by the Baire category theorem, there are $m\in{\mathbb{N}}$, $x_0 \in S$, and $a>0$ such that $S \cap B(x_0,a) \subset \partial_Y B'_m$. Observe that we have two possibilities: $B'_m$ is a singleton or $\overline{B'_m}^Y$ is a rotund body in $Y$. In any case, there exists a closed hyperplane $Y'$ in $Y$ passing through $x_0$ and intersecting $\overline{B'_m}^Y$ only at $x_0$. Then, by applying Lemma \[lemma2\] to the families ${{\mathcal{F}}}= \{ B^\#_n \cap B(x_0,a);\, B^\#_n \in {{\mathcal{B}}}^\#\}$ and $\{B_n\}_{n\in{\mathbb{N}}}$, with respect to the subspace $Y'$, we get a contradiction.
The proof of (ii) is similar but easier. Indeed, observe that if $Y$ and ${\mathcal{B}}'$ are defined as above then we clearly have that ${\mathcal{B}}'$ is countable and then we can proceed as in the previous point.
In the non-separable case, non-existence of coverings by balls satisfying certain condition, were recently proved in the papers [@DEVETIL; @DESOVESTAR]. In [@DESOVESTAR], the authors showed that [*if $X$ is LUR or uniformly smooth then it does not admit star-finite coverings by closed balls, each of positive radius*]{} (we recall that a family of sets is called star-finite if each of its members intersects only finitely many other members of the family). The results contained in [@DEVETIL] imply that [*if $X$ is LUR or Fréchet smooth then it does not admit tilings by closed balls*]{}. However, the following problem remains open, even in the case $X$ is a Hilbert space.
Is it possible to generalize (i) in Theorem \[teo:unifrotundsmooth\], to the case $\mathrm{dens}(X)\geq 2^{\aleph_0}$?
Acknowledgment {#acknowledgment .unnumbered}
==============
The research of the author is partially supported by GNAMPA-INdAM, Project GNAMPA 2020. The author would like to thank J. Somaglia, L. Veselý, and C. Zanco for many discussions on the subject and for useful comments and remarks that helped him in preparing this paper.
[WW]{}
D. Amir, [*Characterizations of inner product spaces*]{}, Operator Theory: Advances and Applications, 20, Birkhauser Verlag, Basel, 1986.
H.H. Corson, [*Collections of convex sets which cover a Banach space,*]{} Fund. Math. **49** (1961), 143–145.
C.A. De Bernardi and L. Veselý, [*Tilings of normed spaces*]{}, Canad. J. Math. **69** (2017), 321–337.
C.A. De Bernardi, J. Somaglia and L. Veselý, [*Star-finite coverings of Banach spaces*]{}, arXiv:2002.04308.
V.P. Fonf, M. Levin, and C. Zanco, [*Covering $L^p$ spaces by balls,*]{} J. Geom. Anal. **24** (2014), 1891–1897.
V.P. Fonf and C. Zanco, [*Covering a Banach space,*]{} Proc. Amer. Math. Soc. **134** (2006), 2607–2611.
V.P. Fonf and C. Zanco, [*Covering the unit sphere of certain Banach spaces by sequences of slices and balls,*]{} Canad. Math. Bull. **57** (2014), 42–50.
V. Klee, [*Dispersed Chebyshev sets and coverings by balls,*]{} Math. Ann. **257** (1981), 251–260.
J. Lindenstrauss and R.R. Phelps, [*Extreme point properties of convex bodies in reflexive Banach spaces*]{}, Israel J. Math. **6** (1968), 39–48.
J. Lindenstrauss, [*Notes on Klee’s paper: “Polyhedral sections of convex bodies”*]{}, Israel J. Math. [**4**]{} (1966), 235–242.
S.M. Srivastava, [*A course on Borel sets*]{}, Graduate Texts in Mathematics, 180, Springer-Verlag, New York, 1998.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In the integer case, the Smarandache function of a positive integer $n$ is defined to be the smallest positive integer $k$ such that $n$ divides the factorial $k!$. In this paper, we first define a natural order for polynomials in ${{\mathbb F}}_q[t]$ over a finite field ${{\mathbb F}}_q$ and then define the Smarandache function of a non-zero polynomial $f \in {{\mathbb F}}_q[t]$, denoted by $S(f)$, to be the smallest polynomial $g$ such that $f$ divides the Carlitz factorial of $g$. In particular, we establish an analogue of a problem of Erd[ő]{}s, which implies that for almost all polynomials $f$, $S(f)=t^d$, where $d$ is the maximal degree of the irreducible factors of $f$.'
address:
- 'School of Mathematical Sciences, Qufu Normal University, Qufu, 273165, China'
- 'School of Mathematics and Statistics, University of New South Wales, Sydney, NSW 2052, Australia'
author:
- Xiumei Li
- Min Sha
title: Polynomial analogue of the Smarandache function
---
Introduction
============
Motivation
----------
In number theory, the Smarandache function of a positive integer $n$ is defined to be the smallest positive integer $k$ such that $n$ divides the factorial $k!$. This function was studied by Lucas [@Lucas] for powers of primes and then by Neuberg [@Neu] and Kempner [@Kem] for general $n$. In particular, Kempner [@Kem] gave the first correct algorithm for computing this function. In 1980 Smarandache [@Sma] rediscoverd this function. It is also sometimes called the Kempner function. This function arises here and there in number theory (for instance, see [@Chen1995; @Kempner1921; @Luca; @Sondow]). Please see [@Liu] for a survey on recent results and [@HS] for a generalization to several variables.
Clearly, the Smarandache function of $n$ is equal to the maximum of those of its prime power factors. For any integer $n \ge 2$, let $P(n)$ be the largest prime factor of $n$; and put $P(1)=1$. For any $x > 1$, denote by $N(x)$ the number of positive integers $n \le x$, whose Smarandache function are not equal to that of $P(n)$ (that is, $P(n)$, this means $n \nmid P(n)!$). In 1991 Erd[ő]{}s [@Erdos] posed a problem answered by Kastanas [@Kas] in 1994 that $N(x)=o(x)$ when $x$ goes to infinity. Later, Akbik [@Akbik] proved that $N(x) = O(x\exp(-\frac{1}{4}\sqrt{\log x}))$, and recently Ivi[' c]{} [@Ivic] showed that $$N(x) = x \exp \left(-\sqrt{2\log x \log\log x} (1+ O(\log\log\log x /\log\log x)) \right);$$ see [@DD; @Ford] for some other previous results. Besides, Ivi[' c]{} [@Ivic] pointed out that the main result in [@DD Equation (1.3)] is not correct.
In this paper, we define and study the Smarandache function for polynomials over a finite field. In particular, we establish an analogue of Erd[ő]{}s’s problem.
Our consideration
-----------------
Let ${{\mathbb F}}_q$ be the finite field of $q$ elements, where $q$ is a power of a prime $p$. Denote by ${{\mathbb A}}={{\mathbb F}}_q[t]$ the polynomial ring of one variable over ${{\mathbb F}}_q$ and ${{\mathbb N}}$ the set of non-negative integers. Let ${{\mathbb N}}^*$ be the set of positive integers. For any non-zero $g \in {{\mathbb A}}$, we denote by ${\text{\rm sign}}(g)$ the leading coefficient of $g$ (which is also called the sign of $g$).
We write ${{\mathbb F}}_q=\{a_{0}=0,a_{1}=1, a_2, \ldots,a_{q-1}\}$ throughout the paper. For any non-zero polynomial $f\in {{\mathbb A}}$ of degree $n$, $f$ can be uniquely written as $$\label{eq:f}
f=a_{i_0}+a_{i_1}t+\ldots+a_{i_n}t^{n}, \quad a_{i_n} \ne 0, \, 0\leq i_j \leq q-1.$$ We define $\delta(f)$ to be the integer: $$\label{eq:delta}
\delta(f) = i_0 + i_1q + \cdots + i_n q^n.$$ In addition, we put $\delta(0)=0$.
Clearly, $\delta$ is a bijective map from ${{\mathbb A}}$ to ${{\mathbb N}}$, and for any $m\in{{\mathbb N}}$, $$\delta^{-1}(m)=a_{i_0}+a_{i_1}t+\ldots+a_{i_k}t^{k},$$ where $i_0 + i_1q + \cdots + i_k q^k$ is the $q$-adic expansion of $m$.
Moreover, we define an order in ${{\mathbb A}}$ based on the map $\delta$: for any $f,g\in {{\mathbb A}}$, $$f>g \quad \mbox{if and only if}\quad \delta(f) > \delta(g);$$ and then $f \ge g$ if and only if $f > g$ or $f=g$.
With these preparations, we define a factorial in ${{\mathbb A}}$.
\[def:factorial\] For any non-zero polynomial $f \in {{\mathbb A}}$, the factorial of $f$ is defined to be $$f!=\prod_{g<f}(f-g).$$ Additionally, we put $0!=1$.
By definition, for any integer $n \ge 1$, $t^n !$ is in fact the product of all the monic polynomials of degree $n$.
This factorial is an analogue of the factorial of the rational integers; see [@LS1] for another analogue. It has been used in [@LS2; @LS3]. Notice that the above factorial of $f$ is equal to the multiplication of the Carlitz factorial of $f$ by a constant (see the comment below Lemma \[lem:factorial\]). For the Carlitz factorial, one can refer to [@Carlitz; @Thakur].
We now can define the Smarandache function for polynomials in ${{\mathbb A}}$. In fact it has been used in [@LS2 Section 4.2] for counting polynomial functions in the residue class rings (see the definition of $\lambda$ there).
\[def:kem\] Given a non-zero polynomial $f \in {{\mathbb A}}$, the Smarandache function $S(f)$ of $f$ is defined to be the smallest polynomial $g$ such that $$f \mid g!.$$ We put $S(0)=0$ by convention.
In Section \[sec:basic\] we establish various basic properties of the Smarandache function $S$, such as the computation, the value set, the inverse images, and fixed points. We emphasize that several of them haven’t be considered in the integer case, such as Proposition \[prop:delta-Kf\] on how the size of a polynomial changes after an action of $S$ and Proposition \[prop:dist\] on the distance to fixed points. We then in Section \[sec:Erdos\] establish an analogue of Erd[ő]{}s’s problem for the function $S$ (see Theorem \[thm:Erdos\]).
Preliminaries
=============
In this section, we gather some results which are used later on.
Some elementary results
-----------------------
We first compute the factorial $f!$ for any $f\in {{\mathbb A}}$.
\[lem:factorial\] For any $f\in {{\mathbb A}}$ of the form , we have $$\label{eq:factorial}
f!=\left(\prod_{j=0}^{n}a_{i_j}! \right) \prod_{j=1}^{n}\prod_{\substack{h \in{{\mathbb A}}, \, \deg h=j \\ {\text{\rm sign}}(h)=1}} h^{i_j}.$$
Denote by $R$ the right hand side of . We rewrite $R$ as $$R=\prod_{j=0}^{n}\prod_{k=0}^{i_{j}-1}\prod_{ \substack{h\in{{\mathbb A}}, \, \deg h =j \\ {\text{\rm sign}}(h)=a_{i_j}-a_{k}}} h
= \prod_{j=0}^{n}\prod_{k=0}^{i_{j}-1}\prod_{ \substack{h\in{{\mathbb A}}, \, \deg h =j \\ {\text{\rm sign}}(h)=a_{i_j}-a_{k}}}(f-(f-h)),$$ where one can see that $f-h$ exactly runs over all the polynomials $g < f$. So, by definition we have $R=f!$.
By definition and Lemma \[lem:factorial\], $f!/ (\prod_{j=0}^{n}a_{i_j}!)$ is exactly the Carlitz factorial of $f$. So, in Definition \[def:kem\] we can replace $g!$ by the Carlitz factorial of $g$.
The following result is a special case of Example 3 in [@Bhargava1997]. We give a proof here.
\[lem:val\] Let $P \in {{\mathbb A}}$ be an irreducible polynomial of degree $d \ge 1$. Then, for any non-zero polynomial $f\in {{\mathbb A}}$ we have $$v_P(f!)=\sum_{j\geq 1}\left\lfloor\frac{\delta(f)}{q^{dj}}\right\rfloor,$$ where $v_P$ is the usual $P$-adic valuation.
Assume that $f$ is of the form . From the formula of $f!$, we see that for any integer $j \ge 1$, if $n= \deg f \ge dj$, then the number of terms in the right hand side of divisible by $P^j$ is exactly equal to $$i_{dj} + i_{dj+1}q + \cdots + i_n q^{n-dj}.$$ Summing up all these estimates we obtain the desired formula.
Clearly, Lemma \[lem:val\] gives the following result.
\[cor:div\] For any $f,g\in {{\mathbb A}}$, if $g \le f$, then $ g! \mid f!$.
We remark that in Corollary \[cor:div\] the converse is not true.
For the proof of Proposition \[prop:comp2\], we need the following lemma, which is in fact a simple generalization of the formula of $\alpha$ in [@Kem page 207] (also the formula in [@Sma Lemma 1]).
\[lem:rep\] Fix a positive integer $n>1$, and define a sequence $\{b_{j}=\frac{n^{j}-1}{n-1}: \, j\in{{\mathbb N}}^*\}$. Then for any $e\in{{\mathbb N}}^*$, $e$ can be uniquely represented as $$\label{eq:rep}
e=c_{1}b_{j_1}+c_{2}b_{j_2}+\cdots+c_{k}b_{j_k},$$ where $ j_1>j_2>\cdots>j_k>0 $ and $1\leq c_i < n$ for $i=1,2,\ldots,k-1, 1\leq c_k\leq n$.
Obviously, ${{\mathbb N}}^*$ is the disjoint union of the sets $[b_{j},b_{j+1})\cap {{\mathbb N}}^*, j\in{{\mathbb N}}^*$, and $b_{j+1} = n b_j +1$ for any $j \in {{\mathbb N}}^*$. So, for any $e\in{{\mathbb N}}^*$, there exists an unique integer $j_{1}\in{{\mathbb N}}^*$ such that $e\in[b_{j_{1}},b_{j_{1}+1})\cap {{\mathbb N}}^* $, then by the division algorithm, we have $$e=c_{1}b_{j_{1}}+r_{1},$$ where $1\leq c_{1}=\lfloor\frac{e}{b_{j_{1}}}\rfloor\leq n$ and $0\leq r_{1}<b_{j_{1}}$. If $r_{1}=0$, as $b_{j_{1}}\leq e <b_{j_{1}+1}$, then $k=1, 1\leq c_{1} \leq n$ and Lemma \[lem:rep\] is proved.
If $r_{1}\neq0$, as $b_{j_{1}}\leq e <b_{j_{1}+1}$, then $1\leq c_{1} < n$. The following procedure is the iterative process that makes use of the division algorithm in the form: $$\begin{aligned}
& r_{1}=c_{2}b_{j_{2}}+r_{2}, \quad 1\leq c_{2} < n, \, 0< r_{2}<b_{j_{2}}, \\
& r_{2}=c_{3}b_{j_{3}}+r_{3}, \quad 1\leq c_{3} < n, \, 0< r_{3}<b_{j_{3}}, \\
& \vdots \\
& r_{k-2}=c_{k-1}b_{j_{k-1}}+r_{k-1}, \quad 1\leq c_{k-1} < n, \, 0< r_{k-1}<b_{j_{k-1}}, \\
& r_{k-1}=c_{k}b_{j_{k}}, \quad 1\leq c_{k}\leq n.\end{aligned}$$ In the above computation the integer $k$ is defined by the condition that $r_{k-1} \neq 0$ and that $r_{k} = 0$. Since $ e\geq b_{j_{1}}> r_{1}\geq b_{j_{2}}>r_{2}\geq \cdots \geq 0$, such a $k$ must exist and $j_{1}>j_{2}>\cdots>j_{k}> 0$.
Collecting all the equalities above, Lemma \[lem:rep\] is proved.
Counting polynomials {#sec:count}
--------------------
For any non-zero $f \in {{\mathbb A}}$, let $\omega(f)$ be the number of distinct monic irreducible factors of $f$, and let $\tau(f)$ be the number of distinct monic factors of $f$.
The following two results should be well-known.
\[lem:count-irre\] For any integer $n \ge 1$, the number of monic irreducible polynomials in ${{\mathbb A}}$ of degree at most $n$ is at most $q^n$.
For any monic irreducible polynomial $f \in {{\mathbb A}}={{\mathbb F}}_q[t]$, if $f$ is of degree $d \le n$, then $f$ corresponds to the monic polynomial $t^rf^s$ of degree $n$, where $n = sd + r$ with $0 \le r < d$ by the division algorithm. Note that this corresponding is injective. So the desired result follows.
\[lem:tau\] For any integer $n \ge 1$, we have $$\sum_{\substack{\textrm{monic $f \in {{\mathbb A}}$} \\ \deg f = n}} \tau(f) = (n+1)q^n.$$
This result has been recorded in [@Rosen Proposition 2.5]. Here we present a different proof. It is easy to see that $$\begin{aligned}
\sum_{\substack{\textrm{monic $f \in {{\mathbb A}}$} \\ \deg f = n}} \tau(f)
& = \sum_{\substack{\textrm{monic $g \in {{\mathbb A}}$} \\ \deg g \le n}} \sum_{\substack{\textrm{monic $h \in {{\mathbb A}}$} \\ \deg h = n - \deg g}} 1
= \sum_{\substack{\textrm{monic $g \in {{\mathbb A}}$} \\ \deg g \le n}} q^{n - \deg g} \\
& = q^n \sum_{j=0}^{n} q^{-j} \cdot q^j = (n+1)q^n.\end{aligned}$$
We now present some counting results for polynomials in ${{\mathbb A}}$ according to the numbers of their monic factors and their maximal monic irreducible factors. These are needed for proving Theorem \[thm:Erdos\].
\[lem:S1\] For any integer $n \ge 1$ and any real $r \ge 1$, let $B = 3r \log \log q^n$ and define $${{\mathcal T}}_1(n,r) = \{\textrm{monic }f \in {{\mathbb A}}:\, \deg f=n, \omega(f) > B\}.$$ Then, we have $$|{{\mathcal T}}_1(n,r)| < \frac{3q^n}{(\log q^n)^r}.$$ If furthermore $n \ge 3$ and $r \ge 2$, we have $$\label{eq:S1}
|{{\mathcal T}}_1(n,r)| < \frac{q^n}{(\log q^n)^r}.$$ Moreover, if $q \ge 3, n \ge 4$ and $r \ge 3$, in ${{\mathcal T}}_1(n,r)$ we can choose $B = 2r\log \log q^n$, and then the estimate still holds.
By definition, we have $\tau(f) \ge 2^{\omega(f)}$ for any non-zero $f \in {{\mathbb A}}$. Using Lemma \[lem:tau\], we deduce that $$\begin{aligned}
(n+1)q^n
&= \sum_{\substack{\textrm{monic $f \in {{\mathbb A}}$} \\ \deg f = n}} \tau(f)
\ge \sum_{\substack{\textrm{monic $f \in {{\mathbb A}}$} \\ \deg f = n}} 2^{\omega(f)}
\ge \sum_{\textrm{$f \in {{\mathcal T}}_1(n,r)$}} 2^{\omega(f)} \\
& > \sum_{\textrm{$f \in {{\mathcal T}}_1(n,r)$}} 2^{3r\log \log q^n}
= 2^{3r\log \log q^n} |{{\mathcal T}}_1(n,r)|.\end{aligned}$$ So, we obtain $$|{{\mathcal T}}_1(n,r)| < \frac{(n+1)q^n}{2^{3r\log \log q^n}}
= \frac{(n+1)q^n}{(\log q^n)^{3r\log 2}}
< \frac{(n+1)q^n}{(\log q^n)^{r+1}}
< \frac{3q^n}{(\log q^n)^r}.$$ If furthermore $n \ge 3$ and $r \ge 2$, we have $$|{{\mathcal T}}_1(n,r)| < \frac{(n+1)q^n}{(\log q^n)^{3r\log 2}}
< \frac{(n+1)q^n}{(\log q^n)^{2r}}
< \frac{q^n}{(\log q^n)^r},$$ where the last inequality comes from $$(n \log q)^r \ge (n \log q)^2 \ge (n \log 2)^2 > n+1, \quad n \ge 3.$$
The final part follows from $$|{{\mathcal T}}_1(n,r)| < \frac{(n+1)q^n}{(\log q^n)^{2r\log 2}}< \frac{q^n}{(\log q^n)^r}$$ when $q \ge 3, n \ge 4$ and $r \ge 3$.
\[lem:S2\] For any integer $n \ge 1$ and any real $r \ge 1$, let $D=2r\log \log q^n$ and define $$\begin{aligned}
{{\mathcal T}}_2(n,r) = \{& \textrm{monic }f \in {{\mathbb A}}:\, \deg f=n, \\
& P^2 \mid f \textrm{ for some irreducible polynomial $P$ with $\deg P > D$}\}.\end{aligned}$$ Then, if $D \ge 4$, we have $$\label{eq:S2}
|{{\mathcal T}}_2(n,r)| < \frac{q^n}{(\log q^n)^r}.$$ Moreover, if $q \ge 3$, in ${{\mathcal T}}_2(n,r)$ we can choose $D=r\log \log q^n \ge 12$, and then the estimate still holds.
For any $f \in {{\mathcal T}}_2(n,r)$, we can write $f=gP^2$ with $D < \deg P \le n/2$ and $\deg g = n - 2 \deg P$. So, we have $$\begin{aligned}
|{{\mathcal T}}_2(n,r)|
& \le \sum_{\substack{\textrm{monic irreducible $P \in {{\mathbb A}}$} \\ D < \deg P \le n/2}} \sum_{\substack{\textrm{monic $g \in {{\mathbb A}}$} \\ \deg g = n - 2\deg P}} 1 \\
& = \sum_{\substack{\textrm{monic irreducible $P \in {{\mathbb A}}$} \\ D < \deg P \le n/2}} q^{n - 2\deg P}
< q^n \sum_{j=\lfloor D \rfloor+1}^{\infty} q^{-2j} \cdot q^j \\
& \le \frac{q^n}{q^{\lfloor D \rfloor}} < \frac{q^n}{q^{-1+ 2r\log\log q^n}} = \frac{q^{n+1}}{(\log q^n)^{2r\log q}} < \frac{q^n}{(\log q^n)^r}, \end{aligned}$$ where the last inequality follows from $(\log q^n)^r \ge \exp(2)$ (due to $D \ge 4$), because it is equivalent to the following inequality $$q < (\log q^n)^{r(2\log q - 1)}.$$
The second part follows similarly.
\[lem:S3\] For any integer $n \ge 1$ and any real $r \ge 1$, let $D=2r\log \log q^n$ and define $$\begin{aligned}
{{\mathcal T}}_3 (n,r) = \{& \textrm{monic }f \in {{\mathbb A}}:\, \deg f=n, P^e \mid f, \\
& e \ge D \textrm{ for some irreducible polynomial $P$ with $\deg P \le D$}\}.\end{aligned}$$ Then, if $D \ge 8$, we have $$\label{eq:S3}
|{{\mathcal T}}_3(n,r)| < \frac{q^n}{(\log q^n)^r}.$$ Moreover, if $q \ge 3$, in ${{\mathcal T}}_3(n,r)$ we can choose $D=r\log \log q^n \ge 19$, and then the estimate still holds.
Let $d = \lceil D \rceil \ge 8$. By definition, for any $f \in {{\mathcal T}}_3(n,r)$, there exists a monic irreducible polynomial $P$ such that $\deg P \le D$ and $P^d \mid f$. As in the proof of Lemma \[lem:S2\], we have $$\begin{aligned}
|{{\mathcal T}}_3(n,r)|
& \le \sum_{\substack{\textrm{monic irreducible $P \in {{\mathbb A}}$} \\ 1 \le \deg P \le D}} \sum_{\substack{\textrm{monic $g \in {{\mathbb A}}$} \\ \deg g = n - d\deg P}} 1 \\
& = \sum_{\substack{\textrm{monic irreducible $P \in {{\mathbb A}}$} \\ 1 \le \deg P \le D}} q^{n - d\deg P}
< q^n \sum_{j=1}^{\infty} q^{-dj} \cdot q^j \\
& \le \frac{2q^n}{q^{D-1}} = \frac{2q^{n+1}}{q^{2r\log \log q^n}} = \frac{2q^{n+1}}{(\log q^n)^{2r\log q}} < \frac{q^n}{(\log q^n)^r},\end{aligned}$$ where the last inequality follows from $(\log q^n)^r \ge \exp(4)$ (due to $D \ge 8$), because it is equivalent to the following inequality $$2q < (\log q^n)^{r(2\log q - 1)}.$$
The second part follows similarly.
Basic properties {#sec:basic}
================
Computing the Smarandache function
----------------------------------
By definition, we directly obtain some simple properties about the Smarandache function.
\[prop:basic\] The following hold:
- for any polynomial $f \in {{\mathbb A}}$ and any $a \in {{\mathbb F}}_q^*$, $S(af)=S(f)$;
- for any non-zero polynomial $f \in {{\mathbb A}}$, $S(f) \le t^{\deg f}$;
- for any irreducible polynomial $f \in {{\mathbb A}}$, $S(f) = t^{\deg f}$.
So, in order to compute the Smarandache function, we only need to consider monic polynomials. By Definition \[def:kem\] and Corollary \[cor:div\], we immediately obtain the following result, which implies that we in fact only need to consider powers of monic irreducible polynomials.
\[prop:comp1\] Suppose that $P_1, P_2,\ldots, P_k$ are distinct monic irreducible polynomials and $e_1,e_2,\ldots,e_k$ are positive integers. Then $$S(P_1^{e_1}P_2^{e_2}\cdots P_k^{e_k})=\max\{S(P_1^{e_1}),S(P_2^{e_2}),\ldots,S(P_k^{e_k})\}.$$
The case of irreducible polynomials is straightforward. We can in fact do more.
Given a polynomial $f \in {{\mathbb A}}$ with $\deg f \ge 1$, assume that either $q \ge 3$ or $f \ne b(t+c)^2$ for any $b,c \in {{\mathbb F}}_q$. Then, $f$ is an irreducible polynomial if and only if $S(f)=t^{\deg f}$.
We only need to prove the sufficiency. Assume that $S(f)=t^{\deg f}$. Without loss of generality, we can further assume that $f$ is monic. By Proposition \[prop:comp1\], we must have $f=P^e$ for some monic irreducible polynomial $P$ of degree $d$ and $e \ge 1$.
We first assume that $q \ge 3$. If $e \ge 2$, since $a_2 P^{e-1}$ and $(a_2-1) P^{e-1}$ are two distinct terms in the factorial $(a_2 P^{e-1})!$ by definition, we have $$v_P((a_2 P^{e-1})!) \ge 2(e-1) \ge e,$$ and so $S(P^e) \le a_2 P^{e-1} < t^{de}$, which contradicts with the assumption $S(P^e) = t^{de}$. Thus, $f=P$ when $q \ge 3$.
We now assume that $q=2$. By assumption, $f \ne (t+c)^2$ for any $c \in {{\mathbb F}}_q$. So, if $d = 1$, we must have $e \ge 3$, and so $v_P(P^{e-1}!) \ge e$, which implies $S(P^e) \le P^{e-1} < t^{e}$ and contradicts with the assumption $S(P^e) = t^{e}$. Thus, we must have $d \ge 2$. If $e \ge 2$, since $tP^{e-1}$ and $(t+1)P^{e-1}$ are two distinct terms in the factorial $t^{d(e-1)+1}!$ by definition, we have $$v_P(t^{d(e-1)+1}!) \ge 2(e-1) \ge e,$$ and so $S(P^e) \le t^{d(e-1)+1} < t^{de}$, which contradicts with the assumption $S(P^e) = t^{de}$. Thus, $f=P$. This completes the proof.
We remark that in the case $q=2$, we have $S(t^2) = S(t^2+1) = t^2$.
We now handle the case of powers of irreducible polynomials by following the strategy used to prove the theorem in [@Kem page 208] (also [@Sma Theorem 1]).
\[prop:comp2\] Suppose that $P \in {{\mathbb A}}$ is an irreducible polynomial of degree $d \ge 1$ and $e$ is a positive integer. Define the sequence $b_{j}=\frac{q^{dj}-1}{q^{d}-1}, j\in{{\mathbb N}}^*$. Then, $e$ is uniquely written as $$e=c_{1} b_{j_1}+ c_{2} b_{j_2} + \cdots+c_{k} b_{j_k},$$ and $$S(P^e)= \delta^{-1}(c_{1}q^{dj_1}+c_{2}q^{dj_2}+\cdots+c_{k}q^{dj_k}),$$ where $ j_1>j_2>\cdots>j_k>0 $ and $1\leq c_i < q^{d}$ for $i = 1,2,\ldots,k-1, 1\leq c_k\leq q^{d}$.
By Lemma \[lem:rep\], we know that $e$ is uniquely written in the form: $$e=c_{1}b_{j_1}+c_{2}b_{j_2}+\cdots+c_{k}b_{j_k},$$ where $ j_1>j_2>\cdots>j_k>0 $ and $1\leq c_i < q^{d}$ for $i =1,2,\ldots,k-1, 1\leq c_k\leq q^{d}$. Denote $$\begin{aligned}
m & =c_{1}q^{dj_1}+c_{2}q^{dj_2}+\cdots+c_{k}q^{dj_k} \\
& =(q^{d}-1)e+(c_{1}+c_{2}+\cdots + c_{k}).\end{aligned}$$ Since $\delta$ is a bijective map from ${{\mathbb A}}$ to ${{\mathbb N}}$, we take $f=\delta^{-1}(m)$. Then, it suffices to prove $S(P^e)=f$.
By Lemma \[lem:val\] and collecting the following equalities and inequalities $$\begin{aligned}
& \left\lfloor\frac{m}{q^{d}}\right\rfloor = c_{1}q^{d(j_{1}-1)}+c_{2}q^{d(j_{2}-1)}+\cdots+c_{k}q^{d(j_{k}-1)}, \\
& \vdots \\
& \left\lfloor\frac{m}{q^{dj_k}}\right\rfloor = c_{1}q^{d(j_1-j_k)}+c_{2}q^{d(j_2-j_k)}+\cdots+c_{k},\\
& \left\lfloor\frac{m}{q^{d(j_k+1)}}\right\rfloor
\geq
c_{1}q^{d(j_1-j_k-1)}+c_{2}q^{d(j_2-j_k-1)}+\cdots+c_{k-1}q^{d(j_{k-1}-j_{k}-1)}, \\
& \vdots \\
& \left\lfloor\frac{m}{q^{dj_1}}\right\rfloor\geq c_{1},\end{aligned}$$ we obtain $$v_P(f!) = \sum_{j \geq 1}\left\lfloor\frac{m}{q^{dj}}\right\rfloor \geq e,$$ which implies that $P^e \mid f!$. Actually, $v_P(f!) = e$ if and only if $c_{k} < q^{d}$.
Now, it remains to prove that for any $g\in{{\mathbb A}}$ and $g<f$, we have $P^e\nmid g!$. In fact, by Corollary \[cor:div\], we only need to prove that for $g = \delta^{-1}(m-1)$, $P^e\nmid g!$; that is, $v_P(g!)<e$. It is easy to obtain the following equalities: $$\begin{aligned}
&\left\lfloor\frac{m-1}{q^{d}}\right\rfloor=c_{1}q^{d(j_{1}-1)}+c_{2}q^{d(j_{2}-1)}+\cdots+c_{k}q^{d(j_{k}-1)}-1, \\
& \vdots \\
& \left\lfloor\frac{m-1}{q^{dj_k}}\right\rfloor=c_{1}q^{d(j_1-j_k)}+c_{2}q^{d(j_2-j_k)}+\cdots+c_{k}-1, \\
& \vdots \\
& \left\lfloor\frac{m-1}{q^{dj_1}}\right\rfloor=c_{1}-1.\end{aligned}$$ Then $v_P(g!) = \sum_{ j \geq 1}\left\lfloor\frac{m-1}{q^{dj}}\right\rfloor = e-j_{1} < e$. This completes the proof.
By Proposition \[prop:comp2\], we directly obtain the following result.
Suppose that $P \in {{\mathbb A}}$ is an irreducible polynomial of degree $d$ and $e\leq q^{d}$ is a positive integer. Then $$S(P^e)= a_{i_0} t^{d} + a_{i_1} t^{d+1} + \cdots + a_{i_k} t^{d+k},$$ where $e=\sum_{j=0}^{k}i_{j}q^{j} $ is the $q$-adic expansion of $e$.
With some more effort we can estimate how the size of a polynomial changes after an action of $S$. We in fact only need to consider reducible polynomials.
\[prop:delta-Kf\] Given a polynomial $f \in {{\mathbb A}}$ with $\deg f \ge 1$, suppose that $f$ is reducible and $f \ne b(t+c)^2$ for any $b, c \in {{\mathbb F}}_q$. Then, we have $$\delta(S(f)) \le \frac{\delta(f)}{q},$$ where the equality holds if and only if $q=2$ or $3$, $f=t^3$.
Without loss of generality, we can assume that $f$ is monic. When $f$ has at least two distinct monic irreducible factors, by Proposition \[prop:basic\] (2) and Proposition \[prop:comp1\], we immediately have $$\delta(S(f)) < \frac{\delta(f)}{q}.$$ So, it remains to consider the following two cases:
- $f=P^e, e \ge 2$ for a monic irreducible polynomial $P$ with $\deg P \ge 2$;
- $f=P^e, e \ge 3$ for a monic linear polynomial $P$.
We first handle the first case. That is, we assume that $f=P^e, e \ge 2$ for a monic irreducible polynomial $P$ with $\deg P \ge 2$.
Let $d= \deg P$, and define $b_j = \frac{q^{dj}-1}{q^d - 1}, j \in {{\mathbb N}}^*$. As before, $e$ can be uniquely written as $$\label{eq:ecb}
e = c_{1} b_{j_1} + c_{2} b_{j_2} + \cdots + c_{k} b_{j_k},$$ where $ j_1>j_2>\cdots>j_k>0 $ and $1\leq c_i < q^{d}$ for $i = 1,2,\ldots,k-1, 1\leq c_k\leq q^{d}$. By Proposition \[prop:comp2\], we have $$\label{eq:Kcq}
\delta(S(f)) = c_{1} q^{dj_1} + c_{2} q^{dj_2} + \cdots + c_{k} q^{dj_k} \le q^{d(j_1 +1)}.$$
If $j_1 \ge 2$, then (using $d \ge 2$) $$e \ge c_{1} b_{j_1} \ge b_{j_1} \ge 1 + 2^2 + \cdots + 2^{2(j_1 -1)} \ge j_1 + 3 ,$$ and thus $$\frac{\delta(f)}{q^d} > \frac{q^{de}}{q^d} = q^{d(e -1)} \ge q^{d(j_1 +2)} > q^{d(j_1+1)} \ge \delta(S(f)).$$
If $j_1 =1$, then $e =c_1b_1 = c_1 \le q^d$ and $\delta(S(f)) = eq^{d}$, and so for $e \ge 3$ $$\frac{\delta(f)}{q^d} > q^{d(e-1)} \ge q^{2d} \ge eq^d = \delta(S(f)).$$ For $e=2$ $$\frac{\delta(f)}{q} > \frac{q^{2d}}{q} = q^{2d-1} \ge 2q^d = \delta(S(f)).$$ This completes the proof for the first case.
We now handle the second case. That is, we assume that $f=P^e, e \ge 3$ for a monic linear polynomial $P$. This means that in and $d=1$.
If $j_1 \ge 3$, then $$e \ge c_{1} b_{j_1} \ge b_{j_1} \ge 1 + 2 + \cdots + 2^{j_1 -1} \ge j_1 + 4 ,$$ and so $$\frac{\delta(f)}{q} \ge q^{e-1} \ge q^{j_1 + 3} > q^{j_1 + 1} \ge \delta(S(f)).$$
If $j_1 =2$, then for $e \ge 5$, we already have $e \ge j_1 +3$, and so $$\frac{\delta(f)}{q} \ge q^{e-1} \ge q^{j_1+2} > q^{j_1 + 1} \ge \delta(S(f)).$$ For $e=4$, we have either $q=2, e=b_1 + b_2, \delta(S(f)) =6$ or $q=3, e=b_2,\delta(S(f))=9$, and then we still obtain $$\frac{\delta(f)}{q} \ge q^3 > \delta(S(f)).$$ For $e=3$, we must have $q=2, e= b_2, \delta(S(f)) =4$, and so $$\frac{\delta(f)}{q} \ge \frac{2^3}{2} = 4 = \delta(S(f)),$$ where the equality holds if and only if $f=t^3$.
If $j_1 =1$, then $e=c_1b_1 =c_1 \le q, \delta(S(f)) = eq$, and thus (using $e \ge 3$) $$\frac{\delta(f)}{q} \ge q^{e-1} \ge q^2 \ge eq = \delta(S(f)),$$ where the equalities hold if and only if $q=3, f=t^3$. This completes the proof.
In the above proof, we in fact have proved the following result.
For any irreducible polynomial $P \in {{\mathbb A}}$ with $\deg P \ge 2$ and any integer $e \ge 3$, we have $$\delta(S(P^e)) < \frac{\delta(P^e)}{q^{\deg P}}.$$
Values of the Smarandache function
----------------------------------
Here, we consider the value set and the inverse image sets of the Smarandache function $S$.
\[prop:val-range\] $ S({{\mathbb A}})= t{{\mathbb A}}$.
By Propositions \[prop:comp1\] and \[prop:comp2\], it is easy to see that $S({{\mathbb A}})\subseteq t{{\mathbb A}}$ (note that $S(b)=0$ for any $b \in {{\mathbb F}}_q$). On the other hand, for any $f=a_{i_1}t + a_{i_2} t^2 + \cdots + a_{i_k} t^{k} \in t{{\mathbb A}}$, we take $e=i_{1}b_{1}+i_{2}b_{2}+\cdots+i_{k}b_{k}$, where $b_{j}=\frac{q^{j}-1}{q-1}, j\in{{\mathbb N}}^{*}$. Then by Proposition \[prop:comp2\], we have $S(t^{e})=f$, and so $t{{\mathbb A}}\subseteq S({{\mathbb A}})$. Thus $ S({{\mathbb A}})= t{{\mathbb A}}$.
We have seen that the Smarandache function $S$ is not injective; see Proposition \[prop:basic\] (1). For any non-zero polynomial $ f\in t{{\mathbb A}}$, denote by $S^{-1}(f)$ the inverse image set of $f$. We now determine all the powers of irreducible polynomials contained in $S^{-1}(f)$.
\[prop:inverse\] Given a non-zero polynomial $f\in t{{\mathbb A}}$ and an integer $d\in{{\mathbb N}}^*$, suppose that $q^d \mid \delta(f)$. Then, $\delta(f)$ is uniquely represented as $$\delta(f)=c_{1}q^{dj_{1}}+c_{2}q^{dj_{2}}+\cdots+c_{k}q^{dj_{k}},$$ with $j_1>j_2>\cdots>j_k>0,1\leq c_i < q^{d}$ for $i =1,2,\ldots,k$. Put $b_{j}=\frac{q^{dj}-1}{q^{d}-1}, j\in{{\mathbb N}}^*$ and $$e_{0}=c_{1}b_{j_{1}}+c_{2}b_{j_{2}}+\cdots+c_{k}b_{j_{k}}.$$ Then $S^{-1}(f)$ contains the subset $$\{P^e: \, \textrm{$P\in{{\mathbb A}}$ is irreducible of degree $d$},\ e\in[ e_{0}-(j_{k}-1),e_{0}]\cap{{\mathbb N}}\}.$$ In particular, when exhausting all the positive integers $d$ satisfying $q^d \mid \delta(f)$, we obtain all the powers of irreducible polynomials contained in $S^{-1}(f)$.
Suppose that $P\in{{\mathbb A}}$ is an irreducible polynomial of degree $d$. By Proposition \[prop:comp2\], we directly have $\delta(S(P^{e_{0}}))=\delta(f)$, and so $S(P^{e_{0}})=f$. When $j_{k}\geq 2$ and $e\in[ e_{0}-(j_{k}-1),e_{0})\cap{{\mathbb N}}$, without loss of generality, we take $e=e_{0}-i, 1\leq i \leq j_{k}-1$, then $e$ is uniquely represented in the form: $$\begin{aligned}
e=c_{1}b_{j_{1}}+\cdots & + c_{k-1} b_{j_{k-1}} + (c_{k}-1)b_{j_{k}} \\
& +(q^{d}-1)b_{j_{k}-1}+\cdots+(q^{d}-1)b_{j_{k}-(i-1)}+q^{d}b_{j_{k}-i}.\end{aligned}$$ By Proposition \[prop:comp2\], we have $S(P^{e})=f$. This in fact completes the proof.
From Proposition \[prop:inverse\], one can guess that the Smarandache function $S$ is not an increasing function. We confirm this by the following result.
For any irreducible polynomials $P, Q \in {{\mathbb A}}$ with $ \deg Q > 1 + \deg P$, there exist positive integers $e_{1}$ and $e_{2}$ such that $P^{e_{1}}>Q^{e_{2}}$ but $S(P^{e_{1}})<S(Q^{e_{2}})$.
For simplicity, denote $d_{1} = \deg P$ and $d_{2} = \deg Q$, and put $b_{j}=\frac{q^{jd_{1}}-1}{q^{d_{1}}-1},j\in{{\mathbb N}}^* $. Since $1\leq d_{1}<d_{2}$, by the division algorithm, there exist $k,r\in{{\mathbb N}}$ such that $$d_{2}= kd_{1}+r,$$ where $k \geq 1$ and $ 0\leq r < d_{1}$. By assumption, we have $d_2 - d_1 \ge 2$.
We first assume $r \neq 0$. Take $e_{1}= b_1 + q^{r-1}b_{k}$ and $e_{2}=1$, then $e_{1} \geq k +1$. So, using Proposition \[prop:comp2\] we have $$\delta(P^{e_{1}})\geq q^{d_{1}e_{1}}\geq q^{d_{1}(k +1)}=q^{d_{2}+d_{1}-r}\geq q^{d_{2}+1}>\delta(Q^{e_{2}})$$ and $$\delta(S(P^{e_{1}}))= q^{d_1} + q^{r-1}q^{kd_{1}}=q^{d_1} + q^{d_2 - 1} < q^{d_{2}} = \delta(S(Q^{e_{2}})).$$ Hence, $P^{e_{1}}>Q^{e_{2}}$ but $S(P^{e_{1}})<S(Q^{e_{2}})$.
We now assume $r=0$. Then $k \geq 2$. We take $e_{1}= b_1+b_{k - 1} + b_{k}$ and $e_{2}=2$, then $e_{1} \geq 2k + 1$. So, using Proposition \[prop:comp2\] we deduce that $$\delta(P^{e_{1}})\geq q^{d_{1}e_{1}}\geq q^{d_{1}(2k +1)}=q^{2d_{2}+d_{1}}\geq q^{2d_{2}+1}>\delta(Q^{e_{2}})$$ and $$\delta(S(P^{e_{1}}))=q^{d_{1}} + q^{d_{1}(k-1)} + q^{kd_{1}}=q^{d_{1}} + q^{d_{2}-d_{1}} + q^{d_{2}}< 2q^{d_{2}} = \delta(S(Q^{e_{2}})).$$ Hence, $P^{e_{1}}>Q^{e_{2}}$ but $S(P^{e_{1}})<S(Q^{e_{2}})$. This completes the proof.
We remark that by Proposition \[prop:comp2\], for any irreducible polynomials $P, Q \in {{\mathbb A}}$ and any positive integer $e$, if $\deg P=\deg Q$, then $S(P^{e})=S(Q^{e})$.
Fixed points
------------
For any $f\in {{\mathbb A}}$, if $S(f)=f$, then we call $f$ a fixed point of $S$. We first determine the fixed points of the Smarandache function $S$.
Given a non-zero polynomial $f \in {{\mathbb A}}$, $ f$ is a fixed point of the Smarandache function $S$ if and only if $$\begin{split}
f
& = \left\{\begin{array}{ll}
t, & \textrm{if $q>2$,}\\
\textrm{$t$ or $t^2$}, & \textrm{if $q=2$.}
\end{array}
\right.
\end{split}$$
If $ f $ is a fixed point, then $S(f)= f$, and by Proposition \[prop:basic\] (1), (2) and Proposition \[prop:comp1\], we must have $f=t^{e},e\in{{\mathbb N}}^*$. So, by definition we obtain the desired result. Indeed, by the definition of factorial (Definition \[def:factorial\]), we have that $t^e \mid t^{e-1}!$ if $e>2$; and if $q>2$, then $t^2 \mid (a_2 t)!$.
We remark that in the integer case all the prime numbers are fixed points of the Smarandache function.
For any integer $n \ge 1$, let $S^{(n)}$ be the $n$-th iteration of $S$. It is easy to see that for any $f \in {{\mathbb A}}$ with $\deg f \ge 1$ there exists some integer $n$ such that $S^{(n)}(f)$ is a fixed point of $S$. We now estimate the number of iterations, which can be viewed as the distance to fixed points.
\[prop:dist\] For any $f \in {{\mathbb A}}$ with $\deg f \ge 1$, there exists a positive integer $n \le 1 + \deg f$ such that $S^{(n)}(f)$ is a fixed point of $S$.
We first note that by Proposition \[prop:val-range\], for any polynomial $f \in {{\mathbb A}}$ with $\deg f \ge 1$, we have $S(f) \in t{{\mathbb A}}$, and so, if $\deg S(f) \ge 2$, then $S(f)$ must be a reducible polynomial.
Now, given $f \in {{\mathbb A}}$ with $\deg f \ge 1$, if $\deg S(f) \ge 3$, then $S(f)$ satisfies the condition in Proposition \[prop:delta-Kf\], and so $$\delta(S^{(2)}(f)) \le \frac{\delta(S(f))}{q}.$$ If again $\deg S^{(2)}(f) \ge 3$, we have $$\delta(S^{(3)}(f)) \le \frac{\delta(S^{(2)}(f))}{q} \le \frac{\delta(S(f))}{q^2}.$$ This process stops when we reach $\deg S^{(j)}(f) \le 2$ for some integer $j$. So, this integer $j$ satisfies $$\delta(S^{(j)}(f)) < q^3.$$ This automatically holds if $$\frac{\delta(S(f))}{q^{j-1}} < q^3,$$ which, together with $\delta(S(f)) \le q^{\deg f}$ by Proposition \[prop:basic\] (2), is implied in $$q^{\deg f - j +1} < q^3.$$ Thus $j \ge \deg f - 1$.
If $\deg S^{(j)}(f) = 2$, since $S^{(j)}(f)$ is reducible, we have that $S^{(j+2)}(f)$ must be a fixed point of $S$.
Then, we always have that $S^{(j+2)}(f)$ is a fixed point of $S$. Hence, for $n = 1 + \deg f$, $S^{(n)}(f)$ is a fixed point of $S$.
We remark that the upper bound in Proposition \[prop:dist\] is optimal. Because when $q\ge 3$, $$S^{(3)}(f) = S^{(2)}(t^2) = S(a_2 t) =t$$ for any irreducible polynomial $f \in {{\mathbb A}}$ of degree $2$.
Analogue of Erd[ő]{}s’s problem {#sec:Erdos}
===============================
In this section, we establish an analogue of Erd[ő]{}s’s problem.
For any non-constant polynomial $f \in {{\mathbb A}}$, let ${{\mathcal P}}(f)$ be the maximal monic irreducible factor of $f$. Following Erd[ő]{}s’s problem, for any integer $n \ge 1$, we define the subset of ${{\mathbb A}}$: $${{\mathcal T}}(n) = \{\textrm{monic }f \in {{\mathbb A}}:\, \deg f=n, S(f) \ne t^{\deg {{\mathcal P}}(f)} \}.$$ One should note that $S({{\mathcal P}}(f)) = t^{\deg {{\mathcal P}}(f)}$.
Using the strategy in [@Akbik], which is in fact a classical approach by considering the number of distinct prime factors and the maximal prime factor, we establish the following analogue of Erd[ő]{}s’s problem.
\[thm:Erdos\] For the sets ${{\mathcal T}}(n)$, we have $$\begin{split}
|{{\mathcal T}}(n)| <
& \left\{\begin{array}{ll}
2^{n+2} \exp(-\sqrt{n}/3), & \textrm{if $q=2$ and $n \ge 3249$,}\\
4q^n \exp(-\sqrt{n}/2), & \textrm{if $q \ge 3$ and $n \ge \max\{13000, \log q\}$.}
\end{array}
\right.
\end{split}$$
Theorem \[thm:Erdos\] implies that for almost all polynomials $f \in {{\mathbb A}}$, $S(f)=t^d$, where $d$ is the maximal degree of the irreducible factors of $f$.
Recall the sets ${{\mathcal T}}_1(n,r), {{\mathcal T}}_2(n,r), {{\mathcal T}}_3(n,r)$ defined in Section \[sec:count\]. To prove Theorem \[thm:Erdos\], we need one more preliminary result.
\[lem:S4\] For any integer $n \ge 1$ and any real $r \ge 1$, define $${{\mathcal T}}_4(n,r) = {{\mathcal T}}(n) \setminus ({{\mathcal T}}_1(n,r) \cup {{\mathcal T}}_2(n,r) \cup {{\mathcal T}}_3(n,r)).$$ Then, for any $f \in {{\mathcal T}}_4(n,r)$ we have $$\deg {{\mathcal P}}(f) < D + \frac{\log D}{\log q}.$$
For any $f \in {{\mathcal T}}_4(n,r)$, we have $S(f) \ne t^{\deg {{\mathcal P}}(f)}$, which implies that $f \nmid t^{\deg {{\mathcal P}}(f)}!$. So, there exists a monic irreducible polynomial $P$ such that $P \mid f$ and $v_P(f) > v_P(t^{\deg {{\mathcal P}}(f)}!)$. Note that $v_P(t^{\deg {{\mathcal P}}(f)}!) \ge 1$ by definition. So, we have $v_P(f) \ge 2$. Then, in view of $f \not\in {{\mathcal T}}_2(n,r)$ and $f \not\in {{\mathcal T}}_3(n,r)$, we must have $\deg P \le D$ and $v_P(f) < D$.
Hence, using Lemma \[lem:val\] we obtain $$\begin{aligned}
D > v_P(f) > v_P(t^{\deg {{\mathcal P}}(f)}!) \ge \frac{q^{\deg {{\mathcal P}}(f)}}{q^{\deg P}}
\ge \frac{q^{\deg {{\mathcal P}}(f)}}{q^D},\end{aligned}$$ which gives the desired result.
We are now ready to prove Theorem \[thm:Erdos\].
For any $f \in {{\mathcal T}}_4(n,r)$, we have $f \not\in {{\mathcal T}}_1(n,r) \cup {{\mathcal T}}_2(n,r) \cup {{\mathcal T}}_3(n,r)$. So, $\omega(f) \le B = 3r \log\log q^n$, and also, if $P^e, e \ge 1,$ is any positive power of a monic irreducible polynomial $P$ such that $P^e \mid f$, then we only have two cases:
- $\deg P \le D, e < D$,
- $\deg P > D, e = 1$,
where $D=2r\log\log q^n$. Case (i) yields at most $\lfloor Dq^D \rfloor$ positive powers of monic irreducible polynomials (using Lemma \[lem:count-irre\]). For Case (ii), since $\deg P \le \deg {{\mathcal P}}(f) < D + \log D / \log q$ by Lemma \[lem:S4\], it also gives at most $\lfloor Dq^D \rfloor$ positive powers of monic irreducible polynomials. Hence, the number of possible powers of monic irreducible polynomials which divides an $f \in {{\mathcal T}}_4(n,r)$ is not greater than $2Dq^D$. However, such an $f$ is the product of at most $B = 3r \log\log q^n$ distinct powers of monic irreducible polynomials. Hence, we have $$\label{eq:S4}
\begin{split}
|{{\mathcal T}}_4(n,r)|
& \le (2Dq^D)^{B} \\
& = (4r \log\log q^n)^{3r \log\log q^n} \cdot q^{6 (r\log\log q^n)^2} \\
& \le q^{7(r\log\log q^n)^2}
\end{split}$$ when $r \log\log q^n \ge 19$. Then, assuming moreover $n \ge 3$ and $r \ge 2$ and using Lemmas \[lem:S1\], \[lem:S2\] and \[lem:S3\], we obtain $$\label{eq:S}
\begin{split}
|{{\mathcal T}}(n)| & \le |{{\mathcal T}}_1(n,r)| + |{{\mathcal T}}_2(n,r)| + |{{\mathcal T}}_3(n,r)| + |{{\mathcal T}}_4(n,r)| \\
& < \frac{3q^n}{(\log q^n)^r} + q^{7(r\log\log q^n)^2}.
\end{split}$$ Now, choosing $$r = \frac{\sqrt{n}}{3\log\log q^n},$$ we obtain $$|{{\mathcal T}}(n)| < 4q^n \exp(-\sqrt{n}/3)$$ when $n \ge 3249$ and $n \ge (6\log\log q^n)^2$. Here the condition on $n$ comes from $\sqrt{n}/3 = r \log\log q^n \ge 19$ and $r \ge 2$.
If $q = 2$, we have $$|{{\mathcal T}}(n)| < 2^{n+2} \exp(-\sqrt{n}/3)$$ when $n \ge 3249$.
We now assume $q \ge 3$. In this case, assuming $n \ge 4, r \ge 3$ and $r\log \log q^n \ge 19$ and using Lemmas \[lem:S1\], \[lem:S2\] and \[lem:S3\], we can choose $B = 2r \log \log q^n$ and $D = r\log \log q^n$, and then becomes $$\begin{split}
|{{\mathcal T}}_4(n,r)|
& \le (2Dq^D)^{B} \\
& = (2r \log\log q^n)^{2r \log\log q^n} \cdot q^{2 (r\log\log q^n)^2} \\
& \le q^{3 (r\log\log q^n)^2}.
\end{split}$$ So, becomes $$|{{\mathcal T}}(n)| < \frac{3q^n}{(\log q^n)^r} + q^{3(r\log\log q^n)^2}.$$ Now, choosing $$r = \frac{\sqrt{n}}{2\log\log q^n},$$ we obtain $$|{{\mathcal T}}(n)| < 4q^n \exp(-\sqrt{n}/2)$$ when $n \ge 1444$ and $n \ge (6\log\log q^n)^2$. Here the condition on $n$ comes from $\sqrt{n}/2 = r\log \log q^n \ge 19$ and $r \ge 3$. Finally, it is easy to see that if $$n \ge \max\{13000, \log q\},$$ then the condition $n \ge (6\log\log q^n)^2$ holds. This completes the proof.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors thank the referee for careful reading of the paper and valuable comments. They also thank George Martin for valuable comments. The first author was supported by the National Science Foundation of China Grant No. 11526119 and the Scientific Research Foundation of Qufu Normal University No. BSQD20130139, and the second author was supported by the Australian Research Council Grant DE190100888.
[99]{}
S. Akbik, *On a density problem of Erd[ő]{}s*, Internat. J. Math. Math. Sci. **22** (1999), 655–658.
M. Bhargava, *P-orderings and polynomial functions on arbitrary subsets of Dedekind rings*, J. Reine Angew. Math. **490** (1997), 101–127.
L. Carlitz, *On certain functions connected with polynomials in a Galois field*, Duke Math. J. **1** (1935), 137–168.
Z. Chen, *On polynomial functions from ${{\mathbb Z}}_n$ to ${{\mathbb Z}}_m$*, Discrete Math. **137** (1995), 137–145.
J.-M. De Koninck and N. Doyon, *On a thin set of integers involving the largest prime factor function*, Internat. J. Math. Math. Sci. **19** (2003), 1185–1192.
P. Erd[ő]{}s, *Problem 6674*, Amer. Math. Monthly. **98** (1991), 965.
K. Ford, *The normal behavior of the Smarandache function*, Smarandache Notions J. **10** (1999), 81–86.
N. Hungerb[ü]{}hler and E. Specker, *A generalization of the Smarandache function to several variables*, Integers: Electronic J. Combin. Number Th. **6**, A23, 2006.
A. Ivi[' c]{}, *On a problem of Erd[ő]{}s involving the largest prime factor of $n$*, Monatsh. Math. **145** (2005), 35–46.
I. Kastanas, *Solution to Problem 6674*, Amer. Math. Monthly **101** (1994), 179.
A.J. Kempner, *Concerning the smallest integer $m!$ divisible by a given integer $n$*, Amer. Math. Monthly **25** (1918), 204–210.
A.J. Kempner, *Polynomials and their residue systems*, Trans. Amer. Math. Soc. **22** (1921), 240–288.
X. Li and M. Sha, *Gauss factorials of polynomials over finite fields*, Int. J. Number Theory **8** (2017), 2039–2054.
X. Li and M. Sha, *Polynomial functions in the residue class ring of Dedekind domain*, Int. J. Number Theory **7** (2019), 1473–1486.
X. Li and M. Sha, *Congruence preserving functions in the residue class rings of polynomials over finite fields*, preprint, 2018, available at <https://arxiv.org/abs/1807.02379>.
H. Liu, *A survey on Smarandache notions in number theory I: Smarandache function*, Scientia Magna, 12 (2017), 132–144.
F. Luca and J. S[' a]{}ndor, *On the composition of a certain arithmetic function*, Funct. Approx. Comment. Math. **41** (2009), 185–209.
E. Lucas, *Question Nr. 288*, Mathesis **3** (1883), 232.
J. Neuberg, *Solutions de questions propos[é]{}es, Question Nr. 288*, Mathesis **7** (1887), 68–69.
M. Rosen, *Number Theory in Function Fields*, GTM **210**, Springer Verlag, New York, 2002.
F. Smarandache, *A function in the number theory*, Analele Univ. Timisoara, Ser. St. Math., Vol. **XVIII** (1980), Fasc. 1, 79–88.
J. Sondow, *A geometric proof that $e$ is irrational and a new measure of its irrationality*, Amer. Math. Monthly **113** (2006), 637–641.
D.S. Thakur, *Binomial and factorial congruences for ${{\mathbb F}}_q[t]$*, Finite Fields Th. App. **18** (2012), 271–282.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Dong Bai
- 'and Yue-Liang Wu'
title: Quadratic Contributions of Softly Broken Supersymmetry in the Light of Loop Regularization
---
Introduction {#Introduction}
============
With the advent of the Higgs boson with a mass of approximately 125 GeV at LHC [@Aad:2012tfa; @Chatrchyan:2012xdj], the last missing piece of the Standard Model (SM) has been found. After hard workings for more than fifty years, eventually we have gotten a mathematically consistent theory at hand, which provides unprecedented agreements with numerous experiments up to the TeV scale.
In SM, the mass of the Higgs is a free parameter, and at the quantum level it receives large contributions from the ultraviolet (UV) physics at some UV scale (say, $M_c$), due to the presence of quadratic contributions in the Higgs self-energy diagrams. In this article, we prefer to use the concept “UV contribution" to refer to what is traditionally called “UV divergence”, as the former is more compatible with the modern effective-field-theory approach to quantum field theories suggested by K. G. Wilson, in which all quantum field theories are defined at some physical UV scale, and the infrared (IR) theories could be obtained from the UV theories by doing renormalization-group transformations [@Wilson:1973jj].
At the one-loop level [@Veltman:1980mj], the effective Higgs mass parameter $m_H^2(M_c/\mu)$ at the low-energy scale $\mu$ is related to the UV parameter $m_H^2(M_c)$ at the UV scale $M_c$ by $$\begin{aligned}
&m_H^2(M_c/\mu)=m_H^2(M_c)-\frac{6}{(4\pi)^2}\left(y^2_t-\frac{1}{4}\lambda_H-\frac{1}{8}g^2_1-\frac{3}{8}g^2_2\right)\left(M_c^2-\mu^2\right)\nonumber\\
&+\mathrm{\ logarithmic\ contributions},
\end{aligned}$$ where $y_t$, $g_1$, $g_2$, $\lambda_H$ are the top-quark Yukawa coupling, the $U(1)_\text{Y}$ gauge coupling, the $SU(2)_{\text{L}}$ gauge coupling and the Higgs quartic coupling, respectively. Here we have ignored the contributions from the rest particles in SM, since their couplings to the Higgs boson are much weaker. It has been shown recently in Ref. [@Bai:2014lea] that, quadratic contributions from the SM Higgs sector can induce spontaneous electroweak symmetry breaking at $\Lambda_{\text{EW}}\simeq 750$ GeV, [ given the SM parameters measured at the low energy as boundary conditions for the renormalization group equations.]{} Such a mechanism is dubbed as quantum electroweak symmetry breaking, as quadratic contributions that play a significant role come from quantum loop effects. The symmetry breaking scale $\Lambda_{\text{EW}}\simeq 750$ GeV could then be treated as another fundamental scale of SM besides the electroweak scale $v=246$ GeV.
[Although SM is a unprecedented triumph of human intelligence, it is generally believed that SM is certainly not the last words we can say about nature. And right now, the most urgent question that confronts us is: *What is the characteristic scale for new physics?* In literature, this is often referred to as the *gauge hierarchy problem*. At present, one of the leading candidates for new physics beyond SM is supersymmetry (SUSY). SUSY introduces bosonic/fermionic partners for each SM fermionic/bosonic particle, and puts stringent constraints on their properties. To describe the real nature, SUSY has to be broken at the low energy. In the high-energy phenomenological studies, usually this is achieved by introducing soft-SUSY-breaking terms into the supersymmetric Lagrangian by hand. The state-of-art constraints on SUSY in the real world could be found in Ref. [@Olive:2016xmw]. Besides the doubling of particle species, SUSY and softly broken SUSY are also characterized by other novel properties, among which the most important one is the absence of quadratic contributions. Traditionally, this issue is handled by the supergraph technique [@Girardello:1981wz]. Although elegant and powerful, supergraph techniques are quite baroque and less useful in phenomenological studies, where, instead, the traditional Feynman diagrammatic calculations are more relevant, and dimensional regularization (DREG) [@tHooft:1972tcz] and dimensional reduction (DRED) [@Siegel:1979wq] are usually adopted to redefine UV divergent Feynman integrals.[^1] However, it is well-known that DREG and DRED cannot track quadratic contributions in the $4-\epsilon$ dimension, which makes them less convenient in studying theoretical aspects of softly broken SUSY such as the aforementioned absence of quadratic contributions.]{}
[In this article, we would like to convince the readers that loop regularization (LORE) proposed in Ref. [@Wu:2002xa; @Wu:2003dd] is an ideal regularization scheme in studying both theoretical and phenomenological properties of SUSY and softly broken SUSY. LORE is believed to be able to preserve various symmetries, including Poincare symmetry, gauge symmetry, SUSY, etc,[^2] and has already been applied in several studies, such as the one-loop renormalization of Non-Abelian gauge theories [@Cui:2008uv], the study of composite Higgs model [@Dai:2003ip], the gravitational corrections to the running of gauge couplings [@Tang:2008ah; @Tang:2010cr; @Tang:2011gz], the renormalization of supersymmetric field theories [@Cui:2008bk], the trace anomaly in quantum electrodynamics (QED) [@Cui:2011za], the diphoton channel of the Higgs decay [@Huang:2011yf], the quadratic running of the effective Higgs mass parameter. In Ref. [@Huang:2011xh; @Huang:2012iu], LORE has been used to calculate two-loop quantum corrections of the $\lambda\phi^4$ theory and QED. In a recent review article [@Wu:2013iga], one of the authors (YLW) makes a comprehensive review of the underlying philosophy and application scenarios of LORE. Noticeably, LORE provides not only useful tools for physicists to study quantum field theories, but also new challenges for mathematicians. Recently, Ref. [@Chapling:2016kpi; @Chapling:2016sfj] prove the three conjectures concerning the asymptotics of sums of products of binomials, powers and logarithms suggested in Ref. [@Wu:2002xa; @Wu:2003dd], and propose closed-form expressions for Irreducible Loop Integrals (ILIs), which are building blocks of LORE. These studies show that LORE is applicable in arbitrary spacetime dimensions, which makes it suitable in studying quantum loop effects of gravitational gauge field theory in six-dimensional spacetime [@Wu:2017xzu] and the unified field theory of basic forces and elementary particles with gravitational origin of gauge symmetry in nineteen-dimensional hyper-spacetime[@Wu:2017rmd]. To demonstrate the power of LORE, we calculate in the later parts of this article the two-loop quadratic contributions of WZ’ model, i.e., Wess-Zumino model with soft SUSY breaking terms, and show explicitly the cancellation of all quadratic contributions.]{}
The rest parts of this article are organized as follows: In Section 2, we give a practical introduction to LORE. In Section 3, we use LORE to calculate quadratic contributions in WZ’ model, up to two loops. In Section 4, we conclude with some final remarks. We also include several Appendices to provide some technical details. Besides the aforementioned motivation to provide new tools to study quadratic contributions in SUSY phenomenology, the results of this article could be interpreted further as follows. As mentioned in Footnote \[Preservation\], the SUSY preservation of LORE has been verified at one loop by checking Slavnov-Taylor identities directly [@Cui:2008bk]. It is desirable to verify explicitly the applicability of LORE to SUSY at two loops and beyond. As complete two-loop calculations are quite complicated, it is wise to first verify some important features of SUSY and softly broken SUSY, such as the absence of quadratic contributions, and approach the final goal step by step.
A Practical Guide to LORE {#LORE}
=========================
In this section, we shall give a practical introduction to LORE. The viewpoint adopted here is slightly different from Ref. [@Wu:2002xa; @Wu:2003dd]. And we mainly concentrate on how to use LORE to do the realistic calculations in phenomenology. Readers who want to know more about LORE are recommended to go to Ref. [@Wu:2013iga] for a more comprehensive introduction.
General Features
----------------
In this subsection, we would like to provide a comparison between LORE and other regularization schemes in literature, making clear the differences between them. The common regularization schemes include sharp cut-off regularization, Pauli-Villars regularization [@Pauli:1949zm], higher derivative regularization [@Iliopoulos:1974zv], DREG, DRED, etc.[^3] The philosophy underlying these regularization schemes can be summarized as follows:
1. First, the Lagrangians of quantum field theories have to be modified in some ways. In the sharp cut-off regularization, the Lagrangians are reformulated on a discrete lattice; in the Pauli-Villars regularization, extra Pauli-Villars ghost fields and interaction vertices are added into the Lagrangian; in the higher derivative regularization, higher derivative interactions are added in; in DREG and DRED, the Lagrangians are reformulated in the $D=4-\epsilon$ spacetime, rather than the ordinary $D=4$ Minkowski spacetime, and one needs to introduce the extra fields called $\epsilon$-scalars when using DRED to regularize gauge field theories.
2. The Feynman rules of the Lagrangian then have to be modified correspondingly. In the sharp cut-off regularization, the integrals over loop momenta are cut off at some energy scale; in Pauli-Villars regularization, extra propagators and vertices of the Pauli-Villars ghosts are introduced; in the higher derivative regularization, no extra fields are introduced, but the propagators of the existing fields are modified; in the DREG and DRED, the dimension of the momentum integrals are changed from 4 to $4-\epsilon$, and new propagators and interaction vertices are needed where the $\epsilon$-scalars mentioned above appear. Also, when using DREG and DRED to handle models containing chiral fermions, one needs extra rules to manipulate the Levi-Civita symbol $\epsilon_{\mu\nu\rho\sigma}$ and $\gamma^5$, which can lead to mathematical inconsistencies [@Siegel:1980qs].
3. Finally, the Feynman integrals corresponding to the Feynman diagrams have to be reformulated using the new Feynman rules derived above. For the regularization schemes mentioned above, at this stage, all the Feynman integrals become finite and thus mathematically well-defined.
The above steps are summarized diagrammatically in Fig. \[MediationScenarioOfRegularization\]. In the following, we shall call this way to construct regularization schemes the *mediation scenario*, just to emphasize the role played by Feynman rules in transmitting the modifications to the Lagrangians into the target Feynman integrals, and thereby make them mathematically well-defined.
![Mediation Scenario for the Regularization of Feynman Integrals[]{data-label="MediationScenarioOfRegularization"}](MediationScenarioOfRegularization.jpg)
LORE is different from all the aforementioned regularization schemes in the sense that its underlying philosophy is different. Instead of utilizing Feynman rules as messengers, it redefines the divergent Feynman integrals directly. So, when using LORE we do not need to modify either the Lagrangians or the Feynman rules directly. We don’t need any unphysical fields, any extra vertices, any modifications to the propagators of the existing fields, or any departures from the ordinary spacetime structures. The essential reason for the fact that LORE preserves SUSY is that SUSY can be regarded as a spacetime symmetry as it extends the usual Minkowski spacetime into the so-called superspace which contains not only the usual commutative coordinates but also anti-commutative coordinates. As emphasized above, unlike the popular DREG and DRED, LORE does not change the spacetime structure at all. So it is physically straightforward to see that LORE should preserve SUSY perfectly.
[For a general divergent Feynman integral, it is shown in Ref. [@Huang:2011xh; @Huang:2012iu] that the structures of UV contributions can be extracted in the following way $$I=I_\text{FP} \otimes I_\text{UVDP} \otimes I_\text{ILI},
\label{LOREstructure}$$ by exploiting Bjorken-Drell’s analogy between Feynman diagrams and electrical circuits [@Bjorken:1965zz].]{} Here, $I$ stands for the divergent Feynman integral, $I_\text{FP}$ for integrals over Feynman parameters, $I_\text{UVDP}$ for integrals over the ultraviolet-divergence-preserving (UVDP) parameters introduced in Ref. [@Huang:2011xh; @Huang:2012iu], and $I_\text{ILI}$ for the ILIs. The $\otimes$ operation here is introduced for heuristic reasons, and $I_x \otimes I_y$ roughly means $\int\!\mathrm{d}x I_x\!\int\!\mathrm{d}y I_y$. Generally, $I_y$ depends on both $x$ and $y$. The following discussions do not rely on the precise definition of the operator $\otimes$. Before giving explicit definitions of these concepts, we want to emphasize some general features of $I_\text{FP}$, $I_\text{UVDP}$, $I_\text{ILI}$ first [@Huang:2011xh; @Huang:2012iu]:
1. Generally, the UV-contribution structures of the Feynman integral $I$ are encoded entirely in $I_\text{UVDP}$ and $I_\text{ILI}$. In other words, the Feynman parameter integrals $I_\text{FP}$ contain no UV contributions.
2. The overall UV contribution of $I$ is solely encoded in $I_\text{ILI}$. If the Feynman integral $I$ has any tensor structure, the tensor structure is also encoded entirely in $I_\text{ILI}$.
3. The UV subcontributions in $I$ are encoded in the UVDP integrals $I_\text{UVDP}$. And it is shown that there is a one-to-one correspondence between UV subcontributions (UV subdivergences) in $I$ and the UV contributions in $I_\text{UVDP}$.
For the one-loop Feynman integrals, there is no UV subcontribution in the Feynman integrals. Thus, in this case we do not need UVDP integrals, and Eq. can be simplified to $$I=I_\text{FP} \otimes I_\text{ILI}.
\label{LOREstructure1loop}$$
ILI and LORE
------------
Let’s start with the definitions of ILIs (i.e. irreducible loop integrals): $$I_{-2\alpha}=\int\!\mathrm{d}^4k \frac{1}{(k^2-\mathcal{M}^2)^{2+\alpha}}.$$ Here $\alpha=-1, 0, 1, 2,\cdots$, and the number ($-2\alpha$) in the subscript labels the superficial degrees of UV contributions of ILIs. $\mathcal{M}$ is generally a function of the physical masses $m_i$, external momenta $p_i$, and other parameters introduced during the calculation, e.g. Feynman parameters $x_i$. Formally, one has $\mathcal{M}=\mathcal{M}(m_i, p_i, x_i, \cdots)$.
One can also introduce extra tensor structures into ILIs: $$\begin{aligned}
&I_{-2\alpha}^{\mu\nu}=\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}}{(k^2-\mathcal{M}^2)^{3+\alpha}},\\
&I_{-2\alpha}^{\mu\nu\rho\sigma}=\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}k^{\rho}k^{\sigma}}{(k^2-\mathcal{M}^2)^{4+\alpha}},\end{aligned}$$ etc. Here $\alpha=-1, 0, 1, 2,\cdots$. These tensor type ILIs are common when doing calculations in gauge field theories.
As mentioned in Section 2.2, at the one-loop level, one can always decompose a Feynman integral $I$ into a “product" of $I_\text{FP}$ and $I_\text{ILI}$. Let’s take the following example to see how this kind of decomposition comes into being.
: Decompose $I=\int\!\mathrm{d}^4k\frac{1}{k^2-m^2_1}\frac{1}{(k+p)^2-m^2_2}$ into the form $I_\text{FP}\otimes I_\text{ILI}$.
Using the standard Feynman parametrization, one can easily show that $$\begin{aligned}
&\:\int\!\mathrm{d}^4k\frac{1}{k^2-m^2_1}\frac{1}{(k+p)^2-m^2_2}\nonumber\\
=&\int^1_0\!\mathrm{d}x\int\!\mathrm{d}^4k\frac{1}{\{(1-x)(k^2-m^2_1)+x[(k+p)^2-m^2_2]\}^2}\nonumber\\
=&\int^1_0\!\mathrm{d}x\int\!\mathrm{d}^4k\frac{1}{\{(k+xp)^2-[(1-x)m^2_1+xm^2_2-x(1-x)p^2]\}^2}\nonumber\\
=&\underbrace{\int^1_0\!\mathrm{d}x}_{I_\text{FP}}\underbrace{\int\!\mathrm{d}^4k\frac{1}{(k^2-\mathcal{M}^2)^2}}_{I_\text{ILI}}.\end{aligned}$$ In the last step, one uses the variable shift[^4] $k\to k-xp$ and $\mathcal{M}^2=(1-x)m^2_1+xm^2_2-x(1-x)p^2$.
In the above, we have introduced the concepts of ILIs and shown explicitly that at the one-loop level, one can decompose Feynman integrals into “products" of $I_\text{FP}$ and $I_\text{ILI}$. As emphasized in Section 2.1, the overall UV contribution and the tensor structure of Feynman integrals should always be encoded in $I_\text{ILI}$. And in LORE, in order to regularize the divergent Feynman integral, one has to give proper redefinitions to ILIs. These redefinitions could be found in Ref. [@Wu:2002xa; @Wu:2003dd], and are reproduced in the following: $$\begin{aligned}
&\underline{Prescription\ 1}:\\
&\int\!\mathrm{d}^4k\frac{1}{k^2-\mathcal{M}^2}:=-i\pi^2\left\{M^2_c-(\mathcal{M}^2+\mu^2_s)\left[\ln\frac{M^2_c}{\mathcal{M}^2+\mu^2_s}-\gamma_{\omega}+1+y_2\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right]\right\},\\
&\underline{Prescription\ 2}:\\
&\int\!\mathrm{d}^4k\frac{1}{(k^2-\mathcal{M}^2)^{n+1}}:=\frac{1}{n}\frac{\partial}{\partial\mathcal{M}^2}\int\!\mathrm{d}^4k\frac{1}{(k^2-\mathcal{M}^2)^n},\quad \quad \quad(n\geq 1)\\
&\underline{Prescription\ 3}\ (Consistency\ Conditions):\\
&I^{\mu\nu}_{-2\alpha}:=\frac{1}{2(\alpha+2)}g^{\mu\nu}I_{-2\alpha}, \quad\ I^{\mu\nu\rho\sigma}_{-2\alpha}:=\frac{1}{4(\alpha+2)(\alpha+3)}(g^{\mu\nu}g^{\rho\sigma}+g^{\mu\rho}g^{\nu\sigma}+g^{\mu\sigma}g^{\nu\rho})I_{-2\alpha}.\\
&\text{Here $\alpha=-1, 0, 1, 2, \cdots$ For $\alpha=-1\ or\ 0$, one has}\\
&I^{\mu\nu}_2:=\frac{1}{2}g^{\mu\nu}I_2, \qquad \qquad \qquad I^{\mu\nu\rho\sigma}_2:=\frac{1}{8}(g^{\mu\nu}g^{\rho\sigma}+g^{\mu\rho}g^{\nu\sigma}+g^{\mu\sigma}g^{\nu\rho})I_2,\\
&I^{\mu\nu}_0:=\frac{1}{4}g^{\mu\nu}I_0, \qquad \qquad \qquad I^{\mu\nu\rho\sigma}_0:=\frac{1}{24}(g^{\mu\nu}g^{\rho\sigma}+g^{\mu\rho}g^{\nu\sigma}+g^{\mu\sigma}g^{\nu\rho})I_0.\end{aligned}$$ Here $M_c$ acts as the UV scale, while $\mu_s$ acts as the infrared (IR) scale. For those theories that are free of IR divergences, $\mu_s$ can be safely set to zero. $\gamma_{\omega}$ equals the Euler constant $\gamma_E$. $y_2(x)=\int^x_0\!\mathrm{d}\sigma\frac{1-e^{-\sigma}}{\sigma}+\frac{1}{x}(1-x-e^{-x})$, and it can be easily shown that when $x\to 0$, $y_2(x)\to 0$. In other words, when $M_c\to \infty$, $y_2$ function in Prescription 1 goes to zero. Prescription 3 plays an extremely important role in preserving gauge symmetries in regularization schemes. It is shown in great details in Ref. [@Wu:2002xa; @Wu:2003dd] that a regularization scheme can preserve gauge symmetries only if these consistency conditions are satisfied. The sharp cut-off regularization scheme, which is well-known to break gauge symmetries, does not satisfy these consistency conditions, while DREG and DRED do.
By using the above prescriptions, one can derive explicit expressions for ILIs other than $I_2$, and many of the useful results are summarized in Appendix. Before moving on to the multi-loop calculations, we want to emphasize the following two points:
1. The above treatment of LORE is practically oriented, and we aim to explain how to use LORE to do realistic calculations. Although many results might seem ad hoc for some readers, for instance, we explain neither why $I_2$ should be defined as that in Prescription 1, nor where the non-trivial $y_2$ function comes from, these are actually well-motivated and we recommend Ref. [@Wu:2002xa; @Wu:2003dd] for further details.
2. In LORE, one can track quadratic contributions along with logarithmic contributions at the same time. This can be seen explicitly in Prescription 1, where the integral on the left-hand side, i.e., $I_2$, is quadratic divergent. This is actually highly nontrivial, when taking into consideration that LORE preserves also gauge symmetries. In the popular DREG and DRED, which preserve gauge symmetries as well, quadratic contributions can only be extracted by tracing pole structures of the Feynman integrals at dimensions lower than 4. Practically, this means that one has to carry out a separate calculation to extract quadratic contributions.
We can also extend the above treatment to multi-loop calculations, although often much more complicated due to the appearance of UV subcontributions. As mentioned in Section 2.1, these subcontributions can be fully captured in the UVDP integrals $I_{UVDP}$, and we recommend Ref. [@Huang:2011xh; @Huang:2012iu] for a discussion of what UVDP integrals are and how to extract subcontributions from it. In this articles, we shall continue taking a practical viewpoint without going into complicated details. Here, instead of doing the decomposition shown in Eq. and regularize the divergent subintegrals one by one, we shall try to give out the final results directly in a way that they can be used repeatedly in practical calculations. The key observation comes from the fact that $I_\text{UVDP}\otimes I_\text{ILI}$ actually comprises the so-called $\alpha\beta\gamma$ integrals introduced by ‘t Hooft and Veltman [@tHooft:1972tcz]. $$I_\text{UVDP}\otimes I_\text{ILI}=I_{\alpha\beta\gamma}.$$ For the two-loop scalar Feynman integrals, a typical $I_{\alpha\beta\gamma}$ is given by[^5] $$I_{\alpha\beta\gamma}=\int\!\mathrm{d}^4l_1\!\int\!\mathrm{d}^4l_2\frac{1}{(l^2_1-m^2_1)^{\alpha}}\frac{1}{(l^2_2-m^2_2)^{\beta}}\frac{1}{[(l_1+l_2+p)^2-m^2_3]^{\gamma}} .$$ So, instead of Eq. , one has $$I=I_\text{FP}\otimes I_{\alpha\beta\gamma}.$$ Ref. [@Huang:2011xh; @Huang:2012iu] contain comprehensive discussions on calculating the $\alpha\beta\gamma$ integrals with the help of the Bjorken-Drell’s electrical-circuit analogy, and show explicitly the one-to-one correspondence between subcontributions in the original $\alpha\beta\gamma$ integrals and those in the UVDP integrals.
The above equation just says that given a general two-loop Feynman integral, we can always using the standard Feynman parametrization to reduce it to the form of a “product" of Feynman parameter integral $I_\text{FP}$ and $I_{\alpha\beta\gamma}$. Since $I_\text{FP}$ does not contain any UV contributions, to regularize $I$, one just needs to regularize the $I_{\alpha\beta\gamma}$ parts. So for practical purposes, instead of doing the hard work of calculating $I_\text{UVDP}$ and $I_\text{ILI}$ case by case, one just needs the regularized results of $I_{\alpha\beta\gamma}$.
For the cases $\alpha$, $\beta$ or $\gamma$ equals zero, $I_{\alpha\beta\gamma}$ can be decomposed into two one-loop $I_\text{ILI}$, and to get the regularized results, all one has to do is to use the one-loop results twice. For $\alpha$, $\beta$, $\gamma$$\ \neq 0$, the only case that is relevant to our calculations of quadratic contributions in WZ’ model in the next section is $I_{111}$ whose quadratic contributions are given by: $$I_{111}=\pi^4M^2_c[3(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega})+1]+\cdots.$$ Here, $q_0$ is an arbitrary mass scale introduced to balance the dimension. The rest cases (e.g. $I_{112}$) just do not contain any quadratic contributions by naive power counting. One can find more discussions about $I_{\alpha\beta\gamma}$ in the Appendix.
Now we have accumulated sufficient information about LORE to finish our calculations. Let’s move on to discuss the WZ’ model and try to calculate the possible quadratic contributions up to two loops.
Quadratic Contributions in WZ’ Model
====================================
WZ’ model, i.e., Wess-Zumino model [@Wess:1973kz; @Wess:1974tw] with soft SUSY breaking terms, is the simplest model for the softly broken SUSY, and is an insightful toy model that shares many important properties of MSSM. The treatment here can be extended straightforwardly to the more complicated models that include gauge bosons since LORE preserves the gauge symmetries as well. The lagrangian of WZ’ model is given by $$\begin{gathered}
\mathcal{L}_{WZ'}=\frac{1}{2}(\partial_{\mu}A)^2-\frac{1}{2}m^2_AA^2+\frac{1}{2}(\partial_{\mu}B)^2-\frac{1}{2}m^2_BB^2+\frac{1}{2}\bar{\psi}(i\cancel{\partial}-m_{\psi})\psi\\
-(m_{\psi}g-\lambda)A^3-(m_{\psi}g+3\lambda)AB^2-g(\bar{\psi}\psi A+i\bar{\psi}\gamma^5\psi B)-\frac{1}{2}g^2(A^2+B^2)^2.
\label{WZ'Lagrangian}\end{gathered}$$ In this article, we adopt the $(+,-,-,-)$ convention and $\gamma^5=i\gamma^0\gamma^1\gamma^2\gamma^3$. A and B are real scalar fields, and $\psi$ is a Majorana fermion. g is a dimensionless real coupling and $\lambda$ is the dimension-one soft SUSY breaking parameter. In Eq. we have integrated out the so-called auxiliary fields, which is a common practice in phenomenology. The conventions of Feynman rules we adopt in the following calculations are quite standard and can be found in Ref. [@Peskin:1995ev] and Ref. [@Denner:1992vza; @Denner:1992me].
[c c]{} Counterterm & Result\
\
\[-2ex\] $\delta Z^{(1)}_A$ & $-\frac{g^2}{4\pi^2}\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_{AA}$ & $\left[\frac{3g^2}{8\pi^2}+\frac{g^2m^2_B}{8\pi^2m^2_A}-\frac{3g^2m^2_{\psi}}{2\pi^2m^2_A}+\frac{9(-\lambda+gm_{\psi})^2}{8\pi^2m^2_A}+\frac{(3\lambda+gm_{\psi})^2}{8\pi^2m^2_A}\right]\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_B$ & $-\frac{g^2}{4\pi^2}\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_{BB}$ & $[\frac{3g^2}{8\pi^2}+\frac{g^2m^2_A}{8\pi^2m^2_B}-\frac{g^2m^2_{\psi}}{2\pi^2m^2_B}+\frac{(3\lambda+gm_{\psi})^2}{4\pi^2m^2_B}]\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_{\psi}$ & $-\frac{g^2}{4\pi^2}\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_{\bar{\psi}\psi}$ & $0$\
\[1ex\] $\delta Z^{(1)}_{AAA}$ & $(\frac{g^2}{4\pi^2}-\frac{3g\lambda}{4\pi^2m_{\psi}})\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_{ABB}$ & $(\frac{g^2}{4\pi^2}+\frac{9g\lambda}{4\pi^2m_{\psi}})\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_{\bar{\psi}\psi A}$ & $0$\
\[1ex\] $\delta Z^{(1)}_{\bar{\psi}\psi B}$ & $0$\
\[1ex\] $\delta Z^{(1)}_{AAAA}$ & $\frac{g^2}{4\pi^2}\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_{BBBB}$ & $\frac{g^2}{4\pi^2}\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] $\delta Z^{(1)}_{AABB}$ & $\frac{g^2}{4\pi^2}\ln\frac{M^2_c}{\mu^2}$\
\[1ex\]
In this section, we shall calculate the quadratic contributions in the self-energy diagrams of the scalar particle A and the pseudoscalar particle B up to two loops. In these calculations, we have used the *Mathematica* package *FeynArts* [@Hahn:2000kx] to generate the relevant Feynman diagrams.
One-loop Calculations\[1LoopCalculation\]
-----------------------------------------
Beyond the tree level, one needs to introduce extra counterterms in order to make the radiative corrections finite. To calculate radiative corrections at two loops, one has to first figure out counterterms at the one-loop level. At the one-loop level, the counterterm lagrangian is given by $$\begin{aligned}
&\mathcal{L}_{ct}=\frac{1}{2}\delta Z^{(1)}_A(\partial_{\mu}A)^2-\frac{1}{2}\delta Z^{(1)}_{AA}m^2_AA^2+\frac{1}{2}\delta Z^{(1)}_B(\partial_{\mu}B)^2-\frac{1}{2}\delta Z^{(1)}_{BB}m^2_BB^2\nonumber\\
&+\frac{1}{2}\delta Z^{(1)}_{\psi}\bar{\psi}i\cancel{\partial}\psi-\frac{1}{2}\delta Z^{(1)}_{\bar{\psi}\psi}m_{\psi}\bar{\psi}\psi-\delta Z^{(1)}_{AAA}m_{\psi}gA^3-\delta Z^{(1)}_{ABB}m_{\psi}gAB^2-\delta Z^{(1)}_{\bar{\psi}\psi A}g\bar{\psi}\psi A\nonumber\\
&-\delta Z^{(1)}_{\bar{\psi}\psi B}ig\bar{\psi}\gamma^5\psi B-\frac{1}{2}\delta Z^{(1)}_{AAAA}g^2A^4-\frac{1}{2}\delta Z^{(1)}_{BBBB}g^2B^4-\delta Z^{(1)}_{AABB}g^2A^2B^2.\end{aligned}$$ At the one-loop level, there are 114 Feynman diagrams that can contribute to the one-loop counterterms. It would be too messy to draw all of these diagrams here. Instead, we choose to present the final results in Table \[OneLoopCounterterms\] directly. Here we have used the MS renormalization scheme and the parameter $\mu$ is the so-called renormalization scale. Apparently, there is no quadratic contribution at the one-loop level.
When taking the supersymmetric limit $m_A=m_B=m_{\psi}=m$, $\lambda=0$, one has $$\begin{aligned}
&\delta Z^{(1)}_A=\delta Z^{(1)}_B=\delta Z^{(1)}_{\psi}=-\frac{g^2}{4\pi^2}\ln\frac{M^2_c}{\mu^2},\\
&\delta Z^{(1)}_{AA}=\delta Z^{(1)}_{BB}=\delta Z^{(1)}_{AAA}=\delta Z^{(1)}_{ABB}\nonumber\\
&=\delta Z^{(1)}_{AAAA}=\delta Z^{(1)}_{BBBB}=\delta Z^{(1)}_{AABB}=\frac{g^2}{4\pi^2}\ln\frac{M^2_c}{\mu^2},\\
&\delta Z^{(1)}_{\bar{\psi}\psi}=\delta Z^{(1)}_{\bar{\psi}\psi A}=\delta Z^{(1)}_{\bar{\psi}\psi B}=0,\end{aligned}$$ which are nothing but the celebrated non-renormalization theorem [@Wess:1974tw; @Wess:1973kz; @Cui:2008bk].
Two-loop Calculations
---------------------
Now we are ready to calculate quadratic contributions in the self-energy diagrams of the scalar particle A and pseudoscalar particle B in WZ’ model at the two-loop level. The relevant Feynman diagrams are given in Fig. \[2LoopAA\] and Fig. \[2LoopBB\]. Here, we only include diagrams that are quadratic divergent according to power counting.
![Two-loop Self-energy Diagrams for Scalar Particle A[]{data-label="2LoopAA"}](2LoopAA5X5.jpg)
[p[5cm]{}c c]{}\
Diagram & Quadratic Contributions\
\
(a1) & $-\frac{27g^2(-\lambda+gm_{\psi})^2}{32\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{m^2_A-x(1-x)k_1^2}$\
\[1ex\] (b1) & $-\frac{3g^2(3\lambda+gm_{\psi})^2}{32\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{m^2_B-x(1-x)k_1^2}$\
\[1ex\] (c1) & $-\frac{9g^2(-\lambda+gm_{\psi})^2}{32\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{m^2_A-x(1-x)k_1^2}$\
\[1ex\] (d1) & $-\frac{g^2(3\lambda+gm_{\psi})^2}{32\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{m^2_B-x(1-x)k_1^2}$\
\[1ex\] (e1) & $-\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (f1) & $\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (g1) & $-\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (h1) & $-\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (i1) & $\frac{9g^2(-\lambda+gm_{\psi})^2}{8\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{m^2_A-x(1-x)k_1^2}$\
\[1ex\] (j1) & $\frac{g^2(3\lambda+gm_{\psi})^2}{8\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{m^2_B-x(1-x)k_1^2}$\
\[1ex\] (k1) & $\frac{9g^4}{64\pi^4}M^2_c\left(\ln\frac{M^2_c}{m^2_A}-\gamma_{\omega}\right)$\
\[1ex\] (l1) & $\frac{3g^4}{64\pi^4}M^2_c\left(\ln\frac{M^2_c}{m^2_B}-\gamma_{\omega}\right)$\
\[1ex\] (m1) & $\frac{3g^4}{64\pi^4}M^2_c\left(\ln\frac{M^2_c}{m^2_A}-\gamma_{\omega}\right)$\
\[1ex\] (n1) & $\frac{g^4}{64\pi^4}M^2_c\left(\ln\frac{M^2_c}{m^2_B}-\gamma_{\omega}\right)$\
\[1ex\] (o1) & $\frac{3g^4}{32\pi^4}M^2_c\left(\ln\frac{M^2_c}{q^2_0}+2\ln\frac{m^2_A}{q^2_0}-\gamma_{\omega}+1\right)$\
\[1ex\] (p1) & $\frac{g^4}{32\pi^4}M^2_c\left(\ln\frac{M^2_c}{q^2_0}+2\ln\frac{m^2_B}{q^2_0}-\gamma_{\omega}+1\right)$\
\[1ex\] (q1) & $\frac{3g^4}{32\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (r1) & $\frac{g^4}{32\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (s1) & $\frac{g^4}{4\pi^4}M^2_c \ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (t1) & $-\frac{3g^4}{32\pi^4}M^2_c \ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (u1) & $-\frac{g^4}{32\pi^4}M^2_c \ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (v1) & $0$\
\[1ex\] (w1) & $-\frac{3g^4}{32\pi^4}M^2_c \ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (x1) & $-\frac{g^4}{32\pi^4}M^2_c \ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (y1) & $0$\
\[1ex\]\
\[-2ex\] Total & $0$\
\[1ex\]
The results for the self-energy diagrams of the scalar particle A are given in Table \[2LoopAA\]. For our current purposes, we only track the quadratic contributions in our calculations.
Here, we see explicitly that the total quadratic contributions vanish. Microscopically, one has $$\begin{aligned}
&(a1)+(c1)+(i1) \sim 0,\\
&(b1)+(d1)+(j1) \sim 0,\\
&(e1)+(f1)+(v1)+(y1) \sim 0,\\
&(g1)+(k1)+(m1)+(o1)+(r1) \sim 0,\\
&(h1)+(l1)+(n1)+(p1)+(q1) \sim 0, \\
&(s1)+(t1)+(u1)+(w1)+(x1) \sim 0,\end{aligned}$$ up to logarithmic contributions and finite terms.
![Two-loop Self-energy Diagrams for Pseudoscalar Particle B[]{data-label="2LoopBB"}](2LoopBB5X5.jpg)
[p[5cm]{} c c]{}\
Diagram & Quadratic Contributions\
\
(a2) & $-\frac{3g^2(3\lambda+gm_{\psi})^2}{32\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{(1-x)m^2_B+xm^2_A-x(1-x)k_1^2}$\
\[1ex\] (b2) & $-\frac{g^2(3\lambda+gm_{\psi})^2}{32\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{(1-x)m^2_A+xm^2_B-x(1-x)k_1^2}$\
\[1ex\] (c2) & $-\frac{g^2(3\lambda+gm_{\psi})^2}{32\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{(1-x)m^2_B+xm^2_A-x(1-x)k_1^2}$\
\[1ex\] (d2) & $-\frac{3g^2(3\lambda+gm_{\psi})^2}{32\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{(1-x)m^2_A+xm^2_B-x(1-x)k_1^2}$\
\[1ex\] (e2) & $\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (f2) & $-\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (g2) & $-\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (h2) & $-\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (i2) & $\frac{g^2(3\lambda+gm_{\psi})^2}{8\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{(1-x)m^2_B+xm^2_A-x(1-x)k_1^2}$\
\[1ex\] (j2) & $\frac{g^2(3\lambda+gm_{\psi})^2}{8\pi^4}M^2_c\int^1_0\!\mathrm{d}x\frac{x}{(1-x)m^2_A+xm^2_B-x(1-x)k_1^2}$\
\[1ex\] (k2) & $\frac{3g^4}{64\pi^4}M^2_c\left(\ln\frac{M^2_c}{m^2_A}-\gamma_{\omega}\right)$\
\[1ex\] (l2) & $\frac{9g^4}{64\pi^4}M^2_c\left(\ln\frac{M^2_c}{m^2_B}-\gamma_{\omega}\right)$\
\[1ex\] (m2) & $\frac{g^4}{64\pi^4}M^2_c\left(\ln\frac{M^2_c}{m^2_A}-\gamma_{\omega}\right)$\
\[1ex\] (n2) & $\frac{3g^4}{64\pi^4}M^2_c\left(\ln\frac{M^2_c}{m^2_B}-\gamma_{\omega}\right)$\
\[1ex\] (o2) & $\frac{g^4}{32\pi^4}M^2_c\left(\ln\frac{M^2_c}{q^2_0}+2\ln\frac{m^2_A}{q^2_0}-\gamma_{\omega}+1\right)$\
\[1ex\] (p2) & $\frac{3g^4}{32\pi^4}M^2_c\left(\ln\frac{M^2_c}{q^2_0}+2\ln\frac{m^2_B}{q^2_0}-\gamma_{\omega}+1\right)$\
\[1ex\] (q2) & $\frac{3g^4}{32\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (r2) & $\frac{g^4}{32\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]$\
\[1ex\] (s2) & $\frac{g^4}{4\pi^4}M^2_c\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (t2) & $-\frac{g^4}{32\pi^4}M^2_c \ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (u2) & $-\frac{3g^4}{32\pi^4}M^2_c \ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (v2) & $0$\
\[1ex\] (w2) & $-\frac{g^4}{32\pi^4}M^2_c\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (x2) & $-\frac{3g^4}{32\pi^4}M^2_c\ln\frac{M^2_c}{\mu^2}$\
\[1ex\] (y2) & $0$\
\[1ex\]\
\[-2ex\] Total & $0$\
\[1ex\]
Similar results can be obtained for the pseudoscalar particle B. The corresponding quadratic contributions in the self-energy diagrams of B are given in Table \[2LoopBB\].
Also, one sees explicitly that the quadratic contributions in self-energy diagrams of B vanish as expected. Microscopically, one has $$\begin{aligned}
&(a2)+(c2)+(i2) \sim 0,\\
&(b2)+(d2)+(j2) \sim 0,\\
&(e2)+(f2) +(v2)+(y2)\sim 0,\\
&(g2)+(k2)+(m2)+(o2)+(r2) \sim 0,\\
&(h2)+(l2)+(n2)+(p2)+(q2) \sim 0, \\
&(s2)+(t2)+(u2)+(w2)+(x2) \sim 0,\end{aligned}$$ up to logarithmic contributions and finite terms.
At the two-loop level, Feynman diagrams other than the self-energy diagrams of A and B, although they cannot be overall quadratic divergent by naive power counting, can also contain quadratic subcontributions. These quadratic subcontributions are resulted from the embedding of the one-loop self-energy diagrams of A and B into the Feynman diagrams and thus should cancel with each other since we have shown explicitly that there is no quadratic contribution at the one loop level in Section \[1LoopCalculation\]. In this way, it is shown by explicit Feynman diagrammatic calculations that up to the two loop level the WZ’ model is free of quadratic contributions.
Conclusions and Remarks
=======================
In this article, we revisit the absence of the quadratic contributions in models with softly broken SUSY using LORE. In previous studies, supergraph techniques have been used to show that models with softly broken SUSY should be free of quadratic contributions. Although elegant, supergraph techniques are less useful in phenomenological studies, where the traditional Feynman diagrammatic approach is more suitable. On the other hand, LORE has been shown in previous studies to be a powerful tool to regularize quantum field theories with gauge symmetries and supersymmetry, and is an ideal regularization scheme for traditional Feynman diagrammatic calculations. We use LORE to calculate the two-loop quadratic contributions in WZ’ model, the simplest model with softly broken SUSY, which contains a scalar particle A, a pseudoscalar B and a Majorana fermion $\psi$, and show that they do cancel with each other perfectly, which is consistent with the results in literature. Moreover, there should be no obstacle to extend our methods to models containing gauge bosons, such as MSSM, thanks to the fact that LORE preserves gauge symmetries as well. [Also, given the fact that quadratic contributions play a crucial role in deriving the gap equations to describe the dynamically generated spontaneous chiral symmetry breaking of QCD [@Dai:2003ip; @Gherghetta:1994cr], the absence of quadratic contributions in SUSY may reveal the fact that its spontaneous breaking has to be carried out in a different manner.]{} We will return to these issues in future publications.
The author(DB) would like to thank Zhuo Liu for enlightening discussions during the preparation of this work, and G. Gnendiger and A. Signer for helpful correspondence after the manuscript was submitted to arXiv. This work was supported in part by the National Science Foundation of China (NSFC) under Grants No. 11690022 & 11475237, and by the Strategic Priority Research Program of the Chinese Academy of Sciences (CAS), Grant No. XDB23030100, and by the CAS Center for Excellence in Particle Physics (CCEPP).
Useful Formulae for LORE
========================
$$\begin{aligned}
&\frac{1}{AB}=\int^1_0\!\mathrm{d}x\frac{1}{[xA+(1-x)B]^2},\\
&\frac{1}{AB^n}=\int^1_0\!\mathrm{d}x\frac{nx^{n-1}}{[(1-x)A+xB]^{n+1}},\end{aligned}$$ $$\begin{aligned}
&\frac{1}{A_1A_2\cdots A_n}=\int^1_0\!\mathrm{d}x_1\cdots\mathrm{d}x_n\delta{(\Sigma x_i-1)}\frac{(n-1)!}{(x_1A_1+x_2A_2+\cdots+x_nA_n)^n},\\
&\frac{1}{A^{m_1}_1A^{m_2}_2\cdots A^{m_n}_n}=\int^1_0\!\mathrm{d}x_1\cdots\mathrm{d}x_n\delta{(\Sigma x_i-1)}\frac{\prod x^{m_i-1}_i}{(\Sigma x_iA_i)^{\Sigma m_i}}\frac{\Gamma(m_1+\cdots+m_n)}{\Gamma(m_1)\cdots\Gamma(m_n)}.\end{aligned}$$ $$\begin{aligned}
&\int\!\mathrm{d}^4k\frac{1}{k^2-\mathcal{M}^2}\nonumber\\
&:=-i\pi^2\left\{M^2_c-(\mathcal{M}^2+\mu^2_s)\left[\ln\frac{M^2_c}{\mathcal{M}^2+\mu^2_s}-\gamma_{\omega}+1+y_2\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right]\right\},\\
&\int\!\mathrm{d}^4k\frac{1}{(k^2-\mathcal{M}^2)^2}=i\pi^2\left[\ln\frac{M^2_c}{\mathcal{M}^2+\mu^2_s}-\gamma_{\omega}+y_0\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right],\\
&\int\!\mathrm{d}^4k\frac{1}{(k^2-\mathcal{M}^2)^3}=-i\pi^2\frac{1}{2(\mathcal{M}^2+\mu^2_s)}\left[1-y_{-2}\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right],\\
&\int\!\mathrm{d}^4k\frac{1}{(k^2-\mathcal{M}^2)^{\alpha}}\nonumber\\
&=(-1)^{\alpha}i\pi^2\frac{\Gamma(\alpha-2)}{\Gamma(\alpha)}\frac{1}{(\mathcal{M}^2+\mu^2_s)^{\alpha-2}}\left[1-y_{-2(\alpha-2)}\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right],
\text{\qquad($\alpha\geq 3$)}\\
&\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}}{(k^2-\mathcal{M}^2)^2}\nonumber\\
&=-\frac{i}{2}g^{\mu\nu}\pi^2\left\{M^2_c-(\mathcal{M}^2+\mu^2_s)\left[\ln\frac{M^2_c}{\mathcal{M}^2+\mu^2_s}-\gamma_{\omega}+1+y_2\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right]\right\},\\
&\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}}{(k^2-\mathcal{M}^2)^3}=\frac{i}{4}g^{\mu\nu}\pi^2\left[\ln\frac{M^2_c}{\mathcal{M}^2+\mu^2_s}-\gamma_{\omega}+y_0\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right],\\
&\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}}{(k^2-\mathcal{M}^2)^4}=-\frac{i}{6}g^{\mu\nu}\pi^2\frac{1}{2(\mathcal{M}^2+\mu^2_s)}\left[1-y_{-2}\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right],\\
&\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}}{(k^2-\mathcal{M}^2)^{\alpha+1}}\nonumber\\
&=(-1)^{\alpha}\frac{i}{2}g^{\mu\nu}\pi^2\frac{\Gamma(\alpha-2)}{\Gamma(\alpha+1)}\frac{1}{(\mathcal{M}^2+\mu^2_s)^{\alpha-2}}\left[1-y_{-2(\alpha-2)}\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right],\\
&\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}k^{\rho}k^{\sigma}}{(k^2-\mathcal{M}^2)^3}\nonumber\\
&:=-\frac{i}{8}g^{\{\mu\nu\rho\sigma\}}\pi^2\Bigg\{M^2_c-(\mathcal{M}^2+\mu^2_s)\Bigg[\ln\frac{M^2_c}{\mathcal{M}^2+\mu^2_s}-\gamma_{\omega}+1\nonumber\\
&+y_2\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\Bigg]\Bigg\},\\
&\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}k^{\rho}k^{\sigma}}{(k^2-\mathcal{M}^2)^4}=\frac{i}{24}g^{\{\mu\nu\rho\sigma\}}\pi^2\left[\ln\frac{M^2_c}{\mathcal{M}^2+\mu^2_s}-\gamma_{\omega}+y_0\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right],\end{aligned}$$ $$\begin{aligned}
&\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}k^{\rho}k^{\sigma}}{(k^2-\mathcal{M}^2)^5}=-\frac{i}{48}g^{\{\mu\nu\rho\sigma\}}\pi^2\frac{1}{2(\mathcal{M}^2+\mu^2_s)}\left[1-y_{-2}\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right],\\
&\int\!\mathrm{d}^4k\frac{k^{\mu}k^{\nu}k^{\rho}k^{\sigma}}{(k^2-\mathcal{M}^{2})^{\alpha+2}}\nonumber\\
&=(-1)^{\alpha}\frac{i}{4}g^{\{\mu\nu\rho\sigma\}}\pi^2\frac{\Gamma(\alpha-2)}{\Gamma(\alpha+2)}\frac{1}{(\mathcal{M}^2+\mu^2_s)^{\alpha-2}}\left[1-y_{-2(\alpha-2)}\left(\frac{\mathcal{M}^2+\mu^2_s}{M^2_c}\right)\right].\end{aligned}$$ Here $$g^{\{\mu\nu\rho\sigma\}}:=g^{\mu\nu}g^{\rho\sigma}+g^{\mu\rho}g^{\nu\sigma}+g^{\mu\sigma}g^{\rho\nu}.$$ $M_c$ and $\mu_s$ in the above formulae regularize the UV contributions and IR divergences respectively. The y functions are given by $$\begin{aligned}
&y_2(x):=\frac{1}{x}(1-x-e^{-x})+\int^x_0\!\mathrm{d}\sigma\frac{1-e^{-\sigma}}{\sigma},\\
&y_0(x)=\int^x_0\!\mathrm{d}\sigma\frac{1-e^{-\sigma}}{\sigma},\\
&y_{-2}(x)=1-e^{-x},\\
&y_{-2(\alpha-1)}(x)=y_{-2(\alpha-2)}(x)-\frac{1}{\alpha-2}x\frac{\partial}{\partial x}y_{-2(\alpha-2)}(x),\end{aligned}$$ for $\alpha \geq 3$. It is easy to show that these y functions approach zero when $x\to 0$. $$\begin{aligned}
&I_{121}:=\int\!\mathrm{d}^4l_1\!\int\!\mathrm{d}^4l_2\frac{1}{l^2_1-m^2_1}\frac{1}{(l^2_2-m^2_2)^2}\frac{1}{(l_1+l_2+p)^2-m^2_3}\nonumber\\
&=-\pi^4\left(\ln\frac{M^2_c}{q^2_0}-\gamma_{\omega}+1\right)\left(\ln\frac{M^2_c}{m^2_2}-\gamma_{\omega}\right)+\text{finite terms},\end{aligned}$$ $$\begin{aligned}
&I_{111}:=\int\!\mathrm{d}^4l_1\!\int\!\mathrm{d}^4l_2\frac{1}{l^2_1-m^2_1}\frac{1}{l^2_2-m^2_2}\frac{1}{(l_1+l_2+p)^2-m^2_3}\nonumber\\
&=\pi^4M^2_c\left[3\left(\ln\tfrac{M^2_c}{q^2_0}-\gamma_{\omega}\right)+1\right]-\pi^4m^2_2\left(\ln\tfrac{M^2_c}{q^2_0}-\gamma_{\omega}+2\right)\left(\ln\tfrac{M^2_c}{m^2_2}-\gamma_{\omega}+1\right)\nonumber\\
&-\tfrac{\pi^4}{56}[-4\pi^2m^2_1+4\pi^2m^2_3+3(m^2_1-m^2_2)\psi^{(1)}(\tfrac{1}{6})+3(-m^2_2+m^2_3)\psi^{(1)}(\tfrac{1}{3})\nonumber\\
&+3(-m^2_1+m^2_2)\psi^{(1)}(\tfrac{2}{3})+3(m^2_2-m^2_3)\psi^{(1)}(\tfrac{5}{6})]\left(\ln\tfrac{M^2_c}{m^2_2}-\gamma_{\omega}+1\right)\nonumber\\
&-\pi^4m^2_1\left(\ln\tfrac{M^2_c}{m^2_1}-\gamma_{\omega}+1\right)\left(\ln\tfrac{M^2_c}{q^2_0}-\gamma_{\omega}-\tfrac{2\pi}{3\sqrt{3}}\right)\nonumber\\
&-\pi^4m^2_3\left(\ln\tfrac{M^2_c}{m^2_3}-\gamma_{\omega}+1\right)\left(\ln\tfrac{M^2_c}{q^2_0}-\gamma_{\omega}-\tfrac{2\pi}{3\sqrt{3}}\right)\nonumber\\
&-\tfrac{\pi^4}{18}[-54m^2_2+36m^2_3+4\sqrt{3}\pi(m^2_1+m^2_3)\nonumber\\
&+\tfrac{4}{3}\pi^2(m^2_1-m^2_2)+(2m^2_2-m^2_1-m^2_3)\psi^{(1)}(\tfrac{1}{6})+(m^2_2-m^2_3)\psi^{(1)}(\tfrac{1}{3})\nonumber\\
&+(m^1_1-2m^2_2+m^2_3)\psi^{(1)}(\tfrac{2}{3})+(-m^2_2+m^2_3)\psi^{(1)}(\tfrac{5}{6})]\left(\ln\tfrac{M^2_c}{m^2_1(m^2_3)}-\gamma_{\omega}+1\right)\nonumber\\
&+\tfrac{\pi^4}{108}p^2[54(\ln\tfrac{M^2_c}{-p^2}-\gamma_{\omega}+1)+81+2\psi^{(1)}(\tfrac{1}{6})+2\psi^{(1)}(\tfrac{1}{3})-2\psi^{(1)}(\tfrac{2}{3})-2\psi^{(1)}(\tfrac{5}{6})]\nonumber\\
&+\text{finite terms} .
\label{I111}\end{aligned}$$ Here $q^2_0$ is an arbitrary mass scale introduced to balance the dimension, and $\psi^{(1)}(z)=\frac{d^2}{dz^2}\ln\Gamma(z)$ is the first order polygamma function. We have assumed $m^2_1 \sim m^2_3$ in Eq. . It is difficult to obtain an explicit analytic expression of $I_{111}$ for general cases.
Derivations of Eq. (A.23)
=========================
Here, we give a brief introduction to the UVDP parametrization, using the Feynman integral $I_{111}$ as an explicit example, the only $\alpha\beta\gamma$ integral relevant to the two-loop quadratic contribution calculations. The UVDP methods are aimed to give proper treatments to subcontributions at two loops. In this approach, the UV contributions arising from large loop momenta are transmitted to the asymptotic regions of UVDP parameter space. $$\begin{aligned}
I_{111}&=\int\!\mathrm{d}^4l_1\int\!\mathrm{d}^4l_2\frac{1}{l_1^2-m_1^2}\frac{1}{l_2^2-m_2^2}\frac{1}{(l_1+l_2+p)^2-m^2_3}\nonumber\\
&=i\pi^2\frac{\Gamma(1)}{\Gamma(1)\Gamma(1)\Gamma(1)}\int_0^\infty\!\prod_{i=1}^3\frac{\mathrm{d}v_i}{(1+v_i)^3}\delta\left(1-\sum_{j=1}^3\frac{1}{1+v_j}\right)\nonumber\\
&\times\frac{(1+v_1)^2(1+v_2)^2(1+v_3)^2}{(3+v_1+v_2+v_3)^2}\int\!\mathrm{d}^4l\frac{1}{l^2-\mathcal{M}^2}\nonumber\\
&=i\pi^2\int_0^\infty\!\prod_{i=1}^3\frac{\mathrm{d}u}{(1+u)^3}\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\int_0^\infty\!\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\nonumber\\
&\times\frac{(1+u)^4(1+w)^2(1+v)^2}{\left[u(1+w)(1+v)+1\right]^2}\int\!\mathrm{d}^4l\frac{1}{l^2-\mathcal{M}^2}\nonumber\\
&=\pi^4\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\int_0^\infty\!\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\int_0^\infty\!\mathrm{d}u\frac{1+u}{\left[u+\frac{1}{(1+w)(1+v)}\right]^2}\nonumber\\
&\times\left\{M_c^2-M^2\left[\ln\frac{M^2_c}{M^2}-\gamma_\omega+1\right]\right\}\nonumber\\
&=I_{111}^{(0)}+I_{111}^{(1)}+I_{111}^{(2)}+I_{111}^{(3)}+I_{111}^{(4)},
\label{I111AB}\end{aligned}$$ with $$\begin{aligned}
\mathcal{M}^2&=\frac{m_1^2}{1+v_1}+\frac{m^2_2}{1+v_2}+\frac{m^2_3}{1+v_3}-\frac{p^2}{3+v_1+v_2+v_3}\nonumber\\
&=\frac{m_1^2}{(1+u)(1+w)}+\frac{m^2_2u}{1+u}+\frac{m^2_3}{(1+u)(1+v)}-\frac{u}{1+u}\frac{p^2}{u(1+w)(1+v)+1}.\end{aligned}$$ The transition between the old UVDP parameters $(v_1,v_2,v_3)$ and the new one $(u,v,w)$ is given by $$\begin{aligned}
&\frac{1}{1+v_1}=\frac{1}{(1+u)(1+w)},\\
&\frac{1}{1+v_2}=\frac{u}{1+u},\\
&\frac{1}{1+v_3}=\frac{1}{(1+u)(1+v)}.\end{aligned}$$ The following identity plays a crucial role in the UVDP treatment of $I_{111}$ integral, $$\begin{aligned}
\frac{1}{A_1A_2\cdots A_n}=\int_0^\infty\!\prod_{i=1}^n\frac{\mathrm{d}v_i}{(1+v_i)^2}\delta\left(\sum^n_{i=1}\frac{1}{1+v_i}-1\right)\frac{(n-1)!}{\left[\sum_{i=1}^{n}\frac{A_i}{1+v_i}\right]^n}.\end{aligned}$$ The integrals $I_{111}^{(0)}$, $I_{111}^{(1)}$, $I_{111}^{(2)}$, $I_{111}^{(3)}$ and $I_{111}^{(4)}$ introduced in the last step of Eq. are analyzed in details as follows. The $I^{(0)}_{111}$ integral contains the quadratic contributions only, $$\begin{aligned}
I_{111}^{(0)}&=\pi^4M_c^2\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\int_0^\infty\mathrm{d}u\frac{1+u}{\left[u+\frac{1}{(1+w)(1+v)}\right]^2}\nonumber\\
&=\pi^4M_c^2\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\nonumber\\
&\times\left\{\ln{\frac{M_c^2}{q^2_0}}+\ln(1+w)(1+v)-\gamma_\omega+(1+w)(1+v)-1\right\}\nonumber\\
&=\pi^4M_c^2\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\left\{\ln\frac{M_c^2}{q^2_0}-\gamma_\omega-1+\ln\left(2+w+\frac{1}{w}\right)+\left(2+\omega+\frac{1}{\omega}\right)\right\}\nonumber\\
&=\pi^4M_c^2\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_\omega\right)+1\right],\end{aligned}$$ while the logarithmic contributions are encapsulated in $I^{(1)}_{111},\cdots,I^{(4)}_{111}$, $$\begin{aligned}
&I^{(1)}_{111}+I^{(2)}_{111}+I^{(3)}_{111}+I^{(4)}_{111}\nonumber\\
&=\pi^4\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\left\{-\mathcal{M}^2\left[\ln\frac{M_c^2}{\mathcal{M}^2}-\gamma_\omega+1\right]\right\}\nonumber\\
&=-\pi^4\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\nonumber\\
&\times\int_0^\infty\mathrm{d}u\frac{1+u}{\left[u+\frac{1}{(1+w)(1+v)}\right]^2}\Bigg[\frac{m_1^2}{(1+u)(1+w)}+\frac{m^2_2u}{1+u}+\frac{m_3^2}{(1+u)(1+v)}\nonumber\\
&-\frac{u}{1+u}\frac{p^2}{u(1+w)(1+v)+1}\Bigg]\times\left[\ln\frac{M^2_c}{\mathcal{M}^2}-\gamma_\omega+1\right].\end{aligned}$$ Here, $I_{111}^{(1)},\cdots,I_{111}^{(4)}$ correspond to logarithmic contributions in the following asymptotic UVDP parameter regions:
- $I^{(1)}_{111}$: $u\to\infty$, $vw=1$, $\mathcal{M}^2\to m^2_2$,
- $I^{(2)}_{111}$: $v\to\infty$, $u\to0$, $w\to0$, $\mathcal{M}^2\to m^2_1$,
- $I^{(3)}_{111}$: $w\to\infty$, $u\to0$, $v\to0$, $\mathcal{M}^2\to m^2_3$,
- $I^{(4)}_{111}$: $-p^2\gg m^2_1, m^2_2, m^2_3$.
Explicitly, we have $$\begin{aligned}
\bullet\ \,I_{111}^{(1)}&=-\pi^4\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\nonumber\\
&\times\int_0^\infty\!\mathrm{d}u\frac{1}{u+\frac{1}{(1+w)(1+v)}}\left[\frac{m^2_1}{(1+u)(1+w)}+\frac{m^2_2u}{1+u}+\frac{m^2_3}{(1+u)(1+v)}\right]\nonumber\\
&\times\left[\ln\frac{M^2_c}{m^2_2}-\gamma_\omega+1\right]\nonumber\\
&=-\pi^4m^2_2\left(\ln\frac{M_c^2}{q^2_0}+2-\gamma_\omega\right)\left(\ln\frac{M^2_c}{m^2_2}-\gamma_\omega+1\right)\nonumber\end{aligned}$$ $$\begin{aligned}
&-\frac{\pi^4}{56}\Big\{-4m_1^2\pi^2+4m_3^2\pi^2+3(m^2_1-m^2_2)\psi^{(1)}(\tfrac{1}{6})+3(-m_2^2+m^2_3)\psi^{(1)}(\tfrac{1}{3})\nonumber\\
&+3(-m^2_1+m^2_2)\psi^{(1)}(\tfrac{2}{3})+3(m^2_2-m^2_3)\psi^{(1)}(\tfrac{5}{6})\Big\}\left(\ln\frac{M^2_c}{m^2_2}-\gamma_\omega+1\right),\end{aligned}$$ $$\begin{aligned}
\bullet\ \,&I^{(2)}_{111}+I^{(3)}_{111}=-\pi^4\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\nonumber\\
&\times\int_0^\infty\!\mathrm{d}u\left[1-\frac{1}{(1+w)(1+v)}\right]\frac{1}{\left[u+\frac{1}{(1+w)(1+v)}\right]^2}\bigg[\frac{m^2_1}{(1+w)(1+u)}+\frac{m^2_2u}{1+u}\nonumber\\
&+\frac{m^2_3}{(1+u)(1+v)}\bigg]\left(\ln\frac{M^2_c}{\mathcal{M}^2}-\gamma_\omega+1\right)\nonumber\\
&=-\pi^4m^2_1\left(\ln\frac{M^2_c}{m_1^2}-\gamma_\omega+1\right)\int_0^\infty\!\mathrm{d}v\frac{v^2}{(1+v)(1+v+v^2)}\nonumber\\
&-\pi^4m^2_3\left(\ln\frac{M^2_c}{m_3^2}-\gamma_\omega+1\right)\int_0^\infty\!\mathrm{d}w\frac{w^2}{(1+w)(1+w+w^2)}\nonumber\\
&-\pi^4\int_0^\infty\frac{\mathrm{d}w}{(1+w)^2}\frac{1}{1+w+\frac{1}{w}}\bigg\{(m^2_1-m^2_2+m^2_3)\left(w+\frac{1}{w}\right)+(2m^2_1-m^2_2+2m^2_3)\nonumber\\
&+\left[2m^2_2-m^2_3-m^2_1\left(1+\frac{1}{w}\right)+\frac{m^2_2}{w}\right]\ln\left(w+\frac{1}{w}+2\right)\bigg\}\nonumber\\
&\times\left[\ln M^2_c-\gamma_\omega+1-\ln\left(\frac{m^2_3}{1+\frac{1}{w}}+\frac{m^2_1}{1+w}\right)\right]\nonumber\\
&=-\pi^4m^2_1\left(\ln\frac{M^2_c}{m^2_1}-\gamma_\omega+1\right)\left(\ln\frac{M^2_c}{q^2_0}-\gamma_\omega-\frac{2\pi}{3\sqrt{3}}\right)\nonumber\\
&-\pi^4m^2_3\left(\ln\frac{M^2_c}{m^2_3}-\gamma_\omega+1\right)\left(\ln\frac{M^2_c}{q^2_0}-\gamma_\omega-\frac{2\pi}{3\sqrt{3}}\right)\nonumber\\
&-\frac{\pi^4}{18}\left(\ln\frac{M^2_c}{\mathcal{M}^2}-\gamma_\omega+1\right)\bigg\{-54m^2_2+36m^2_3+4\sqrt{3}(m^2_1+m^2_3)\pi\nonumber\\
&+\frac{4}{3}(m^2_1-m^2_2)\pi^2-(m^2_1-2m^2+m^2_3)\psi^{(1)}(\tfrac{1}{6})+(m^2_2-m^2_3)\psi^{(1)}(\tfrac{1}{3})\nonumber\\
&+(m^2_1-2m^2+m^2_2)\psi^{(1)}(\tfrac{2}{3})+(-m^2_2+m^2_3)\psi^{(1)}(\tfrac{5}{6})\bigg\},\end{aligned}$$ $$\begin{aligned}
\bullet\ \,I_{111}^{(4)}&=\pi^4\int_0^\infty\!\frac{\mathrm{d}w}{(1+w)^2}\frac{\mathrm{d}v}{(1+v)^2}\delta\left(1-\frac{1}{1+w}-\frac{1}{1+v}\right)\nonumber\\
&\times\int_0^\infty\!\mathrm{d}u\frac{u}{(1+w)(1+v)\!\!\left[u+\frac{1}{(1+w)(1+v)}\right]^3}\,p^2\left(\ln\frac{M_c^2}{\mathcal{M}^2}-\gamma_\omega+1\right)\nonumber\\
&=\frac{\pi^4p^2}{108}\bigg\{54\left(\ln\frac{M_c^2}{-p^2}-\gamma_\omega+1\right)+81+\psi^{(1)}(\tfrac{1}{6})+2\psi^{(1)}(\tfrac{1}{3})-2\psi^{(1)}(\tfrac{2}{3})-2\psi^{(1)}(\tfrac{5}{6})\bigg\}.\end{aligned}$$ When putting $I_{111}^{(0)},\cdots,I_{111}^{(4)}$ together, one obtains Eq. .
Details for Feynman Diagrammatic Calculations
=============================================
Here, we present some technical details of Feynman diagrammatic calculations by choosing Diagram (e1), (f1), (v1) and (y1) as sample diagrams and calculating their quadratic contributions explicitly. The rest diagrams could be calculated in a similar way. $$\begin{aligned}
(e1)&=-\frac{(-2ig)^4}{512\pi^8}\int\!\mathrm{d}^4q_1\!\int\!\mathrm{d}^4q_2\frac{1}{q_1^2-m_{\psi}^2}\frac{1}{q_2^2-m_{\psi}^2}\frac{1}{(q_1+q_2)^2-m_A^2}\frac{1}{(q_1-k_1)^2-m_{\psi}^2}\nonumber\\
&\times\frac{1}{(q_1+k_1)^2-m_{\psi}^2}\text{tr}\big\{(m_{\psi}+\cancel{q}_2-\cancel{k}_1)(m_\psi+\cancel{q}_2)(m_{\psi}-\cancel{q}_1)(m_\psi-\cancel{q}_1-\cancel{k}_1)\big\}\nonumber\\
&=-\frac{g^4}{8\pi^8}\int_0^1\!\mathrm{d}x\!\int_0^1\!\mathrm{d}y\!\int\!\mathrm{d}^4l_1\!\int\!\mathrm{d}^4l_2\nonumber\\
&\times\frac{1}{\left\{l^2_1+x(1-x)k^2_1-m_{\psi}^2\right\}\left\{l_2^2+y(1-y)k^2_1-m_\psi^2\right\}\left\{[l_1+l_2+(y-x)k_1]^2-m^2_A\right\}}\nonumber\\
&+\text{logarithmic contributions}+\text{finite terms}\nonumber\\
&=-\frac{g^4}{8\pi^8}\int_0^1\!\mathrm{d}x\!\int_0^1\!\mathrm{d}y\pi^4M_c^2\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_\omega\right)+1\right]+\text{logarithmic contributions}\nonumber\\
&+\text{finite terms}\nonumber\\
&=-\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_\omega\right)+1\right]+\text{logarithmic contributions}+\text{finite terms}.\end{aligned}$$ $$\begin{aligned}
(f1)&=\frac{(2g)^4}{32\pi^8}\int\!\mathrm{d}^4q_1\!\int\!\mathrm{d}^4q_2\frac{1}{q_1^2-m_\psi^2}\frac{1}{q^2_2-m_\psi^2}\frac{1}{(q_1+q_2)^2-m_B^2}\frac{1}{(q_2-k_1)^2-m_\psi^2}\nonumber\\
&\times\frac{1}{(q_1+k_1)^2-m_\psi^2}\text{tr}\bigg\{(m_\psi+\cancel{q}_2-\cancel{k}_1)(m_\psi+\cancel{q}_2)\gamma^5(m_\psi-\cancel{q}_1)(m_\psi-\cancel{q}_1-\cancel{k}_1)\gamma^5\bigg\}\nonumber\\
&=\frac{g^4}{8\pi^8}\int_0^1\!\mathrm{d}x\!\int_0^1\!\mathrm{d}y\!\int\!\mathrm{d}^4l_1\!\int\!\mathrm{d}^4l_2\nonumber\\
&\times\frac{1}{\left\{l^2_1+x(1-x)k^2_1-m_\psi^2\right\}\left\{l_2^2+y(1-y)k_1^2-m_\psi^2\right\}\left\{[l_1+l_2+(y-x)k_1]^2-m_B^2\right\}}\nonumber\\
&+\text{logarithmic contributions}+\text{finite terms}\nonumber\end{aligned}$$ $$\begin{aligned}
&=\frac{g^4}{8\pi^8}\int_0^1\!\mathrm{d}x\!\int_0^1\!\mathrm{d}y\,\pi^4M_c^2\left[3\left(\ln\frac{M_c^2}{q^2_0}-\gamma_\omega\right)+1\right]+\text{logarithmic contributions}\nonumber\\
&+\text{finite terms}\nonumber\\
&=\frac{g^4}{8\pi^4}M^2_c\left[3\left(\ln\frac{M^2_c}{q^2_0}-\gamma_\omega\right)+1\right]+\text{logarithmic contributions}+\text{finite terms}.\end{aligned}$$ The counterterm diagrams (v1) and (y1) are given by $$\begin{aligned}
(v1)=(y1)=0,\end{aligned}$$ as the counterterm vertices $\delta Z^{(1)}_{\bar{\psi}\psi A}=\delta Z^{(1)}_{\bar{\psi}\psi B}=0$. In this article, we are interested in the quadratic contributions only, and have not tracked the logarithmic (sub)contributions and finite terms. The technical subtleties associated with LORE’s treatment of overlapping contributions have already been demonstrated in Ref. [@Huang:2011xh; @Huang:2012iu], and we recommend interested readers to go to those references for a comprehensive treatment.
[999]{}
G. Aad [*et al.*]{} \[ATLAS Collaboration\], “Observation of a new particle in the search for the Standard Model Higgs boson with the ATLAS detector at the LHC,” Phys. Lett. B [**716**]{}, 1 (2012) \[arXiv:1207.7214 \[hep-ex\]\]. S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], “Observation of a new boson at a mass of 125 GeV with the CMS experiment at the LHC,” Phys. Lett. B [**716**]{}, 30 (2012) \[arXiv:1207.7235 \[hep-ex\]\]. K. G. Wilson and J. B. Kogut, “The Renormalization group and the epsilon expansion,” Phys. Rept. [**12**]{}, 75 (1974). M. J. G. Veltman, “The Infrared - Ultraviolet Connection,” Acta Phys. Polon. B [**12**]{}, 437 (1981). D. Bai, J. W. Cui and Y. L. Wu, “Quantum Electroweak Symmetry Breaking Through Loop Quadratic Contributions,” Phys. Lett. B [**746**]{}, 379 (2015) \[arXiv:1412.3562 \[hep-ph\]\]. C. Patrignani [*et al.*]{} \[Particle Data Group\], “Review of Particle Physics,” Chin. Phys. C [**40**]{}, no. 10, 100001 (2016). L. Girardello and M. T. Grisaru, “Soft Breaking of Supersymmetry,” Nucl. Phys. B [**194**]{}, 65 (1982). G. ’t Hooft and M. J. G. Veltman, “Regularization and Renormalization of Gauge Fields,” Nucl. Phys. B [**44**]{}, 189 (1972). W. Siegel, “Supersymmetric Dimensional Regularization via Dimensional Reduction,” Phys. Lett. [**84B**]{}, 193 (1979).
S. P. Martin and M. T. Vaughn, “Regularization dependence of running couplings in softly broken supersymmetry,” Phys. Lett. B [**318**]{}, 331 (1993) \[hep-ph/9308222\]. Y. L. Wu, “Symmetry principle preserving and infinity free regularization and renormalization of quantum field theories and the mass gap,” Int. J. Mod. Phys. A [**18**]{}, 5363 (2003) \[hep-th/0209021\]. Y. L. Wu, “Symmetry preserving loop regularization and renormalization of QFTs,” Mod. Phys. Lett. A [**19**]{}, 2191 (2004) \[hep-th/0311082\]. J. W. Cui and Y. L. Wu, “One-Loop Renormalization of Non-Abelian Gauge Theory and beta Function Based on Loop Regularization Method,” Int. J. Mod. Phys. A [**23**]{}, 2861 (2008) \[arXiv:0801.2199 \[hep-ph\]\]. Y. B. Dai and Y. L. Wu, “Dynamically spontaneous symmetry breaking and masses of lightest nonet scalar mesons as composite Higgs bosons,” Eur. Phys. J. C [**39**]{}, S1 (2005) \[hep-ph/0304075\]. Y. Tang and Y. L. Wu, “Gravitational Contributions to the Running of Gauge Couplings,” Commun. Theor. Phys. [**54**]{}, 1040 (2010) \[arXiv:0807.0331 \[hep-ph\]\]. Y. Tang and Y. L. Wu, “Quantum Gravitational Contributions to Gauge Field Theories,” Commun. Theor. Phys. [**57**]{}, 629 (2012) \[arXiv:1012.0626 \[hep-ph\]\]. Y. Tang and Y. L. Wu, “Gravitational Contributions to Gauge Green’s Functions and Asymptotic Free Power-Law Running of Gauge Coupling,” JHEP [**1111**]{}, 073 (2011) \[arXiv:1109.4001 \[hep-ph\]\].
J. W. Cui, Y. Tang and Y. L. Wu, “Renormalization of Supersymmetric Field Theories in Loop Regularization with String-mode Regulators,” Phys. Rev. D [**79**]{}, 125008 (2009) \[arXiv:0812.0892 \[hep-ph\]\]. J. W. Cui, Y. L. Ma and Y. L. Wu, “Explicit derivation of the QED trace anomaly in symmetry-preserving loop regularization at one-loop level,” Phys. Rev. D [**84**]{}, 025020 (2011) \[arXiv:1103.2026 \[hep-ph\]\]. D. Huang, Y. Tang and Y. L. Wu, “Note on Higgs Decay into Two Photons $H\to \gamma\gamma$,” Commun. Theor. Phys. [**57**]{}, 427 (2012) \[arXiv:1109.4846 \[hep-ph\]\]. D. Huang and Y. L. Wu, “Consistency and Advantage of Loop Regularization Method Merging with Bjorken-Drell’s Analogy Between Feynman Diagrams and Electrical Circuits,” Eur. Phys. J. C [**72**]{}, 2066 (2012) \[arXiv:1108.3603 \[hep-ph\]\]. D. Huang, L. F. Li and Y. L. Wu, “Consistency of Loop Regularization Method and Divergence Structure of QFTs Beyond One-Loop Order,” Eur. Phys. J. C [**73**]{}, no. 4, 2353 (2013) \[arXiv:1210.2794 \[hep-ph\]\]. Y. L. Wu, “Quantum Structure of Field Theory and Standard Model Based on Infinity-free Loop Regularization/Renormalization,” Int. J. Mod. Phys. A [**29**]{}, 1430007 (2014) \[arXiv:1312.1403 \[hep-th\]\]. R. Chapling, “Asymptotics of Certain Sums Required in Loop Regularisation,” Mod. Phys. Lett. A [**31**]{}, no. 04, 1650030 (2016) \[arXiv:1601.04966 \[hep-th\]\]. R. Chapling, “Note on Closed-Form Expressions for Irreducible Loop Integrals,” arXiv:1608.05311 \[hep-th\]. Y. L. Wu, “Maximal symmetry and mass generation of Dirac fermion and gravitational gauge field theory in six-dimensional spacetime,” arXiv:1703.05436 \[hep-th\]. Y. L. Wu, “Unified field theory of basic forces and elementary particles with gravitational origin of gauge symmetry in hyper-spacetime,” arXiv:1705.06365 \[hep-th\]. W. Pauli and F. Villars, “On the Invariant regularization in relativistic quantum theory,” Rev. Mod. Phys. [**21**]{}, 434 (1949). J. Iliopoulos and B. Zumino, “Broken Supergauge Symmetry and Renormalization,” Nucl. Phys. B [**76**]{}, 310 (1974). C. Gnendiger [*et al.*]{}, “To ${d}$, or not to ${d}$: recent developments and comparisons of regularization schemes,” Eur. Phys. J. C [**77**]{}, no. 7, 471 (2017) \[arXiv:1705.01827 \[hep-ph\]\]. O. A. Battistel, A. L. Mota and M. C. Nemes, “Consistency conditions for 4-D regularizations,” Mod. Phys. Lett. A [**13**]{}, 1597 (1998). A. P. Baeta Scarpelli, M. Sampaio, B. Hiller and M. C. Nemes, “Chiral anomaly and CPT invariance in an implicit momentum space regularization framework,” Phys. Rev. D [**64**]{}, 046013 (2001) \[hep-th/0102108\]. A. P. Baeta Scarpelli, M. Sampaio and M. C. Nemes, “Consistency relations for an implicit n-dimensional regularization scheme,” Phys. Rev. D [**63**]{}, 046004 (2001) \[hep-th/0010285\]. R. Pittau, “A four-dimensional approach to quantum field theories,” JHEP [**1211**]{}, 151 (2012) \[arXiv:1208.5457 \[hep-ph\]\]. W. Siegel, “Inconsistency of Supersymmetric Dimensional Regularization,” Phys. Lett. [**94B**]{}, 37 (1980). J. D. Bjorken and S. D. Drell, “Relativistic quantum fields.”
J. Wess and B. Zumino, “Supergauge Transformations in Four-Dimensions,” Nucl. Phys. B [**70**]{}, 39 (1974). J. Wess and B. Zumino, “A Lagrangian Model Invariant Under Supergauge Transformations,” Phys. Lett. [**49B**]{}, 52 (1974). M. E. Peskin and D. V. Schroeder, “An Introduction to quantum field theory,” Reading, USA: Addison-Wesley (1995) 842 p A. Denner, H. Eck, O. Hahn and J. Kublbeck, “Feynman rules for fermion number violating interactions,” Nucl. Phys. B [**387**]{}, 467 (1992). A. Denner, H. Eck, O. Hahn and J. Kublbeck, “Compact Feynman rules for Majorana fermions,” Phys. Lett. B [**291**]{}, 278 (1992). T. Hahn, “Generating Feynman diagrams and amplitudes with FeynArts 3,” Comput. Phys. Commun. [**140**]{}, 418 (2001) \[hep-ph/0012260\]. T. Gherghetta, “Regularization in the gauged Nambu-Jona-Lasinio model,” Phys. Rev. D [**50**]{}, 5985 (1994) \[hep-ph/9408225\].
[^1]: Noticeably, it is shown in Ref. [@Martin:1993yx], DREG and DRED can lead to different running couplings in softly broken SUSY, due to the fact the latter preserves SUSY, while the former doesn’t.
[^2]: These symmetry-preserving properties of LORE are believed to hold at arbitrary loops. At present, the preservation of Non-Abelian gauge symmetries and SUSY has been verified at one loop by checking the Slavnov-Taylor identities [@Cui:2008uv] and SUSY Ward identities [@Cui:2008bk], while the preservation of Abelian gauge symmetries has been verified at two loops by checking the Ward identities [@Huang:2012iu].\[Preservation\]
[^3]: For an introduction to recent developments and comparisons of regularization scheme, we recommend Ref. [@Gnendiger:2017pys], which also contains a discussion on other four-dimensional regularization schemes such as the implicit regularization (IREG) [@Battistel:1998sz; @BaetaScarpelli:2001ix; @BaetaScarpelli:2000zs] and four-dimensional renormalization (FDR) [@Pittau:2012zd]. Although not included in our current goals, comparing LORE with these regularization schemes would be interesting and important. Hopefully, we could handle this issue in future studies.
[^4]: As shown in Ref. [@Cui:2008bk], such kind of momentum shift is perfectly legal in LORE, even though at this stage, one has not introduced any regularization scheme yet. One can show explicitly that the order of doing momentum shift and LORE does not matter at all, and different orders should give the same results. So here we choose to do the momentum shift first and then use LORE to regularize the integrals.
[^5]: In the real calculations, one encounters extra complications from nontrivial numerators other than 1. The discussion here only deals with the simplest case of $I_{\alpha\beta\gamma}$, and the general cases can be treated in a similar way.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Medium-resolution optical spectroscopy of the candidate YSOs associated with the small, nearby molecular cloud Lynds 1333 revealed four previously unknown classical T Tauri stars, two of which are components of a visual double, and a Class I source, *IRAS* 02086+7600. The spectroscopic data, together with new $V,R_{C},I_{C}$ photometric and 2MASS $J$, $H$, and $K_s$ data allowed us to estimate the masses and ages of the new T Tauri stars. We touch on the possible scenario of star formation in the region. L1333 is one of the smallest and nearest known star forming clouds, therefore it may be a suitable target for studying in detail the small scale structure of a star forming environment.'
author:
- |
M. Kun$^{1}$[^1], S. Nikolić$^{2}$, L. E. B. Johansson$^{3}$, Z. Balog$^{4}$[^2] and A. Gáspár$^{5}$\
$^{1}$Konkoly Observatory, H-1525 Budapest, P.O. Box 67, Hungary\
$^{2}$Departamento de Astronomía Universidad de Chile, Casilla 36-D, Santiago, Chile\
$^{3}$Onsala Space Observatory, S-439 92 Onsala, Sweden\
$^{4}$Steward Observatory, University of Arizona, 933 N. Cherry Av., Tucson AZ, USA 85721\
$^{5}$University of Szeged, Dóm tér 9, Szeged, H-6720 Hungary
date: 'Received / Accepted '
title: 'Low-mass star formation in Lynds 1333'
---
\[firstpage\]
ISM: clouds; ISM: individual: L1333; stars: formation; stars: pre-main-sequence
Introduction {#Sect_1}
============
Filamentary molecular clouds with embedded dense cores form a remarkable subset of star forming clouds in our galactic environment [e.g. @Onishi96; @NC05; @HTPG05]. Young stellar objects (YSOs) are associated with several cores along the filaments. The formation scenario of the filaments and stars within them, however, are not well understood. The filaments may be parts of shells, swept up by powerful stellar winds or supernovae [e.g. @KMT04], or may result from fragmentation of sheet-like structures [@Hartmann02], or may be shaped by large-scale flows like the galactic rotation [@Koda]. Detailed studies of their density and velocity structures, as well as the properties of the YSOs born in them may help understand their formation and evolution.
Lynds 1333, a small dark cloud of opacity class 6 (Lynds 1962) in Cassiopeia, at (l,b)=(12888,+1371) is part of a filamentary complex. According to the available observations L1333 is starless, and thus has been included in several studies of starless cores [e.g. @LMT99; @LMT01; @Lee04]. @OKS [hereinafter referred to as Paper I], studied first this cloud. They derived a distance of $180\pm20$pc from the Sun using Wolf diagram method. Their $^{13}$CO and C$^{18}$O observations have shown L1333 to be part of a long, filamentary molecular structure, stretching from $l \sim 126\degr$ to $133\degr$ and from $b \sim +13\degr$ to $+15\degr$, and referred to this molecular complex as [*L1333 molecular cloud*]{}. The angular extent of the molecular complex corresponds to a length of some 30pc at a distance of 180pc. @KMT04 found that the L1333 complex is part of a giant far infrared loop GIRL G126+10.
Recent star formation in the L1333 molecular cloud complex has been indicated by the presence of the *IRAS* source *IRAS* 02086+7600, whose *IRAS* colour indices are indicative of a Class I protostar, nevertheless it coincides with a faint star in the *Digitized Sky Survey* image. Due to its appearance as an optically visible star with large far-infrared excess several authors considered this object as a possible evolved star. @FNP included *IRAS* 02086+7600 in a multiband photometric survey for candidate post-AGB stars. They could not confirm the post-AGB nature of the star, and noted that it may be an ultracompact region, or a post-AGB star, or a YSO. *IRAS* 02086+7600 appeared as a possible planetary nebula in the target lists of @PM88 and @VSP95.
Based on its *IRAS* colours, @Slysh included this object, as a candidate ultracompact region, in their search for OH maser emission. They detected it as a thermal OH source at the velocity of 3.1kms$^{-1}$. The molecular maps presented in Paper I revealed that this *IRAS* source is projected on a dense C$^{18}$O core of a nearby molecular cloud whose radial velocity is +3.0kms$^{-1}$, same as that of the OH source, suggesting that *IRAS* 02086+7600 most probably is a low-mass YSO. The C$^{18}$O spectrum observed at its position exhibited a wing-like feature, indicative of molecular outflow (Paper I). No known Herbig–Haro object is associated with this source.
In addition to *IRAS* 02086+7600, 18 H$\alpha$ emission stars have been detected in objective prism Schmidt plates in the region of L1333 (Paper I). Three of these stars are associated with the *IRAS* point sources *IRAS* F02084+7605, 02103+7621, and 02368+7453.
The aim of our present study is to establish an elementary data base on the star forming activity of the L1333 complex. We observed the optical spectra of the candidate YSOs in order to establish their pre-main-sequence nature and their spectral types. We also performed optical photometry of the objects in order to determine their luminosities and positions in the HRD. We describe our observational data in Sect. \[Sect\_2\]. Our results on the properties of the observed stars, a short description of the large-scale environment of the cloud, and the possible star formation scenario are presented in Sect. \[Sect\_3\]. Sect. \[Sect\_4\] gives a short summary.
Observations and results {#Sect_2}
========================
Spectroscopy {#Sect_2.1}
------------
All the PMS star candidate H$\alpha$ emission objects and *IRAS* sources listed in Paper I were observed on 4th January 2001, using the *ALFOSC* spectrograph installed on the 2.5-m Nordic Optical Telescope in the Observatorio del Roque de los Muchachos in La Palma. The spectra were taken through grism 8, giving a dispersion of 1.5Å/pixel over the wavelength region 5800–8350Å. Using a 1-arcsec slit the spectral resolution was $\lambda / \Delta \lambda \approx 1000$ at $\lambda=6560$Å. The exposure times of 900s for the H$\alpha$ emission stars resulted in $S / N \ga 100$. For the much fainter *IRAS* 02086+7600 the exposure time was 2400s, resulting in $S / N \approx 20$. Spectra of helium and neon lamps were observed before and after each stellar observation for wavelength calibration. We observed a series of spectroscopic standards for spectral classification purposes. *IRAS* 02086+7600 was also observed on 13 September 2005, using the CAFOS instrument on the 2.2-m telescope of Calar Alto Observatory. Using the grism R-100, the observed part of the spectrum covered the wavelength interval 5800–9000Å. The spectral resolution of CAFOS observation, using a 1.5-arcsec slit, was $\lambda / \Delta \lambda \approx 1000$ at $\lambda=8500$Å. The exposure time 2400s resulted in $S / N \approx 7$ at 8500Å. We reduced and analysed the spectra using standard [iraf]{} routines.
We confirmed the pre-main-sequence nature of three candidates listed in Paper I: OKSH$\alpha$ 5, 6, and 16, all coinciding with *IRAS* point sources and projected on the molecular clouds. The other candidate H$\alpha$ objects listed in Paper I proved to be field stars without prominent H$\alpha$ emission and LiI absorption. We found by chance during the observations that a faint star some 1.8 arcsec south–southeast of OKSH$\alpha$ 6, associated with *IRAS* 02103+7621, was also a pre-main-sequence star. We refer to the two components as OKSH$\alpha$6N and OKSH$\alpha$6S, respectively.
The wavelength range of *ALFOSC* spectra was suitable for determining several flux ratios defined as tools for spectral classification by @Kirk, @MK, and @PGZ. We measured these spectral features on the spectra of our stars, and calibrated them against the spectral type and luminosity class by measuring them in a series of standard stars observed during the same run. The accuracy of the two-dimensional spectral classification, estimated from the range of spectral types obtained from different flux ratios, is $\pm1$ subclass @KPNJW.
Results of the spectroscopy are presented in Table \[Tab1\]. In addition to the derived spectral types we present the equivalent widths of the H$\alpha$ and lines in Å, the 10%-width of the H$\alpha$ line in kms$^{-1}$, as well as list the additional emission lines observed in the spectra. The uncertainties given in parentheses have been derived from the repeatability of the measurements. The real uncertainties of the equivalent widths may be higher due to the blending of the line with neighbouring absorption or emission features $\lambda$6718, $\lambda$6717). The spectra, normalized to the continua, are shown in Fig. \[Fig1\].
---------------- ------------------ -------- ----------------- ------------------- --------------- -------------------------------
Star *IRAS* name Sp.T. $EW$(H$\alpha$) W(10%)(H$\alpha$) $EW$(Li[i]{}) Other emission lines
(Å) (kms$^{-1}$) (Å)
$\cdots$ 02086+7600$^{1}$ $<$ M0 $-$44.8(0.5) 570(10) $\cdots$ 6300, 6363; 6717, 6731; 6548,
6584; 7155; 7323; 7378
$\cdots$ 02086+7600$^{2}$ $-$34.0(1.0) 760(20) $\cdots$ 6300, 6363; 6717, 6731;
7155; 7319; 8446;
8498, 8542, 8662
OKSH$\alpha$5 F02084+7605 M0.5IV $-$25.1(0.5) 610(12) 0.47(0.02)
OKSH$\alpha$6N 02103+7621 K7V $-$51.5(0.5) 710(10) 0.40(0.02)
OKSH$\alpha$6S 02103+7621 M2IV $-$42.3(1.0) 465(15) 0.13(0.02) ; 5873, 6678; 6717, 6731
OKSH$\alpha$16 02368+7453 K7III $-$8.0(0.2) 520(15) 0.71(0.02)
---------------- ------------------ -------- ----------------- ------------------- --------------- -------------------------------
$^{1}$ ALFOSC spectrum, 2001; $^{2}$ CAFOS spectrum, 2005
Both Fig. \[Fig1\] and Table \[Tab1\] show that OKSH$\alpha$5, OKSH$\alpha$6N, OKSH$\alpha$6S, and OKSH$\alpha$16 are classical T Tauri stars (CTTS). Their spectral types are K or M, and their H$\alpha$ emission lines fulfil the criteria established for the CTTSs by @Martin (i. e. $W$(H$\alpha$) exceeds the threshold value of 5Å for K-type stars and 10Å for M-type stars) and by @WB03 (the width of the H$\alpha$ emission line 10% above the continuum level is significantly larger than 300kms$^{-1}$). The spectral types were converted into effective temperatures $T_\rmn{eff}$ following @KH95 for luminosity class V, and @deJager87 for luminosity classes IV and III. The adopted $T_\rmn{eff}$ values are listed in Table \[Tab\_result\].
*IRAS* 02086+7600 displays an emission spectrum, containing strong H$\alpha$ and several forbidden lines, as well as the triplet in the CAFOS spectrum. The spectral resolution and $S/N$ of the spectra are insufficient for identifying absorption features, suitable for spectral classification. In particular, no TiO band, conspicuous in M-type spectra, can be seen, suggesting that its spectral type is probably earlier than M0. The large number of forbidden lines resembles Class I objects [@KBTB98; @WH04], thought to be either younger than, or identical with the youngest CTTSs. The high far-infrared excess of Class I objects suggests that the central star and its accretion disc are embedded in a dusty envelope. Optical photons from such a source escape through the polar cavities of the envelope, cleared by the protostellar wind. Several Class I sources cannot be detected at optical wavelengths, and several others are extended, suggesting that starlight, scattered from the circumstellar dust, contributes to their optical flux [e.g. @Eisner]. The star-like appearance of *IRAS* 02086+7600 suggests that one of its polar cavities lies close to our line of sight. The extremely broad H$\alpha$ emission line supports this assumption.
, $_{\boldmath{C}}$, $_{\boldmath{C}}$ imaging and photometry {#Sect_2.2}
-------------------------------------------------------------
Photometric observations of the young stars in L1333 in the $V$, $R_{C}$ and $I_{C}$ bands were undertaken on 13 October 2001, 21 September 2003, and 11 December 2004 using the 1-m RCC-telescope of Konkoly Observatory. In 2001 a Wright Instruments EEV CCD05-20 CCD camera was used, whose pixel size of 22.5$\mu$m corresponded to 0.35 arcsec on the sky. In 2003 and 2004 we used a Princeton Instruments VersArray:1300B camera, that utilizes a back-illuminated, 1300$\times$1340 pixel Roper Scientific CCD. The pixel size is 20$\mu$m, corresponding to 0.31 arcsec on the sky. Integration times were between 180s and 600s. The open cluster NGC7790 was observed each night several times, at various airmasses, for calibrating the photometry. We reduced the images in [IRAF]{}. After bias subtraction and flatfield correction PSF-photometry was performed using the [DAOPHOT]{} package. The transformation formulae between the instrumental and standard magnitudes and colour indices as a function of the airmass were established each night by measuring the instrumental magnitudes of some 30 photometric standard stars published by @Stetson for NGC7790. OKSH$\alpha$6S was invisible in our $V$ images, thus only its $R_{C}$ and $I_{C}$ could be determined.
The results of the photometry for the three epochs are presented in Table \[Tab2\]. The photometric errors, given in parentheses, are quadratic sums of the formal errors of the instrumental magnitudes and those of the coefficients of the transformation equations. In some cases magnitudes measured at various epochs differ from each other by more than 0.1 mag. Comparison of the magnitudes of other stars within the field of the target objects has not shown such large discrepancies. Therefore we conclude that part of these deviations is due to the variability of the stars. In order to determine the interstellar extinction suffered by the stars and their luminosities we used the averages of the magnitudes presented in Table \[Tab2\].
Star band 13 Oct. 2001 21 Sept. 2003 11 Dec. 2004
---------------- ------------- -------------- --------------- -------------- --
*IRAS* 02086 V $\cdots$ $\cdots$ 19.65(0.07)
R$_\rmn{C}$ 18.28(0.06) 17.89(0.04) 17.95(0.04)
I$_\rmn{C}$ 16.65(0.03) 16.36(0.04) 16.55(0.04)
OKSH$\alpha$5 V 15.54(0.05) $\cdots$ 15.38(0.04)
R$_\rmn{C}$ 14.12(0.04) 14.02(0.03) 14.03(0.03)
I$_\rmn{C}$ 12.57(0.03) 12.47(0.03) 12.61(0.03)
OKSH$\alpha$6N V 13.17(0.03) 13.21(0.03) $\cdots$
R$_\rmn{C}$ 12.59(0.03) 12.50(0.03) $\cdots$
I$_\rmn{C}$ 11.55(0.02) 11.67(0.02) $\cdots$
OKSH$\alpha$6S V $\cdots$ $\cdots$ $\cdots$
R$_\rmn{C}$ $\cdots$ 15.20(0.10) $\cdots$
I$_\rmn{C}$ 13.86(0.05) 13.77(0.05) $\cdots$
OKSH$\alpha$16 V $\cdots$ 17.42(0.04) 17.55(0.05)
R$_\rmn{C}$ $ \cdots$ 15.52(0.03) 15.48(0.03)
I$_\rmn{C}$ $\cdots$ 13.56(0.03) 13.55(0.03)
: Results of optical photometry[]{data-label="Tab2"}
Our images have revealed *IRAS* 02086+7600 to be slightly extended. Small reflection nebulae can be seen next to both OKSH$\alpha$5 and *IRAS* 02086+7600. We chose PSF-photometry in order to minimize the contribution of the extended emission to the resulting magnitudes listed in Table \[Tab2\].
Discussion {#Sect_3}
==========
Interstellar extinction and SEDs {#Sect_3.1}
--------------------------------
We supplemented our observational data with the near infrared data of the 2MASS All Sky Catalog [@2MASS] and the far infrared data of the *IRAS* PSC and FSC in order to characterize the circumstellar environments of our target objects. Other infrared, e.g. *Spitzer* data are not available for these objects. A single object is associated with OKSH$\alpha$6 in both catalogues. The fluxes of the counterparts, 2MASSJ02152532+7635196 and *IRAS* 02103+7621 are thus combined from those of both components of the visual double.
We derived the interstellar extinction suffered by our stars based on the assumption that their total emission in the $I_{C}$ band originates from the photosphere [see e.g. @Meyer; @Cieza05], and the total $E_{R_{C}-I_{C}}$ colour excess results from interstellar reddening. The unreddened colour indices were adopted from @KH95 for luminosity classes V and IV, and from @Bessell for OKSH$\alpha$16, whose spectral features indicated a luminosity class III. The extinction $A_{V}$ was derived from $E_{R_{C}-I_{C}}$ as $A_{V} = 4.76 \times E_{R_{C}-I_{C}}$ [@Cohen]. For determining the extinction in the other photometric bands, we used the relations $A_{R_{C}} = 0.78\,A_{V}$, $A_{I_{C}} = 0.59\,A_{V}$ for the optical [@Cohen], and $A_{J} = 0.26\,A_{V}$, $A_{H} = 0.15\,A_{V}$, $A_{K_{s}} = 0.10\,A_{V}$ for the near-infrared. These latter values, slightly different from the standard @RL extinction law, are based on the relations presented for the 2MASS-bands by @NC05.
This method cannot be applied for *IRAS* 02086+7600, whose spectral type and thus photospheric colour indices are unknown. We derived the interstellar extinction suffered by this object by assuming that its unreddened position in $J-H$ vs. $H-K_{s}$ colour-colour diagram is on the [*T Tauri locus*]{} defined by @Meyer, i.e. its colour indices satisfy the relationship $(J-H)_0 = 0.58 (H-K)_0 + 0.52$. We also determined $A_{V}$ by this method for OKSH$\alpha$5 and OKSH$\alpha$16 and found that the results were compatible with those derived from $E_{R_{C}-I_{C}}$. The extinctions adopted are listed in column 2 of Table \[Tab\_result\]. We note that the dereddened colour index $R_{C}-I_{C}$ of *IRAS* 02086+7600 is $(R_{C}-I_{C})_{0} = 0.73$, suggesting a $\sim$ K7-type star, in accordance with the absence of TiO bands in the spectrum.
We constructed the spectral energy distributions of the stars using their magnitudes corrected for the interstellar extinction. The fluxes corresponding to zero magnitude were obtained from @Glass for the $V R_{C} I_{C}$ bands, and from the 2MASS All Sky Data release web document for the $J H K_s$ bands. The resulting dereddened SEDs are shown in Fig. \[Fig\_sed\]. In the plots of OKSH$\alpha$ stars, dashed lines show the SEDs of the photospheres, determined from the dereddened $I_{C}$ magnitudes and from the colour indices corresponding to the spectral types. The SEDs confirm the CTTS-nature of the OKSH$\alpha$ stars: their SEDs display significant far-infrared excesses and negative slopes between 2 and 25$\mu$m, characteristic of Class II infrared sources [@Lada], i.e. stars surrounded by dusty accretion disks. The plot of OKSH$\alpha$6 shows the sum of both components.
Contrary to the OKSH$\alpha$ stars, the slope of the SED of *IRAS* 02086+7600 is $d \log (\lambda F_{\lambda}) / d \log \lambda = 0.80$ between 2 and 25$\mu$m, characteristic of Class I sources. The shape of the SED over the wavelength interval 0.55–25$\mu$m can be well matched with the sum of three blackbodies, indicated in Fig. \[Fig\_sed\], and suggesting three dominant temperatures in the inner regions of the system. The hottest component, fitted to the optical part of the SED (dashed line), has $T \approx 4000\,K$ and corresponds to the photosphere of a K7-type central star. Ignorance of the contribution of veiling and scattered light to the optical fluxes make this temperature estimate somewhat uncertain. We rely on this value, in view of the observational results that the veiling is constant redward of $\sim$ 5000Å [e.g. @BB90; @WH04], and assuming that we were able to exclude a considerable part of the scattered light by performing PSF-photometry. The uncertainty, estimated from the goodness of the fit, is $\pm200$K. The next component of the SED is a blackbody with $T = 1400\,K$, quite similar to the spectra of the NIR excesses of classical T Tauri stars [@Muzerolle], which have been succesfully modelled as the ‘photosphere’ of the inner rim of the disc, emitting like a blackbody near the dust sublimation temperature. The third dominant temperature, suggested by the shape of the SED, is $\sim 130$K. Beyond 25$\mu$m the SED turns flat, suggesting the peak position between 60 and 100$\mu$m.
Positions of the pre-main-sequence stars in the HRD {#Sect_3.2}
---------------------------------------------------
Bolometric luminosities of the OKSH$\alpha$ stars were derived by applying the bolometric corrections $BC_{I_C}$, tabulated by @Hartigan to the dereddened $I_{C}$ magnitudes, and adopting a distance of 180pc. The distribution of the stars in the HRD is displayed in Fig. \[Fig\_hrd\]. Errors of $\log\,T_\rmn{eff}$ were derived from the accuracy of $\pm\,1$ spectral subclass of the spectral classification. The error of the luminosity comes from the quadratic sum of $\delta I_{C} \approx 0.03$, $\delta A_{I_{C}} \approx 0.42$, $\delta BC_{I_{C}} \approx 0.01$ and $\delta (5 \log D) \approx 0.08$. The effect of distance uncertainty on the relative positions of the stars was estimated with the assumption that the scatter of the distances of stars is same as their largest projected separation, i.e. $\sim$7pc.
The luminosity of *IRAS* 02086+7600 over the wavelength interval 0.55–135$\mu$m was calculated by integrating the available dereddened fluxes. Applying a long-wavelength correction as proposed by @Kenyon90 for objects with SED peaks near $100\,\mu$m, $L_{>135} \approx 0.86\,L_{100}$, and neglecting the luminosity below 0.55$\mu$m, we obtained a total luminosity of $L_{tot} = 1.04\,L_{\sun}$. In addition to the bolometric luminosity of the central star ($L^{*}_{bol}$), this total luminosity includes the luminosity of the accretion shock ($L_{acc,shock}$), and those generated and reprocessed in the disc ($L_{acc,disc}$ and $L_{rep,disc}$). In order to obtain $L^{*}_{bol}$, contributions of $L_{acc,shock}$, $L_{acc,disc}$, and $L_{rep,disc}$ have to be subtracted from $L_{tot}$. To this end we utilized two empirical relationships, resulted from comprehensive studies of large YSO samples. First, it was shown by @WH04, that the total luminosity of Class I sources can be approximated as $L_{tot} = 1.08\,L^{*}_{bol} + 1.58\,L_{acc,shock}$, where $L_{acc,shock} = (G M_{*} \dot{M}) / R_{*}$. The second relationship is that established by @Muzerolle between $\dot{M}$ and the luminosity of the $\lambda$8542 line. Following their method, the measured $EW(\mbox{Ca\,{\sc ii}}\,\lambda\,8542) = 10$Å resulted in $\dot{M} \approx 5.7 \times 10^{-9} M_{\sun} / yr$, comparable to the median value $\dot{M} = 7.9\times 10^{-9} M_{\sun} / yr$, obtained by @WH04 for the Class I objects of Taurus. With the assumptions $M_{*} = 0.8\,M_{\sun}$, and $R_{*} = 2\,R_{\sun}$ the obtained mass accretion rate led to $L_{acc,shock} = 0.07\,L_{\sun}$ and $L^{*}_{bol} = 0.86\,L_{\sun}$. The resulting temperature and luminosity of *IRAS* 02086+7600 is plotted as the open circle in Fig. \[Fig\_hrd\]. The uncertainty was calculated as in the case of OKSH$\alpha$ stars.
Evolutionary tracks and isochrones, as well as the position of the birthline and zero-age main-sequence [@PS99] are also shown in Fig. \[Fig\_hrd\]. We obtained masses 0.8M$_{\sun}$ and 0.2M$_{\sun}$ for OKSH$\alpha$6N and OKSH$\alpha$6S, respectively. Our data suggest that within the accuracy both components are coeval, $\sim$3–5 million years old. OKSH$\alpha$5, OKSH$\alpha$16, and *IRAS* 02086+7600 lie close to the $10^{6}$-yr isochrone. The weak-line T Tauri stars identified by @TNKF05 near L1333 are also plotted in Fig. \[Fig\_hrd\] (see Sect. \[Sect\_3.4\]). Table \[Tab\_result\] summarizes our main results: $A_{V}$, $T_\rmn{eff}$, $L / L_{\sun}$, as well as the masses in $M_{\sun}$ and ages in $10^{6}$yrs, read from Fig. \[Fig\_hrd\].
-0.4cm \[Fig\_hrd\]
---------------- ------------- --------------- -------------- -------------- ---------------
Star $A_\rmn{V}$ $T_\rmn{eff}$ $L$ Mass Age
(mag) (K) ($L_{\sun}$) ($M_{\sun}$) (10${^6}$ yr)
*IRAS* 02086 3.76(1.2:) 4000 0.86 0.8 1.5
OKSH$\alpha$5 2.57(0.42) 3570 0.58 0.3 1.0
OKSH$\alpha$6N 0.71(0.42) 4060 0.49 0.8 5.0
OKSH$\alpha$6S 1.29(0.55) 3410 0.11 0.15 3.0
OKSH$\alpha$16 5.28(0.50) 3870 0.93 0.4 1.0
---------------- ------------- --------------- -------------- -------------- ---------------
: Properties of the pre-main-sequence stars associated with L1333, derived from spectroscopic and photometric data[]{data-label="Tab_result"}
Large-scale environment of L1333 {#Sect_3.3}
--------------------------------
The map of the visual extinction of the region $117\degr \lid l \lid 135\degr$ and $+10\degr \lid b \lid +17\degr$, taken from the [*Atlas and Catalog of Dark Clouds*]{} by @DUK and displayed in Fig. \[Fig\_ext\], shows that L1333 is near the middle of a long, diffuse filamentary cloud complex spanning from $l \sim 120\degr$ to $l\sim134\degr$, far beyond the limits of the molecular observations presented in Paper I. The dark cloud seen at $127\degr \la l \la 131\degr$ is catalogued as DUK 853 by @DUK and contains eight clumps ([*P1–P8*]{} in the order of decreasing mass). L1333 as catalogued by @Lynds corresponds to the largest clump [*P1*]{}. *IRAS* 02086+7600 and OKSH$\alpha$5 are located at the high-latitude edge of the second largest clump [*P2*]{}, and OKSH$\alpha$6 is projected on the edge of the small clump [*P7*]{} at the highest latitude side of DUK 853. OKSH$\alpha$16 is projected near the centre of clump [*P4*]{} of the same dark cloud. @DUK’s catalogue provides an opportunity to derive the masses of the clouds and their clumps. The total mass of the clouds within the diffuse filament between $l \sim 120\degr$ and $l \sim 134\degr$, derived from the visual extinction, is $\sim$2300M$_{\odot}$. Clump masses range between 2 and 30M$_{\odot}$.
In order to assess the star forming history of the whole region we also plotted in Fig. \[Fig\_ext\] the weak-line T Tauri stars identified by @TNKF05, and lying far from any dark cloud. @TNKF05 suggest that the parent clouds of these stars might have been connected to the L1333 complex. In order to properly compare the ages of these WTTSs with those of our CTTSs, we plotted their data, taken from @TNKF05’s Table 3, in Fig. \[Fig\_hrd\] (@TNKF05 used isochrones of @DM94, giving somewhat different results.). The ages of the WTTSs, assuming a distance of 200pc, are between 3 and 10 million years, with the youngest one at the highest longitude end of the chain, and the oldest on the low-latitude end.
A possible scenario of star formation {#Sect_3.4}
-------------------------------------
Comparison of the properties of dense C$^{18}$O cores of L1333 with those of other nearby star forming clouds have shown these cores to be smaller and less massive than the similar regions of Taurus, Ophiuchus, Lupus and Chamaeleon clouds [Paper I; @MHY; @TOMF02]. Star formation in such an environment is thought to be assisted by some external trigger, and the filamentary clouds themselves have probably been created by large-scale motions of the interstellar gas. The most plausible scenario, suggested by the arc-like structure is, that energetic stellar winds and/or supernova explosions of high-mass stars at lower galactic latitudes lifted the gas above the galactic plane and compressed it to form stars. In this case the apparent filament is a projection of a shell, and its line-of-sight extent may be comparable to its length. The distribution of YSOs relative to the clouds does not support this scenario. The young stars of L1333 are located at the high-latitude side of the cloud, with the oldest member, OKSH$\alpha$6, lying farthest from the cloud. The lack of H$\alpha$ emission stars, as well as YSO-like *IRAS* and 2MASS point sources on the low-latitude side of the filament suggests star formation propagating toward lower latitudes, and a source of trigger at higher galactic latitudes.
A possible candidate trigger source is the collision of high velocity gas with the giant radio continuum emitting region [*Loop III*]{}, described by @Verschuur. Our target objects are located near the far side of Loop III [@Berkhuijsen; @Spoelstra]. @Verschuur has shown that Loop III collided with high velocity gas originating from a galactic supershell some $7 \times 10^{5}$ ago. The collision has been well modelled for latitudes $b > 20\degr$. At lower latitudes, however, the behaviour of the supershell and its collision with the local interstellar matter has not yet been studied. In order to reveal the geometry of the possible collision in the latitude range $10\degr$–$20\degr$, the velocity distributions of both molecular and atomic gas have to be studied in detail. Closer to the galactic plane the high velocity gas of the supershell might have decelerated before reaching our galactic neighbourhood.
In this scenario the high-density regions, created by the colliding surfaces, have small line-of-sight extent. The ages obtained for *IRAS* 02086+7600, OKSH$\alpha$5, and OKSH$\alpha$16, taking into account their accuracies, support this scenario. OKSH$\alpha$6 was, however, born apparently before the collision. The weak-line T Tauri stars to the west of the cloud complex make the pattern of star formation of this region even more complicated. They indicate a prolonged star formation in the region. The age distribution of the WTTSs suggests star formation propagating from lower to higher galactic longitudes. More accurate age determinations and more detailed mapping of molecular velocity distribution are needed to clarify the picture.
Conclusions {#Sect_4}
===========
We identified five low-mass YSOs in the small filamentary molecular complex associated with the dark cloud Lynds 1333. Their masses are in the interval 0.15–0.8M$_{\sun}$, and they are 1–5 million years old. We confirmed that *IRAS* 02086+7600 is a Class I YSO associated with the L1333 complex, and found its age to be comparable to those of the CTTSs born in the same cloud. The relative distribution of YSOs and clouds suggests that the star formation might have been triggered by the collision of high velocity gas with Loop III.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is partly based on observations with Nordic Optical Telescope operated on the island of La Palma jointly by Denmark, Finland, Iceland, Norway, and Sweden, in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofísica de Canarias. The data presented here have been taken using ALFOSC, which is owned by the Instituto de Astrofísica de Andaluc[í]{}a (IAA) and operated at the Nordic Optical Telescope under agreement between IAA and the NBIfAFG of the Astronomical Observatory of Copenhagen. Our results are partly based on observations obtained at the Centro Astronómico Hispano Alemán (CAHA) at Calar Alto, operated jointly by the Max-Planck Institut für Astronomie and the Instituto de Astrofísica de Andalucía (CSIC). We are indebted to Francesco Palla for sending his data set on pre-main sequence evolution, and to László Szabados for careful reading of the manuscript. Financial support from the Hungarian OTKA grants T034584, T037508, TS049872, T042509, and T049082 is acknowledged. SN acknowledges support from the Chilean *Centro de Astrofísica* FONDAP No. 15010003 and Serbian Ministry of Science and Environmental Protection grant No. 146016.
Basri G., Batalha C., 1990, ApJ, 363, 654
Berkhuijsen E.M., 1971, A&A, 14, 359
Bessell M. S., 1979, PASP, 91, 589
Cieza L. A., Kessler-Silacci J. E., Jaffe D., Harvey P. M., Evans N. J. II., 2005, ApJ, 635, 422
Cohen J. G., Frogel J. A., Persson S. E., Elias J. H., 1982, ApJ, 249, 481
Cutri R. M., Skrutskie M. F., van Dyk S., et al., 2003, VizieR On-line Data Catalog: II/246
D’Antona F., Mazzitelli I., 1994, ApJS, 90, 467
de Jager C., Nieuwenhuijzen H., 1987, A&A, 177, 217
Dobashi K., Uehara H., Kandori R., Sakurai T., Kaiden M., Umemoto T., Sato F., 2005, PASJ, 57, S1
Eisner J. A., Hillenbrand L. A., Carpenter J. M., Wolf S., 2005, ApJ, 635, 396
Fujii T., Nakada Y., Parthasarathy M., 2002, A&A, 385, 884
Glass I. S., 1999, Handbook of Infrared Astronomy, Cambridge Univ. Press, p. 63
Hartigan P., Strom K. M., Strom S. E., 1994, ApJ, 427, 961
Hartmann L., 2002, ApJ, 578, 914
Heithausen A., Thaddeus P., 1990, ApJ, 353, 49
Hily-Blant P., Teyssier D., Phillip S., Güsten R., 2005, A&A, 440, 909
Kenyon S. J., Hartmann L., 1995, ApJS, 101, 117
Kenyon S. J., Hartmann L., Strom K. M., Strom S. E., 1990, AJ, 99, 869
Kenyon S. J., Brown D. I., Tout C. A., Berlind P. 1998, AJ, 115, 2491
Kirkpatrick J. D., Henry T. J., McCarthy D. W., 1991, ApJS, 77, 417
Kiss Cs., Moór A., Tóth L.V., 2004, A&A, 418, 131
Koda J., Sawada T., Hasegawa T., Scoville N. Z., 2006, ApJ, 638, 191
Kun M., Prusti T., Nikolić S., Johansson L. E. B., Walton N. A., 2004, A&A, 418, 89
Lada C. J., 1991, in Lada C.J., Kylafis N.D., eds., The Physics of Star Formation and Early Stellar Evolution. Kluwer, p.329
Lee C. W., Myers P. C., Tafalla M., 1999, ApJ, 531, 366
Lee C. W., Myers P. C., Tafalla M., 2001, ApJS, 136, 703
Lee C. W., Myers P. C., Plume R., 2004, ApJS, 153, 523
Lynds B. T., 1962, ApJS, 7, 1
Martín E. L., 1997, A&A, 321, 492
Martín E. L., Kun M., 1996, A&AS, 116, 467
Meyer M. R., Calvet N., Hillenbrand L. A., 1997, AJ, 114, 288
Mizuno A., Hayakawa T. Yamaguchi N., et al., 1998, ApJ, 507, L83
Muzerolle J., Calvet N., Hartmann L., D’Alessio P., 2003, ApJ, 597, L49
Nielbock M., Chini R., 2005, A&A, 434, 585
Obayashi A., Kun M., Sato F., Yonekura Y., Fukui Y., 1998, AJ, 115, 274 (Paper I)
Onishi T., Mizuno A., Kawamura A., Ogawa H., Fukui Y., 1996, ApJ, 465, 851
Palla F., Stahler S. W., 1999, ApJ, 525, 772
Preibisch T., Guenther E., Zinnecker H., 2001, AJ, 121, 1040
Preite-Martinez A., 1988, A&AS, 76, 317
Rieke G. H., Lebofsky M. J., 1985, ApJ, 288, 618
Slysh V. I., Dzura A. M., Valtts I. E., Gerard, E., 1994, A&AS, 106, 87
Spoelstra T. A. T., 1972, A&A, 21, 61
Stetson P. B., 2000, PASP, 112, 925
Tachihara K., Neuhäuser R., Kun M., Fukui Y., 2005, A&A, 437, 919
Tachihara K., Onishi T., Mizuno A., Fukui Y., 2002, A&A, 385, 909
Van de Steene G. C., Pottasch S. R., 1995, A&A, 299, 238
Verschuur G. L., 1993, ApJ, 409, 205
White R. J., Basri G., 2003, ApJ, 582, 1109
White R. J., Hillenbrand L. A., 2004, ApJ, 616, 998
[^1]: E-mail: [email protected]
[^2]: On leave from University of Szeged, Dept. of Optics and Quantum Electronics, Dóm tér 9, Szeged, H-6720 Hungary
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Given a set of points in the plane, a *crossing family* is a collection of segments, each joining two of the points, such that every two segments intersect internally. Aronov et al. \[Combinatorica, 14(2):127-134, 1994\] proved that any set of $n$ points contains a crossing family of size $\Omega(\sqrt{n})$. They also mentioned that there exist point sets whose maximum crossing family uses at most $\frac{n}{2}$ of the points. We improve the upper bound on the size of crossing families to $5\lceil \frac{n}{24} \rceil$. We also introduce a few generalizations of crossing families, and give several lower and upper bounds on our generalized notions.'
author:
- William Evans and Noushin Saeedi
bibliography:
- 'references.bib'
title: On problems related to crossing families
---
Introduction
============
Let $P$ be a set of $n$ points in general position in the plane. A collection of line segments, each joining two of the points, is called a *crossing family* if every two segments intersect internally. Let $\text{crf}(P)$ denote the size of the maximum crossing family in $P$, and let $\text{crf}(n)=\min_{|P|=n} \text{crf}(P)$, where the minimum is taken over all $n$-point sets $P$ in general position in the plane. @Pach-crf studied the size of $\text{crf}(n)$. They noted that a set of $n$ points chosen at random in a unit disc, “almost surely” has a linear-sized crossing family, and that there are point sets whose maximum crossing family uses at most $\frac{n}{2}$ of the points. They proved that any set of $n$ points contains a crossing family of size at least $\Omega(\sqrt{n})$. It is conjectured that $\text{crf}(P)=\Theta(n)$. We improve the upper bound on the size of crossing families. We consider more general variants of “crossings” and restricted classes of point sets. We study some generalized notions of crossing families, where we consider both combinatorial and geometric generalizations.
Point sets $A$ and $B$ are *separable* if they can be separated by a line. A point set $A$ *separates* point set $B$ from $C$ if $A$ and $B \cup C$ are separable and every line through two points in $A$ has all of $B$ on one side and all of $C$ on the other side. We show that any crossing family of a point set $A \cup B \cup C$ such that $A$ separates $B$ from $C$ has all its segments incident to one set (i.e. $A$ or $B$ or $C$). We exploit this “separating property” between subsets of points and design a template for constructing $n$-point sets whose crossing family is of size at most $5\lceil \frac{n}{24}\rceil$ (see Section \[sec:up\]).
A point set $A$ *avoids* a point set $B$ if no line formed by a pair of points in $A$ intersects the convex hull of $B$. $A$ and $B$ are *mutually avoiding* if $A$ avoids $B$ and $B$ avoids $A$. A point set that can be partitioned into separable point sets $A$ and $B$ such that $A$ avoids $B$ is *$1$-avoiding*. We study the combinatorial properties of crossing families in $1$-avoiding point sets. We relax the combinatorial properties of crossing families and introduce the notion of “side compatible subsets”. As a step towards finding maximum crossing families in $1$-avoiding point sets, we give linear bounds on side compatible subsets for some simplified versions of the problem (see Section \[sec:sc\]).
We also study some geometric generalizations of crossing families. A *spoke set* for $P$ is a set $\mathcal{L}$ of non-parallel lines such that each open unbounded region in the arrangement $\mathcal{L}$ has at least one point of $P$. The size of a spoke set $\mathcal{L}$ is the number of lines in $\mathcal{L}$. If $\text{crf}(P)=k$, then the size of the largest spoke set for $P$ is at least $k$. This is because by rotating the supporting line of each segment in a crossing family clockwise infinitesimally about the midpoint of the segment, we obtain a spoke set. @gcrf17 studied the properties of spoke sets in the dual plane. He claimed any $1$-avoiding point set of size $n$ has a spoke set of size $\frac{n}{4}$, but as we explain in Section \[sec:ss\], the proof does not seem to be correct.
We introduce a generalized notion for the dual of a spoke set, called an “M-semialternating path” (see Section \[sec:m\]). M-semialternating paths are related to pseudolines in two-coloured line arrangements that intersect (in order) lines of alternate colours in the arrangement. We show that the sizes of different types of M-semialternating paths are connected and give upper bounds on the size of certain M-semialternating paths. We also improve the upper bound on the size of spoke sets from $\frac{9n}{20}$ to $\frac{n}{4}+1$. Lastly, given a point set, we study the family of segments (each joining two of the points) such that for every pair, the supporting line of one intersects the interior of the other. We call such a family of segments a *stabbing family*. We show that the size of the largest stabbing family in an $n$-point set is $\frac{n}{2}$ (see Section \[sec:stab\]).
During the final preparation of this manuscript, we became aware of a recent concurrent work by @pach-new, showing that any set of $n$ points in general position has a crossing family of size at least $\dfrac{n}{2^{O(\sqrt{\log{n}})}}$. The methods used in our work and theirs are different, and even though we study generalized notions of crossing families or restricted classes of point sets, our results are different from theirs and not implied by their work.
Related work
============
Crossing families
-----------------
@Pach-crf introduced the notion of crossing families and studied the size of $\text{crf}(n)$. They noted that for non-convex point sets, the maximum crossing family may contain at most $\frac{n}{2}$ points. Two equal-sized disjoint sets $A$ and $B$ can be *crossed* if there exists a crossing family exhausting $A$ and $B$ such that each line segment connects a point in $A$ to a point in $B$. @Pach-crf characterized the properties of pairs of point sets that can be crossed. Consider point sets $A$ and $B$ which are separated by a line. A point $a$ in $A$ sees a point $b$ in $B$ at rank $i$ if $b$ is the $i$-th point seen counterclockwise from $a$ (starting from the direction of a separating line). Two sets $A$ and $B$ with cardinality $s$ obey the *rank condition* if there exist labelings $a_1,a_2,\dots,a_s$ and $b_1, b_2, \dots, b_s$ such that for all $i$, $a_i$ sees $b_i$ at rank $i$ and vice versa. If the labelings are such that for all $i$ and $j$, $a_i$ sees $b_j$ at rank $j$ and vice versa, then the sets obey the *strong rank condition*. An $n$-point set is *dense* if the ratio of the maximum distance between any pair of points to the minimum distance (between any pair of points) is $O(\sqrt{n})$.
@Pach-crf proved that two sets $A$ and $B$ can be crossed if and only if they obey the rank condition. $A$ and $B$ are mutually avoiding if and only if they obey the strong rank condition. They showed that any set of $n$ points contains a pair of mutually avoiding subsets of size $\Theta{\sqrt{n}}$, and hence $\text{crf}(n)=\Omega(\sqrt{n})$. @Valtr constructed a dense $n$-point set that contains no pair of mutually avoiding subsets of size more than $O(\sqrt{n})$. Since the strong rank condition is much stronger than the condition needed for having a crossing family (i.e. rank condition), it is believed that the lower bound can be improved. In fact, the conjecture is that $\text{crf}(n)=\Theta(n)$.
@Pach-crf noted that a set of $n$ points chosen at random in a unit disc, “almost surely” has a linear-sized crossing family. @Valtr-dense showed that the maximum crossing family in a dense point set is almost linear. @sol-hl proved that given a set $P$ of $2n$ points in general position, $\text{crf}(P)=n$ if and only if $P$ has exactly $n$ halving lines. A *halving line* in a point set $P$ with an even number of points, is a line through two points of $P$ such that both its half-planes contain exactly the same number of points.
Generalizations of crossing families
------------------------------------
Several generalizations of the notion of crossing families have been studied.
In recent years, there has been growing interest in studying pairwise crossings among other objects (rather than segments) formed by a given point set. A *geometric graph* is a graph drawn in the plane so that the vertices are represented by points in general position and the edges are represented by straight line segments connecting the corresponding points. A geometric graph is *complete* if there is an edge between every pair of points. An *H-crossing family* in a complete geometric graph is a set of vertex disjoint isomorphic copies of $H$ that are pairwise crossing, where two geometric subgraphs $h$ and $h'$ *cross* if there is an edge is $h$ that crosses an edge of $h'$. A crossing family can be defined as a $K_2$-crossing family (in a complete geometric graph). @tverberg showed that every complete geometric graph contains a $K_3$-crossing family of size $\lfloor \frac{n}{3}\rfloor$ (which is tight). @cr-ham showed the tight bound of $\frac{n}{4}$ for $P_4$-crossing families. @ghc18 showed that every complete geometric graph contains a $P_3$-crossing family of size $O(\sqrt{\frac{n}{2}})$, a $K_{1,3}$-crossing family of size $\frac{n}{6}$, and a $K_4$-crossing family of size $\frac{n}{4}$. Crossing families have also been studied in *topological graphs* where the edges are represented by Jordan curves (rather than straight line segments). Note that if two edges share an interior point, they must properly cross at that point. A topological graph is *simple* if every pair of edges intersect at most once (i.e. at a common endpoint or at a proper crossing). A graph is *$k$-quasi-planar* if it can be drawn as a topological graph with no $k$ pairwise crossing edges. It is conjectured that for any fixed $k\ge 2$, there exists $c_k$ such that every $k$-quasi-planar graph on $n$ vertices has at most $c_kn$ edges, where $c_k$ is a constant that depends on $k$. @quasi-planar proved the conjecture for $k=3$. @quasi-4 proved the conjecture for $k=4$. For $k>4$, the best known upper bound on the maximum number of edges in $k$-quasi-planar graphs in which no pair of edges intersect in more than $t$ points is $2^{\alpha(n)^{c}}n \log{n}$, where $\alpha(n)$ denotes the inverse Ackermann function and $c$ depends only on $k$ and $t$ [@suk-k-quasi]. For simple $k$-quasi-planar graphs, the best known upper bound is $c_k n \log(n)$ [@suk-k-quasi]. @valtr-kp showed that for any fixed $k \ge 3$, any geometric graph on $n$ vertices with no $k$ pairwise parallel edges contains at most $c_k n$ edges, where two edges are *parallel* if the intersection point of their supporting lines is not on either of the edges. He proved that any geometric graph on $n$ vertices with no $k$ pairwise crossing edges contains at most $c_k n \log(n)$ edges. He also showed that the same bound holds when the edges are drawn as $x$-monotone curves [@valtr-kc]. @k10 gave a simple $3$-quasi-planar drawing of $K_{10}$ (i.e. no three edges are pairwise crossing). It is known that the maximum crossing family of any geometric graph on ten vertices is at least three [@g-k10]. A (topological) graph is *$k$-planar* if it can be drawn in the plane such that no edge is crossed more than $k$ times. @k-quasi-rel proved that for $k \ge 3$, every simple topological $k$-planar graph can be redrawn (by rerouting some edges) to become $(k+1)$-quasi-planar.
@gcrf17 studied another generalization of crossing families called spoke sets (which was first introduced by @bose). Any crossing family of size $k$ guarantees a spoke set of size $k$, but the reverse is not true. @gcrf17 characterized the family of “spoke matchings”, which are geometric matchings in $2k$-point sets admitting a spoke set $\mathcal{L}$ of size $k$, where each matching edge connects the points in antipodal regions of the arrangement $\mathcal{L}$. He also studied the properties of spoke sets in the dual plane. He claimed that any $1$-avoiding $n$-point set has a spoke set of size $\frac{n}{4}$; however, the proof does not seem to be correct. Lastly, he proved that there exist $n$-point sets whose largest spoke set is of size at most $\frac{9}{20}n$ [@sch; @gcrf17].
A *cell-path* of a line arrangement is a sequence of cells in the arrangement such that any two consecutive cells of the sequence share a boundary edge and no cell appears more than once. A cell-path is *alternating* if the common edges of consecutive cells alternate in color. The dual of a spoke set corresponds to a cell-path with certain properties. @cell-path-Linda and @cell-path-Aich studied the existence of long cell-paths in line arrangements. They also studied the bicolored version of the problem, in which they look for long alternating cell-paths. These problems have close connections to problems on point sets in the plane (through duality). @rb-points give a survey on combinatorial problems on bicolored points in the plane.
Upper bound on crossing families {#sec:up}
================================
In this section, we show that there exist $n$-point sets whose crossing family is of size at most $5\lceil\frac{n}{24}\rceil$.
Let $A$, $B$, and $C$ be three disjoint point sets. We say $A$ *separates* $B$ from $C$ if
1. $A$ and $B \cup C$ are separable by a line\[sep-c1\], and
2. every line through two points in $A$ has all of $B$ on one side and all of $C$ on the other.\[sep-c2\]
\[def:separate\]
Let $P_1$, $P_2$, and $P_3$ be three disjoint point sets. Let $i \in \{1,2,3\}$. The three sets have the *separating property* if, for some $i$, $P_i$ separates $P_{i-1}$ from $P_{i+1}$, where indices are arithmetic modulo $3$. If for all $i$, $P_i$ separates $P_{i-1}$ from $P_{i+1}$, then the three sets satisfy the *full separating* property. Let $L(P_i)$ denote the set of all lines through two points in $P_i$. For three point sets $P_1$, $P_2$, and $P_3$ with the full separating property, we define the *core* to be the intersection of all bounded regions formed by any three lines $l_1$, $l_2$, and $l_3$, where $l_1 \in L(P_1)$, $l_2 \in L(P_2)$, and $l_3 \in L(P_3)$. We refer to a segment connecting a point in $P_i$ to a point in $P_j$ as a *$P_iP_j$-segment* or a segment of *type* $P_iP_j$. A segment is *incident* to a point set $P$ if it has an endpoint in $P$. We say a set $\mathcal{S}$ of segments *emanate* from a point set $P$ if all segments in $\mathcal{S}$ have an endpoint in $P$.
Let $A$, $B$, and $C$ be three disjoint point sets such that $A$ separates $B$ from $C$. Any crossing family in $A \cup B \cup C$ emanates from $A$ or $B$ or $C$. \[lem:split\]
Since $A$ separates $B$ from $C$, no $AB$-segment crosses an $AC$-segment; otherwise, the line through the the endpoints of these segments in $A$ would have at least one point from $B$ and one point from $C$ on the same side, which violates \[sep-c2\]. Thus, no crossing family can contain an $AB$-segment and an $AC$-segment.
Let $\mathtt{CH}(S)$ denote the convex hull of $S$, where $S$ is a point set. Since $A$ separates $B$ from $C$, by Definition \[def:separate\], $\mathtt{CH}(B \cup C) \cap \mathtt{CH}(A)=\emptyset$ (otherwise, $A$ and $B\cup C$ are not separable). Moreover, $\mathtt{CH}(A \cup B) \cap \mathtt{CH}(C)=\emptyset$ and $\mathtt{CH}(A \cup C) \cap \mathtt{CH}(B)=\emptyset$ (otherwise, property \[sep-c2\] is violated). This implies that no crossing family contains an $XY$-segment and a $ZZ$-segment, where $X,Y,Z \in \{A,B,C\}$ and $X,Y \neq Z$. Therefore, any crossing family emanates from $A$ or $B$ or $C$.
Lemma \[lem:split\] immediately implies that there exist sets of $n$ points whose crossing family is of size at most $\lceil\frac{n}{3}\rceil$. We extend the idea of having the separating property among certain subsets of the point set to obtain a better upper bound. First, we describe a template for constructing point sets with maximum crossing family of size at most $\lceil\frac{n}{4}\rceil$. We then modify the construction to improve the upper bound to $5\lceil\frac{n}{24}\rceil$.
We start with a set of four points in non-convex position. We denote the points on the convex hull by $p_1$, $p_2$, and $p_3$, and the point inside by $q$. Note that there is no crossing between the segments joining any two of these points. Now we grow a small disc around each point. Call the discs $D_1$, $D_2$, $D_3$, and $D_q$ accordingly. We replace each point $p$ with $\frac{n}{4}$ points such that
1. the new points are inside the disc around $p$,
2. for all $i \in [3]$, the new points in $D_i$ are almost on a line, and
3. for all $i \in [3]$, $D_i$ separates $D_{i+1}$ from $D_{i-1} \cup D_q$. The indices are arithmetic modulo $3$.
\[scale=.75\] (D) at (0,0); (D) – +(90: 5 cm) coordinate (C); (D) – +(330: 5 cm) coordinate (B); (D) – +(210:5 cm) coordinate (A);
\(A) – +(.2,-.35) coordinate (N1); at (N1) [[$k$]{}]{}; (B) – +(-.3,-.3) coordinate (N2); at (N2) [[$k$]{}]{}; (C) – +(.3,-.2) coordinate (N3); at (N3) [[$k$]{}]{}; at (D) [[$k$]{}]{};
in [A,B,C,D]{} () circle (0.75cm);
let 1=(A) in (1,1) – +(195:.45cm) coordinate (a); let 1=(a) in (1,1) – +(105:.07cm) coordinate (a1); let 1=(a) in (1,1) – +(285:.07cm) coordinate (a2); (a) – +(15:.9cm); (a1) – +(15:10cm); (a2) – +(15:10cm);
let 1=(B) in (1,1) – +(315:.45cm) coordinate (b); let 1=(b) in (1,1) – +(45:.07cm) coordinate (b1); let 1=(b) in (1,1) – +(225:.07cm) coordinate (b2); (b) – +(135:.9cm); (b1) – +(135:10cm); (b2) – +(135:10cm);
let 1=(C) in (1,1) – +(75:.45cm) coordinate (c); let 1=(c) in (1,1) – +(165:.07cm) coordinate (c1); let 1=(c) in (1,1) – +(345:.07cm) coordinate (c2); (c) – +(255:.9cm); (c1) – +(255:10cm); (c2) – +(255:10cm);
at (a) [[$D_1\hspace{.2cm}$]{}]{}; at (b) [[$D_2$ ]{}]{}; at (c) [[$\hspace{.25cm}D_3$ ]{}]{}; let 1=(D) in (1,1) – +(.6,0) coordinate (d); at (d) [[$\hspace{.1cm}D_q$ ]{}]{};
See Figure \[fig:nover4\]. The following lemma implies that the segments forming a crossing family in this configuration, should all have an endpoint in the same disk (thus, the maximum crossing family of such a point configuration is of size at most $\frac{n}{4}$).
Let $A,B,C$, and $D$ be four point sets such that the set $\{A, B, C\}$ satisfies full separating property and $D$ is inside the corresponding core. Any crossing family in $A \cup B \cup C \cup D$ emanates from one of $A,B,C,$ or $D$. \[lem:nover4\]
Let $P=A \cup B \cup C \cup D$. Note that for any $X,Y \in \{A,B,C,D\}$, $\mathtt{CH}(X \cup Y)$ and $\mathtt{CH}(P \setminus (X \cup Y))$ are disjoint. Thus, no crossing family can contain an $XY$-segment and a $WZ$-segment if $\{X,Y\} \cap \{W,Z\} =\emptyset$, where $X,Y,Z,W \in \{A,B,C,D\}$.
Let $\mathcal{D}=\{A,B,C,D\}$. Note that (since $A,B$, and $C$ satisfy the full separating property and $D$ is inside the corresponding core) every triple of sets in $\mathcal{D}$ satisfies the separating property. Thus, for any triple of sets, the segments of a crossing family induced by them have an endpoint in the same set. As a result, a crossing family not emanating from a set $X\in \mathcal{D}$ should be incident to all four sets. However, any collection of segments that is incident to all four sets without emanating from one of them must contain an $XY$-segment and a $WZ$-segment, where $\{X,Y\} \cap \{W,Z\}=\emptyset$, and hence cannot form a crossing family.
Note that in the point configuration illustrated in Figure \[fig:nover4\], there is no restriction on the orientation of the points in $D_q$, and all disks contain the same number of points. We modify this configuration such that a more specific positioning of the points in $D_q$ allows us to place more points in $D_q$ without increasing the size of the maximum crossing family. In the following, we describe how to construct such a point configuration in more detail. See Figure \[fig:5nover24\]. We start by putting three discs $D_1$, $D_2$, and $D_3$, each containing $5k$ almost collinear points, such that they satisfy the full separating property. In the core defined by them, we put six more discs $\{A_i,S_i \mid i \in [3]\}$, where each $S_i$ is a “super” disk containing two disks $B_i$ and $C_i$. Each disk $A_i, B_i$, and $C_i$ contains $k$ points. An *$X$-disk*, where $X \in \{A,B,C,D,S\}$ refers to any of the disks $X_1,X_2$, or $X_3$. Each disk $X_i$ (where $X \in \{A,B,C,D,S\}$) is obtained by rotating $X_{i-1}$ counterclockwise $\frac{2\pi}{3}$ around the origin (marked with $\times$ in Figure \[fig:5nover24\]). The indices are arithmetic modulo $3$. Let $P$ denote the set of all points. $P$ is *$3$-fold symmetric* with respect to the origin, that is, it can be partitioned into three sets (called *wings*) of size $\frac{|P|}{3}$ such that each wing rotated by $\frac{2\pi}{3}$ and $\frac{4\pi}{3}$ (around the origin) gives the other two wings. Each wing is composed of $A_i,B_i,C_i,D_i$ (for some $i \in [3]$) and the order of the disks along each wing in increasing distance from the origin is $A_i,B_i,C_i,D_i$. For each disk $X$, let $c(X)$ denote the center of $X$. If a disk $X$ contains a point $p$, we write $p \in X$. (For simplicity in writing, we may treat a disk $X$ as the set of points inside disk $X$ at times.) Let $\mathcal{D} =\{A_i,B_i,C_i,D_i \mid i \in [3]\}$. Every pair of disks in $\mathcal{D}$ are disjoint, and each disk is infinitesimally small so that for every triple $\langle X,Y,Z \rangle$ of disks (where $X,Y,Z \in \mathcal{D}$), the orientation of $\langle x,y,z \rangle$, where $x \in X$, $y\in Y$, $z \in Z$, is the same as the orientation of $\langle c(X),c(Y),c(Z) \rangle$. (Note that $S_i$ contains $B_i \cup C_i$, hence $S_i \notin \mathcal{D}$.) A triple $\langle X,Y,Z \rangle$ of disks, where $X,Y,Z \in \mathcal{D}$, has a *positive orientation* if the circle through points $\langle c(X),c(Y),c(Z) \rangle$ is traversed counterclockwise when we encounter the points in cyclic order $ c(X),c(Y),c(Z),c(X)$. A triple of disks in $\mathcal{D}$ has a *negative orientation* if the circle through $\langle c(X),c(Y),c(Z)\rangle$ is traversed clockwise. The set of disks $\{A_i,B_i,C_i \mid i \in [3]\}$ and the points inside them are such that
- the set $\{c(D_i),c(A_i) \mid i \in[3] \}$ has a crossing family of size three (the set of segments $\{\overline{c(D_i)c(A_{i+1})} \mid i \in[3]\}$ forms a crossing family),
- the maximum crossing family of $\{c(D_i),c(B_i) \mid i \in[3] \}$ is of size two,
- the maximum crossing family of $\{c(D_i),c(C_i) \mid i \in[3] \}$ is of size two,
- $\langle D_i,A_{i+1},B_i \rangle$ and $\langle D_i,A_{i+1},C_i \rangle$ have negative and positive orientations, respectively,
- $\langle B_i,A_{i+1},A_i \rangle$ and $\langle C_i,A_{i+1},A_{i+2} \rangle$ have negative and positive orientations, respectively,
- $\langle D_i,A_i,B_i \rangle$ and $\langle D_i,A_{i+2},C_i \rangle$ have positive and negative orientations, respectively,
- $\langle C_i,B_{i+1},A_i \rangle$ has a positive orientation, and
- the points in $S_i$ are almost collinear, and every line through two points in $S_i$ separates $D_i$ from $P \setminus (D_i \cup S_i)$.
![A $3$-fold symmetric set of $24$ points with maximum crossing family of size $5$.[]{data-label="fig:5nover24"}](5over24-3.pdf)
We show that the maximum crossing family of the point configuration described above is at most $5k$.
There exist sets of $n$ points whose crossing family is of size at most $5\lceil\frac{n}{24}\rceil$. \[thm:up-5nover24\]
Before proving Theorem \[thm:up-5nover24\], we draw attention to the following lemma.
If there exists an $m$-point set with maximum crossing family of size $f > 1$, then there exists an $n$-point set with maximum crossing family of size $ f\cdot\lceil \frac{n}{m} \rceil $. \[lem:grow\]
Let $P$ be an $m$-point set with maximum crossing family of size $f$. We use $P$ as a base set to construct an $n$-point set $P'$ as follows: each point in $P$ is replaced with $\lfloor \frac{n}{m} \rfloor$ or $\lceil \frac{n}{m} \rceil$ imperceptibly perturbed copies (each copy of a point is distinct). Let $\mathcal{F}'$ be a maximum crossing family in $P'$. Given $\mathcal{F}'$, contract each copy (of a point in $P$) to the original point (in $P$) and let $Q$ be the new set of segments obtained. Note that the segments in $Q$ are all pairwise touching (that is, they either intersect or have a common endpoint). Let $\mathcal{F} \subseteq Q$ be a maximum crossing family. For a set of segments $S$, let $p(S)$ denote the set of points induced by $S$. If $|\mathcal{F}|>1$, then $Q$ admits no triangle that is formed by a point in $p(Q) \setminus p(\mathcal{F})$ and a segment in $\mathcal{F}$ (otherwise, at least one pair of segments in $Q$ are disjoint). Therefore, in this case, the maximum crossing family in $P'$ is at most $f\cdot\lceil\frac{n}{m}\rceil$. If $|\mathcal{F}|=1$, the size of the crossing family in $P'$ may be up to $\frac{3}{2}\lceil\frac{n}{m}\rceil$.
Let $P$ be the point configuration described earlier when $k=1$ (depicted in Figure \[fig:5nover24\]). We show that the maximum crossing family of $P$ has size at most five. This, together with Lemma \[lem:grow\], concludes the proof.
Recall that in the point configuration we described, the orientation of every triple of points from three different disks is the same as the orientation of their disks. This implies that our configuration (with $k>1$) is a subclass of the point sets obtained from appealing to Lemma \[lem:grow\]. A *disk-contraction* contracts all points in a disk to the center of the disk. For a set $\mathcal{S}$ of segments, $\widetilde{\mathcal{S}}_{\mathcal{X}}$ denotes the set of segments obtained after performing a disk-contraction on $\mathcal{S}$, where each disk in $\mathcal{X}$ is contracted. We refer to a segment $\overline{c(X)c(Y)}$ in $\widetilde{\mathcal{S}}_{\mathcal{X}}$, where $X,Y \in \mathcal{X}$, as an $XY$-segment or a segment of type $XY$. An $XY$-segment is incident to disks $X$ and $Y$. Let $\mathcal{D}=\{A_i,B_i,C_i,D_i \mid i\in[3]\}$. Let $\mathcal{F}$ be a maximum crossing family of $P$. We first show that if $|\mathcal{F}| >5$, then $\mathcal{F}$ may only consist of segments of type $D_iX_j$, where $X \in \{A,B,C\}$. We then prove that any crossing family formed by these type of segments is indeed of size at most five. We consider the following cases.
1. \[c1\]$\mathcal{F}$ is not incident to any $D$-disks. The number of points in $P$ that are not in $D$-disks is nine. Hence, in this case, the crossing family is of size at most four.
2. \[c2\]$\mathcal{F}$ contains a $D_iD_j$-segment (where $i,j \in [3]$). Recall that the three sets $D_1,D_2$, and $D_3$ satisfy the full separating property and that $P \setminus \bigcup\limits_{i=1}^{3} D_i$ lies in the core defined by the $D$-disks. As a result, by Lemma \[lem:nover4\], all segments in $\mathcal{F}$ are incident to $D_i$ or all are incident to $D_j$. Thus, the size of the crossing family is at most five.
3. \[c3\]$\mathcal{F}$ is incident to no $A$-disks. Let $\widetilde{\mathcal{F}}=\widetilde{\mathcal{F}}_{\{S_i,D_i \mid i \in[3]\}}$. Recall that the maximum crossing family of $\{c(D_i),c(S_i) \mid i \in[3] \}$ is of size two. Suppose $|\mathcal{F}|>5$. There are three possibilities:
- $\widetilde{\mathcal{F}}$ admits a triangle. By Case \[c2\], $\mathcal{F}$ contains no $D_iD_j$-segments. Hence, the triangle is incident to at least two $S$-disks. Note that $c(S_i)$ can be incident to at most two segments in $\widetilde{\mathcal{F}}$. If $\widetilde{\mathcal{F}}$ contains more than three segments, then a vertex of the triangle is incident to at least three segments (because the segments in $\widetilde{\mathcal{F}}$ are pairwise touching), and hence the triangle should be incident to one $D$-disc. Since $\widetilde{\mathcal{F}}$ is not incident to any $A$-disks, $\widetilde{\mathcal{F}}$ cannot contain a segment touching all three segments of the triangle. Hence, $\widetilde{\mathcal{F}}$ consists of exactly three segments, and $|\mathcal{F}| \le \frac{3}{2}\cdot 2=3$.
- $\widetilde{\mathcal{F}}$ contains a pair of segments that are crossing. Note that $\widetilde{\mathcal{F}}$ admits no triangles and that $\mathcal{F}$ contains no $D_iD_j$-segments. As a result $|\mathcal{F}|\le 2 \cdot 2=4$.
- $\widetilde{\mathcal{F}}$ forms a star (i.e. has a point that is incident to all segments). Thus, $|\mathcal{F}| \le 5$.
4. \[c4\]$\mathcal{F}$ contains an $A_iA_j$-segment (where $i,j \in [3]$). Let $\widetilde{\mathcal{F}}=\widetilde{\mathcal{F}}_{\mathcal{D}}$. Let $e \in \widetilde{\mathcal{F}}$ be an $A_iA_j$-segment. All segments in $\widetilde{\mathcal{F}}$ are pairwise touching. The endpoints of $e$ cannot be incident to any other segments. Thus, each segment in $\widetilde{\mathcal{F}} \setminus e$ should intersect $e$. Moreover, any segment that crosses $e$ is incident to $c(A_k)$, where $k\neq i,j$. Thus, $|\mathcal{F}|=2$.
5. \[c5\]$\mathcal{F}$ contains an $S_iS_j$-segment (where $i,j \in [3]$). Note that any segment crossing an $S_iS_i$-segment is of type $D_iA_{i+1}$. Thus, if $\mathcal{F}$ contains an $S_iS_i$-segment, $|\mathcal{F}|=2$. Let $\widetilde{\mathcal{F}}=\widetilde{\mathcal{F}}_{\{A_i,S_i,D_i \mid i \in[3]\}}$. Let $e \in \widetilde{\mathcal{F}}$ be an $S_iS_j$-segment, where $i \neq j$. Either $\widetilde{\mathcal{F}}$ admits a triangle or not. We consider the two cases below.
- $\widetilde{\mathcal{F}}$ admits a triangle. Suppose $|\mathcal{F}|>5$. By Case \[c2\], $\widetilde{\mathcal{F}}$ contains no $D_xD_y$-segments (where $x,y \in [3]$). For any $x \in[3]$, $c(A_x)$ can be incident to at most one segment in $\widetilde{\mathcal{F}}$. Thus, any triangle contains an $S_iS_j$-segment. Assume that the triangle is formed by $e$ and a point $v$. Recall that for $x \in[3]$, $S_x$ separates $D_x$ from $P \setminus (D_x \cup S_x)$. Thus $v \notin \{D_i,D_j\}$. Note that no segment incident to $v$ crosses $e$. Moreover, $c(S_i)$ and $c(S_j)$ can be incident to at most two segments. Therefore, $\widetilde{\mathcal{F}}$ consists of exactly three segments, and hence $|\mathcal{F}|\le \frac{3}{2}\cdot 2=3$.
- $\widetilde{\mathcal{F}}$ does not admit a triangle. Any segment in $\widetilde{\mathcal{F}}$ that crosses $e$ is incident to an $A$-disk. Hence, $e$ can be crossed by at most three segments. At least one of $c(S_i)$ or $c(S_j)$ is incident to only one segment (if both are incident to two segments, then a triangle is formed). Therefore, $|\mathcal{F}|\le 5$.
6. \[c6\]$\mathcal{F}$ contains an $A_iS_j$-segment (where $i,j \in [3]$). Let $e \in \mathcal{F}$ be an $A_iS_j$-segment. If $i=j$, then $|\mathcal{F}|=2$ because at most one segment can cross $e$. If $i \neq j$, then $|\mathcal{F}|\le 4$ because at most three segments can cross $e$.
Suppose $|\mathcal{F}|>5$. By Cases \[c1\]-\[c6\] , $\mathcal{F}$ may only contain segments of type $D_iX_j$, where $X \in \{A,B,C\}$ and $i,j \in [3]$. Thus, by Case \[c3\], there exists a segment of type $D_iA_j$. Let $\widetilde{\mathcal{F}}=\widetilde{\mathcal{F}}_{\mathcal{D}}$. We consider two cases.
- There exists a segment $e \in \widetilde{\mathcal{F}}$ that is of type $D_iA_i$ or $D_iA_{i+2}$. Without loss of generality, assume $e$ is of type $D_iA_i$. (The other case is symmetric.) Let $Q \subseteq \widetilde{\mathcal{F}}$ be a maximum crossing family. Recall that all segments in $\widetilde{\mathcal{F}}$ are incident to $D$-disks. Note that all segments in $\widetilde{\mathcal{F}}$ that cross $e$ (if any) should be incident to the same $D$-disk. Thus, $|Q| \le 2$. All segments in $\widetilde{\mathcal{F}}$ are pairwise touching. Therefore, $\widetilde{\mathcal{F}}$ consists of either a star or two crossing stars (two stars are crossing if each segment of a star crosses all segments of the other star). If $\widetilde{\mathcal{F}}$ forms a star, then $|\mathcal{F}| \le 5$. So, we assume that there exists a segment in $\widetilde{\mathcal{F}}$ that crosses $e$. Note that at most two such segments exist. Let $\mathcal{S}_1 \subset \widetilde{\mathcal{F}}$ denote the segments of the star containing $e$. Let $\mathcal{S}_2$ denote the segments of the star crossing $e$. $|\mathcal{S}_2| \le 2$. We consider the two cases below.
- $|\mathcal{S}_2| = 1$. Thus, there exists a $D_{i+1}S_i$-segment. Recall that $S_i$ separates $D_i$ from $D_{i+1}$. Hence, $\mathcal{S}_1$ is not incident to $S_i$. By Case \[c2\], there are no $D_iD_{i+1}$-segments. Hence $|\mathcal{S}_1| \le 4$, which implies $|\mathcal{F}| \le 5$.
- $|\mathcal{S}_2| = 2$. Thus $|\mathcal{S}_1| \le 3$ (using Case \[c2\]). Therefore, $|\mathcal{F}| \le 5$.
- All segments in $\widetilde{\mathcal{F}}$ that connect a $D$-disk to an $A$-disk are of type $D_iA_{i+1}$. $\widetilde{\mathcal{F}}$ may be incident to one, two or three $A$-disks. We consider each of these cases below.
- $\widetilde{\mathcal{F}}$ is incident to exactly one $A$-disk. Let $e$ denote the (only) $D_iA_{i+1}$-segment in $\widetilde{\mathcal{F}}$. If no segment in $\widetilde{\mathcal{F}}$ crosses $e$, then $|\mathcal{F}|\le 5$. So, we assume that there exists a segment in $\widetilde{\mathcal{F}}$ that crosses $e$. Note that at most one such segment exists. This implies that $\widetilde{\mathcal{F}}$ consists of a star together with a segment crossing all the star segments. Therefore, by Case \[c2\] and the fact that $S_i$ separates $D_i$ from $D_{i+1}$, $|\mathcal{F}| \le 4$ (recall that $\widetilde{\mathcal{F}}$ is incident to only one $A$-disk).
- $\widetilde{\mathcal{F}}$ is incident to exactly two $A$-disk. This implies that $\widetilde{\mathcal{F}}$ is incident to exactly two $D$-disks (otherwise, either $\widetilde{\mathcal{F}}$ is not pairwise touching or it is incident to all $A$-disks). Thus, $\widetilde{\mathcal{F}}$ consists of two crossing stars. Note that the number of segments in each star is at most two. Thus, $|\mathcal{F}|\le 4$.
- $\widetilde{\mathcal{F}}$ is incident to three $A$-disk. Let $E$ denote the set of three segments in $\widetilde{\mathcal{F}}$ incident to $A$-disks. If $|\mathcal{F}|>3$, there exists a segment in $\widetilde{\mathcal{F}}$ that crosses at least two of the segments in $E$. Any such segment is incident to an $A$-disk. Recall that no $A$-disk can be incident to more than one segment. Therefore, $|\mathcal{F}|=3$.
Therefore, the maximum crossing family of $P$ cannot be greater than five.
$1$-avoiding point sets {#sec:sc}
=======================
In this section, we restrict our attention to $1$-avoiding point sets. We study the combinatorial properties of crossing families in $1$-avoiding point sets. We then introduce a relaxation on these combinatorial properties. As a step towards finding maximum crossing families in $1$-avoiding point sets, we study the size of our relaxed notion of crossing families called “side compatible subsets”. We give linear bounds on side compatible subsets for some simplified versions of the problem.
Combinatorial properties of crossing families
---------------------------------------------
Let $P=P_{B \vdash R}$ be a $1$-avoiding point set $B \cup R$ where $B$ and $R$ are two separable equal-sized sets of points such that $B$ avoids $R$. Without loss of generality, for point set $P_{B \vdash R}$, we assume $B$ and $R$ are separable by a vertical line and $R$ lies to the left of $B$. The *dual* of a point $p=(a,b)$ is the line $p^\star=\{(x,y) \mid y=ax-b\}$ and the dual of a line $l=\{(x,y) : y=ax+b\}$ is the point $l^\star=(a,-b)$. The dual of a $1$-avoiding point set $P_{B \vdash R}$ consists of two sets of lines $B^\star$ (blue lines) and $R^\star$ (red lines) such that each line in $B^\star$ has a larger slope than any line in $R^\star$ and each line in $R^\star$ intersects (or sees) the lines of $B^\star$ in the same order. We call such a line arrangement a *$1$-avoiding* line arrangement.
We study characteristics of crossing families in $1$-avoiding point sets in combinatorial terms and in the dual plane.
An *allowable sequence* is a sequence $\pi_1, \dots, \pi_l$ of permutations of $[n]$ satisfying the following properties:
1. $\pi_1$ is the identity permutation (i.e. $\pi_1=1,\dots,n$) and $\pi_l$ is the reverse of $\pi_1$;
2. The move from $\pi_i$ to $\pi_{i+1}$ consists of reversing one or more non-overlapping substrings (each of size at least two);
3. Any two elements in $[n]$ reverse their order exactly once.
An allowable sequence is *simple* if each move from $\pi_i$ to $\pi_{i+1}$ consists of reversing just one pair of elements.
All combinatorial information of a point configuration or a line arrangement in general position is encoded by a simple allowable sequence. A simple allowable sequence associated with a point configuration encodes the total slope order of lines joining pairs of points. An allowable sequence associated with a line arrangement $L$ encodes the order in which a sweeping vertical line from left to right sees the intersection points of arrangement $L$. Using allowable sequences for studying a geometric problem on a point configuration or a line arrangement is equivalent to generalizing the problem to a generalized point configuration[^1] or a pseudoline arrangement.
Let $|B^\star|=|R^\star|=n$. Say each line in $R^\star$ sees the blue lines in the order $b^\star_1,b^\star_2,\dots,b^\star_n$; and $b^\star_1$ intersects the red lines in the order $r^\star_1,r^\star_2,\dots,r^\star_n$. We can represent the order in which each blue line intersects the red lines in a table, where row $i$ contains the indices of red lines in the order that they are seen by $b^\star_i$. We denote such a table by $T({B^\star},{R^\star})$. For subsets ${B^\star}' \subset B^\star$ and ${R^\star}' \subset R^\star$, we can *reduce* the table to ${B^\star}'$ and ${R^\star}'$ if we only keep the rows representing ${B^\star}'$ and indices representing ${R^\star}'$. We refer to the table reduced to ${B^\star}'$ and ${R^\star}'$ as a *subtable*.
We know $B$ and $R$ can be crossed if and only if they obey the rank condition. This implies that $P_{B \vdash R}$ has a crossing family of size $k$ whose segments connect a point in $B$ to a point in $R$, if and only if there are subsets ${B^\star}' \subseteq B^\star$ and ${R^\star}'\subseteq R^\star$, each of size $k$, such that the elements of the diagonal of the table reduced to ${B^\star}'$ and ${R^\star}'$ are all distinct.
Let $P=P_{B \vdash R}$. A subtable of $T(B^\star,R^\star)$ whose diagonal $\mathrm{d}_1 \mathrm{d}_2 \dots \mathrm{d}_k$ consists of distinct elements has the property that for all rows $j$, $\mathrm{d}_i$ is before $\mathrm{d}_j$ if and only if $i<j$. See Figure \[fig:tab\]. \[lem:tab\]
[c|c|c|c|c|c|c|c|c|]{} & & & &\
& & & &\
& & & &\
&&\
& &\
& &\
& &\
& & & &\
& & &\
We can think of $P^\star$ as a set of lines $R^\star$ which are seen by a set of parallel lines $B^\star$. Hence the permutations obtained from the rows of $T(B^\star,R^\star)$ satisfy the property that each permutation is either identical to the preceding permutation or can be obtained from it by reversing one or more non-overlapping increasing substrings. Note that no two elements reverse their order more than once in these permutations (i.e. no two lines cross more than once).
We prove by contradiction. Let row $b$ be the first row where there exists $a<b$ such that $\mathrm{d}_a$ is after $\mathrm{d}_b$. Note that $b>1$. Let row $c$ be the last row where $\mathrm{d}_a$ is after $\mathrm{d}_c$. Note that $b \le c <k$. Since there are $c-1$ elements before $\mathrm{d}_c$ in row $c$ and $\mathrm{d}_a$ is after $\mathrm{d}_c$, there must be an element $\mathrm{d}_d$, with $d>c$, before $\mathrm{d}_c$ in row $c$. Thus,
- in row $a$, $\mathrm{d}_a$ is before $\mathrm{d}_d$;
- in row $c$, $\mathrm{d}_a$ is after $\mathrm{d}_d$; and
- in row $d$, $\mathrm{d}_a$ is before $\mathrm{d}_d$;
which is not possible since $\mathrm{d}_a$ and $\mathrm{d}_d$ cannot reverse their order more than once.
Lemma \[lem:tab\] implies the following.
Let $P=P_{B \vdash R}$. Let $T'$ be a subtable of $T(B^\star,R^\star)$ with distinct diagonal entries. Let $T'$ be obtained by reducing $T(B^\star,R^\star)$ to ${B^\star}'$ and ${R^\star}'$, where $|{B^\star}'|=|{R^\star}'|$. Let ${B^\star}'={{b^\star}'_1,\dots {b^\star}'_k}$. Let $\mathrm{d}_i$ and $\mathrm{d}_j$ (where $i \neq j$) be two elements on the diagonal of $T'$. The intersection of red lines $r^\star_{\mathrm{d}_i}$ and $r^\star_{\mathrm{d}_j}$ is either above both blue lines ${b^\star}'_i$ and ${b^\star}'_j$, or below both of them. \[cor:crf-tab\]
Figure \[fig:tab-line\] illustrates an example of a $1$-avoiding line arrangement together with its table representation.
; ;
iin [1,2,3,4]{} (0,[6-i]{}) – (-,[6-i]{}) node \[label=left:[$b^\star_\i$]{}\] ;
(-1.1,-1.96) – (-3.33, 6.44) node \[label=above:[$r^\star_1$]{}\] ; (-4,-2) – (-1.68,6.46) node \[label=above:[$r^\star_2$]{}\] ; (-4.08,-1.29) – (-0.9,5.96) node \[label=above:[$~r^\star_3$]{}\] ; (-3.21,-1.34) – (-0.45,5.6) node \[label=above right:[$r^\star_4$]{}\] ;
; ; ; ;
iin [1,2,3,4]{}[ (Di) circle (2.5\*pt); ]{}
(D1) – (D2) – (D3) –(D4);
iin [1,2,3,4]{}[ at (,[6-i]{}) [[$b^\star_\i:$]{}]{}; ]{}
at ([1.4\*]{},[5]{}) [[$1 ~ 2 ~ 3 ~ 4$]{}]{}; at ([1.4\*]{},[4]{}) [[$1 ~ 2 ~ 3 ~ 4$]{}]{}; at ([1.4\*]{},[3]{}) [[$2 ~ 1 ~ 3 ~ 4$]{}]{}; at ([1.4\*]{},[2]{}) [[$2 ~ 3 ~ 1 ~ 4$]{}]{};
(1.2\*,5.2)–(1.6\*,1.8);
Using the table representation $T(B^\star,R^\star)$, it is easy to see that any $1$-avoiding $2n$-point set $P=P_{B \vdash R}$ has a crossing family of size at least $\sqrt{n}$. The middle row of the table has $n$ elements, which implies there is a subsequence of size at least $\sqrt{n}$ in the middle row which is either increasing or decreasing. Denote this monotone subsequence by $S=\{s_1,s_2,\dots, s_l\}$. Let $R^\star_S=\{r^\star_{s_i} \mid s_i \in S\}$. If the subsequence is increasing, then all the red lines in $R^\star_S$ intersect after the middle blue line, and hence $T(B^\star,R^\star)$ can be reduced to $R^\star_S$ and any $l$-subset of $\{b^\star_1,b^\star_2,\dots,b^\star_{\frac{n}{2}}\}$ to form a subtable with distinct diagonal entries. If the subsequence is decreasing, all red lines in $R^\star_S$ intersect before the middle blue line, and hence there exists a subtable with distinct diagonal entries among the lower half of $T(B^\star,R^\star)$.
In the following, we study a generalized variant of a table, called a combinatorial table, where each row is a permutation of $[n]$ and no two elements reverse their order more than once. The sequence of the permutations of a combinatorial table may contain identical permutations and any subsequence containing distinct permutations of a combinatorial table is a subsequence of an allowable sequence. A combinatorial table corresponds to a table representation of a $1$-avoiding arrangement which may not be realizable using straight lines (i.e. $1$-avoiding arrangement of pseudolines).
Let $\Pi=\Pi_1\Pi_2\cdots\Pi_l$ be a sequence of $l$ permutations of $[n]$ such that no two elements reverse their order more than once. Let $S=\{s_1,\ldots,s_m\}$ be an $m$-subset of $[l]$ where $s_1<\dots<s_m$. Let $E$ be an $m$-subset of $[n]$. A *combinatorial* table $\mathtt{T}(S,E)$ is a table where row $i$ is the subsequence of $\Pi_{s_i}$ containing only the elements that are in $E$. We may refer to a combinatorial table as a table for simplicity.
Let $\Pi=\Pi_1\Pi_2\cdots\Pi_{{n \choose 2} +1}$ be a simple allowable sequence with $\Pi_1=1,2,\dots,n$. There exits $S=\{s_1,\dots, s_{n-1}\}$ where $S \subset [{{n \choose 2}+1}]$ and $s_1<\dots<s_{n-1}$, such that table $\mathtt{T}(S,[2..n])$ has distinct entries on its diagonal.
Let $S$ be the set of all numbers $i < {n \choose 2}+1$ where $\Pi_i$ differs from $\Pi_{i+1}$ in having $1$ and its right neighbour flipped. Let $\mathrm{d}_1,\dots,\mathrm{d}_{n-1}$ be the diagonal entries of $\mathtt{T}(S,[2..n])$. Note that $1$ is the leftmost element in $\Pi_1$ and no pair of elements flip more than once. Thus, $\mathrm{d}_i$ is the $i$-th element flipped by $1$. If an entry $\mathrm{d}_j$ on the diagonal is repeated, then $1$ and $\mathrm{d}_j$ have to flip more than once in the allowable sequence, which is not possible.
Let $\Pi=\Pi_1\Pi_2\cdots \Pi_l$ be a sequence of permutations of $[n]$ where $\Pi$ may contain repeated permutations but all repetitions of the same permutation are consecutive, and the largest subsequence containing all distinct permutations is a subsequence of an allowable sequence. There exists $\Pi$ with $l=\Theta(n^2)$, such that for no $n$-subset $S$ from $[l]$, table $T(S,[n])$ has distinct elements on its diagonal.
Let $l=\lfloor \frac{n}{2} \rfloor^2+n-1$. Construct $\Pi=\Pi_1 \Pi_2 \cdots \Pi_l$ such that $\Pi_1= 1 2 \cdots n$, and for all $1 \le i < l$, $\Pi_{i+1}$ is the same as $\Pi_i$ except for the following cases:
- if $i=\lceil \frac{n}{2} \rceil +k(n+1)$, where $k \in \{0,1,\dots,\lceil \frac{n-1}{4} \rceil-1\}$, the permutation $\Pi_{i+1}$ differs from $\Pi_{i}$ in having $\frac{n}{2}-k$ and $\frac{n}{2}+1+k$ flipped, and
- if $i=n+k(n+1)$, where $k \in \{0,1,\dots,\lfloor \frac{n}{4} \rfloor-1\}$, the permutation $\Pi_{i+1}$ differs from $\Pi_{i}$ in having $1+k$ and $n-k$ flipped.
Relaxation on the combinatorial properties
------------------------------------------
As a step towards finding maximum crossing families in $1$-avoiding point sets, we study a relaxed notion of crossing families.
Let $\text{crf}(P_1,P_2)$ denote the size of the maximum crossing family whose segments connect a point in $P_1$ to a point in $P_2$. Let $B=\{\mathtt{b}_1,\dots,\mathtt{b}_n\}$ and $R=\{\mathtt{r}_1,\dots,\mathtt{r}_n\}$ be two separable sets of $n$ points such that $B$ avoids $R$. Let $B^\star=\{\mathtt{b}^\star_1,\dots,\mathtt{b}^\star_n\}$ and $R^\star=\{\mathtt{r}^\star_1,\dots,\mathtt{r}^\star_n\}$ denote the duals of $B$ and $R$, respectively. Recall that $\text{crf}(B,R)=k$ if and only if there are $k$-subsets ${B^\star}' \subset B^\star$ and ${R^\star}' \subset R^\star$ such that $T'=T({B^\star}',{R^\star}')$ has distinct diagonal entries. A necessary property for $T'$ (by Corollary \[cor:crf-tab\]) is that for any two diagonal entries $\mathrm{d}_i$ and $\mathrm{d}_j$ of $T'$, the intersection of red lines $r^\star_{\mathrm{d}_i}$ and $r^\star_{\mathrm{d}_j}$ is either above both blue lines ${b^\star}'_i$ and ${b^\star}'_j$, or below both of them. We refer to this property as the “sidedness” property in crossing families. In the following, we introduce a relaxation on crossing families, called “side compatibility”, which preserves the sidedness property.
We start by defining some terminology. (See Figure \[fig:s-vs-os\] for illustrations.) Given a line arrangement $\mathcal{L}=\{l_1,\dots,l_n\}$, the *bar representation* of $\mathcal{L}$ is composed of $n \choose 2$ horizontal *bars* (i.e. horizontal segments) arranged such that the $i$-th bar from below represents the $i$-th intersection point $p_i$ from below in $\mathcal{L}$. We assume no intersection points in $\mathcal{L}$ have the same $y$-coordinate (we rotate $\mathcal{L}$ if necessary). For intersection point $p_i$ of lines $l_{a_i}$ and $l_{b_i}$ in $\mathcal{L}$, the corresponding bar is a segment from $(a_i,i)$ to $(b_i,i)$.
A *bar stack* $\mathcal{B}_{l,n}$ is an arrangement of $l$ bars $B_1 B_2 \cdots B_l$ where ($\expandafter{\romannumeral 1}$) each bar $B_i$ extends from $(a_i,i)$ to $(b_i,i)$ where $a_i$ and $b_i$ are integers and $1 \le a_i < b_i \le n$, and ($\expandafter{\romannumeral 2}$) for two bars $B_i$ and $B_j$ either $a_i \neq a_j$ or $b_i \neq b_j$ or both. We say $B_i$ is at *height* $i$. Note that $\mathcal{B}_{l,n}$ may not come from the representation of a line arrangement. This makes bar stacks more expressive than line arrangements at representing “intersections”. However, bar stacks may represent fewer “intersections” compared to line arrangements. Having fewer than $n \choose 2$ bars (“intersections”) makes bar stacks easier to work with. Let $\mathcal{W}_{n,l}=\mathit{w}_1 \mathit{w}_2 \dots \mathit{w}_n$ be a sequence of $n$ integers from zero to $l$ (the numbers may not be distinct). Let $\mathcal{C}$ be a subset of $[n]$ (containing distinct numbers) for which there exists an injective function $f: \mathcal{C} \rightarrow \mathcal{W}_{n,l}$ such that for any bar $B_i$ whose endpoints’ $x$-coordinates both belong to $\mathcal{C}$, either
1. $f(a_i) < i$ and $f(b_i) <i$\[bar-P1\], or
2. $f(a_i) \ge i$ and $f(b_i) \ge i$\[bar-P2\].
We say $\mathcal{C}$ is a *side compatible subset* for $\mathcal{W}_{n,l}$ in $\mathcal{B}_{l,n}$. We refer to the function $f$ as the *mapping function* for $\mathcal{C}$. If for a side compatible subset $\mathcal{C}$, we have the additional property that in case of \[bar-P1\], $f(a_i)<f(b_i)$ and in case of \[bar-P2\], $f(a_i)>f(b_i)$, we say $\mathcal{C}$ is an *ordered* side compatible subset.
The mapping function $f$ of a side compatible subset $\mathcal{C}$ matches every element $c \in \mathcal{C}$ to $f(c)$; thus, the pair $(\mathcal{C},f)$ defines a matching with edges $\{(c,f(c)) \mid c \in \mathcal{C}\}$. A side compatible subset satisfies the sidedness property, in the sense that for $c_i,c_j \in \mathcal{C}$, where $c_i<c_j$, the height of the bar with horizontal interval $[c_i,c_j]$ (representing the “intersection” of $c_i$ and $c_j$) is less than both $f(c_i)$ and $f(c_j)$, or greater than or equal to both of them.
We can visualize a side compatible subset in the following way. Think of each $\mathit{w}_i \in \mathcal{W}_{n,l}$ as a distinct horizontal wire above $B_{\mathit{w}_i}$ and below $B_{\mathit{w}_i+1}$. If $\mathit{w}_i =0$ the wire is below $B_1$, and if $\mathit{w}_i = l$ the wire is above $B_l$. ($\mathcal{W}_{n,l}$ is represented by $n$ different horizontal wires.) We refer to the vertical lines going through endpoints of bars as *pillars*. The pillar through $e \in [n]$ is the vertical line $x=e$. We may think of function $f:\mathcal{C} \rightarrow \mathcal{W}_{n,l}$ as assigning a marble for each $\mathit{c} \in \mathcal{C}$ on the intersection point of the pillar through $c$ and wire $f(c)$. A *marbling* of *size* $m$ is a set of $m$ marbles such that all marbles lie on the intersection points of the pillars and the wires. A *valid* marbling is a marbling such that each pillar or wire contains at most one marble. We say an endpoint of a bar is *associated* with a marble if the vertical line (pillar) through it contains a marble. (A pillar may go through the endpoints of a number of bars.) A *side compatible marbling* is a valid marbling such that for every bar both of whose endpoints are associated with marbles, both marbles are above or below the bar. An ordered side compatible subset corresponds to a side compatible marbling such that for every bar both of whose endpoints are associated with marbles, the marble associated with the right endpoint is closer to the bar. We refer to such a marbling as an *ordered side compatible marbling*.
Let $P_{B \vdash R}$ be a $1$-avoiding $2n$-point set, where $B=\{\mathtt{b}_1,\dots,\mathtt{b}_n\}$ and $R=\{\mathtt{r}_1,\dots,\mathtt{r}_n\}$. $B^\star=\{\mathtt{b}^\star_1,\dots,\mathtt{b}^\star_n\}$ and $R^\star=\{\mathtt{r}^\star_1,\dots,\mathtt{r}^\star_n\}$ denote the duals of $B$ and $R$, respectively. Let $\mathcal{W}=\mathit{w}_1,\dots,\mathit{w}_n$, where $\mathit{w}_i$ is the number of intersection points among $R^\star$ that are below line $\mathtt{b}^\star_i$. Consider a side compatible marbling corresponding to a side compatible subset $\mathcal{C}$ for $\mathcal{W}$ in the bar representation of $R^\star$ using the mapping function $f$. A marble at the intersection point of the pillar through $c$ and wire $f(c)$ corresponds to the segment $\mathtt{r}_c\mathtt{B}_{f(c)}$. This correspondence implies that any crossing family defines a marbling that is side compatible (by Corollary \[cor:crf-tab\]). So clearly, $\text{crf}(B,R)=k$ implies that there is a side compatible subset of size $k$ for $\mathcal{W}$ in the bar representation of $R^\star$. However, the reverse is not true. See Figure \[fig:s-vs-os\] for an example.
; ;
iin [1,2,3]{} (.5\*,i) -++ (-,0) node \[label=left:[$b^\star_\i$]{}\] ;
(-0.04, -0.99) – (1.22,4.92)node \[label=above left:[$r^\star_1$]{}\] ; (-1.73,-1) – (1.57,4.97) node \[label=above:[$~r^\star_2$]{}\] ; (-2.65,-1.04) – (2.12,4.9) node \[label=above right:[$r^\star_3$]{}\] ;
iin [1,2,3]{}[ (2\*,[i]{}) – (,[i]{}) ; ([+i]{},0) – ([+i]{},5) ; ]{}
(+1,3) circle (3.5\*pt); (+2,1) circle (3.5\*pt); (+3,2) circle (3.5\*pt);
(+2,2.5) – (+3,2.5) ; (+3,3.5) – (+1,3.5) ; (+2,4.5) – (+1,4.5) ;
([-1]{},2.5) -++ ([+2]{},0); at ([2\*+3]{},2.5) [$y=1$]{}; ([-1]{},3.5) -++ ([+2]{},0); at ([2\*+3]{},3.5) [$y=2$]{}; ([-1]{},4.5) -++ ([+2]{},0); at ([2\*+3]{},4.5) [$y=3$]{};
at (1.5\*,-1) [$\mathcal{W}=\langle 0, 0, 1 \rangle$]{};
The sidedness property implies that the intersection point of the supporting lines of the two segments corresponding to two marbles in a side compatible marbling is either on both segments (i.e., the segments are crossing) or on neither of them. This is easy to verify under duality. See Figure \[fig:s-os-dual\].
; ;
in [1,2,3,4]{}
iin [1,2]{}[ (,i) -++ (.5\*,0); ]{}
=0 (,3) – (+2,0); (4+,3) – (+1.5,0);
>0 (+2,3) – (,0); (+1.5,3) – (+4,0);
; ; ; ;
in [121,112,221,212,311,322,411,422]{} (M) circle (4pt);
(M121) – (,3) – (,2) – (M121); (M121) – (1.5\*,2) – (1.5\*,0) – (+2,0) – (M121); (M112) – (1.5\*,3) – (,3) – (,1) – (M112); (M112) – (1.5\*,1) – (1.5\*,0) – (+1.5,0) – (M112);
(M221) – (2\*+2,3) – (2\*,3) – (2\*,2) – (M221); (M221) – (2.5\*,2) – (2.5\*,0) – (2\*,0) – (M221); (M212) – (2\*+1.5,3) – (2\*,3) – (2\*,1) – (M212); (M212) – (2.5\*,1) – (2.5\*,0) – (M212);
(M311) – (3\*,3) – (3\*,1) – (M311); (M311) – (3.5\*,1) – (3.5\*,0) – (3\*+2,0) – (M311); (M322) – (3.5\*,3) – (3\*,3) – (3\*,2) – (M322); (M322) – (3.5\*,2) – (3.5\*,0) – (3\*+1.5,0) – (M322);
(M411) – (4\*+2,3) – (4\*,3) – (4\*,1) – (M411); (M411) – (4.5\*,1) – (4.5\*,0) – (4\*,0) – (M411); (M422) – (4\*+1.5,3) – (4\*,3) – (4\*,2) – (M422); (M422) – (4.5\*,2) – (4.5\*,0) – (M422);
Note that the pair $(\mathcal{B}_{l,n}, \mathcal{W}_{n,l})$ defines a sequence of permutations of $[n]$. In particular, we can assign a permutation to each wire $w$ as follows: we start with the identity permutation, and consider a horizontal sweep line that moves top-down until it hits wire $w$; whenever the sweep line hits a bar with horizontal interval $[i,j]$, we swap the positions of $i$ and $j$ in the permutation. For a side compatible subset $\mathcal{C}$, let $\mathcal{M}_\mathcal{C}$ denote the corresponding side compatible marbling for $\mathcal{C}$. Let $T(\mathcal{C})$ denote the table where row $i$ is the permutation, restricted to elements of $\mathcal{C}$, assigned to the $i$-th wire (numbered top-down) containing a marble. In order for a side compatible subset $\mathcal{C}$ to correspond to a crossing family, the diagonal entries of $T(\mathcal{C})$ should all be distinct. It is easy to verify that the diagonal entries are distinct only if $\mathcal{C}$ is an ordered side compatible subset; or in other words, for the segments corresponding to a side compatible marbling to cross pairwise, the marbling should be ordered side compatible. This is easy to see under duality because if two marbles satisfy side compatibility but not ordered side compatibility, then the supporting lines of the corresponding two segments intersect outside both of them. Thus, $\text{crf}(B,R)=k$ if and only if the largest *ordered* side compatible subset for $\mathcal{W}$ in the bar representation of $R^\star$ is of size $k$. Hence, $\min_B \text{crf}(B,R) = k$ (where the minimum is taken over all point sets $B$ that avoid $R$) if and only if for any $\mathcal{W}_{n,{n \choose 2}}$, there exists an ordered side compatible subset $\mathcal{C}$ in the bar representation of $R^\star$, where $|\mathcal{C}|=k$.
\[obs:bar\] Let $\mathcal{W}$ be a sequence of $n$ integers from zero to $n \choose 2$. For any arrangement $\mathcal{L}$ of $n$ lines, there exists $\mathcal{W}$ such that the largest side compatible subset $\mathcal{C}$ for $\mathcal{W}$ in the bar representation of $\mathcal{L}$ is of size at most $\frac{n}{2}$.
Let $\mathcal{W}=\{0\}^\frac{n}{2}\{{n \choose 2}\}^\frac{n}{2}$. If $|\mathcal{C}|>\frac{n}{2}$, then there exits $c_1,c_2 \in \mathcal{C}$ such that $f(c_1)=0$ and $f(c_2)= {n \choose 2}$ (where $f$ is the mapping function for $\mathcal{C}$). However, this implies that the bar whose horizontal interval extends from $c_1$ to $c_2$ does not satisfy \[bar-P1\] or \[bar-P2\].
Given any bar stack $\mathcal{B}_{n,n}$, there exists an $n$-subsequence $\mathcal{W}$ of $\langle 0~1~\cdots~n\rangle$, such that the largest side compatible subset for $\mathcal{W}$ in $\mathcal{B}_{n,n}$ is of size $n$. \[lem:all-wires\]
Let $Y=\{ i+0.5 \mid i \in [0..n] \}$. Assume for each $y \in Y$, there is a wire at height $y$. We claim, for some $y$, there exists a side compatible marbling of size $n$ such that a marble associated with an endpoint of a bar is below the bar if and only if the bar is at height greater than $y$.
The endpoints of bars that lie on a pillar partition the pillar into a number of vertical intervals. Let $I^y_x$ denote the vertical interval of the pillar through $x$ that contains $y$, where $x \in [n]$ (See Figure \[fig:interval-bar\]). Let $X \subseteq [n]$ and $U^y_X=\bigcup\limits_{x\in X} I^y_x$. The union of two vertical intervals is simply the union of the $y$-values of the two intervals. Note that since $y \in I^y_x$ for all $x$, $U^y_{X}$ is also an interval. For an interval $I$, let $\sharp(I)$ denote the number of elements in $Y$ that are in $I$. Let $E^y_X$ denote the set of endpoints of bars whose $x$-coordinates are in $X$ and $y$-coordinates are in $U^y_{X}$.
iin [0,1,...,8]{}[ (0,i+.5) -++ (8,0); at (12,) [$y=\pgfmathresult$]{}; ]{} (-3,5.5) -++ (13,0);
(7,1) – (8,1); (2,2) – (7,2); (1,3) – (8,3); (1,4) – (2,4); (3,5) – (4,5); (5,6) – (6,6); (4,7) – (5,7); (3,8) – (6,8);
(P1) at (7,1); (P2) at (8,1); (P3) at (2,2); (P4) at (7,2); (P5) at (1,3); (P6) at (8,3); (P7) at (1,4); (P8) at (2,4); (P9) at (3,5); (P10) at (4,5); (P11) at (5,6); (P12) at (6,6); (P13) at (4,7); (P14) at (5,7); (P15) at (3,8); (P16) at (6,8);
(P6) – (8,8.5); (P4) – (7,8.5); (P12) – (6,0.5); (P11) – (5,0.5); (P10) – (P13); (P9) – (P15); (P8) – (2,8.5); (P7) – (1,8.5);
iin [1,...,16]{} (Pi) circle (3.5pt);
If (for some $y$) a marbling of size $n$ is such that every marble associated with an endpoint of a bar is below the bar if and only if the bar is at height greater than $y$, then for all $x \in [n]$, the $y$-position of the marble on the pillar through $x$ is in $I^y_x$. Clearly, such a marbling satisfies \[bar-P1\] or \[bar-P2\]. In order to prove the claim, we need to show there exists such a marbling that is also valid (that is, no two marbles have the same $y$-position). It is easy to see that if for all $X \subseteq [n]$, $\sharp(U^y_X) \ge |X|$ then there is enough room so that each marble can have a distinct $y$-position. Therefore, if for some $y$, we have the property that for all $X \subseteq [n]$, $\sharp(U^y_X) \ge |X|$, then the claim follows.
Assume, for the sake of contradiction, that for all $y$, there exists $X_y \subseteq [n]$ such that $\sharp(U^y_{X_y}) < |X_y|$.
1. \[Bprop1\] $U^y_{X_y}$ is unbounded on at most one side.
If $U^y_{X_y}$ is unbounded on both sides, then $\sharp(U^y_{X_y})=n+1$. However, $|X_y| \le n$. Thus $\sharp(U^y_{X_y}) < |X_y|$ is contradicted.
2. \[Bprop2\] If $U^y_{X_y}$ is bounded from above, then at least $\lceil \frac{|X_y|}{2} \rceil -1$ elements of $Y \cap U^y_{X_y}$ are greater than $y$. If $U^y_{X_y}$ is bounded from below, then at least $\lceil \frac{|X_y|}{2} \rceil -1$ elements of $Y \cap U^y_{X_y}$ are less than $y$.
Let $U^y_{X_y}$ be bounded from above. Recall that $U^y_{X_y}=\cup_{x\in X_y} I^y_x$, and for all $x \in X_y$, $I^y_x$ contains $y$. This implies there are at least $|X_y|$ endpoints in $E^y_{X_y}$ whose heights are greater than $y$. Since no two bars lie at the same height, every integer height contains exactly two endpoints, and hence $E^y_{X_y}$ has at most two endpoints at every integer height. The vertical range between every two consecutive bars contains a distinct element of $Y$. Therefore, we infer that there are at least $\lceil \frac{|X_y|}{2} \rceil -1$ elements of $Y$ in $U^y_{X_y}$ that are greater than $y$. A similar argument works when $U^y_{X_y}$ is bounded from below.
3. \[Bprop3\]If $U^y_{X_y}$ is bounded on both sides, then
1. \[Bp1\]$|X_y|$ is even,
2. \[Bp2\]$\sharp(U^y_{X_y}) = |X_y|-1$,
3. \[Bp3\]$y$ is the (unique) median in $Y \cap U^y_{X_y}$,
4. \[Bp4\] $E^y_{X_y}$ has exactly two endpoints at every integer height in $U^y_{X_y}$, and
5. \[Bp5\] no two endpoints in $E^y_{X_y}$ that have different heights and are both above or below $y$ lie on the same pillar.
\[Bprop2\] together with $|X_y| > \sharp(U^y_{X_y})$ implies $|X_y| > 2 \bigl\lceil \frac{|X_y|}{2} \bigr\rceil -1$. Thus, $|X_y| = \bigl\lceil \frac{|X_y|}{2} \bigr\rceil +\bigl\lfloor \frac{|X_y|}{2} \bigr\rfloor \ge 2 \bigl\lceil \frac{|X_y|}{2} \bigr\rceil$, and hence $\bigl\lfloor \frac{|X_y|}{2} \bigr\rfloor \ge \bigl\lceil \frac{|X_y|}{2} \bigr\rceil$. This immediately implies \[Bp1\] and \[Bp2\]. Moreover, the number of elements of $Y$ in $U^y_{X_y}$ that are greater (or less) than $y$ is exactly $\frac{|X_y|}{2} -1$. This consequently implies \[Bp3\], \[Bp4\] and \[Bp5\].
4. \[Bprop4\] If $U^y_{X_y}$ is bounded, then neither $U^{y-1}_{X_{y-1}}$ nor $U^{y+1}_{X_{y+1}}$ can be bounded.
Let $p$ be an endpoint whose height is below $y$ and above $y-1$ (i.e. at height $y-0.5$). Suppose $U^{y-1}_{X_{y-1}}$ is bounded. Note that \[Bp4\] implies that $p $ is in both $E^{y-1}_{X_{y-1}}$ and $E^{y}_{X_{y}}$ (since $p$ is immediately above or below $y-1$ and $y$). Thus, $E^{y-1}_{X_{y-1}}$ has an endpoint $p'$ that lies on the same pillar as $p$, and whose height is less than $y-1$. Recall that $U^y_{X_y}$ is bounded, and hence \[Bp5\] implies that $p' \notin E^{y}_{X_{y}}$. Thus $\sharp(U^{y-1}_{X_{y-1}}) \ge \sharp(U^y_{X_y})$. However, this implies that the pillar through $p$ contains two endpoints in $E^{y-1}_{X_{y-1}}$ that are both above $y-1$ and have different heights. This contradicts \[Bp5\]. A similar argument works for $U^{y+1}_{X_{y+1}}$.
We consider two cases.
1. $n$ is even.\
Let $y$ be the unique median value in $Y$ (i.e. $y=\frac{n}{2}+0.5$). By \[Bprop1\] we know $U^y_{X_y}$ is bounded on at least one side. If $U^y_{X_y}$ is unbounded, then \[Bprop2\] implies $\sharp(U^y_{X_y}) \ge \frac{n}{2}+1+\bigl\lceil \frac{|X_{y}|}{2} \bigr\rceil -1$; and since $|X_y| > \sharp(U^y_{X_y})$, we infer $|X_y| >n$, which is not possible. Therefore, $U^y_{X_y}$ is bounded (on both sides). Using \[Bprop4\], we know both $U^{y-1}_{X_{y-1}}$ and $U^{y+1}_{X_{y+1}}$ are unbounded.
Consider $U^{y-1}_{X_{y-1}}$. Note that $|X_{y-1}| \le n$ and hence $\sharp(U^{y-1}_{X_{y-1}}) \le n-1$. If $U^{y-1}_{X_{y-1}}$ is unbounded from above, then $|X_{y-1}| > \sharp(U^{y-1}_{X_{y-1}}) \ge \frac{n}{2}+2+\bigl\lceil \frac{|X_{y-1}|}{2} \bigr\rceil -1$, and hence $|X_{y-1}| > n+2$, which is not possible. Therefore, $U^{y-1}_{X_{y-1}}$ is unbounded from below (and bounded from above). As a result, $|X_{y-1}| > \sharp(U^{y-1}_{X_{y-1}}) \ge \frac{n}{2}+\bigl\lceil \frac{|X_{y-1}|}{2} \bigr\rceil -1$, and hence $\bigl\lfloor \frac{|X_{y-1}|}{2} \bigr\rfloor > \frac{n-2}{2}$. Therefore (since $n$ is even) $|X_{y-1}|=n$ and $\sharp(U^{y-1}_{X_{y-1}})=n-1$. Note that $|X_{y-1}|=n$ implies $E^{y-1}_{X_{y-1}}$ has exactly two endpoints at every integer height in $U^{y-1}_{X_{y-1}}$. Moreover, two endpoints in $E^{y-1}_{X_{y-1}}$ that are both above $y-1$ and have different heights need to lie on different pillars.
Similarly, we infer that $U^{y+1}_{X_{y+1}}$ is unbounded from above (and bounded from below); $|X_{y+1}|=n$ and $\sharp(U^{y+1}_{X_{y+1}})=n-1$; and consequently, $E^{y+1}_{X_{y+1}}$ has exactly two endpoints at every integer height in $U^{y+1}_{X_{y+1}}$; and two endpoints in $E^{y+1}_{X_{y+1}}$ that are both below $y+1$ and have different heights need to lie on different pillars.
Let $F$ denote the four endpoints at heights $\frac{n}{2}$ and $\frac{n}{2}+1$ (i.e. the endpoints with heights immediately above or below $y$). No pillar going through an endpoint $p \in F$ can contain any other endpoint with height between $2$ and $n-1$. Recall that $U^y_{X_y}$ is bounded. This implies $\sharp(U^y_{X_y})=n-1$ and $|X_y|=n$. Moreover, $B_n$ and $B_{\frac{n}{2}}$ need to have identical horizontal intervals (and so do $B_1$ and $B_{\frac{n}{2}+1}$), which is not possible.
2. $n$ is odd.\
Let $y$ be the smaller median value in $Y$ (i.e. $y=\frac{n-1}{2}+0.5$). If $U^y_{X_y}$ is unbounded from above, then $|X_y| > \sharp(U^y_{X_y}) \ge \frac{n+1}{2}+1+\bigl\lceil \frac{|X_y|}{2} \bigr\rceil -1$. Hence $|X_y| > n+1$, which is not possible. If $U^y_{X_y}$ is unbounded from below, then $|X_y| > \sharp(U^y_{X_y}) \ge \frac{n+1}{2}+\bigl\lceil \frac{|X_y|}{2} \bigr\rceil -1$. Hence $2 \bigl\lfloor \frac{|X_y|}{2} \bigr\rfloor > n-1$, which implies $|X_y| > n$. But again, this is not possible. Therefore, $U^y_{X_y}$ is bounded.
Let $y'$ be the bigger median value in $Y$. Using a similar argument, we conclude that $U^{y'}_{X_{y'}}$ is also bounded. But this is a contradiction to \[Bprop4\].
Given any bar stack $\mathcal{B}_{n,n}$, there exists an $\lfloor \frac{n}{2} \rfloor$-subsequence $\mathcal{W}$ of $\langle 0~1~\cdots~n\rangle$, such that the largest ordered side compatible subset for $\mathcal{W}$ in $\mathcal{B}_{n,n}$ is of size $\lfloor\frac{n}{2}\rfloor$.
Let $\mathcal{B}'$ be the set of bars in $\mathcal{B}_{n,n}$ with heights at most $\lfloor \frac{n}{2} \rfloor$. We claim there exists an injective function $f:[n] \rightarrow [n]$ such that for any bar $B_i=[(a_i,i),(b_i,i)]$ in $\mathcal{B}'$, $f(a_i)>f(b_i) \ge i$ (where $i \in [\lfloor \frac{n}{2} \rfloor]$). We can visualize this claim in terms of marbling: Let $Y=\{i +0.5 \mid i \in [0..n]\}$. Assume for each $y \in Y$, there is a wire at height $y$. We claim there exists a valid marbling of size $n$ such that for every bar $B_i \in \mathcal{B}'$, both marbles associated with the bar are above the bar; and moreover, the marble associated with the left endpoint of $B_i$ is higher than the marble associated with the other endpoint.
Note that if the claim is true, every bar $B_i \in \mathcal{B}'$ forms a partial order on the heights of marbles lying on pillars through $a_i$ and $b_i$; that is, it implies the inequalities $f(a_i)>f(b_i) \ge i$, where $f(x)+0.5$ is the height of the marble that is on pillar through $x$. We construct a directed graph representing all partial orders obtained from $\mathcal{B}'$, and prove that there exists a total order (consistent with all partial orders) on the heights of marbles such that each marble has a distinct height in $Y$.
We construct a directed 2-coloured graph $G=(V,E)$ as follows:
- $V=V_p \cup V_h$, where $V_p = p_1 \cdots p_n$ and $V_h=h_1 \cdots h_{\lfloor \frac{n}{2} \rfloor}$ are sets of white and black vertices, respectively.
- For every $i \in [\lfloor \frac{n}{2} \rfloor]$, there is a directed edge from $p_{a_i}$ to $p_{b_i}$, and another edge from $p_{b_i}$ to $h_i$.
Note that $G$ has no directed cycles. For every $v \in V_p$, let $r(v)$ denote the number of white vertices that are reachable from $v$. We refer to $r(v)$ as the *r-value* of vertex $v$. Decompose $G$ into weakly connected components . A *weakly connected component* in a directed graph is a maximal connected component in the underlying undirected graph (that is, if replacing all directed edges with undirected edges). Let $$H(C) =
\begin{cases}
\max\{i \mid h_i \in V(C)\} & \text{if } |V(C)| >1\\
0 & \text{otherwise,}
\end{cases}$$ where $C$ is a weakly connected component and $V(C)$ is the set of vertices in $C$. We refer to $H(C)$ as the *h-value* of $C$. The h-value of a component is zero if the component is *trivial* (i.e. it consists of only one vertex). Note that a trivial component does not contain a black vertex. If $p_v$ is the only vertex in a component then the pillar through $v$ does not contain any endpoints of bars in $\mathcal{B}'$.
Let $c$ be the number of weakly connected components in $G$. Let $G_1 G_2 \cdots, G_c$ be the ordering of the weakly connected components of $G$ in non-increasing order of their h-values. Let $P(G_i)$ denote the number of white vertices in $G_i$. For each component $G_i$, sort its white vertices in non-increasing order of their r-values. Let $p_{\pi_i(0)}, p_{\pi_i(1)},\cdots$ be this ordering. We define function $f$ so that $f(\pi_i(j))= -j+n-\sum_{k=1}^{i-1} P(G_k)$. Note that for every pair of white vertices $p_v$ and $p_u$, if $p_v$ can reach $p_u$, then $f(v) > f(u)$. Recall that for any white vertex $p_v \in G_i$, $n-\sum_{k=1}^{i}P(G_k)+1 \le f(v) \le n-\sum_{k=1}^{i-1}P(G_k)$. In order to prove our claim, we need to show that $f$ is an injective function with range $[n]$. This immediately follows if for every component $G_i$, $H(G_i) \le n-\sum_{k=1}^{i}P(G_k)+1$.
Note that for every $i \in [c]$, $0 \le H(G_i) \le \lfloor \frac{n}{2} \rfloor$. If $H(G_i)=0$, then $H(G_i) \le n-\sum_{k=1}^{i}P(G_k)+1$ because $\sum_{k=1}^c P(G_k) \le n$. Recall that $H(G_i)=0$ if and only if $|V(G_i)|=1$. The h-values of the components that contain more than one vertex are all distinct. Therefore, if $H(G_i) >0$, then $H(G_i)\le \lfloor \frac{n}{2} \rfloor - (i-1)$. Let $E^i_p$ denote the number of edges in the subgraph induced by the white vertices in $G_i$. We know $P(G_i) \le E^i_p +1$, and hence $\sum_{k=1}^i P(G_k) \le i+ \sum_{k=1}^i E^k_p$. Note that $\sum_{k=1}^c E^k_p \le \lfloor \frac{n}{2} \rfloor$. Therefore, $n-\sum_{k=1}^{i}P(G_k)+1 \ge \lceil \frac{n}{2} \rceil -i +1$, and subsequently $n-\sum_{k=1}^{i}P(G_k)+1 \ge H(G_i)$. This concludes the proof of our claim.
In the following, we use the function $f$ defined in our claim to prove the lemma. Initialize $\mathcal{C}=[n]$. Using our claim, we know that for every bar $B_i=[(a_i,i),(b_i,i)]$ in $\mathcal{B}_{n,n}$, where $i \in [\lfloor \frac{n}{2} \rfloor]$, $f(a_i)>f(b_i) \ge i$. For all $j \in [\lfloor \frac{n}{2} \rfloor +1 .. n]$ such that neither $f(a_j)>f(b_j) \ge j$ nor $f(a_j)<f(b_j) < j$ holds, we remove $a_j$ or $b_j$ from $\mathcal{C}$. It is easy to verify that at the end $|\mathcal{C}| \ge n -\lceil \frac{n}{2} \rceil = \lfloor \frac{n}{2} \rfloor$. Using function $f$ on $\mathcal{C}$, we guarantee that $\mathcal{C}$ is an ordered side compatible subset for a subsequence of size $|\mathcal{C}| \ge \lfloor \frac{n}{2} \rfloor$ obtained from $\langle 1~\cdots~n\rangle$.
If for any $n$, any bar stack $\mathcal{B}_{n,n}$, and any sequence $\mathcal{W}_1$ of $2n+2$ integers, there is a side compatible subset of size $n$ for $\mathcal{W}_1$ in $\mathcal{B}_{n,n}$, then for any sequence $\mathcal{W}_2$ of $n$ integers, there is a side compatible subset of size $\lfloor\frac{n}{2}\rfloor-2$ for $\mathcal{W}_2$ in $\mathcal{B}_{n,n}$. \[obs:2nTOn\]
For any bar stack $\mathcal{B}_{n,n}$, there exists a subset $\mathcal{C} \subseteq [n]$ of size $\lfloor\frac{n}{2}\rfloor$ such that the number of bars in $\mathcal{B}_{n,n}$ whose endpoints’ $x$-coordinates both belong to $\mathcal{C}$ is at most $\lfloor\frac{n}{2}\rfloor$ (note that for any $\lfloor\frac{n}{2}\rfloor$-subset $C$, either $C$ or a $\lfloor\frac{n}{2}\rfloor$-subset from $[n]\setminus C$ satisfies this property). Let $\mathcal{B}'$ denote the set of bars in $\mathcal{B}_{n,n}$ induced by $\mathcal{C}$ ($\mathcal{B}'$ is a set of at most $\lfloor\frac{n}{2}\rfloor$ bars whose heights range from $1$ to $n$). If the premise of the statement in Observation \[obs:2nTOn\] is true, then $\mathcal{C}$ is a side compatible subset for any sequence of $2\lfloor \frac{n}{2}\rfloor+2$ integers in $\mathcal{B}'$ (here a number $\mathit{w}$ in the sequence corresponds to a wire that is below any bar whose height is greater than $\mathit{w}$ and is above any bar whose height is equal or less than $\mathit{w}$). As a result, there is a side compatible subset of size at least $\lfloor\frac{n}{2}\rfloor-2$ for any sequence of $n$ integers in $\mathcal{B}_{n,n}$. This concludes the proof.
Let $\mathcal{W}$ be any sequence of $2n+2$ integers from zero to $n$. Given any bar stack $\mathcal{B}_{n,n}$, there exits a side compatible subset $\mathcal{C}$ for $\mathcal{W}$ that is of size $\frac{n}{2}$. \[lem:n-bar-n-wire\]
We use the notation used in the proof of Lemma \[lem:all-wires\]. Let $X_y$ be the smallest subset for which $\sharp(U^y_{X_y})<|X_y|$. First note that $U^y_{X_y}$ is bounded for $y \in \{n-0.5, n+0.5, n+1.5, n+2.5\}$ (i.e. any of the four middle wires). There cannot exist more than $n$ consecutive wires whose $U^y_{X_y}$ is bounded. Thus, there exists $y$ such that $U^y_{X_y}$ in unbounded and $\sharp(U^y_{X_y}) \ge \frac{n}{2}+1$. This implies that there exists a side compatible subset of size $\frac{n}{2}$.
There exist a bar stack $\mathcal{B}_{{n \choose 2},n}$ and a sequence $\mathcal{W}_{l,{n \choose 2}}$, with $l = \Theta(n^2)$, for which the largest side compatible subset is less than $n$.
First note that we can partition the edges of $K_n$ into $\frac{n}{2}$ paths of size $n-1$. See Figure \[fig:kn-paths\]. We construct a bar stack using these paths. Let $P_0,\dots, P_{\frac{n}{2}-1}$ denote the set of these paths. Assume that the vertices of $K_n$ are labeled $1,\dots,n$. Let $e^i_k = (u^i_k,v^i_k)$ denote the $k$-th edge in $P_i$. $e^i_k$ corresponds to $B_{i*(n-1)+k}$, which is a segment from $(u^i_k,i*(n-1)+k )$ to $(v^i_k,i*(n-1)+k )$. Let $\mathcal{B}_i$ denote the set of bars $\{B_{i*(n-1)+k} \mid k \in [n-1]\}$. We refer to $\mathcal{B}_i$ as a *block* (of segments). Recall that $\mathcal{W}_{l,{n \choose 2}}$ corresponds to a set of wires. Let $l=(\frac{n}{2}+1)(n-1)$. We construct $\mathcal{W}_{l,{n \choose 2}}$ so that there are $n-1$ wires between every two consecutive blocks (i.e. there are $n-1$ wires between $B_{i*(n-1)+n-1}$ and $B_{(i+1)*(n-1)+1}$ for all $i$ in $\{0, \dots,\frac{n}{2}-2\}$); $n-1$ wires below $\mathcal{B}_0$; and $n-1$ wires above $\mathcal{B}_{\frac{n}{2}-1}$. Note that there is no wire between the segments that belong to the same block. A set of wires $W$ *encompasses* a block if the block is between two wires in $W$. Any set of $n$ wires encompasses a block, and each block of segments is incident to all $n$ pillars. Thus, the size of the largest side compatible subset is at most $n-1$.
\[scale=.7\] ;
in [0,1,...,4]{}[ (O) at (3\*,0); iin [0,1,...,10]{}[ (O) – +(36\*i: cm) coordinate (Di); (Di) circle (2pt); ]{} ]{}
(D00) – (D10) – (D90) – (D20) – (D80) – (D30) –( D70) – (D40) – (D60) – (D50); (D11) – (D21) – (D01) – (D31) – (D91) – (D41) –( D81) – (D51) – (D71) – (D61); (D22) – (D32) – (D12) – (D42) – (D02) – (D52) –( D92) – (D62) – (D82) – (D72); (D33) – (D43) – (D23) – (D53) – (D13) – (D63) –( D03) – (D73) – (D93) – (D83); (D44) – (D54) – (D34) – (D64) – (D24) – (D74) –( D14) – (D84) – (D04) – (D94);
Does there exist a bar stack $\mathcal{B}_{l,n}$ together with a sequence $\mathcal{W}_{n,l}$, where $l=\Theta(n^2)$, such that the largest side compatible subset for $\mathcal{W}_{n,l}$ in $\mathcal{B}_{l,n}$ is of size $o(n)$?
Given a bar stack $\mathcal{B}$, let $G(\mathcal{B})$ denote the geometric graph whose vertices are the endpoints of segments in $\mathcal{B}$ such that two vertices $u$ and $v$ are connected by an edge if $(u,v)$ is a segment in $\mathcal{B}$ or if $u$ and $v$ belong to the same pillar. A *simple* cycle in a graph is a cycle that does not self-intersect.
Let $\mathcal{B}$ be the bar stack obtained from the bar representation of a line arrangement. $G(\mathcal{B})$ does not admit any simple cycles.
Let $\mathcal{W}=\mathit{w}_1 \mathit{w}_2 \dots \mathit{w}_n$ be any sequence of $n$ integers from zero to ${n \choose 2}$. Let $\mathcal{B}$ be any bar representation of a line arrangement. Is the largest side compatible subset for $\mathcal{W}$ in $\mathcal{B}$ linear-sized? \[lem:bar-rep\]
Generalizations of crossing families
====================================
In this section, we study two (geometric) generalizations of crossing families.
The first generalization may be viewed as a generalization of spoke sets in the dual plane. In Section \[sec:ss\], we summarize the definitions and results on spoke sets in the primal and dual planes. We then introduce a more generalized notion for the dual of spoke sets, which we call “M-semialternating paths”. M-semialternating paths are related to pseudolines in two-coloured line arrangements that intersect (in order) lines of alternate colours in the arrangement. We give a few upper bounds on the sizes of certain M-semialternating paths. As a step towards showing that spoke sets are linear-sized, we prove that the sizes of certain M-semialternating paths are connected. This also helps us improve the upper bound on the size of spoke sets from $\frac{9n}{20}$ to $\frac{n}{4}+1$.
The second generalization generalizes a crossing family to a family of segments such that for every pair of segments, the supporting line of one intersects the interior of the other. We call such a family of segments a *stabbing family*. We show that the size of the largest stabbing family in an $n$-point set is $\frac{n}{2}$.
Spoke sets {#sec:ss}
----------
@bose studied partitions of complete geometric graphs into plane trees and introduced the notion of spoke sets, which is closely related to crossing families.
\[def:ss\] Let $\mathcal{P}$ be a set of points in general position. A set $\mathcal{L}$ of pairwise non-parallel lines such that each open unbounded region in the arrangement of $\mathcal{L}$ has at least one point of $\mathcal{P}$ is called a *spoke set* for $\mathcal{P}$. The size of a spoke set $\mathcal{L}$ is the number of lines in $\mathcal{L}$.
Note that if we extend the segments of a crossing family to lines and perturb them infinitesimally with a clockwise rotation so that the endpoints of each segment are on different sides of its perturbed line, the resulting line arrangement has exactly one point in every unbounded region and hence forms a spoke set. While any crossing family of size $k$ guarantees a spoke set of size $k$, the reverse is not true. It is not known whether the order of magnitude of the sizes of crossing families and spoke sets are the same or not.
@gcrf17 studied spoke sets in the dual plane and introduced the notion of spoke paths:
\[def:sp\] A *cell-path* in an arrangement $\mathcal{L}$ of lines (or pseudolines) is a sequence of cells in the arrangement such that consecutive cells share an edge. The *length* of a cell-path is one less than the number of cells involved. A cell-path $\mathcal{C}=\langle C_0,C_1,\dots, C_k\rangle$ is *AB-semialternating* if for every even $i<k-1$, $C_i$ is above the line separating $C_i$ and $C_{i+1}$ if and only if $C_{i+2}$ is above the line separating $C_{i+1}$ and $C_{i+2}$. A cell-path is *line-monotone* if the lines extending the edges shared by two consecutive cells are all distinct. A cell-path in a subarrangement of $\mathcal{L}$ that is line-monotone and AB-semialternating is called a *spoke path* for $\mathcal{L}$. A cell-path is *admitted* by the set of lines that extend the edges shared by two consecutive cells in the cell-path.
Let $P$ be a point set and $P^\star$ denote the dual of $P$. Let $\mathcal{C}=\langle C_0,C_1,\dots, C_{2k}\rangle$ be a spoke path for $P^\star$. $\mathcal{L}=\{p^\star_i \mid p_i \in C_{2i}, i \in[k]\}$ is a spoke set for $P$ (where $p^\star_i$ is the dual of $p_i$).
\[@gcrf17, Lemma 1\] Let $\mathcal{P}$ be a point set and $\mathcal{P^{\star}}$ be the dual line arrangement for $\mathcal{P}$. $\mathcal{P}$ contains a spoke set of size $k$ if and only if $\mathcal{P^\star}$ contains a spoke path of length $2k$.
\[@gcrf17, Lemma 2\] Let $\mathcal{P}=\mathcal{P}_{\mathcal{B} \vdash \mathcal{R}}$ be a $1$-avoiding $2k$-point set. The dual line arrangement $\mathcal{P^\star}$ contains a spoke path of length $k+2$, if $k$ is even, and $k+3$, if $k$ is odd. \[schnider-claim\]
Conjecture \[schnider-claim\] is claimed to be proved by @gcrf17 but the proof given does not seem to be correct. The sketch of their proof is the following: Given the line arrangement $\mathcal{P}^\star=\mathcal{B}^\star \cup \mathcal{R}^\star$, they construct an “extended diagram” for $\mathcal{P}^\star$, which is a horizontal wiring diagram for $\mathcal{R}^\star$ and a vertical wiring diagram for $\mathcal{B}^\star$, with the same intersection order along every pseudoline as in $\mathcal{P}^\star$. A *wiring diagram* is an arrangement of pseudolines consisting of piecewise linear “wires”, where the wires (i.e. pseudolines) are horizontal except for small neighbourhoods of their crossings with other wires. They change the extended diagram through a number of steps to reach a certain type of an extended diagram, which has two spoke paths of size $2k$. They then reverse their moves to get back to the initial extended diagram, and in each step modify those spoke paths accordingly so that they become spoke paths of the new extended diagram obtained. Note that the size of the spoke paths may shrink at each step. They show that at least one of those spoke paths is large enough when they return to the initial diagram. However, the problem with their proof is that their proposed rules for getting new diagrams from old ones may cause the pseudolines to double cross and hence what is obtained is not guaranteed to be an extended diagram. It is easy to change the rules so that at each step we can guarantee that no two pseudolines double cross, however, with the new rules, the process of modifying the spoke paths when reversing the moves becomes problematic. That is, either it is not easy to maintain a spoke path or to guarantee having a large enough one.
M-semialternating paths {#sec:m}
-----------------------
Here, we generalize the notion of spoke paths. We start by some terminologies.
We call a polygonal chain whose line segments connect points in consecutive cells of a spoke path admitted by $\mathcal{A} \subseteq \mathcal{L}$ an *AB-semialternating path* for $\mathcal{A}$. The definition of spoke paths (Definition \[def:sp\]) implies that any AB-semialternating path for $\mathcal{A}$ starts and ends in median cells of $\mathcal{A}$. A *median cell* is a cell of the arrangement that has an equal number of lines above and below it. A cell is of *level* $k$ if each point in its interior is above exactly $k$ lines. Figure \[fig:sp-m\] shows an AB-semialternating path for a subarrangement.
(0.09,0.28)–(4,-3.36); (0.92,1.25)–(2.5,-3.43); (1.44,1.62)–(0.93,-4.14); (2.29,1.7)–(-0.66,-4);
(-1.08,-1.4)–(3.08,1.71); (-1.09,-1.98)–(3.04,0.76); (-1.21,-3.43)–(2.94,-0.66); (-1.29,-3.72)–(4,-0.75);
\(A) at (-0.31,-0.17); (B) at (0.53,0.48); (C) at (1,0); (D) at (1.35,-0.27); (E) at (1.44,-0.61); (F) at (2.09,-0.46); (G) at (3,-1);
in [A,B,...,G]{} () circle (1pt);
\(A) –(B); (B) –(C); (C) –(D); (D) –(E); (E) –(F); (F) –(G);
Let $\mathcal{C}=\langle C_0, C_1,\cdots,C_{|\mathcal{A}|}\rangle$ be a line-monotone cell-path admitted by $\mathcal{A}$, where $|\mathcal{A}|$ is even. An *even* cell in $\mathcal{C}$ is a cell whose level has the same parity as that of $C_0$. Let the *level-sequence* of $\mathcal{C}$ be $seq(\mathcal{C})=\langle s_0,s_1,\cdots,s_{\frac{|\mathcal{A}|}{2}}\rangle$, where $s_i$ is the level of $C_{2i}$ in $\mathcal{C}$. Note that every even cell in a spoke path admitted by $\mathcal{A}$ is a median cell of $\mathcal{A}$; hence the level-sequence of a spoke path admitted by $\mathcal{A}$ is $\{\frac{\mathcal{A}}{2}\}^{\frac{\mathcal{A}}{2}+1}$.
We generalize the notion of an AB-semialternating path admitted by a subarrangement $\mathcal{A}$ so that it intersects each line in $\mathcal{A}$ once and the level-sequence of the underlying cell-path is non-decreasing. A pair of unbounded regions $U_1$ and $U_2$ in a line arrangement are *antipodal* if for every line of the arrangement, the points in $U_1$ and $U_2$ are on opposite sides. The fact that the underlying cell-path intersects each line in $\mathcal{A}$ exactly once, implies that our generalized path needs to start and end in antipodal unbounded regions (of $\mathcal{A}$).
Let $\mathcal{L}=\mathcal{L_B} \cup \mathcal{L_R}$ be a two-coloured line arrangement, where $|\mathcal{L}_\mathcal{B}|=|\mathcal{L}_\mathcal{R}|$. Let $\ell$ be a pseudoline intersecting each line in $\mathcal{L}$ once. Let $\ell(\mathcal{L})=l_1, l_2,\dots, l_{|\mathcal{L}|}$ define the order in which $\ell$ intersects the lines of $\mathcal{L}$. We say $\ell$ is monotonically semialternating, or M-semialternating for short, for $\mathcal{L}$ if
1. \[semialt\]for every odd $i<|\mathcal{L}|$, $l_i$ and $l_{i+1}$ are of different colours, and
2. the level-sequence of the underlying cell-path for $\ell$ is non-decreasing.
If $\mathcal{A} \subseteq \mathcal{L}$ admits an M-semialternating pseudoline, we say $\mathcal{L}$ contains an *M-semialternating path* of *size* $|\mathcal{A}|$. We say $\ell$ is *semialternating* if it satisfies \[semialt\].
Let $\mathcal{P}_{\mathcal{B} | \mathcal{R}}$ be a point set $\mathcal{B} \cup \mathcal{R}$, where $\mathcal{B}$ is a blue point set and $\mathcal{R}$ is a red point set, such that $\mathcal{B}$ and $\mathcal{R}$ are two equal-sized sets of points that are separable by a vertical line. We call $\mathcal{P}_{\mathcal{B} | \mathcal{R}}$ a *color-separable* point set. Let $\mathcal{L}_\mathcal{B}$ and $\mathcal{L}_\mathcal{R}$ be the dual line arrangements for $\mathcal{B}$ and $\mathcal{R}$, respectively. Let $\mathcal{L}=\mathcal{L}_\mathcal{B} \cup \mathcal{L}_\mathcal{R}$, where $\mathcal{L}_\mathcal{B}$ is blue and $\mathcal{L}_\mathcal{R}$ is red. We call $\mathcal{L}$ a *color-separable* line arrangement. Recall that a line arrangement is *$1$-avoiding* if its dual point configuration is $1$-avoiding.
We show that the sizes of certain M-semialternating paths in color-separable line arrangements are connected. M-semialternating paths that start in different levels of a color-separable line arrangement correspond to different concepts in the dual plane. We exploit this correspondence to improve the upper bound on the size of spoke sets. We also give a linear upper bound on the size of subarrangements admitting semialternating pseudolines. We show that if we consider lines rather than pseudolines, there exist $1$-avoiding line arrangements whose largest subarrangement admitting an M-semialternating line is of constant size.
For a point set $P$ and a line $l$, let $A(P,l)$ denote the set of points in $P$ that are above $l$. A *parallel set* of size $k$ in a two-coloured point set $\mathcal{P}=\mathcal{P}_{\mathcal{B} | \mathcal{R}}$ is a set of lines $L={L_0,L_1,\dots,L_k}$ for which there exist $B \subseteq \mathcal{B}$ and $R \subseteq \mathcal{R}$, where $|B|=|R|=k$, such that for any $1 \le i \le k$, $A(B,L_{i-1}) \subsetneq A(B,L_i)$ and $A(R,L_{i-1}) \subsetneq A(R,L_i)$. A *focal parallel set* is a parallel set whose lines all intersect at the same point.
Let $\mathcal{P}=\mathcal{P}_{\mathcal{B} | \mathcal{R}}$ be a color-separable point set. Let $\mathcal{P}^\star$ be the dual line arrangement for $\mathcal{P}$. $\mathcal{P}$ contains a parallel set (or focal parallel set) of size $k$ if and only if a $2k$-subset $\mathcal{L}$ of $\mathcal{P}^\star$ admits an M-semialternating pseudoline (or line) that starts and ends in cells of $\mathcal{L}$ that are of levels zero and $2k$.
Let $\mathcal{P}=\mathcal{P}_{\mathcal{B}|\mathcal{R}}$ be a color-separable $2n$-point set, and let $\mathcal{L}$ be the dual line arrangement for $\mathcal{P}$. If $\mathcal{L}$ contains an M-semialternating path of size $2k$, then either the spoke set or the parallel set for $\mathcal{P}$ is of size at least $\frac{k}{2}$. \[lem:ssORps\]
Let $\ell$ denote an M-semialternating pseudoline for $\mathcal{L}' \subseteq \mathcal{L}$, where $|\mathcal{L}'|=2k$. Let $\mathcal{C}$ denote the underlying cell-path for $\ell$ in $\mathcal{L}'$. Let $\mathcal{S}= \langle s_0,s_1,\cdots,s_k \rangle$ denote the level-sequence of $\mathcal{C}$. Note that $s_0+s_k=2k$ and for every $i \in [k]$, $s_i-s_{i-1} \in \{0,2\}$. Assume $s_0=\mathtt{x}$ (thus $s_k=2k-\mathtt{x}$), where $0 \le \mathtt{x} \le k$. $\mathcal{S}$ contains $\frac{2k-\mathtt{x}-\mathtt{x}}{2}+1=k-\mathtt{x}+1$ distinct numbers and $\mathtt{x}$ repetitions. Let $\mathcal{I}_1= \{ i \mid s_i-s_{i-1}=2\}$ and $\mathcal{I}_2=[k] \setminus \mathcal{I}_1$. Note that $|\mathcal{I}_1|=k-\mathtt{x}$ and $|\mathcal{I}_2|=\mathtt{x}$. Let $B(i)$ denote the set of lines that are below the $(i+1)$-th even cell in $\mathcal{C}$ (note that $|B(i)|=s_i$). Let $\mathcal{L}_1=\bigcup \limits_{i \in \mathcal{I}_1} B(i) \setminus B(i-1)$, and $\mathcal{L}_2=\bigcup \limits_{i \in \mathcal{I}_2} B(i) \triangle B(i-1)$, where $\triangle$ represents the symmetric difference. Recall that $\ell$ intersects each line in $\mathcal{L}'$ exactly once. $\ell$ forms an M-semialternating pseudoline for $\mathcal{L}_1$ that starts and ends in cells of levels zero and $|\mathcal{L}_1|=2|\mathcal{I}_1|=2(k-\mathtt{x})$ (Note that the level-sequence of the underlying cell-path for $\ell$ in $\mathcal{L}_1$ is strictly increasing). Similarly, $\ell$ forms a spoke path of size $|\mathcal{L}_2|=2|\mathcal{I}_2|=2\mathtt{x}$ for $\mathcal{L}_2$ (The level-sequence of the underlying cell-path for $\ell$ in $\mathcal{L}_2$ is a constant sequence). Hence the dual point set for $\mathcal{L}_1$ has a parallel set of size $k-\mathtt{x}$ and the dual point set for $\mathcal{L}_2$ has a spoke set of size $\mathtt{x}$. This consequently implies that either the parallel set or the spoke set for $\mathcal{P}$ is of size at least $\frac{k}{2}$.
Let $\ell$ be an M-semialternating pseudoline for the color-separable line arrangement $\mathcal{L}$, where $|\mathcal{L}|=2k$. Let $\mathcal{C}=\langle C_0,C_1,\cdots,C_{2k} \rangle$ denote the underlying cell-path for $\ell$ in $\mathcal{L}$. Let $o$ be a point in $C_0 \cap \ell$. Let $Rot_\alpha(\cdot)$ be the function that rotates the input $\alpha$ degrees clockwise about $o$. Clearly, for any $\alpha$, $Rot_\alpha(\ell)$ remains semialternating for $Rot_\alpha(\mathcal{L})$. However, the level-sequence of the underlying cell-path for $Rot_{\alpha}(\ell)$ in $Rot_{\alpha}(\mathcal{L})$ may become non-monotone. Moreover, the lines below $Rot_\alpha(C_0)$ are not (necessarily) of the same color. It is easy to see (proof below) that given any M-semialternating pseudoline, there exists an angle $\alpha$ such that the level-sequence of the underlying cell-path for $Rot_\alpha(\ell)$ in $Rot_\alpha(\mathcal{L})$ is strictly increasing. However, there may not exist an angle $\alpha$ such that the level-sequence of the underlying cell-path for $Rot_\alpha(\ell)$ in $Rot_\alpha(\mathcal{L})$ is a constant sequence.
Let $\mathcal{S}=\langle s_0,s_1,\cdots, s_k \rangle$ denote the level-sequence of $\mathcal{C}$. $s_i-s_{i-1} \in \{0,2\}$. Since $\ell$ crosses each line of $\mathcal{L}$ once and the level-sequence is non-decreasing, we infer that $0 \le s_0 \le k$; thus, since $\mathcal{L}$ is color-separable, all the lines below $C_0$ are of the same color. (Recall that a vertical line to the left of all intersection points of $\mathcal{L}$ intersects all the red and all the blue lines consecutively.) Let $\theta^-$ be such that the set of lines below $o$ in $Rot_{\theta^-}(\mathcal{L})$ is the same as the set below $o$ in $\mathcal{L}$ except one less. ($\theta^-$ is undefined if the level of $o$ in $\mathcal{L}$ is zero.) Similarly, let $\theta^+$ be such that the set of lines below $o$ in $Rot_{\theta^+}(\mathcal{L})$ is the same as the set below $o$ in $\mathcal{L}$ except one more. Let $\overline{\mathcal{S}}=\langle \overline{s}_0,\overline{s}_1,\cdots,\overline{s}_k \rangle$ and $\overset{+}{\mathcal{S}}=\langle \overset{+}{s}_0,\overset{+}{s}_1,\cdots,\overset{+}{s}_k \rangle$ be the level-sequences of the underlying cell-paths for $Rot_{\theta^-}(\ell)$ in $Rot_{\theta^-}(\mathcal{L})$ and $Rot_{\theta^+}(\ell)$ in $Rot_{\theta^+}(\mathcal{L})$, respectively. There exists $d \in[k]$ such that for all $j<d$, $\overset{+}s_j=s_j+1$ and for all $j \ge d$, $\overset{+}s_j=s_j-1$. Similarly, there exists $u \in[k]$ such that for all $j<u$, $\overline{s}_j=s_j-1$ and for all $j \ge u$, $\overline{s}_j=s_j+1$.
$\overline{\mathcal{S}}$ is monotone. The difference between two consecutive elements in $\overline{\mathcal{S}}$ can either be zero or two (since $\overline{\mathcal{S}}$ represents the level-sequence of a cell-path). This implies that $u$ is such that $s_{u}=s_{u-1}$, and hence $\overline{\mathcal{S}}$ contains more distinct numbers compared to $\mathcal{S}$. The lines below $o$ in $\mathcal{L}$ are all of the same color, and the lines below $o$ in the rotated arrangement (with rotation angle $\theta^-$) are still of the same color. Therefore, we can continue rotating by $\theta^-$ degrees until the (rotated) pseudoline starts from the zero level and we get a strictly increasing level-sequence.
$\overset{+}{\mathcal{S}}$ is monotone only if $s_d \neq s_{d-1}$. Starting with the pair $\mathcal{L}$ and $\ell$ and rotating them $\theta^+$ degrees, $Rot_{\theta^+}(\ell)$ remains M-semialternating with respect to $Rot_{\theta^+}(\mathcal{L})$. However, if we continue rotating this way, the level-sequence of the underlying cell-path may become non-monotone. Hence, by rotating an M-semialternating pseudoline (together with the arrangement), it may not be possible to get a level-sequence that is constant.
The minimum size of the largest spoke set among all configurations of color-separable $n$-point sets is the same as the minimum size of the largest parallel set in a color-separable $n$-point set. The statement still holds if we replace “color-separable” with “$1$-avoiding”. \[lem:pset-ss\]
Let $\mathcal{P}=\mathcal{P}_{\mathcal{B} | \mathcal{R}}$ be a color-separable point set. Let $ss(\mathcal{P})$ and $ps(\mathcal{P})$ denote the sizes of the maximum spoke set and parallel set for $\mathcal{P}$, respectively. We show that for every color-separable point set $\mathcal{P}$ there exists a color-separable point set $\mathcal{P}'$ such that $ss(\mathcal{P'})=ps(\mathcal{P})$ and $ps(\mathcal{P}')=ss(\mathcal{P})$. We construct $\mathcal{P}'$ as follows:
- Assume by rotation and translation that $\mathcal{B}$ and $\mathcal{R}$ lie on the right and left sides of the $y$-axis, respectively.
- Let $\mathcal{L}$ be the dual line arrangement for ${\mathcal{P}}$.
- Let $\mathcal{L}'$ be the line arrangement that is obtained by rotating $\mathcal{L}$ $90$ degrees clockwise.
- $\mathcal{P}'$ is the dual point configuration for $\mathcal{L}'$.
We know there exists $L_s \subseteq \mathcal{L}$, where $|L_s|=2 \cdot ss(\mathcal{P})$, such that $L_s$ admits an M-semialternating pseudoline $\ell_s$ that starts and ends in median cells of $L_s$. Similarly, there exists $L_p \subseteq \mathcal{L}$, where $|L_p|=2 \cdot ps(\mathcal{P})$, such that $L_p$ admits an M-semialternating pseudoline $\ell_p$ that starts and ends in cells of $L_p$ that are of levels zero and $|L_p|$. Let $Rot(\cdot)$ denote the function that rotates the input $90$ degrees clockwise about the origin. Recall that $\mathcal{L}'=Rot(\mathcal{L})$. Clearly, $Rot(\ell_s)$ is a semialternating pseudoline for $Rot(L_s)$. Similarly, $Rot(\ell_p)$ is semialternating for $Rot(L_p)$. Recall that any M-semialternating pseudoline $\ell$ admitted by an arrangement $\mathcal{A}$ intersects every line of $\mathcal{A}$ once. Let $B_\ell$ denote the set of lines below the first cell in the underlying cell-path for $\ell$ in $\mathcal{A}$. If the level-sequence of the underlying cell-path for $Rot(\ell)$ in $Rot(\mathcal{A})$ becomes non-monotone, $B_\ell$ must contain both blue and red lines (note that $Rot(\ell)$ remains semialternating and cannot intersect any line in $Rot(\mathcal{A})$ more than once). $B_{\ell_p}$ in $L_p$ is an empty set, and $B_{\ell_s}$ in $L_s$ consists of lines that are of the same color. Thus, $Rot(\ell_s)$ and $Rot(\ell_p)$ are M-semialternating pseudolines for $Rot(L_s)$ and $Rot(L_p)$, respectively. $Rot(\ell_s)$ starts and ends in cells of levels zero and $|L_s|$ in $Rot(L_s)$, and $Rot(\ell_p)$ starts and ends in median cells in $Rot(L_p)$. Therefore, $\mathcal{P}'$ contains a parallel set of size $ss(\mathcal{P})$, and a spoke set of size $ps(\mathcal{P})$.
In the following, we prove (in two ways) that if $\mathcal{P}$ has the additional property that $\mathcal{B}$ avoids $\mathcal{R}$ (i.e. $\mathcal{P}$ is $1$-avoiding), then $\mathcal{P}'$ is also $1$-avoiding.
Let ${\mathcal{P}}=\{p_1,p_2,\dots,p_n\}$, where $p_i=(a_i,b_i)$. We have $$\begin{aligned}
\mathcal{L} &= \{p^\star_i=\{(x,y) : y=a_ix-b_i\} \mid i \in [n] \}, \\
\mathcal{L}'&=\{ \{(x,y) : y=-\frac{x}{a_i}-\frac{b_i}{a_i}\} \mid i \in [n] \}, \text{~and} \\
\mathcal{P}'&=\{ p'_i = (-\frac{1}{a_i},\frac{b_i}{a_i}) \mid i \in [n] \}.\end{aligned}$$ Note that since no points in $\mathcal{P}$ are on the $y$-axis, $\mathcal{L}'$ and $\mathcal{P}'$ are well-defined. Let $o(p,q,r)$ denote the orientation of points $p$, $q$, and $r$. That is, $o(p,q,r)=+1$ if the circle through $p,q,r$ is clockwise; $o(p,q,r)=-1$ if the circle through $p,q,r$ is counterclockwise; and $o(p,q,r)=0$ if the three points are collinear. Clearly, three points in ${\mathcal{P}}$ are collinear if and only if their corresponding points in $\mathcal{P}'$ are collinear. Let $p_i$, $p_j$, and $p_k$ be three points in ${\mathcal{P}}$. Without loss of generality, we assume $p_i$ and $p_j$ are on the same side of the $y$-axis. Note that the ordering of the $x$-coordinates of $p_i$ and $p_j$ is the same as that of $p'_i$ and $p'_j$. Suppose $p_k$ is above the line through $p_i$ and $p_j$. We show that $p'_k$ is above the line through $p'_i$ and $p'_j$ if and only if an even number of points in $\{p_i,p_j,p_k\}$ have negative $x$-coordinates. Similarly, if $p_k$ is below the line through $p_i$ and $p_j$, then $p'_k$ is below the line through $p'_i$ and $p'_j$ if and only if an even number of points in $\{p_i,p_j,p_k\}$ have negative $x$-coordinates. This is because
$$\begin{aligned}
o(p_i,p_j,p_k) &= a_1b_2-a_2b_1+a_2b_3-a_3b_2+a_3b_1-a_1b_3 \text{, and}\\
o(p'_1,p'_j,p'_k)&=-\frac{b_2}{a_1a_2}+\frac{b_1}{a_1a_2}-\frac{b_3}{a_2a_3}+\frac{b_2}{a_2a_3}-\frac{b_1}{a_1a_3}+\frac{b_3}{a_1a_3},\end{aligned}$$
which implies $o(p_i,p_j,p_k)=a_1a_2a_3 \cdot o(p'_i,p'_j,p'_k)$. Therefore, $$\begin{aligned}
o(p_i,p_j,p_k) =
\begin{cases}
o(p'_i,p'_j,p'_k) & \text{if } \text{neg}_x(p_i,p_j,p_k) \equiv 0 \pmod{2},\\
-o(p'_i,p'_j,p'_k) & \text{otherwise},
\end{cases}\end{aligned}$$ where $\text{neg}_x(p_i,p_j,p_k)$ is the number of points in $\{p_i,p_j,p_k\}$ that are on the left side of the $y$-axis. This immediately implies that if ${\mathcal{P}}={\mathcal{P}}_{{\mathcal{B}} \vdash {\mathcal{R}}}$ is $1$-avoiding and ${\mathcal{B}}$ and ${\mathcal{R}}$ are on different sides of the $y$-axis, then $\mathcal{P}'=\mathcal{B}' \cup \mathcal{R}'$ is a $1$-avoiding point set where $\mathcal{B}'$ avoids $\mathcal{R}'$, and $\mathcal{B}'$ and $\mathcal{R}'$ are on different sides of the $y$-axis.
Alternatively, we can prove $\mathcal{P}'$ is $1$-avoiding using the dual plane directly. Recall that ${\mathcal{P}}={\mathcal{P}}_{{\mathcal{B}} \vdash {\mathcal{R}}}$, and $\mathcal{L}={\mathcal{P}^\star}={\mathcal{B}^\star} \cup {\mathcal{R}^\star}$ is such that ($\uppercase\expandafter{\romannumeral 1}$) each line in ${\mathcal{R}}^\star$ intersects the lines of ${\mathcal{B}}^\star$ in the same order, and ($\uppercase\expandafter{\romannumeral 2}$) the lines above (or below) each unbounded median cell of $\mathcal{L}$ are all of the same color. Note that both properties ($\uppercase\expandafter{\romannumeral 1}$) and ($\uppercase\expandafter{\romannumeral 2}$) hold for $Rot(\mathcal{L}')$, and hence the dual of $\mathcal{L}'$ is a $1$-avoiding point set.
A *parallel family* is a set of segments such the supporting lines of every pair of segments intersect outside both segments. The transformation we use for proving Lemma \[lem:pset-ss\] turns out to be the same, upto a negation in the $x$-coordinate, as the transformation used by @Pach-crf in proving that the problems of finding the maximum crossing family and maximum parallel family are equivalent.
Let $\mathcal{P}=\mathcal{P}_{\mathcal{B} \vdash \mathcal{R}}$ be a $1$-avoiding $2n$-point set. The sizes of the largest spoke set and largest parallel set for $\mathcal{P}$ are of the same order of magnitude. \[q:ssTOps\]
\[lem:semiTOm\] Let $\mathcal{L}$ be a color-separable line arrangement. The size of the largest M-semialternating path in $\mathcal{L}$ is at least half the size of the largest semialternating path in $\mathcal{L}$.
Let $\ell$ be a semialternating pseudoline for subarrangement $\mathcal{L}' \subseteq \mathcal{L}$. Let $\mathcal{C}=\langle C_0, C_1,\cdots,C_{2k}\rangle$ be the underlying cell-path for $\ell$ in $\mathcal{L}'$. Let $\mathcal{S}=\langle s_0,s_1,\cdots, s_k \rangle$ denote the level-sequence of $\mathcal{C}$. Clearly, for any $i \in [k]$, $s_i-s_{i-1} \in \{0,2,-2\}$. If for all $i \in [k]$, $s_i-s_{i-1} \in \{0,2\}$ then the level-sequence is monotone. Let $\mathcal{I}=\{ i \mid s_i-s_{i-1}=-2\}$. Let $B(i)$ denote the set of lines that are below $C_{2i}$. Note that $\ell$ intersects each line of $\mathcal{L}'$ once. This implies that for $i \in \mathcal{I}$, $B(i-1) \setminus B(i) \subseteq B(0)$. Therefore, $|\mathcal{I}| \le \dfrac{s_0}{2}$. Let $\mathcal{L}''$ be the arrangement $\mathcal{L}'$ without $ \bigcup \limits_{i \in \mathcal{I}} B(i-1) \setminus B(i)$. The underlying cell-path for $\ell$ in $\mathcal{L}''$ is monotone. Recall that $s_0 \le k$. Hence $|\mathcal{L}''| \ge \dfrac{|\mathcal{L}'|}{2}$. This concludes the proof.
\[conj:large-m\] The size of the largest semialternating path for any $1$-avoiding line arrangement $\mathcal{L}$ is linear (in the number of lines in $\mathcal{L}$).
Conjectures \[conj:large-m\] and \[q:ssTOps\] (if answered affirmatively), together with Lemmas \[lem:ssORps\] and \[lem:semiTOm\], imply that any $1$-avoiding point set has a linear-sized spoke set.
\[conj:max-over-min\] Let $\mathcal{L}$ be a $1$-avoiding line arrangement. Let $\ell$ be a directed pseudoline that intersects each line in $\mathcal{L}$ once. For $i=0,1,\dots,n$, let $\mathcal{A}_i$ be the largest sub-arrangement that admits an M-semialternating $\ell$ that starts in level $i$ of sub-arrangement $\mathcal{A}_i$. Let $\mathcal{A}=\{|\mathcal{A}_i| \mid i \in [0..n],\mathcal{A}_i \neq \emptyset\}$. The ratio of the maximum element in $\mathcal{A}$ to the minimum element in $\mathcal{A}$ is constant.
Conjectures \[q:ssTOps\] and \[conj:max-over-min\] are equivalent.
Let $\mathcal{L}$ be a $1$-avoiding line arrangement consisting of $n$ blue lines and $n$ red lines. There exists $\mathcal{L}$ whose largest semialternating path is of size $\frac{10}{11}\cdot 2n$.
First, we construct a $1$-avoiding line arrangement consisting of $11$ blue lines and $11$ red lines such that it admits no semialternating pseudoline. See Figure \[fig:noFullPath\]. Each blue line has a greater slope than that of a red line. The blue lines are almost parallel. Let $b_i$ denote the $i$-th leftmost blue line. Let $r_i$ denote the red line with the $i$-th largest slope. For the sake of contradiction, suppose that there exists a semialternating pseudoline $l$. Note that the intersection of $l$ and $b_6$ may only appear on either the downward ray originating at $V$ or the upward ray originating at $U$ (otherwise, $l$ is not semialternating). We consider the following cases:
1. $l$ intersects the downward ray originating at $V$. In order for $l$ to intersect each blue line in $\{b_1,\dots,b_5\}$, it needs to intersect $r_4$ prior to $b_6$. If the intersection of $l$ and $b_6$ is below $r_4$, then in order for $l$ to be semialternating, it needs to cross $r_4$ again, which is not possible. Thus, the intersection of $l$ and $b_6$ is above $r_4$ (and $l$ starts below $r_4$). Let $l_1,\dots, l_{22}$ denote the order in which $l$ intersects the lines of the arrangement. For an odd $i$, we say that $l_i$ and $l_{i+1}$ are paired. Recall that $l$ needs to intersect each blue line in $\{b_1,\dots,b_5\}$. However, there are at most four red lines that may be paired with $b_1,\dots b_5$. Hence, either $l$ cannot be semialternating or it needs to double cross some red lines.
2. $l$ intersects the upward ray originating at $U$. In order for $l$ to intersect each blue line in $\{b_1,\dots,b_5\}$, it needs to intersect $r_{11}$ prior to $b_6$. This implies that, in order for $l$ to intersect every line, it needs to intersect red lines $\{r_5,\dots,r_{10}\}$ prior to $r_{11}$. Therefore, $l$ needs to intersect at least seven red lines prior to $b_6$, which is not possible.
By replacing each line of the arrangement depicted in Figure \[fig:noFullPath\] with $\frac{n}{11}$ lines of the same color such that they are all almost parallel to the initial line and their distances to each other are all infinitesimally small, we obtain the lemma.
\[lem:fps-c\] There exist $1$-avoiding point sets whose largest focal parallel set is of constant size.
We construct an arrangement of blue lines $\mathcal{L_{\mathcal{B}}}$ and red lines $\mathcal{L_{\mathcal{R}}}$ whose dual point sets $\mathcal{B}$ and $\mathcal{R}$ are mutually avoiding and the largest focal parallel set of $\mathcal{B} \cup \mathcal{R}$ is of constant size.
1. Let $\{b_1,b_2,\dots, b_n\}$ be a set of blue vertical lines where $b_i$ is $x=i$.\[b-unit\]
2. Let $\{r_1,r_2,\dots,r_n\}$ be a set of red horizontal lines, where $r_1$ and $r_2$ are $y=1$ and $y=2$, respectively.
3. We define the remainder of the red lines incrementally. For $1< i < n$, we define $r_{i+1}$ from $\{r_1,r_i\}$ and $\{b_1,b_2,b_n\}$. Let $\mathtt{x}(l_1,l_2)$ denote the intersection point of lines $l_1$ and $l_2$. Let $d_i$ be the line joining $\mathtt{x}(r_1,b_1)$ and $\mathtt{x}(r_i,b_2)$. Let $x_i=\mathtt{x}(d_i,b_n)$. We define $r_{i+1}$ to be a horizontal line that is slightly above $x_i$. See Figure \[fig:fps\].\[construct-r\]
; ; ;
iin [1,2,...,]{} (i-,0) –+ (0,) node \[label=above:[i<3 $b_\i$ i=6 $b_n$ i=4 $\ldots$ ]{}\] ;
(-/2,1) –+(,0) node \[label=right:[$r_1$]{}\] ; (-/2,1.5) –+(,0) node \[label=right:[$r_2$]{}\] ;
; ;
(X1) to +($(X1)-(X2)$); ; at (X3) ;
let 1=(X3) in (-/2,1.1\*1) –+(6\*,0) node \[label=right:[$r_3$]{}\] ;
; (X1) to +($(X1)-(X4)$); ; at (X5) ;
let 1=(X5) in (-/2,1.05\*1) –+(6\*,0) node \[label=right:[$r_4$]{}\] ;
iin [1,2,3,4,5]{} (Xi) circle (2.5\*pt);
Each red line intersects the blue lines in the same order. Similarly, each blue line intersects the red lines in the same order. Hence $\mathcal{B} \cup \mathcal{R}$ is mutually avoiding. Let $\mathcal{L}=\mathcal{L_{\mathcal{R}}} \cup \mathcal{L_{\mathcal{B}}}$. To make sure that the largest focal parallel set of $\mathcal{B} \cup \mathcal{R}$ is of constant size, we need to show that no line added to $\mathcal{L}$ can be M-semialternating for a subset of $\mathcal{L}$ that is of size $\omega(1)$.
Note that any line $l$ intersects the blue lines monotonically (in order of their indices). Likewise, any line $l$ intersects the red lines monotonically. Let $s(l)$ denote the order in which the line $l$ intersects the lines in $\mathcal{L}$. If a line $l$ admits an M-semialternating path of size $m$, then there exists a subset $\mathcal{I} \subset [n]$ of size $\Theta(m)$ such that for every $i \in \mathcal{I}$, $r_i$ and $r_{i+1}$ do not appear consecutively in $s(l)$. For all $i \in \mathcal{I}$ except for at most one, the difference between the $x$-coordinates of $\mathtt{x}(l,r_i)$ and $\mathtt{x}(l,r_{i+1})$ is greater than or equal to one. Let $i$ be the smallest such index. The slope of the line through $\mathtt{x}(l,r_i)$ and $\mathtt{x}(l,r_{i+1})$ is less than the slope of $d_{i+1}$ (by our construction). Hence $b_n$ appears before $r_{i+2}$ in $s(l)$, which implies that $|\mathcal{I}|$ is constant.
In order to make sure that the dual point set $\mathcal{B} \cup \mathcal{R}$ is in general position, we perturb the blue and red lines imperceptibly so that all lines of the same colour become “almost” parallel. A *grid* point of $\mathcal{L}=\mathcal{L}_\mathcal{R} \cup \mathcal{L}_\mathcal{B}$ is an intersection point of a red line in $\mathcal{L}_\mathcal{R}$ and a blue line in $\mathcal{L}_\mathcal{B}$ . Let $\mathcal{R}_i =\{r_1,r_2,\dots,r_i\}$ and let $\mathcal{L}_i=\mathcal{R}_i \cup \mathcal{L}_\mathcal{B}$. Note that the modification of the arrangement should not violate the property that every line through two grid points of $\mathcal{L}_i$, that is not a line in $\mathcal{L}_i$, intersects all blue lines prior to intersecting $r_{i+1}$. Lastly, we may rotate the resulting arrangement $45$ degrees clockwise so that the blue and red points (of the dual point configuration) are to the left and right sides of the $y$-axis.
A point set $P$ is a *wheel* if there exists a dummy point $q \notin P$ in the plane such that if we start with a vertical line through $q$ and rotate it clockwise about $q$ until it gets vertical again, the rotating line would encounter points of $P$ on alternating sides of the vertical line through $q$. A two-coloured point set is an *alternating wheel* if it is a wheel where the rotating line through the dummy point sees the points with alternating colours. The size of a wheel is the number of points it has.
It is easy to see the following:
A wheel of size $k$ implies a spoke set of size $\frac{k}{2}$.\[obs:pin-sp\]
The largest alternating wheel in a bicolored $2k$-point set that contains a crossing family of size $k$ consisting of bicolored segments (segments whose endpoints are not of the same color) may be of constant size.\[obs:crf-pin\]
Observation \[obs:pin-sp\] immediately follows from the definitions of wheels and spoke sets. Note that an alternating wheel $P$ corresponds to an M-semialternating line in $P^\star$ that starts and ends in the median cells of $P^\star$. Lemma \[lem:fps-c\], together with Lemma \[lem:pset-ss\] (when considering M-semialternating lines rather than pseudolines), proves Observation \[obs:crf-pin\].
There exists a $1$-avoiding $n$-point set whose largest parallel set is of size at most $\frac{n}{4}+1$. \[lem:p-set-upex\]
We construct a two-coloured point set $\mathcal{P}=\mathcal{P}_{\mathcal{B} | \mathcal{R}}$. Let $\mathcal{B}$ and $\mathcal{R}$ be the set of blue and red points, respectively. We construct $\mathcal{B}$ such that the blue points are almost collinear, that is, they are on an arc of a very large circle. We construct $\mathcal{R}$ in the same way and make sure that every line joining two red points is a halving line for $\mathcal{B}$. See Figure \[fig:psUPex\]. Assume $\mathcal{B}$ and $\mathcal{R}$ are almost on vertical and horizontal lines, respectively, and let $\mathcal{B}$ be to the right of $\mathcal{R}$. Let $\mathcal{R}=\mathcal{R}_1 \cup \mathcal{R}_2$, where $\mathcal{R}_1$ and $\mathcal{R}_2$ are the first and second halves of $\mathcal{R}$ when traversed from left to right. Let $\mathcal{B}=\mathcal{B}_1 \cup \mathcal{B}_2$, where $\mathcal{B}_1$ and $\mathcal{B}_2$ are the first and second halves of $\mathcal{B}$ when traversed from top to bottom. Let $L$ be a maximum parallel set exhausting $R \cup B$, where $R \subseteq \mathcal{R}$ and $B \subseteq \mathcal{B}$. For each $L_i \in L$, where $i>0$, let $r_i=A(R,L_i) \setminus A(R,L_{i-1})$ and $b_i=A(B,L_i) \setminus A(B,L_{i-1})$.
ł[15]{}; ;
\(O) at (0,0); in [,-]{} iin [0,1,2,3]{}[ (A) at (180+i +:łcm); (A) circle (1.5pt); ]{} (180+4.5\* :łcm) arc (180+4.5\* :180-4.5\* :łcm); at (180+5\* :łcm) ; at (180+2.5\* :łcm) ; at (180-2.5\* :łcm) ;
in [,-]{} iin [0,1,2,3]{}[ (A) at (180+i +:łcm); (A) circle (1.5pt); ]{} (180+4.5\* :łcm) arc (180+4.5\* :180-4.5\* :łcm); at (180-5\* :łcm) ; at (180+2.5\* :łcm) ; at (180-2.5\* :łcm) ;
If $R \cap \mathcal{R}_1 = \emptyset$, then the parallel set is of size at most $\frac{n}{4}$. Assume $r_i \in R \cap \mathcal{R}_1$. We consider two cases.
1. $b_i \in \mathcal{B}_1$. Note that for any $L_j \in L$, where $b_j \in \mathcal{B}_2$, $A(R,L_i) \subset A(R,L_j)$ (as $b_i$ is above $b_j$), and hence every red point in $R$ that is to the right of $r_i$ is in $ A(R,L_j)$. Therefore $|B \cap \mathcal{B}_2| \le 1$.
2. $b_i \in \mathcal{B}_2$. If $B \cap \mathcal{B}_1 = \emptyset$, then the parallel set is of size at most $\frac{n}{4}$. So we assume that there exists $L_j$ such that $b_j \in \mathcal{B}_1$. Since $b_j$ is above $b_i$, we know $j<i$. Thus $r_i \notin A(R,L_j)$, which implies $r_j$ is to the left of $r_i$. Note that $A(R,L_{j})$ contains $r_j$ and all red point in $R$ to the left of $r_j$. This implies that for any line $L_k$ such that $b_k \in \mathcal{B}_2$, $A(R,L_k)=R$. Therefore, $|B \cap \mathcal{B}_2|=1$.
As a result, the size of the parallel set is at most $\max\{|\mathcal{B}_1|,|\mathcal{B}_2|\}+1 = \frac{n}{4}+1$.
Note that Lemma \[lem:p-set-upex\] together with Lemma \[lem:pset-ss\] improves on the $\frac{9n}{20}$ upper bound known on the size of spoke sets due to @sch.
There exists a $1$-avoiding $n$-point set whose largest spoke set is of size at most $\frac{n}{4}+1$. \[cor:ssUP\]
Corollary \[cor:ssUP\] closes the gap between lower and upper bounds on the size of spoke sets for $1$-avoiding point sets if Conjecture \[schnider-claim\] is true.
The following observation compares the notions of crossing family, spoke set, parallel family and parallel set (for point sets in general position).
For a given point set $P$,
- if $P$ has a crossing family of size $k$, then there exists $P' \subseteq P$ where $|P'|=2k$ and $P'$ has $k$ pairwise crossing halving edges.
- if $P$ has a spoke set of size $k$, then there exists $P' \subseteq P$ where $|P'|=2k$ and $P'$ has a set of $k$ halving lines $L$ such that the corresponding halving edge of each halving line is distinct and crosses all the halving lines in $L$.
- if $P$ has a parallel family of size $k$, then there exists $P' \subseteq P$ where $|P'|=2k$ and $P'$ has a set of pairwise parallel $0-,2-,\dots,(2k-2)-$edges.
- if $P$ has a parallel set of size $k$, then there exists $P' \subseteq P$ such that $|P'|=2k$ and $P'$ has a set of $0-,2-,\dots,2k-$sets, where for any $0\le i < k$, the $2i-$set is contained in the $2(i+1)-$set. A *$k$-set* of point set $P$ is a subset $S$ of $P$ containing $k$ points that is separable from its complement $P \setminus S$ by a straight line.
Stabbing families {#sec:stab}
-----------------
Here, we study a more generalized notion than spoke sets, for which we can easily prove a linear lower bound. We start by some definitions.
A *stabbing family* is a set of segments such that for every pair of segments, the line extension of one intersects the interior of the other one.
Given two segments $e_1$ and $e_2$, extend the segments to obtain two lines. If the intersection of these lines lies
- on both segments, we say the segments are *crossing*.
- on $e_1$ but not on $e_2$, we say $e_2$ *stabs* $e_1$.
- outside both segments, we say the segments are *parallel*.
Two segments are *non-crossing* if they are either parallel or one stabs the other.
Note that any pair of segments in a stabbing family is either crossing or one stabs the other. Let $\mathcal{L}_\mathcal{P}$ denote a spoke set for $\mathcal{P}$. A *spoke matching*, with respect to $\mathcal{L}_\mathcal{P}$, is a matching where each segment (i.e. matching edge) connects two points in $\mathcal{P}$ that lie in antipodal unbounded regions of the arrangement of $\mathcal{L}_\mathcal{P}$; and no two segments start (or end) in the same unbounded region. It is easy to see that a spoke matching does not contain any parallel segments, and hence forms a stabbing family. Thus, any point set with a spoke set of size $k$, has a stabbing family of size at least $k$. However, the reverse is not true. @gcrf17 characterizes the family of spoke matchings and describes certain other properties that need to be satisfied by spoke matchings (See Theorem 2 in [@gcrf17]).
Let $\mathcal{P}=\mathcal{P}_{\mathcal{B} \vdash \mathcal{R}}$ be a $1$-avoiding point set. There is a perfect matching in $\mathcal{P}$ such that every edge of the matching connects a point in $\mathcal{R}$ to a point in $\mathcal{B}$ and the matching edges are all pairwise non-crossing. \[lem:non-cross\]
Without loss of generality, assume $\mathcal{R}$ and $\mathcal{B}$ lie to the left and right of the $y$-axis, respectively. Label the points of $\mathcal{B}$ by $b_1b_2\dots b_n$ so that for all $i<j$, $\mathcal{R}$ lies to the right of the directed line from $b_i$ to $b_j$ (since $\mathcal{P}_{\mathcal{B} \vdash \mathcal{R}}$ is $1$-avoiding such a labeling exists). Start with $k=1$. For each $k$, rotate a vertical line through $b_k$, counterclockwise about $b_k$, until it hits an unmatched red point, say $r_k$. Match $b_k$ with $r_k$. Increment $k$ by one and repeat the last step as long as $k \le n$. It is easy to see that the matching obtained has the desired property. For the sake of contradiction suppose $b_ir_i$ and $b_jr_j$ cross, where $i<j$. Note that since the matching is constructed incrementally, $r_j$ is unmatched when $b_i$ and $r_i$ are matched. However, since both $r_i$ and $r_j$ are to the right of the directed line from $b_i$ to $b_j$ (and the assumption that $\mathcal{B}$ and $\mathcal{R}$ are to the left and right sides of the $y$-axis), we know that if we rotate a vertical line through $b_i$ counterclockwise about $b_i$, it sees $r_j$ prior to $r_i$. This implies that the algorithm picks $r_j$ (over $r_i$) for $b_i$, and hence $b_ir_i$ cannot be a matching edge.
Let $\mathcal{P}=\mathcal{P}_{\mathcal{B} \vdash \mathcal{R}}$ be a $1$-avoiding point set. There is a perfect matching in $\mathcal{P}$ such that every edge of the matching connects a point in $\mathcal{R}$ to a point in $\mathcal{B}$ and the matching edges form a stabbing family.
Assume (by rotation and transformation if necessary) that $\mathcal{B}$ and $\mathcal{R}$ lie on the right and left sides of the $y$-axis, respectively. We transform ${\mathcal{P}}$ to a new point set $\mathcal{P}'$ so that the dual line arrangement for $\mathcal{P}'$ is a rotation of the dual line arrangement for $\mathcal{P}$ by $90\degree$ (similar to the transformation used in Lemma \[lem:pset-ss\]). Recall that a segment $s$ in the primal plane transforms to a double wedge $W_s$ in the dual plain. We refer to the point representing the dual of the supporting line of $s$ as the *apex* of $W_s$. Two segments in the primal plane cross if and only if in the dual plane, the apex of each double wedge is inside the other double wedge. Two segments in the primal plane are parallel if and only if in the dual plane, the apex of neither double wedge is inside the other double wedge. Segment $e$ stabs segment $f$ if and only if in the dual plane, the apex of $W_e$ is inside $W_f$, and the apex of $W_f$ is outside $W_e$. Let the segment $e_i$ within ${\mathcal{P}}$ transform to the segment $e'_i$ within $\mathcal{P}'$. Note that for a bicolored segment $e_i$, $W_{e_i}$ contains a horizontal line. Hence, the complementary double wedge of $W_{e_i}$ when rotated $90\degree$ is the dual of $e'_i$. Therefore, for a pair of bicolored segments $e_i$ and $e_j$,
- $e_i$ and $e_j$ are crossing if and only if $e'_i$ and $e'_j$ are parallel,
- $e_i$ stabs $e_j$ if and only if $e'_j$ stabs $e'_i$, and
- $e_i$ and $e_j$ are parallel if and only if $e'_i$ and $e'_j$ are crossing.
We prove in Lemma \[lem:pset-ss\] that if $\mathcal{P}$ is $1$-avoiding, so is $\mathcal{P}'$. Lemma \[lem:non-cross\] implies that $\mathcal{P}'$ has a perfect bicolored matching whose segments are pairwise non-crossing. Therefore, ${\mathcal{P}}$ has a perfect bicolored matching that forms a stabbing family.
We can easily generalize this result to general point sets.
The largest stabbing family for any $2n$-point set $\mathcal{P}$ in general position is of size $n$.
Translate $\mathcal{P}$ so that the $y$-axis becomes a halving line in $\mathcal{P}$. Assume the right and left halves in $\mathcal{P}$ are blue and red respectively. Similar to what we did before, we transform ${\mathcal{P}}$ to a new point set $\mathcal{P}'$ so that the dual line arrangement for $\mathcal{P}'$ is a rotation of the dual line arrangement for $\mathcal{P}$ by $90\degree$. It is a well-known result that every two-colored point set admits a line, called a ham-sandwich cut, that simultaneously bisects each color class. We say that a matching is *non-crossing* if the matching edges are pairwise non-crossing. We find a non-crossing bicolored matching in $\mathcal{P}'$ by induction on $|\mathcal{P}'|$. If $|\mathcal{P}'|=2$, we match the two points. Otherwise, we find a ham-sandwich cut splitting $\mathcal{P}'$ into two subsets each containing half the red and half the blue points. We match the points lying on a ham-sandwich cut (if any) and find non-crossing bicolored matchings in each of the two (smaller) subsets. A non-crossing bicolored matching in $\mathcal{P}'$ corresponds to a bicolored matching forming a stabbing family in $\mathcal{P}$.
#### **[Acknowledgments.]{}**
An initial upper bound of $\frac{n}{4}$ on the size of crossing families was achieved during the second author’s visit at EPFL. We thank Gábor Tardos and János Pach for helpful discussion at EPFL.
[^1]: A *generalized configuration* of points in general position is a finite set of points in the Euclidean plane, together with an arrangement of pseudolines, such that every pair of points lie on exactly one pseudoline and each pseudoline contains exactly two points.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recent character and phoneme-based parametric TTS systems using deep learning have shown strong performance in natural speech generation. However, the choice between character or phoneme input can create serious limitations for practical deployment, as direct control of pronunciation is crucial in certain cases. We demonstrate a simple method for combining multiple types of linguistic information in a single encoder, named *representation mixing*, enabling flexible choice between character, phoneme, or mixed representations during inference. Experiments and user studies on a public audiobook corpus show the efficacy of our approach.'
address: |
MILA\
Université de Montréal\
Montréal, Canada
title: Representation Mixing for TTS Synthesis
---
Text-to-speech, deep learning, recurrent neural network, attention, sequence-to-sequence learning.
Introduction {#sec:format}
============
TTS synthesis [@hunt1996unit] focuses on building systems which can generate speech (often in the form of an audio feature sequence $a$), given a set of linguistic sequences, $l$. These sequences, $l$ and $a$, are of different length and dimensionality thus it is necessary to find the alignment between the linguistic information and the desired audio. We approach the alignment problem by jointly *learning* to align these two types of information [@bahdanau2014neural], effectively translating the information in the linguistic sequence(s) into audio through learned transformations for effective TTS synthesis [@sotelo2017char2wav; @wang2017tacotron; @shen2017natural; @tachibana2018efficiently; @ping2018clarinet].
Data Representation
-------------------
We employ *log mel spectrograms* as the audio feature representation, a well-studied time-frequency representation [@smith2011spectral] for audio sequence $a$. Various settings used for this transformation can be seen in Table \[table:hyperparameters\]. Linguistic information, $l$, can be given at the abstract level as graphemes (also known as characters when using English) or at a more detailed level which may include pronunciation information, such as phonemes. Practically, character-level information is widely available in open data sets though this representation allows ambiguity in pronunciation [@librivox10].
Motivating Representation Mixing
--------------------------------
In some cases, it may be impossible to fully realize the desired audio without being given direct knowledge of pronunciation. Take as a particular example the sentence *“All travelers were exhausted by the wind down the river”*. The word *“wind”* in this sentence can be either noun form such as *“The wind blew swiftly on the plains”*, or verb form for traveling such as *“... and as we wind on down the road”*, both of which have different pronunciation.
Without external knowledge (such as additional context, or an accompanying video) of which pronunciation to use for the word, a TTS system which operates on character input will always have ambiguity in this situation. Alternate approaches such as grapheme to phoneme methods [@rao2015grapheme] will also be unable to resolve this problem. Such cases are well-known in both TTS and linguistics [@black1998issues; @eddington2015meaning], motivating our desire to flexibly combine grapheme and phoneme inputs in a single encoder. Our method allows per example control of pronunciation during inference *without requiring* this information for all desired inputs, and similar methods have also been important for other systems [@ping2018deep].
![[]{data-label="fig:embedding"}](embedding_module_cropped.pdf){width="8cm" height="4.1cm"}
Representation Mixing Description
=================================
The input to the system consists of one data sequence, $l_j$, and one mask sequence, $m$. The data sequence, $l_j$, consists of a mixture between a character sequence, $l_c$, and a phonetic sequence, $l_p$. The mask sequence describes which respective sequence the symbol came from with an integer ($0$ or $1$). Each time a training sequence is encountered, it is randomly mixed at the word level and we assign all spaces and punctuation as characters.
For example, a pair “$the$ $cat$”, “$@d@ah$ $@k@ah@t$” (where @ is not a used symbol and serves only as an identifier to mark phoneme boundaries) can be resampled as “$the$ $@k@ah@t$”, with a corresponding mask of $[0, 0, 0, 0, 1, 1, 1]$. On the next occurrence, it may be resampled as “$@d@ah$ $cat$”, with mask $[1, 1, 0, 0, 0, 0]$. This resampling procedure can be seen as data augmentation, as well as training the model to smoothly process both character and phoneme information without over-reliance on either representation. We demonstrate the importance of each aspect in Section \[sec:experiments\].
Combining Embeddings
--------------------
The full mixed sequence, $l_j$, separately passes through two embedding matrices, $e_c$ and $e_p$, and is then combined using the mask sequence to form a joint mixed embedding, $e_j$. For convenience, $e_c$ and $e_p$, are set to a vocabulary size that is the *max* of the character and phoneme vocabulary sizes. This mixed embedding is further combined with an embedding of the mask sequence itself, $e_m$, for a final combined embedding, $e_f$. This embedding, $e_f$, is treated as the standard input for the rest of the network. A diagram describing this process is shown in Figure \[fig:embedding\]. $$\begin{aligned}
e_j =&\: (1 - m) * e_c + m * e_p \\
e_f =&\: e_m + e_j\end{aligned}$$
Stacked Multi-scale Residual Convolution
----------------------------------------
In the next sections, we describe the network architecture for transforming the mixed representation into a spectrogram and then an audio waveform. A full system diagram of our neural network architecture can be seen in Figure \[fig:network\].
![[]{data-label="fig:network"}](network_diagram_cropped.pdf){width="8.0cm" height="5.5cm"}
The final embedding, $e_f$, from the previous section is used as input to a stacked multi-scale residual convolutional subnetwork (SMRC). The SMRC consists of several layers of multi-scale convolutions, where each multi-scale layer is in turn a concatenation of $1\times1$, $3\times3$, and $5\times5$ layers concatenated across the channel dimension. The layers are connected using residual bypass connections [@heres2016] and batch normalization [@Ioffe+Szegedy-2015] is used throughout. After the convolutional stage, the resulting activations are input to a bidirectional LSTM layer [@schuster1997bidirectional; @Hochreiter+Schmidhuber-1997]. This ends the *encoding* part of the network.
The Importance of Noisy Teacher Forcing
---------------------------------------
All audio information in the network passes through a multilayer pre-net with dropout [@srivastava2014dropout; @wang2017tacotron; @shen2017natural], in both training and evaluation. We found that using pre-net corrupted audio in every layer (including the attention layer) greatly improves the robustness of the model at evaluation, while corruption of linguistic information made the generated audio sound less natural.
Attention-based RNN Decoder
---------------------------
The encoding activations are attended using a Gaussian mixture (GM) attention [@Graves2013] driven by an LSTM network (conditioned on both the text and pre-net activations), with a softplus activation for the step of the Gaussian mean as opposed to the more typical exponential activation. We find that a softplus step substantially reduces instability during training. This is likely due to the relative length differences between linguistic input sequences and audio output.
Subsequent LSTM decode layers are conditioned on pre-net activations, attention activations, and the hidden state of the layer before, forming a series of skip connections [@Graves2013; @zhang2016architectural]. LSTM decode layers utilize cell dropout regularization [@semeniuta2016recurrent]. The final hidden state is projected to match the dimensionality of the audio frames and a mean squared error loss is calculated between the predicted and true next frames.
Truncated Backpropagation Through Time (TBPTT)
----------------------------------------------
The network uses truncated backpropagation through time (TBPTT) [@Hochreiter+Schmidhuber-1997] in the decoder, only processing a subsection of the relevant audio sequence while reusing the linguistic sequence until the end of the associated audio sequence. TBPTT training allows for a particular iterator packing scheme not available with full sequence training, which continuously packs subsequences resetting only the particular elements of the sequence minibatches (and the associated RNN state) when reaching the end of an audio sequence. A simplified diagram of this approach can be seen in Figure \[fig:minibatch\].
![[]{data-label="fig:minibatch"}](single_mb_cropped.pdf){width="7.0cm"}
![[]{data-label="fig:minibatch"}](tbptt_mb_cropped.pdf){width="8.0cm"}
Converting Features Into Waveforms
----------------------------------
After predicting a log mel spectrogram sequence from linguistic input, further transformation must be applied to get an audible waveform. A variety of methods have been proposed for this inversion such as the Griffin-Lim transform [@griffin1984signal], and modified variants [@slaney1994auditory]. Neural decoders such as the conditional WaveNet decoder [@vandenoordwavenet2016; @shen2017natural] or conditional SampleRNN [@mehri2016samplernn] can also act effectively as an inversion procedure to transform log mel spectrograms into audible waveforms.
Optimization methods also work for inversion and we utilize an L-BFGS based routine. This process optimizes randomly initialized parameters (representing the desired waveform) passed through a log mel spectrogram transform, to make the transform of these parameters match a given log mel spectrogram [@Thome-personal-communication].
This results in a set of parameters (which we now treat as a fixed waveform) with a lossy transform that closely matches the given log mel spectrogram. The overall procedure closely resembles effective techniques in other generative modeling settings [@gatys2016image].
*Inversion Method* *Seconds per Sample* *STRTF*
---------------------- ---------------------- --------- --
100 L-BFGS 1.49E-4 3.285
1000 L-BFGS 1.32E-3 29.08
Modified Griffin-Lim 2.81E-4 6.206
100 L + GL 4.11E-4 9.063
1000 L + GL 1.6E-3 35.28
WaveNet 7.9E-3 174.7
100 L + GL + WN 8.3E-3 183.7
1000 L + GL + WN 9.5E-3 210.0
: []{data-label="table:compute"}
Inversion Pipeline
------------------
This work uses a combined pipeline of L-BFGS based inversion, followed by modified Griffin-Lim for waveform estimation [@slaney1994auditory]. The resulting waveform is of moderate quality, and usable for certain applications including speech recognition. Using either L-BFGS or Griffin-Lim in isolation did not yield usable audio in our experiments. We also found that this ordering (L-BFGS then Griffin-Lim) in our two-stage pipeline works much better than the reverse setting.
To achieve the highest quality output, we then use the moderate quality waveform output from the two stage pipeline, converted back to log mel spectrograms, for conditional WaveNet synthesis. Though other work [@shen2017natural] clearly shows that log mel spectrogram conditioned WaveNet can be used directly, we find the two stage pipeline allows for quicker quality checking during development, as well as slightly higher quality in the resulting audio after WaveNet sampling. We suspect this is because the pre-trained WaveNet [@yamamoto18wavenet] that we use is trained on ground-truth log mel spectrograms rather than predicted outputs. The synthesis speed of each setting is shown in Table \[table:compute\].
Related Work
============
Representation mixing is closely related to the “mixed-character-and-phoneme” setting described in Deep Voice 3 [@ping2018deep], with the primary difference being our addition of the mask embedding $e_m$. We found utilizing a mask embedding alongside the character and phoneme embeddings further improved quality, and was an important piece in the text portion of the network. The focused user study in this paper also highlights the advantages of this kind of mixing independent of high-level architecture choices, since our larger system is markedly different from that used in Deep Voice 3.
The SMRC subnetwork is closely related to the CBHG encoder [@wang2017tacotron], differing in the number and size of convolutional scales used, no max-pooling, use of residual connections, and multiple stacks of multi-scale layers. We use a Gaussian mixture attention, first described by Graves [@Graves2013] and used in several speech related papers [@sotelo2017char2wav; @skerry2018towards; @taigman2017loop], though we utilize *softplus* activation for the mean-step parameter finding that it improved stability in early training.
Our decoder LSTMs utilize cell dropout [@semeniuta2016recurrent] as seen in prior work [@ha2016hypernetworks; @ha2017neural]. Unlike Tacotron [@wang2017tacotron], we do not use multi-step prediction, convolutional layers in the mel-decoding stage, or Griffin-Lim inside the learning pathway. Audio processing closely resembles Tacotron 2 [@shen2017natural] overall, though we precede conditional WaveNet with a computationally efficient two stage L-BFGS and Griffin-Lim inference pipeline for quality control, and as an alternative to neural inversion.
Experiments {#sec:experiments}
===========
The model is trained on LJSpeech [@ljspeech17], a curated subset of the LibriVox corpus [@librivox10]. LJSpeech consists of $13,100$ audio files (comprising a total time of approximately $24$ hours) of read English speech, spoken by Linda Johnson. The content of these recordings is drawn from scientific, instructional, and political texts published between $1893$ and $1964$. The recordings are stored as $22.05$ kHz, $16$ bit WAV files, though these are conversions from the original $128$ kbps MP3 format.
Character level information is extracted from audio transcriptions after a normalization and cleaning pipeline. This step includes converting all text to lowercase along with expanding acronyms and numerical values to their spelled-out equivalents. Phonemes are then extracted from the paired audio and character samples, using a forced alignment tool such as Gentle [@gentle17]. This results in character and phoneme information aligned along word level boundaries, so that randomly *mixed* linguistic sequences can be sampled repeatedly throughout training. When using representation mixing for training, we choose between characters and phonemes with probability $.5$ for each word in all experiments.
Audio Processing
------------------ --
Embeddings
SMRC
Enc Bidir LSTM
Pre-net
Attention LSTM
Decoder LSTMs
Optimizer
Other
: []{data-label="table:hyperparameters"}
Log Mel Spectrogram Inversion Experiments
-----------------------------------------
We use a high quality implementation of WaveNet, including a pre-trained model directly from Yamamoto et. al. [@yamamoto18wavenet]. This model was also trained on LJSpeech, allowing us to directly use it as a neural inverse to the log mel spectrograms predicted by the attention-based RNN model, or as an inverse to log mel spectrograms extracted from the network predictions after further processing. Ultimately combining the two stage L-BFGS and modified Griffin-Lim pipeline with a final conditional WaveNet sampling pass demonstrated the best quality, and is what we used for the user study shown in Table \[table:study\].
Preference Testing
------------------
Given the overall system architecture, we pose three primary questions (referenced in column Q in Table \[table:study\]) as a user study:
1\) Does representation mixing (RM) improve character-based inference over a baseline which was trained only on characters?
2\) Does RM improve phoneme with character backoff for unknown word (PWCB) inference over a baseline trained on fixed PWCB? 3) Does PWCB inference improve over character-based inference in a model trained with representation mixing?
To answer these questions, we pose each as a preference test by presenting study participants with $20$ paired choices. Each user was instructed to choose the sample they preferred [^1] . The $20$ tests were chosen randomly in a pool of $123$ possible tests covering all three question types, from $41$ possible sentences. Presentation order was randomized for each question and every user. The overall study consisted of $22$ users and $429$ responses across all categories (some users didn’t select any preference for some questions).
*Q* ***Model A*** *Model B* *C. A* *Total* *% A*
----- --------------- ----------- -------- --------- -------
1 [RM (char)]{} Char 101 137 73.7%
2 [RM (PWCB)]{} PWCB 106 145 73.1%
3 [RM (PWCB)]{} RM (char) 86 147 58.5%
: []{data-label="table:study"}
We see that users clearly prefer models trained with representation mixing, even when using identical information for inference. This highlights the data augmentation aspects of representation mixing, as regardless of information type representation mixing (RM) gives clearly preferable results compared to static representations (Char, PWCB). The preference of representation mixing over static PWCB also means that introducing phoneme and character information without mixing is less beneficial than full representation mixing.
It is also clear that for representation mixing models, pronunciation information (PWCB) at inference gives preferable samples compared to character information. This is not surprising, but further reinforces the importance of using pronunciation information where possible. Representation mixing enables choice in input format, allowing the possibility to use many different inference representations for a given sentence with a single trained model.
Conclusion {#sec:print}
==========
This paper shows the benefit of *representation mixing*, a simple method for combining multiple types of linguistic information for TTS synthesis. Representation mixing enables inference conditioning to be controlled independently of training representation, and also results in improved quality over strong character and phoneme baselines trained on a publically available audiobook corpus.
[^1]: [ <https://s3.amazonaws.com/representation-mixing-site/index.html>]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $n$ be a positive integer. The aim of this paper is to study two local moves $V(n)$ and $V^{n}$ on welded links, which are generalizations of the crossing virtualization. We show that the $V(n)$-move is an unknotting operation on welded knots for any $n$, and give a classification of welded links up to $V(n)$-moves. On the other hand, we give a necessary condition for which two welded links are equivalent up to $V^{n}$-moves. This leads to show that the $V^{n}$-move is not an unknotting operation on welded knots except for $n=1$. We also discuss relations among $V^{n}$-moves, associated core groups and the multiplexing of crossings.'
address:
- 'Institute for Mathematics and Computer Science, Tsuda University,2-1-1 Tsuda-Machi, Kodaira, Tokyo, 187-8577, Japan'
- 'Faculty of Education and Integrated Arts and Sciences, Waseda University, 1-6-1 Nishi-Waseda, Shinjuku-ku, Tokyo, 169-8050, Japan'
- 'Faculty of Commerce, Waseda University, 1-6-1 Nishi-Waseda, Shinjuku-ku, Tokyo, 169-8050, Japan'
author:
- 'Haruko A. Miyazawa'
- Kodai Wada
- Akira Yasuhara
title: Generalized virtualization on welded links
---
[^1]
[^2]
[^3]
Introduction
============
A $\mu$-component [*virtual link diagram*]{} is the image of an immersion of ordered and oriented $\mu$ circles in the plane, whose transverse double points admit not only [*classical crossings*]{} but also [*virtual crossings*]{} illustrated in Figure \[xing\].[^4] We emphasize that a virtual link diagram is always ordered and oriented unless otherwise specified.
[xing.eps]{} (-19,-15)[classical crossing]{} (78,-15)[virtual crossing]{}
A [*virtual link*]{} is an equivalence class of virtual link diagrams under [*generalized Reidemeister moves*]{}, which consist of Reidemeister moves R1–R3 and virtual moves VR1–VR4 illustrated in Figure \[GRmoves\] [@Kauffman]. In virtual context, there are two forbidden local moves OC and UC (meaning [*over-crossings*]{} and [*under-crossings commute*]{}, respectively) illustrated in Figure \[Fmoves\]. The extension of the generalized Reidemeister moves which also allows the OC-move is called [*welded Reidemeister moves*]{}, and a sequence of welded Reidemeister moves is called a [*welded isotopy*]{}. A [*welded link*]{} is an equivalence class of virtual link diagrams under welded isotopy [@FRR].
[GRmoves.eps]{} (24.5,24.5)[VR1]{} (90.5,24.5)[VR2]{} (186.5,24.5)[VR3]{} (287,24.5)[VR4]{} (47.5,88)[R1]{} (75,88)[R1]{} (169.2,88)[R2]{} (272.5,88)[R3]{}
[Fmoves.eps]{} (43.4,20.5)[OC]{} (183.5,20.5)[UC]{}
A virtual link diagram is [*classical*]{} if it has no virtual crossings, and a welded link is [*classical*]{} if it has a classical link diagram. M. Goussarov, M. Polyak and O. Viro [@GPV] essentially proved that welded isotopic classical link diagrams can be related by Reidemeister moves R1–R3. Therefore, welded links can be viewed as a natural extension of classical links. We remark that any virtual knot diagram can be unknotted by UC-moves and welded Reidemeister moves [@GPV; @Kanenobu; @N]. This result is one reason why the UC-move is still forbidden in welded context.
In classical knot theory, local moves have played important roles and hence have been studied widely; see for example [@M; @MN; @HNT; @A; @MiyaY; @DP02; @DP04]. Recently, some “classical” local moves, which exchange classical tangle diagrams, have been studied for welded knots and links [@ABMW-JMSJ; @ABMW; @S; @NNSY]. In this paper, we will study “non-classical” local moves for welded links. A typical non-classical local move is the crossing virtualization. The [*crossing virtualization*]{} V is a local move on a virtual link diagram replacing a classical crossing with a virtual one; see the left-hand side of Figure \[virtualization\]. We remark that any virtual link diagram can be deformed into a diagram of the trivial link by applying the crossing virtualization repeatedly. The crossing virtualization is equivalent to the local move illustrated in the right-hand side of Figure \[virtualization\]. Here, two local moves are [*equivalent*]{} if each move is realized by a sequence of the other move and welded Reidemeister moves.
[virtualize.eps]{} (41,20)[V]{} (104,20)[V]{}
Let $n$ be a positive integer. The aim of this paper is to study two oriented local moves $V(n)$ and $V^{n}$ illustrated in the upper and lower sides of Figure \[GV\], respectively. They are considered as generalizations of the crossing virtualization. In fact, both $V(1)$- and $V^{1}$-moves are equivalent to the crossing virtualization. Note that if $n$ is even then the $V(n)$-move may change the number of components. Two welded links are [*$V(n)$-equivalent*]{} (resp. [*$V^{n}$-equivalent*]{}) if their diagrams are related by $V(n)$-moves (resp. $V^{n}$-moves) and welded Reidemeister moves.
[GV.eps]{} (83,79)[$V(n)$]{} (152,59)[[$1$]{}]{} (179.5,59)[[$2$]{}]{} (237,59)[[$n$]{}]{} (88,20)[$V^{n}$]{} (152,2)[[$1$]{}]{} (208,2)[[$2$]{}]{} (293,2)[[$n$]{}]{}
We have that the $V(n)$-move is an unknotting operation on welded knots for any positive integer $n$ because a UC-move is realized by a sequence of the $V(n)$-move and welded Reidemeister moves (Proposition \[prop-UC\]). Moreover, we give a classification of welded links up to $V(n)$-equivalence in the sense of Theorems \[th-even\] and \[th-odd\].
\[th-even\] Let $n$ be an even number. Any welded link is $V(n)$-equivalent to the unknot.
Let $D$ be a virtual link diagram. For any $i,j$ $(i\neq j)$, let $\lambda_{ij}(D)$ denote the sum of the signs of all classical crossings of $D$ whose overpasses and underpasses belong to the $i$th and $j$th component, respectively. The integer $\lambda_{ij}(D)$ is a welded link invariant and also preserved by UC-moves. For a welded link $L$, the [*ordered linking number $\lambda_{ij}(L)$*]{} is defined to be $\lambda_{ij}(D)$ for a diagram $D$ of $L$.
It is not hard to see that if $n$ is odd then the modulo-$n$ reduction of $\lambda_{ij}(L)+\lambda_{ji}(L)$ is preserved by $V(n)$-moves. Using these invariants we have the following.
\[th-odd\] Let $n$ be an odd number. Two $\mu$-component welded links $L$ and $L'$ are $V(n)$-equivalent if and only if $\lambda_{ij}(L)+\lambda_{ji}(L)\equiv \lambda_{ij}(L')+\lambda_{ji}(L')$ $\pmod{n}$ for any $i,j$ $(1\leq i<j\leq\mu)$.
On the other hand, $V^{n}$-moves preserve the modulo-$n$ reduction of $\lambda_{ij}(L)$ for any positive integer $n$. However, these invariants are not strong enough to classify welded links up to $V^{n}$-equivalence because the $V^{n}$-move is not an unknotting operation on welded knots except for $n=1$ (Proposition \[prop-trefoil\]). Considering the UC-move, which is an unknotting operation for welded knots, we have the following.
\[th-V\^n\] Let $n$ be a positive integer. Two $\mu$-component welded links $L$ and $L'$ are $(V^{n}+{\rm UC})$-equivalent if and only if $\lambda_{ij}(L)\equiv\lambda_{ij}(L')$ $\pmod{n}$ for any $i,j$ $(1\leq i\neq j\leq\mu)$.
Here, two welded links are [*$(V^{n}+{\rm UC})$-equivalent*]{} if their diagrams are related by $V^{n}$-moves, UC-moves and welded Reidemeister moves.
Theorem \[th-V\^n\] easily follows from the classification of welded links up to UC-moves, given in [@O Theorem 8] and [@Nasybullov Theorem 4.7]. In this paper, we will prove Theorem \[th-V\^n\] without using their classification. Our proof of the theorem contains a simple proof for their classification.
We also discuss relations among [*unoriented*]{} $V^{n}$-moves, associated core groups and the multiplexing of crossings. The [*associated core group*]{} is known as an unoriented classical link invariant [@J; @Kelly; @FR; @W]. This group is naturally extended to an unoriented welded link invariant, and furthermore, it is preserved by unoriented $V^{n}$-moves for any even number $n$ (Proposition \[prop-V2\]). In [@MWY17] the authors introduced the notion of [*multiplexing*]{} of crossings for an unoriented $\mu$-component welded link $L$, which yields a new unoriented welded link $L(m_{1},\ldots,m_{\mu})$ associated with a $\mu$-tuple $(m_{1},\ldots,m_{\mu})$ of integers. For any $\mu$-tuple $(m_{1},\ldots,m_{\mu})$ of even numbers, $L(m_{1},\ldots,m_{\mu})$ is deformed into the $\mu$-component trivial link by unoriented $V^{2}$-moves (Proposition \[th-V2\]). As a consequence, we have that there are infinitely many nontrivial welded knots whose associated core groups are isomorphic to that of the unknot (Theorem \[prop-infinitely\]).
Arrow calculus {#sec-arrow}
==============
To show Theorems \[th-even\], \[th-odd\] and \[th-V\^n\], we will use [*arrow calculus*]{} introduced by J.-B. Meilhan and the third author in [@MY]. In this section, we will briefly recall the basic definitions of arrow calculus from [@MY]. We only need the notion of [w]{}-arrow, and refer the reader to [@MY] for more details of arrow calculus.
Let $D$ be a virtual link diagram. A [*[w]{}-arrow $\gamma$*]{} for $D$ is an oriented arc immersed in the plane of the diagram such that;
1. the endpoints of $\gamma$ are contained in $D\setminus\{\text{crossings of $D$}\}$,
2. all singularities of $\gamma$ are virtual crossings,
3. all singularities between $D$ and $\gamma$ are virtual crossings, and
4. $\gamma$ has a number (possibly zero) of decorations $\bullet$ on the interior of $\gamma$, called [*twists*]{}, which are disjoint from all crossings.
The initial and terminal points of $\gamma$ are called the [*tail*]{} and the [*head*]{}, respectively. For a union of [w]{}-arrows for $D$, all crossings among [w]{}-arrows are assumed to be virtual.
Hereafter, diagrams are drawn with bold lines while [w]{}-arrows are drawn with thin lines.
Let $A$ be a union of [w]{}-arrows for $D$. Now we define [*surgery along $A$*]{} on $D$ which yields a new virtual link diagram, denoted by $D_{A}$, as follows. Suppose that there exists a disk in the plane which intersects $D\cup A$ as illustrated in Figure \[surgery\]. Then the figure indicates the result of surgery along a [w]{}-arrow of $A$ on $D$. We emphasize that the surgery move depends on the orientation of the strand of $D$ containing the tail of the [w]{}-arrow.
[surgery.eps]{} (5,-12)[$D\cup A$]{} (85,-12)[$D_{A}$]{} (153,-12)[$D\cup A$]{} (233,-12)[$D_{A}$]{}
If a [w]{}-arrow of $A$ intersects a (possibly the same) [w]{}-arrow (resp. $D$), then the result of surgery is essentially the same as above but each intersection introduces virtual crossings illustrated in the left-hand side (resp. center) of Figure \[surgery2\]. Moreover, if a [w]{}-arrow of $A$ has some twists, then each twist is converted to a half-twist whose crossing is virtual; see the right-hand side of Figure \[surgery2\].
[surgery2.eps]{} (-5,21)[$A$]{} (12,-13)[$A$]{} (70,-13)[$D_{A}$]{} (119,21)[$A$]{} (136,-13)[$D$]{} (193,-13)[$D_{A}$]{} (261,-13)[$A$]{} (318,-13)[$D_{A}$]{}
An [*arrow presentation*]{} for a virtual link diagram $D$ is a pair $(T,A)$ of a virtual link diagram $T$ without classical crossings and a union $A$ of [w]{}-arrows for $T$ such that $T_{A}$ is welded isotopic to $D$. Every virtual link diagram has an arrow presentation because any classical crossing can be replaced by a virtual one with a [w]{}-arrow; see Figure \[Aprst\].
[Aprst.eps]{} (41,22)[VR2]{}
Two arrow presentations are [*equivalent*]{} if the surgeries yield welded isotopic virtual link diagrams. [*Arrow moves*]{} consist of virtual moves VR1–VR3 involving [w]{}-arrows and/or strands of $D$ and the local moves AR1–AR10 on arrow presentations illustrated in Figure \[Amoves\]. Here, each vertical strand in AR1–AR3 is either a strand of $D$ or a [w]{}-arrow, and the symbol $\circ$ on a [w]{}-arrow in AR8 and AR10 denotes that the [w]{}-arrow may or may not contain a twist. Two arrow presentations are equivalent if and only if they are related by arrow moves [@MY Theorem 4.5].
[Amoves.eps]{} (37,212)[AR1]{} (152,212)[AR2]{} (267,212)[AR3]{} (37,148.5)[AR4]{} (152,148.5)[AR5]{} (267,148.5)[AR6]{} (37,84.5)[AR7]{} (152,84.5)[AR8]{} (267,84.5)[AR9]{} (58,21)[AR10]{} (241.5,21)[AR10]{}
In the rest of this section, we will introduce several local moves on arrow presentations. We first consider two [*allowable moves*]{} AR11 and AR12 as illustrated in Figures \[AR11\] and \[AR12\], respectively. Each of the moves is realized by a sequence of arrow moves. Figure \[pf-AR112\] shows that the left-hand side moves in Figures \[AR11\] and \[AR12\] are realized by arrow moves, where $\overset{\text{AR}}{\sim}$ in the figure denotes a sequence of arrow moves. The other cases are similarly shown.
[AR11.eps]{} (58,20.5)[AR11]{} (230,20.5)[AR11]{}
[AR12.eps]{} (58,20.5)[AR12]{} (230,20.5)[AR12]{}
[pf-AR112.eps]{} (-30,75)[[AR11 :]{}]{} (-30,12)[[AR12 :]{}]{} (51,81)[[AR]{}]{} (118,81)[[AR10]{}]{} (192.5,81)[[AR]{}]{} (261.5,81)[[AR]{}]{}
(51,19)[[AR]{}]{} (118,19)[[AR10]{}]{} (192.5,19)[[AR]{}]{} (261.5,19)[[AR]{}]{}
The [*heads exchange move*]{}[^5] is a local move on an arrow presentation exchanging positions of consecutive two heads of [w]{}-arrows; see Figure \[Hexch\]. While there are several kinds of heads exchange moves depending on the orientation of the strand containing the tail and existence or nonexistence of a twist for a [w]{}-arrow, we have the following.
[Hexch.eps]{}
\[lem-Hexch\] A heads exchange move is realized by a sequence of the [H]{}-move illustrated in Figure $\ref{H}$ and arrow moves.
[H.eps]{} (67,20)[H]{}
We demonstrate that two of the heads exchange moves are realized by sequences of the H-move and arrow moves. The upper side of Figure \[pf-Hexch\] indicates the proof when the orientation of the strand containing the tail of a single [w]{}-arrow is opposite to that of the H-move. The lower side of Figure \[pf-Hexch\] indicates the proof for the case where a [w]{}-arrow has a twist. It is not hard to show the other cases.
[pf-Hexch.eps]{} (59.5,77)[AR]{} (145.5,76)[H]{} (224.5,77)[AR]{} (59.5,19.5)[AR]{} (145,19)[H]{} (224.5,19.5)[AR]{}
The [*head-tail exchange move*]{}[^6] is a local move on an arrow presentation exchanging positions of consecutive a head and a tail of [w]{}-arrows; see Figure \[HTexch\].
[HTexch.eps]{}
\[lem-HTexch\] A head-tail exchange move is realized by a sequence of the heads exchange move and arrow moves.
See Figure \[pf-HTexch\].
[pf-HTexch.eps]{} (52,22)[AR]{} (125,22)[AR10]{} (204,22)[[heads]{}]{} (198,8)[[exchange]{}]{} (286,22)[AR]{}
Three kinds of moves, AR7, heads exchange and head-tail exchange moves, are called [*ends exchange moves*]{}. From Sublemmas \[lem-Hexch\] and \[lem-HTexch\] we have the following.
\[lem-ends\] An ends exchange move is realized by a sequence of the [H]{}-move and arrow moves.
$V(n)$-moves and UC-moves {#sec-moves}
=========================
In this section, we will show that the $V(n)$-move is an unknotting operation on welded knots. We start with the following lemma concerning the UC-move.
\[lem-UC\] An arrow presentation for a [UC]{}-move is realized by a sequence of the heads exchange move and arrow moves. Conversely, surgery along a heads exchange move is realized by a sequence of the [UC]{}-move and welded Reidemeister moves.
Figure \[pf-UC\] shows that an arrow presentation for a UC-move is realized by a sequence of the heads exchange move and arrow moves. In the figure, we choose certain orientations of two strands at the virtual crossing. The other cases are similarly shown.
[pf-UC.eps]{} (134,91)[AR]{} (214.5,91)[AR]{} (267,49)[AR]{} (212,23)[[heads]{}]{} (206,9)[[exchange]{}]{} (134,23)[AR]{}
Conversely, Figure \[UC-Hexch\] shows that surgery along an H-move is realized by a sequence of the UC-move and welded Reidemeister moves, where $\overset{\text{w}}{\sim}$ in the figure denotes a welded isotopy. This and Sublemma \[lem-Hexch\] complete the proof.
[UC-Hexch.eps]{} (137,101)[w]{} (219,101)[w]{} (137,25.5)[w]{} (218,25.5)[w]{} (265,57)[UC]{}
We define the [*$A(n)$-move*]{} as a local move on an arrow presentation depending on the parity of $n$. The $A(n)$-move is illustrated in Figure \[A(n)odd\] (resp. Figure \[A(n)even\]) when $n$ is odd (resp. even).
[A\_n-odd.eps]{} (44,39)[$A(n)$]{} (106,5)[[$1$]{}]{} (106,14)[[$2$]{}]{} (106,57)[[$n$]{}]{}
(222,39)[$A(n)$]{} (285.5,5)[[$1$]{}]{} (285.5,14)[[$2$]{}]{} (285.5,57)[[$n$]{}]{}
[A\_n-even.eps]{} (44,39)[$A(n)$]{} (106,14)[[$1$]{}]{} (106,23)[[$2$]{}]{} (106,57)[[$n$]{}]{}
(222,39)[$A(n)$]{} (285.5,14)[[$1$]{}]{} (285.5,23)[[$2$]{}]{} (285.5,57)[[$n$]{}]{}
\[lem-HvsA(n)\] Let $n$ be a positive integer. An ends exchange move is realized by a sequence of the $A(n)$-move and arrow moves.
By Lemma \[lem-ends\], it suffices to show that an H-move is realized by a sequence of the $A(n)$-move and arrow moves for any $n$. The upper (resp. lower) side of Figure \[pf-propUC\] indicates the proof for the case where $n$ is odd (resp. even), while the figure describes only when $n=3$ (resp. $n=2$).
[pf-propUC.eps]{} (68,184)[$A(n)$]{} (154,184)[AR11,12]{} (256,184)[$A(n)$]{} (67,104)[$A(n)$]{} (168,103)[AR]{} (266,89)[AR11,12]{} (271,36)[AR]{} (166,27)[$A(n)$]{} (69.5,26)[AR]{}
\[lem-A(n)V(n)\] An arrow presentation for a $V(n)$-move is realized by a sequence of the $A(n)$-move and arrow moves. Conversely, surgery along an $A(n)$-move is realized by a sequence of the $V(n)$-move and welded Reidemeister moves.
It is not hard to see that the right-hand side move in Figure \[A(n)odd\] (resp. Figure \[A(n)even\]) is realized by a sequence of the left-hand side move in Figure \[A(n)odd\] (resp. Figure \[A(n)even\]) and arrow moves. See, for example, Figure \[pf-A(n)odd\] in the case when $n$ is odd. Furthermore, an arrow presentation for a $V(n)$-move is realized by a sequence of the left-hand side move in Figure \[A(n)odd\] (resp. Figure \[A(n)even\]) and arrow moves when $n$ is odd (resp. even). Conversely, it is obvious that surgery along the left-hand side move in Figure \[A(n)odd\] (resp. Figure \[A(n)even\]) is realized by a sequence of the $V(n)$-move and welded Reidemeister moves when $n$ is odd (resp. even). Therefore, we have the conclusion.
[pf-A\_n-odd.eps]{} (49,45)[AR]{} (118,45)[$A(n)$]{} (192,45)[AR]{} (40,20)[[$1$]{}]{} (40,29)[[$2$]{}]{} (40,56)[[$n$]{}]{} (112,20)[[$1$]{}]{} (112,29)[[$2$]{}]{} (112,56)[[$n$]{}]{}
As a consequence of Lemmas \[lem-UC\], \[lem-HvsA(n)\] and \[lem-A(n)V(n)\], we have the following.
\[prop-UC\] Let $n$ be a positive integer. A [UC]{}-move is realized by a sequence of the $V(n)$-move and welded Reidemeister moves. Hence, the $V(n)$-move is an unknotting operation for welded knots.
Here, we define the [*$A^{n}$-move*]{} as a local move on an arrow presentation illustrated in Figure \[A\^n\].
[An.eps]{} (49,29)[$A^{n}$]{} (106,5)[[$1$]{}]{} (106,14)[[$2$]{}]{} (106,40)[[$n$]{}]{}
(227,29)[$A^{n}$]{} (285,5)[[$1$]{}]{} (285,14)[[$2$]{}]{} (285,40)[[$n$]{}]{}
\[lem-realizingA\^[n]{}\] Let $n$ be an odd number. An $A^{n}$-move is realized by a sequence of the $A(n)$-move and arrow moves.
We consider the [*head-tail reversal move*]{} illustrated in Figure \[HTrev\], which is realized by a sequence of the move AR9 and $A(n)$-move. (Figure \[pf-HTrev\] shows that one of the head-tail reversal moves is realized by a sequence of the move AR9 and $A(n)$-move. The other case is similarly shown.) Combining the head-tail reversal moves with an $A(n)$-move, we can realize an $A^{n}$-move.
[HTrev.eps]{}
[pf-HTrev.eps]{} (39.5,83)[AR9]{} (104,84)[$A(n)$]{} (98,71)[[$n$]{}]{} (98,48)[[$2$]{}]{} (98,41)[[$1$]{}]{} (98,34.5)[[$n$]{}]{} (98,27)[[$n-1$]{}]{} (98,5)[[$1$]{}]{} (170.5,83)[AR9]{} (163,113)[[$n$]{}]{} (163,91.5)[[$2$]{}]{} (163,84)[[$1$]{}]{} (163,71)[[$n$]{}]{} (163,48)[[$2$]{}]{} (163,41)[[$1$]{}]{}
By arguments similar to those in the proof of Lemma \[lem-A(n)V(n)\] we have the following.
\[lem-A\^nV\^n\] An arrow presentation for a $V^{n}$-move is realized by a sequence of the $A^{n}$-move and arrow moves. Conversely, surgery along an $A^{n}$-move is realized by a sequence of the $V^{n}$-move and welded Reidemeister moves.
Proofs of theorems {#sec-proofs}
==================
In this section, we will give the proofs of Theorems \[th-even\], \[th-odd\] and \[th-V\^n\].
If $n$ is even then any welded link can be deformed into some welded knot by $V(n)$-moves, since $V(n)$-moves can change the number of components of the welded link. Therefore, Theorem \[th-even\] follows from Proposition \[prop-UC\].
Let $\mathbf{1}$ be the ordered oriented $\mu$-component trivial string link diagram without crossings such that all strands are oriented upwards. For an integer $a$, let $(\mathbf{1},H_{ij}(a))$ denote the arrow presentation of Figure \[sigma\], that is, $H_{ij}(a)$ consists of $|a|$ horizontal [w]{}-arrows whose tails (resp. heads) are attached to the $i$th (resp. $j$th) strand of $\mathbf{1}$ $(1\leq i<j\leq\mu)$ such that each [w]{}-arrow has exactly one twist if $a\geq 0$, and no twist otherwise. Note that, for arrow presentations $(\mathbf{1},H_{ij}(a))$ and $(\mathbf{1},H_{kl}(a'))$, the stacking products $(\mathbf{1},H_{ij}(a))*(\mathbf{1},H_{kl}(a'))$ and $(\mathbf{1},H_{kl}(a'))*(\mathbf{1},H_{ij}(a))$ are related by ends exchange moves and arrow moves, hence, by $A(n)$-moves and arrow moves. Here, the [*stacking product $(\mathbf{1},A)*(\mathbf{1},B)$*]{} of arrow presentations $(\mathbf{1},A)$ and $(\mathbf{1},B)$ is the arrow presentation corresponding to the diagram $\mathbf{1}_{A}*\mathbf{1}_{B}$. Let $\prod_{1\leq i<j\leq \mu}(\mathbf{1},H_{ij}(a_{ij}))$ denote the stacking products of $(\mathbf{1},H_{ij}(a_{ij}))$ for integers $a_{ij}$ $(1\leq i<j\leq\mu)$. We remark that the ordered linking numbers $\lambda_{ij}$ and $\lambda_{ji}$ of the closure of the string link diagram $\prod_{1\leq i<j\leq \mu}\mathbf{1}_{H_{ij}(a_{ij})}$ are equal to $a_{ij}$ and $0$, respectively $(1\leq i<j\leq\mu)$.
[sigma.eps]{} (37,-23)[$(a\geq 0)$]{} (218,-23)[$(a<0)$]{} (-3,-9)[[$1$st]{}]{} (31,-9)[[$i$th]{}]{} (63,-9)[[$j$th]{}]{} (95,-9)[[$\mu$th]{}]{} (71,22)[[$1$]{}]{} (71,31)[[$2$]{}]{} (71,56)[[$a$]{}]{}
(177,-9)[[$1$st]{}]{} (212,-9)[[$i$th]{}]{} (244,-9)[[$j$th]{}]{} (275,-9)[[$\mu$th]{}]{} (251,22)[[$1$]{}]{} (251,31)[[$2$]{}]{} (251,56)[[$|a|$]{}]{}
\[lem-A(n)\] Let $n$ be an odd number. For any $\mu$-component virtual string link diagram $D$, there are integers $a_{ij}$ with $0\leq a_{ij}<n$ $(1\leq i<j\leq\mu)$ such that an arrow presentation for $D$ and $\prod_{1\leq i<j\leq \mu}(\mathbf{1},H_{ij}(a_{ij}))$ are related by $A(n)$-moves and arrow moves.
Let $(\mathbf{1},\bigcup_{1\leq i,j\leq\mu}W_{ij})$ be an arrow presentation for a $\mu$-component virtual string link diagram, where $W_{ij}$ is a set of [w]{}-arrows for $\mathbf{1}$ whose tails (resp. heads) are attached to the $i$th (resp. $j$th) strand ($1\leq i,j\leq\mu$, possibly $i=j$). We show that $(\mathbf{1},\bigcup_{1\leq i,j\leq\mu}W_{ij})$ can be deformed into the desired form by $A(n)$-moves and arrow moves (including ends exchange, head-tail reversal moves and $A^{n}$-moves). First, the ends of each [w]{}-arrow in $W_{ii}$ $(1\leq i\leq\mu)$ can be moved into position to be removed by a single AR8. Hence, all [w]{}-arrows in $W_{ii}$ are removed for any $i$. Next, $(\mathbf{1},\bigcup_{1\leq i\neq j\leq\mu}W_{ij})$ can be deformed into $\prod_{1\leq i<j\leq\mu}(\mathbf{1},H_{ij}(a_{ij}))$ for some integers $a_{ij}$ by combining head-tail reversal, ends exchange moves and AR9. Finally, we obtain the desired form by performing $A^{n}$-moves and AR9.
It suffices to show the “if” part. Let $D$ and $D'$ be virtual link diagrams of $L$ and $L'$, respectively. For any virtual link diagram, there exists a virtual string link diagram whose closure is welded isotopic to the virtual link diagram. Hence, by Lemma \[lem-A(n)\], two arrow presentations $(T,A)$ for $D$ and $(T',A')$ for $D'$ can be related to the closures of $\prod_{1\leq i<j\leq \mu}(\mathbf{1},H_{ij}(a_{ij}))$ and $\prod_{1\leq i<j\leq \mu}(\mathbf{1},H_{ij}(a'_{ij}))$, respectively, for some non-negative integers $a_{ij},a'_{ij}$ $(<n)$, by $A(n)$-moves and arrow moves. Then, for any $i,j$ $(1\leq i<j\leq\mu)$, we have $$a_{ij}\equiv\lambda_{ij}(D)+\lambda_{ji}(D)\equiv\lambda_{ij}(D')+\lambda_{ji}(D')\equiv a'_{ij}
\pmod{n}.$$ Since $0\leq a_{ij},a'_{ij}<n$ it follows that $a_{ij}=a'_{ij}$. Therefore, $(T, A)$ and $(T', A')$ are related by $A(n)$-moves and arrow moves. Consequently, $D(=T_{A})$ and $D'(=T'_{A'})$ are related by $V(n)$-moves and welded Reidemeister moves.
For an integer $b$, let $(\mathbf{1},\overline{H}_{ij}(b))$ denote the arrow presentation of Figure \[sigma2\], that is, $\overline{H}_{ij}(b)$ consists of $|b|$ horizontal [w]{}-arrows whose heads (resp. tails) are attached to the $i$th (resp. $j$th) strand of $\mathbf{1}$ $(1\leq i<j\leq\mu)$ such that each [w]{}-arrow has no twist if $b\geq 0$, and exactly one twist otherwise. We remark that, for integers $a_{ij}$ and $b_{ij}$, the numbers $\lambda_{ij}$ and $\lambda_{ji}$ of the closure of the string link diagram $\prod_{1\leq i<j\leq\mu}(\mathbf{1}_{H_{ij}(a_{ij})}*\mathbf{1}_{\overline{H}_{ij}(b_{ij})})$ are equal to $a_{ij}$ and $b_{ij}$, respectively $(1\leq i<j\leq\mu)$.
[sigma2.eps]{} (38,-23)[$(b\geq 0)$]{} (219,-23)[$(b<0)$]{} (-3,-9)[[$1$st]{}]{} (31,-9)[[$i$th]{}]{} (63,-9)[[$j$th]{}]{} (95,-9)[[$\mu$th]{}]{} (71,22)[[$1$]{}]{} (71,31)[[$2$]{}]{} (71,56)[[$b$]{}]{}
(177,-9)[[$1$st]{}]{} (212,-9)[[$i$th]{}]{} (244,-9)[[$j$th]{}]{} (275,-9)[[$\mu$th]{}]{} (251,22)[[$1$]{}]{} (251,31)[[$2$]{}]{} (251,56)[[$|b|$]{}]{}
\[lem-A\^n\] Let $n$ be a positive integer. For any $\mu$-component virtual string link diagram $D$, there are integers $a_{ij},b_{ij}$ with $0\leq a_{ij},b_{ij}<n$ $(1\leq i<j\leq\mu)$ such that an arrow presentation for $D$ and $\prod_{1\leq i<j\leq \mu}((\mathbf{1},H_{ij}(a_{ij}))*(\mathbf{1},\overline{H}_{ij}(b_{ij})))$ are related by $A^{n}$-moves, ends exchange moves and arrow moves.
The proof of the lemma above can be done by a similar way to the proof of Lemma \[lem-A(n)\]. Note that we are not permitted to use the head-tail reversal move. This is the reason why we need not only $H_{ij}(a)$ but also $\overline{H}_{ij}(b)$.
It suffices to show the “if” part. Let $D$ and $D'$ be virtual link diagrams of $L$ and $L'$, respectively. By Lemma \[lem-A\^n\], two arrow presentations $(T,A)$ for $D$ and $(T',A')$ for $D'$ can be related to the closures of $\prod_{1\leq i<j\leq \mu}((\mathbf{1},H_{ij}(a_{ij}))*(\mathbf{1},\overline{H}_{ij}(b_{ij})))$ and $\prod_{1\leq i<j\leq \mu}((\mathbf{1},H_{ij}(a'_{ij}))*(\mathbf{1},\overline{H}_{ij}(b'_{ij})))$, respectively, for some non-negative integers $a_{ij},b_{ji},a'_{ij},b'_{ji}$ $(<n)$ by $A^{n}$-moves, ends exchange moves and arrow moves. Then, for any $i,j$ $(1\leq i<j\leq\mu)$, we have $$a_{ij}\equiv\lambda_{ij}(D)\equiv\lambda_{ij}(D')\equiv a'_{ij}\pmod{n}$$ and $$b_{ij}\equiv\lambda_{ji}(D)\equiv\lambda_{ji}(D')\equiv b'_{ij}
\pmod{n}.$$ Since $0\leq a_{ij},b_{ij},a'_{ij}b'_{ij}<n$ it follows that $a_{ij}=a'_{ij}$ and $b_{ij}=b'_{ij}$. Therefore, $(T, A)$ and $(T', A')$ are related by $A^n$-moves, ends exchange moves and arrow moves. Lemmas \[lem-ends\], \[lem-UC\] and \[lem-A\^nV\^n\] imply that $D(=T_{A})$ and $D'(=T'_{A'})$ are related by $V^{n}$-moves, UC-moves and welded Reidemeister moves.
$V^{n}$-moves and UC-moves {#sec-UC}
==========================
As mentioned in Section 1, the $V^{n}$-move is not an unknotting operation except for $n=1$. To prove this, we use the Alexander polynomials which are obtained from the group of a welded link by using the Fox free derivative. Here, the [*group*]{} of a virtual link diagram is known to be a welded link invariant [@Kauffman Section 4], and hence (the elementary ideals in the sense of [@CF] and) the Alexander polynomials are naturally extended to welded link invariants. By a similar way to the proof of Theorem 1 in [@Kinoshita], we can show the following.
\[prop-Alex\] Let $n$ be a positive integer. If two welded links $L$ and $L'$ are $V^{n}$-equivalent, then for a non-negative integer $k$ and for the $k$th elementary ideals $E^{k}_{L}(t)$ and $E^{k}_{L'}(t)$ of $L$ and $L'$, respectively, we have $$E^{k}_{L}(t)\equiv E^{k}_{L'}(t)\mod{I(1-t^{n})},$$ where $I(1-t^{n})$ is the ideal generated by $1-t^{n}$ in $\mathbb{Z}[t^{\pm 1}]$. In particular, for the $($$1$-variable$)$ $k$th Alexander polynomials $\Delta_{L}^{k}(t)$ and $\Delta_{L'}^{k}(t)$ of $L$ and $L'$, respectively, we have $$\Delta_{L}^{k}(t)\equiv{\varepsilon}t^{r}\Delta_{L'}^{k}(t)\mod{I(1-t^{n})}$$ for some ${\varepsilon}\in\{\pm1\}$ and $r\in\mathbb{Z}$.
Let $D$ and $ D'$ be virtual link diagrams of $L$ and $L'$, respectively. It suffices to show that if $D$ and $D'$ are related by a single $V^{n}$-move then for their properly chosen Alexander matrices $A_{D}(t)$ and $A_{D'}(t)$, $$A_{D}(t)\equiv A_{D'}(t)\mod{I(1-t^{n})}.$$
Suppose that $D'$ is obtained from $D$ by a single R1 and a single $V^{n}$-move, and put labels $x_{1},x_{2}$ and $x_{3}$ on arcs of $D$ and $D'$ as illustrated in Figure \[pf-Alex\] and labels $x_{4},\ldots,x_{l}$ on the other arcs outside the figure.
[pf-Alex.eps]{} (33,-18)[$D$]{} (200,-18)[$D'$]{} (81,20.5)[R1,$V^{n}$]{} (154,2)[[$1$]{}]{} (210.5,2)[[$2$]{}]{} (293,2)[[$n$]{}]{} (-5,35)[$x_{1}$]{} (72,35)[$x_{2}$]{} (-5,-5.5)[$x_{3}$]{} (72,-5.5)[$x_{3}$]{} (107,35)[$x_{1}$]{} (107,-5.5)[$x_{3}$]{} (311,35)[$x_{2}$]{} (311,-5.5)[$x_{3}$]{}
Then, we obtain group presentations of the groups $G(D)$ and $G(D')$ of $D$ and $D'$, respectively, as follows: $$\begin{array}{ll}
G(D)=\langle x_{1},x_{2},x_{3},x_{4},\ldots,x_{l}\mid x_{1}x_{2}^{-1},\{r_{i}\}\rangle, \\
G(D')=\langle x_{1},x_{2},x_{3},x_{4},\ldots,x_{l}\mid x_{1}x_{3}^{n}x_{2}^{-1}x_{3}^{-n},\{r_{i}\}\rangle,
\end{array}$$ where $\{r_{i}\}$ is the set of relations corresponding to the other crossings. Using the Fox free derivative [@CF], we have the Alexander matrices $A_{D}(t)$ and $A_{D'}(t)$ of $D$ and $D'$, respectively, as follows: $$A_{D}(t)=\begin{pmatrix}
1 &-1 &0 &0 &\cdots &0 \\ \cdashline{1-6}[3pt/3pt]
\multicolumn{6}{c}{} \\
\multicolumn{6}{c:}{\raisebox{0pt}[0pt][0pt]{$\mathfrak{a}\gamma\left(\cfrac{r_{i}}{x_{j}}\right)$}} \\
\multicolumn{6}{c}{} \\
\end{pmatrix},
A_{D'}(t)=\begin{pmatrix}
1 &-t^{n} &t^{n}-1 &0 &\cdots &0 \\ \cdashline{1-6}[3pt/3pt]
\multicolumn{6}{c}{} \\
\multicolumn{6}{c:}{\raisebox{0pt}[0pt][0pt]{$\mathfrak{a}\gamma\left(\cfrac{r_{i}}{x_{j}}\right)$}} \\
\multicolumn{6}{c}{} \\
\end{pmatrix}.$$ Therefore, $A_{D}(t)-A_{D'}(t)$ is a zero matrix modulo ${I(1-t^{n})}$.
Now we have the following.
\[prop-trefoil\] The $V^{n}$-move is not an unknotting operation on welded knots for $n\geq 2$.
We show that the trefoil knot is not $V^{n}$-equivalent to the unknot for $n\geq 2$. The first Alexander polynomial of the trefoil knot is $1-t+t^{2}$, and that of the unknot is $1$. It suffices to show that $1-t+t^{2}-{\varepsilon}t^{r}\not\in I(1-t^{n})$ for any $n\geq 2$ by Proposition \[prop-Alex\] $({\varepsilon}\in\{\pm 1\},r\in\mathbb{Z})$.
Suppose that $n\geq 2$. We define a map $f_{n}:\mathbb{Z}[t^{\pm 1}]\rightarrow \mathbb{Z}$ by $f_{n}(\sum_{i}a_{i}t^{i})=\sum_{i\equiv 0,2\pmod{n}}a_{i}$, where $a_{i}\in\mathbb{Z}$. Since $f_{n}({\delta}t^{s}(1-t^{n}))=0$ $({\delta}\in\{\pm 1\}, s\in\mathbb{Z})$ it follows that $f(b)=0$ for any element $b\in I(1-t^{n})$. On the other hand, $f_{n}(1-t+t^{2}-{\varepsilon}t^{r})$ is not equal to $0$ for any $n\geq 2$. This completes the proof.
The proposition above immediately implies the following corollary.
\[cor-UC\] A [UC]{}-move is realized by a sequence of the $V^{n}$-move and welded Reidemeister moves if and only if $n=1$.
Unoriented $V^{n}$-moves and associated core groups
===================================================
In this section, we will discuss relations among [*unoriented*]{} $V^{n}$-moves, associated core groups and the multiplexing of crossings.
For an unoriented classical link diagram $D$, the [*associated core group $\Pi^{(2)}_{D}$*]{} is defined as follows. Each arc of $D$ yields a generator, and each classical crossing gives a relation $yx^{-1}yz^{-1}$, where $x$ and $z$ correspond to the underpasses and $y$ corresponds to the overpass at the crossing. This group $\Pi^{(2)}_{D}$ is known as a classical link invariant [@J; @Kelly; @FR; @W].
\[rem-MWada\] Let $L$ be an unoriented classical link in the $3$-sphere and $D$ a classical diagram of $L$. M. Wada [@W] proved that $\Pi_{D}^{(2)}$ is isomorphic to the free product of the fundamental group of the double branched cover $M_{L}^{(2)}$ of the $3$-sphere branched along $L$ and the infinite cyclic group $\mathbb{Z}$. That is, $\Pi_{D}^{(2)}\cong\pi_{1}\left(M_{L}^{(2)}\right)*\mathbb{Z}$.
We similarly define the associated core group $\Pi^{(2)}_{D}$ of an unoriented [*virtual*]{} link diagram $D$ by generators and relations as described above. (Note that virtual crossings do not produce any generator or relation.) It is not hard to see that $\Pi^{(2)}_{D}$ is preserved by welded Reidemeister moves, and hence we define the [*associated core group $\Pi^{(2)}_{L}$*]{} of an unoriented welded link $L$ to be $\Pi^{(2)}_{D}$ of a diagram $D$ of $L$. Moreover we have the following.
\[prop-V2\] If $n$ is even, then $\Pi^{(2)}_{L}$ is preserved by unoriented $V^{n}$-moves.
$\Pi^{(2)}_{L}$ is preserved by unoriented $V^{2}$-moves as illustrated in Figure \[V2\], and furthermore, an unoriented $V^{n}$-move is realized by unoriented $V^{2}$-moves for any even number $n$.
[V2.eps]{} (-7,28)[$x$]{} (-8,-1)[$y$]{} (86.5,20)[$V^{2}$]{} (104,28)[$x$]{} (104,-1)[$y$]{} (138.5,34)[$y$]{} (138.5,-8)[$x$]{} (158,34)[$yx^{-1}y$]{} (187,-8)[$yx^{-1}y$]{} (232,28)[$y(y^{-1}xy^{-1})y=x$]{} (232,-1)[$y$]{}
There are welded knots whose associated core groups are nontrivial, for example, all knots having nontrivial Fox colorings; see [@P Proposition 4.5]. Therefore, the proposition above gives an alternative proof for that the $V^{n}$-move is not an unknotting operation on welded knots for any even number $n$.
In [@MWY17], the authors introduced the [*multiplexing*]{} of crossings for an unoriented virtual link diagram, which yields a new unoriented virtual link diagram. Let $(m_{1},\ldots,m_{\mu})$ be a $\mu$-tuple of integers, and let $D=\bigcup_{i=1}^{\mu}D_{i}$ be an unoriented $\mu$-component virtual link diagram. For a classical crossing of $D$ whose overpass belongs to $D_{j}$, we define the [*multiplexing*]{} of the crossing associated with $m_{j}$ as a local change illustrated in Figure \[multiplexing\]. When $m_{j}=0$, the multiplexing of the crossing is the the crossing virtualization of it. The number of classical crossings that appear in the multiplexing of the crossing is the absolute value of $m_{j}$. Let $D(m_{1},\ldots,m_{\mu})$ denote the virtual link diagram obtained from $D$ by the multiplexing of all classical crossings of $D$ associated with $(m_{1},\ldots,m_{\mu})$. For welded isotopic virtual link diagrams $D$ and $D'$, $D(m_{1},\ldots,m_{\mu})$ and $D'(m_{1},\ldots,m_{\mu})$ are also welded isotopic for any $(m_{1},\ldots,m_{\mu})\in\mathbb{Z}^{\mu}$ [@MWY17 Theorem 2.1]. For an unoriented $\mu$-component welded link $L$, we define $L(m_{1},\ldots,m_{\mu})$ to be $D(m_{1},\ldots,m_{\mu})$ of a diagram $D$ of $L$.
It is not hard to see that $L(m_{1},\ldots,m_{\mu})$ can be deformed into $L(0,\ldots,0)$ by unoriented $V^{2}$-moves for any $\mu$-tuple $(m_{1},\ldots,m_{\mu})$ of even numbers. Since $L(0,\ldots,0)$ is trivial, we have the following.
[multiplexing.eps]{} (14,-15)[$m_{j}>0$]{} (125,-15)[$m_{j}=0$]{} (235,-16)[$m_{j}<0$]{} (160,150)[$\leftarrow$ $D_{j}$]{} (51,57)[[$1$]{}]{} (-4,11)[[$m_{j}$]{}]{} (265,51)[[$1$]{}]{} (227,27)[[$|m_{j}|$]{}]{}
\[th-V2\] Let $(m_{1},\ldots,m_{\mu})$ be a $\mu$-tuple of even numbers. For any unoriented $\mu$-component welded link $L$, $L(m_{1},\ldots,m_{\mu})$ is deformed into the $\mu$-component trivial link by unoriented $V^{2}$-moves.
In [@MWY17 Theorem 3.2], the authors proved that unoriented classical knots $K$ and $K'$ are equivalent up to mirror image if and only if $K(m)$ and $K'(m)$ are welded isotopic up to mirror image for any fixed non-zero integer $m$. Hence, it follows that if a classical knot $K$ is nontrivial then $K(m)$ is also nontrivial. By Propositions \[prop-V2\] and \[th-V2\], if $m$ is even then $\Pi^{(2)}_{K(m)}$ is isomorphic to that of the unknot, that is, $\Pi^{(2)}_{K(m)}\cong\mathbb{Z}$. Therefore, we have the following theorem although by Remark \[rem-MWada\] the associated core groups seem to be very strong invariants.
\[prop-infinitely\] Let $m$ $(\neq 0)$ be an even number. For any nontrivial unoriented welded knot $K$, $K(m)$ is nontrivial and $\Pi^{(2)}_{K(m)}\cong\mathbb{Z}$.
$\overline{V}(n)$-moves and $\overline{V}^{n}$-moves
====================================================
$\overline{V}(n)$-moves
-----------------------
When $n$ is odd, one may consider the $V(n)$-move involving two strands being oriented antiparallel. We call such a move the [*$\overline{V}(n)$-move*]{}. In this subsection, we will show that $V(n)$- and $\overline{V}(n)$-moves are equivalent.
For an odd number $n$, we define the [*$\overline{A}(n)$-move*]{} as a local move on an arrow presentation illustrated in Figure \[A(n)opposite\]. By arguments similar to those in the proof of Lemma \[lem-A(n)V(n)\], we have the following.
[A\_n-opposite.eps]{} (44,43)[$\overline{A}(n)$]{} (106,10)[[$1$]{}]{} (106,19)[[$2$]{}]{} (106,61)[[$n$]{}]{}
(222,43)[$\overline{A}(n)$]{} (284.5,10)[[$1$]{}]{} (284.5,19)[[$2$]{}]{} (284.5,61)[[$n$]{}]{}
Let $n$ be an odd number. An arrow presentation for a $\overline{V}(n)$-move is realized by a sequence of the $\overline{A}(n)$-move and arrow moves. Conversely, surgery along an $\overline{A}(n)$-move is realized by a sequence of the $\overline{V}(n)$-move and welded Reidemeister moves.
By deformations similar to those in Figure \[pf-Hexch\], we have the following.
\[lem-Hdash\] An [H]{}-move is realized by a sequence of the ${\rm H}'$-move illustrated in Figure $\ref{Hdash}$ and arrow moves.
[Hdash.eps]{} (67,20)[${\rm H}'$]{}
Here, we consider two [*allowable moves ${\rm AR11}'$ and ${\rm AR12}'$*]{} illustrated in Figure \[AR1112dash\], each of which is realized by a sequence of arrow moves similar to that in Figure \[pf-AR112\]. Using the moves ${\rm AR11}'$ and ${\rm AR12}'$, we have the following.
[AR1112dash.eps]{} (58,20.5)[${\rm AR11}'$]{} (230,20.5)[${\rm AR12}'$]{}
\[lem-A(n)V(n)-bar\] Let $n$ be an odd number. An ends exchange move is realized by a sequence of the $\overline{A}(n)$-move and arrow moves.
By Lemmas \[lem-ends\] and \[lem-Hdash\], it is enough to show that an ${\rm H}'$-move is realized by a sequence of the $\overline{A}(n)$-move and arrow moves. Figure \[pf-A\_n-bar\] indicates the proof. While the proof describes only when $n=3$ in the figure, it is essentially the same in all cases.
[pf-A\_n-bar.eps]{} (68,25)[$\overline{A}(n)$]{} (152.5,25)[${\rm AR11}', 12'$]{} (257,25)[$\overline{A}(n)$]{}
Now we can show the following.
Let $n$ be an odd number. $V(n)$- and $\overline{V}(n)$-moves are equivalent.
By Lemmas \[lem-A(n)V(n)\] and \[lem-A(n)V(n)-bar\], it is enough to show that $A(n)$- and $\overline{A}(n)$-moves are equivalent. Figure \[pf-A(n)\] shows that an $\overline{A}(n)$-move is realized by a sequence of the $A(n)$-move and arrow moves, while the figure describes only when $n=3$. (Note that we can use ends exchange moves by Lemma \[lem-HvsA(n)\].) Conversely, by deformations similar to those in the figure, it is not hard to see that an $A(n)$-move is realized by a sequence of the $\overline{A}(n)$-move and arrow moves. This completes the proof.
[pf-A\_n-opposite.eps]{} (49,27)[AR]{} (125,27)[[ends]{}]{} (117,14)[[exchange]{}]{} (199,27)[$A(n)$]{} (277.5,27)[AR]{}
$\overline{V}^{n}$-moves
------------------------
For a positive integer $n$, we define the [*$\overline{V}^{n}$-move*]{} as the $V^{n}$-move involving two strands being oriented antiparallel. Also, we define the [*$A^{n}$-move*]{} as a local move on an arrow presentation illustrated in Figure \[A\^n-bar\]. By arguments similar to those in the proof of Lemma \[lem-A(n)V(n)\], we have the following.
[An-bar.eps]{} (48,32)[$\overline{A}^{n}$]{} (105,9)[[$1$]{}]{} (105,18)[[$2$]{}]{} (105,43)[[$n$]{}]{}
(227,32)[$\overline{A}^{n}$]{} (284.5,9)[[$1$]{}]{} (284.5,18)[[$2$]{}]{} (284.5,43)[[$n$]{}]{}
\[lem-A\^nV\^n-bar\] Let $n$ be a positive integer. An arrow presentation for a $\overline{V}^{n}$-move is realized by a sequence of the $\overline{A}^{n}$-move and arrow moves. Conversely, surgery along an $\overline{A}^{n}$-move is realized by a sequence of the $\overline{V}^{n}$-move and welded Reidemeister moves.
It is not hard to see that $A^{n}$- and $\overline{A}^{n}$-moves are equivalent. This together with Lemmas \[lem-A\^nV\^n\] and \[lem-A\^nV\^n-bar\] implies the following.
Let $n$ be a positive integer. $V^{n}$- and $\overline{V}^{n}$-moves are equivalent.
[99]{}
H. Aida, [*Unknotting operations of polygonal type*]{}, Tokyo J. Math. [**15**]{} (1992), 111–121.
B. Audoux, P. Bellingeri, J.-B. Meilhan, E. Wagner, [*On usual, virtual and welded knotted objects up to homotopy*]{}, J. Math. Soc. Japan [**69**]{} (2017), 1079–1097.
B. Audoux, P. Bellingeri, J.-B. Meilhan, E. Wagner, [*Extensions of some classical local moves on knot diagrams*]{}, Michigan Math. J. [**67**]{} (2018), 647–672.
R. H. Crowell, R. H. Fox, [*Introduction to knot theory*]{}, Reprint of the 1963 original, Graduate Texts in Mathematics [**57**]{} (Springer-Verlag, New York-Heidelberg, 1977).
M. K. Dabkowski, J. H. Przytycki, [*Burnside obstructions to the Montesinos-Nakanishi $3$-move conjecture*]{}, Geom. Topol. [**6**]{} (2002), 355–360.
M. K. Dabkowski, J. H. Przytycki, [*Unexpected connections between Burnside groups and knot theory*]{}, Proc. Natl. Acad. Sci. USA [**101**]{} (2004), 17357–17360.
R. Fenn, R. Rimányi, C. Rourke, [*The braid-permutation group*]{}, Topology [**36**]{} (1997), 123–135.
R. Fenn, C. Rourke, [*Racks and links in codimension two*]{}, J. Knot Theory Ramifications [**1**]{} (1992), 343–406.
M. Goussarov, M. Polyak, O. Viro, [*Finite-type invariants of classical and virtual knots*]{}, Topology [**39**]{} (2000), 1045–1068.
J. Hoste, Y. Nakanishi, K. Taniyama, [*Unknotting operations involving trivial tangles*]{}, Osaka J. Math. [**27**]{} (1990), 555–566.
D. Joyce, [*A classifying invariant of knots, the knot quandle*]{}, J. Pure Appl. Algebra [**23**]{} (1982), 37–65.
T. Kanenobu, [*Forbidden moves unknot a virtual knot*]{}, J. Knot Theory Ramifications [**10**]{} (2001), 89–96.
L. H. Kauffman, [*Virtual knot theory*]{}, European J. Combin. [**20**]{} (1999), 663–690.
A. J. Kelly, [*Groups from link diagrams*]{}, Ph.D. Thesis, University of Warwick (1990). Available at http://wrap.warwick.ac.uk/63616/
S. Kinoshita, [*On the distribution of Alexander polynomials of alternating knots and links*]{}, Proc. Amer. Math. Soc. [**79**]{} (1980), 644–648.
J.-B. Meilhan, A. Yasuhara, [*Arrow calculus for welded and classical links*]{}, Algebr. Geom. Topol. [**19**]{} (2019) 397–456.
K. Miyazaki, A. Yasuhara, [*Generalized $\#$-unknotting operations*]{}, J. Math. Soc. Japan [**49**]{} (1997), 107–123.
H. A. Miyazawa, K. Wada, A. Yasuhara, [*Link invariants derived from multiplexing of crossings*]{}, Algebr. Geom. Topol. [**18**]{} (2018), 2497–2507.
H. Murakami, Y. Nakanishi, [*On a certain move generating link-homology*]{}, Math. Ann. [**284**]{} (1989), 75–89.
H. Murakami, [*Some metrics on classical knots*]{}, Math. Ann. [**270**]{} (1985), 35–45.
T. Nakamura, Y. Nakanishi, S. Satoh, A. Yasuhara, [*The pass move is an unknotting operation for welded knots*]{}, Topology Appl. [**247**]{} (2018), 9–19.
T. Nasybullov, [*Classification of fused links*]{}, J. Knot Theory Ramifications [**25**]{} (2016), 1650076, 21 pp.
S. Nelson, [*Unknotting virtual knots with Gauss diagram forbidden moves*]{}, J. Knot Theory Ramifications [**10**]{} (2001), 931–935.
T. Okabayashi, [*Forbidden moves for virtual links*]{}, Kobe J. Math. [**22**]{} (2005), 49–63.
J. H. Przytycki, [*On Slavik Jablan’s work on $4$-moves*]{}, J. Knot Theory Ramifications [**25**]{} (2016), 1641014, 26 pp.
S. Satoh, [*Crossing Changes, Delta Moves and Sharp Moves on Welded Knots*]{}, Rocky Mountain J. Math. [**48**]{} (2018), 967–979.
M. Wada, [*Group invariants of links*]{}, Topology [**31**]{} (1992), 399–406.
[^1]:
[^2]: The second author was supported by a Grant-in-Aid for JSPS Research Fellow (\#17J08186) of the Japan Society for the Promotion of Science.
[^3]: The third author was partially supported by a Grant-in-Aid for Scientific Research (C) (\#17K05264) of the Japan Society for the Promotion of Science and a Waseda University Grant for Special Research Projects (\#2018S-077).
[^4]: For simplicity, we do not use here the usual drawing convention for virtual crossings, which is a small circle around the corresponding double point.
[^5]: Note that our definition slight differs from the one given in [@MY Lemma 5.14].
[^6]: Note that our definition slight differs from the one given in [@MY Lemma 5.16].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We revisit the effects of the Hubbard repulsion on quantum spin Hall effects (QSHE) in two-dimensional quantum lattice models. We present both unbiased exact diagonalization and density-matrix renormalization group simulations with numerical evidences for a continuous quantum phase transition (CQPT) separating QSHE from the topologically trivial antiferromagnetic phase. Our numerical results suggest that, the nature of CQPT exhibits distinct finite-size scaling behaviors, which may be consistent with either Ising or XY universality classes for different time-reversal symmetric QSHE systems.'
author:
- 'Tian-Sheng Zeng'
- 'W. Zhu'
- 'Jian-Xin Zhu'
- 'D. N. Sheng'
title: Nature of continuous phase transitions in interacting topological insulators
---
Introduction
============
Landau continuous phase transitions classified by spontaneous symmetry breaking of local order parameters [@Landau1937] are important concepts in condensed matter physics that lie at the heart of our understanding of various aspects such as quantum magnetism and superconductivity. However topological phase transitions among topological phases of matter which are generally indistinguishable by any local order parameters [@Wen1990], should require a change in topological invariant. For a symmetry-protected topological phase, its topological characterization is only well-defined in the presence of symmetry, like fermionic QSHE with time-reversal symmetry can be characterized by a $Z_2$ topological index, which is the main concern of our paper. In the interacting systems with topologically non-trivial structure, the interplay between topology, symmetry and interaction, may lead to complex nature for the quantum phase transition, possibly a first-order transition [@Varney2011; @Amaricci2015; @Roy2016]. Nevertheless, whether or not a continuous phase transition can be accompanied by a change of topological order, is an intricate open question [@Samkharadze2016; @Lee2015; @Castelnovo2008; @Hamma2008], which motivates us to reinvestigate the strong correlation effects on the interacting topological insulators with $Z_2$ topological index.
Recent studies on topological insulators have indicated such a concrete example of interaction-driven continuous quantum phase transition (CQPT) from $Z_2$ topological order to antiferromagnetism, where the universal continuous evolutions of physical quantities are expected [@Ran2006; @Moon2012; @Whitsitt2016]. Within the Kane-Mele-Hubbard (KMH) model [@Kane2005a; @Kane2005b], Xu and Moore proposed a CQPT from the QSHE to the trivial Mott insulator driven by interactions [@Xu2006]. In the strong coupling limit, Rachel and Le Hur first derived its effective spin Hamiltonian up to second order perturbation, and concluded that the Mott antiferromagnetism (AFM) is in the transverse $xy$-plane, instead of in the longitudinal $z$-direction [@Rachel2010]. And this scenario including the CQPT from the QSHE to the trivial $xy$-AFM was supported by numerical studies including quantum Monte Carlo (QMC) methods [@Hohenadler2011; @Zheng2011; @Hohenadler2012; @Assaad2013a; @Assaad2013b; @Hohenadler2014], the variational cluster approach [@Yu2011; @Budich2012], and the mean field theory [@Pesin2010; @Murakami2007; @Rachel2010; @Vaezi2012; @Chen2015; @Fiete2012; @Reuther2012; @Wu2012; @Liu2013; @Laubach2014]. In QMC simulations of spin order, a finite size analysis shows that the transverse long-range spin correlation $\langle S_{{{\bf r}}}^xS_{{{\bf r}}'}^x\rangle$ remains a robust finite value as the distance $|{{\bf r}}-{{\bf r}}'|$ increases in the antiferromagnetic regime, while the longitudinal long-range spin correlation $\langle S_{{{\bf r}}}^zS_{{{\bf r}}'}^z\rangle$ vanishes as the system size increases [@Zheng2011; @Hohenadler2012; @Assaad2013a; @Hohenadler2014]. In Ref. [@Assaad2013a], a Curie-law signature in the magnetic susceptibility is identified by adiabatically inserting a $\pi$ flux. Early studies based on mean-field theory [@Pesin2010; @Murakami2007; @Rachel2010; @Vaezi2012; @Chen2015] predicted the existence of intermediate topological antiferromagnetic phases at certain moderate Hubbard repulsion, making the nature of CQPT more [*intricate*]{}. However, there is no signal of an intermediate topological phase being detected by recent numerical QMC simulations of $Z_2$ invariant [@Lang2013; @Hung2014]. Furthermore, as to the phase diagram of KMH model, the transition from QSHE to antiferromagnetic Mott insulator has been theoretically predicted to belong to the three-dimensional $XY$ universality class [@Lee2011; @Griset2012; @Hohenadler2013], and this continuous transition nature with the universal critical exponents $\beta=0.3486,\nu=0.6717$ is rigorously demonstrated by the finite-size scaling of the $xy$-transverse spin structure factor in the QMC numerical simulations [@Hohenadler2011; @Hohenadler2012; @Assaad2013a; @Hohenadler2014]. Taking into account the rich class of CQPT, it is natural and important to ask whether the CQPT nature is common to different time-reversal symmetric quantum spin Hall systems realized on different lattice geometries.
In this work, we study this interaction-driven transition nature in two representative topological lattice models with time-reversal symmetry through the state-of-the-art density-matrix renormalization group (DMRG) and exact diagonalization (ED) techniques. In Sec. \[model\], we introduce the time-reversal symmetric spinful fermionic Hamiltonian in two typical $\pi$-flux checkerboard and Haldane-honeycomb lattices. In Sec. \[ground\], by tuning the Hubbard repulsion, we demonstrate a CQPT from $Z_2$ QSHE at weak interactions to a trivial Mott antiferromagnetic insulator at strong interactions, with the evidences from Chern number matrix and spin structure factors. In particular, we identify the classification of CQPT is not unique. Specifically, the transition matches with three-dimensional XY universality class in the Haldane-honeycomb lattice, while for the typical $\pi$-flux checkerboard lattice, the transition is possibly in the universality class of the 2D Ising model. Finally, in Sec. \[summary\], we summarize our results and compare the difference between QSHE and integer quantum Hall effect.
Theoretical Models {#model}
==================
We consider the spinful fermions in two representative topological lattice models with time-reversal symmetry: (i) the Haldane-honeycomb (HC) lattice [@Wang2011] $$\begin{aligned}
&H_{HC}^{\uparrow}=-t'\sum_{\langle\langle{{\bf r}},{{\bf r}}'\rangle\rangle}[c_{{{\bf r}}',\uparrow}^{\dag}c_{{{\bf r}},\uparrow}\exp(i\phi_{{{\bf r}}'{{\bf r}}})+H.c.]\nonumber\\
&-t\!\sum_{\langle{{\bf r}},{{\bf r}}'\rangle}\!\!c_{{{\bf r}}',\uparrow}^{\dag}c_{{{\bf r}},\uparrow}
-t''\!\sum_{\langle\langle\langle{{\bf r}},{{\bf r}}'\rangle\rangle\rangle}\!\!\!\! c_{{{\bf r}}',\uparrow}^{\dag}c_{{{\bf r}},\uparrow}+H.c.,\end{aligned}$$ and (ii) the $\pi$-flux checkerboard (CB) lattice [@Sun2011] $$\begin{aligned}
&H_{CB}^{\uparrow}=-t\!\sum_{\langle{{\bf r}},{{\bf r}}'\rangle}\!\big[c_{{{\bf r}}',\uparrow}^{\dag}c_{{{\bf r}},\uparrow}\exp(i\phi_{{{\bf r}}'{{\bf r}}})+H.c.\big]\nonumber\\
&-\!\sum_{\langle\langle{{\bf r}},{{\bf r}}'\rangle\rangle}\!\! t_{{{\bf r}},{{\bf r}}'}'c_{{{\bf r}}',\uparrow}^{\dag}c_{{{\bf r}},\uparrow}
-t''\!\sum_{\langle\langle\langle{{\bf r}},{{\bf r}}'\rangle\rangle\rangle}\!\!\!\! c_{{{\bf r}}',\uparrow}^{\dag}c_{{{\bf r}},\uparrow}+H.c.\end{aligned}$$ Due to time-reversal symmetry, we take $H_{CB}^{\downarrow}=\mathcal{T}H_{CB}^{\uparrow}\mathcal{T}^{-1}$ and $H_{HC}^{\downarrow}=\mathcal{T}H_{HC}^{\uparrow}\mathcal{T}^{-1}$ with $\mathcal{T}$ the time-reversal operation. Here $c_{{{\bf r}},\sigma}^{\dag}$ is the particle creation operator of spin $\sigma=\uparrow,\downarrow$ at site ${{\bf r}}$, $\langle\ldots\rangle$,$\langle\langle\ldots\rangle\rangle$ and $\langle\langle\langle\ldots\rangle\rangle\rangle$ denote the nearest-neighbor, the next-nearest-neighbor, and the next-next-nearest-neighbor pairs of sites, respectively. Typically, we choose $t''=0,\phi=\pi/2$ for honeycomb lattice which reduces to the famous Kane-Mele (KM) model [@Kane2005a; @Kane2005b], and $t''=0,\phi=\pi/4$ for checkerboard lattice [@Sun2011]. In the flat band limit, we take the parameters $t'=0.6t,t''=-0.58t,\phi=2\pi/5$ for honeycomb lattice and $t'=0.3t,t''=-0.2t,\phi=\pi/4$ for checkerboard lattice.
Taking into account the on-site Hubbard repulsions $V_{int}=U\sum_{{{\bf r}}}n_{{{\bf r}},\uparrow}n_{{{\bf r}},\downarrow}$ where $n_{{{\bf r}},\sigma}$ is the particle number operator of spin-$\sigma$ at site ${{\bf r}}$, the model Hamiltonian becomes $H=H_{CB}^{\downarrow}+H_{CB}^{\uparrow}+V_{int}$ ($H=H_{HC}^{\downarrow}+H_{HC}^{\uparrow}+V_{int}$). In the following we explore the many-body ground state of $H$ at half-filling $N_{\uparrow}/N_s=N_{\downarrow}/N_s=1/2$ in a finite system of $N_x\times N_y$ unit cells (the total number of sites is $N_s=2\times N_x\times N_y$) with particle conservation $U(1)\times U(1)$-symmetry. In the ED study, with the translational symmetry, the energy states are labeled by the total momentum $K=(K_x,K_y)$ in units of $(2\pi/N_x,2\pi/N_y)$ in the Brillouin zone. For larger systems we exploit DMRG on the cylindrical geometry, and keep the number of state basis up to 3000 to obtain accurate results.
Interaction-driven phase transitions {#ground}
====================================
In this section, we present the numerical analysis of the interaction-driven phase transition from two-component QSHE to antiferromagnetism at half-filling. The two-component QSHE can be identified by the Chern number matrix with featureless spin structure factors, and the corresponding charge (spin) pumpings are complementary to and consistent with the Chern number matrix.
ED analysis
-----------
 Numerical ED results for two-component fermions at half-filling $N_s=2\times2\times4=16,N_{\uparrow}=N_{\downarrow}=8$ in the Haldane-honeycomb lattice with $t'=0.3t,t''=0,\phi=\pi/2$. (a) The low energy spectrum as a function of onsite repulsion $U$. (b) The energy spectrum gap for the lowest two energy states in the whole parameter plane $(\theta^{x}_{\uparrow}=\theta^{x}_{\downarrow},\theta^{y}_{\uparrow}=\theta^{y}_{\downarrow})$, keeping $Z_2$ symmetry. (c) The antiferromagnetic spin structure factors $S_{AF}^{zz},S_{AF}^{xy}$ of the ground state as a function of $U$. (d) The topological transition signature of the ground state obtained from its many-body Chern number $C_{\uparrow,\uparrow}$ and the standard deviations of Berry curvature as a function of $U$. ](energy1.eps){height="2.25in" width="3.4in"}
We first present an ED study of the ground state properties for HC lattice with two different lattice sizes $N_s=16,12$. In Fig. \[energyhc\](a), we plot the low energy evolution as a function of on-site repulsion $U$. For weak interactions, there always exists a stable unique ground state at $K=(0,0)$ with a large gap separated from higher levels. By tuning $U$ from weak to strong, there is not any level crossing between the ground state and high-level excited states. Also this ground state does not undergo the level crossing with excited levels in the $Z_2$-symmetric parameter plane $(\theta^{x}_{\uparrow}=\theta^{x}_{\downarrow},\theta^{y}_{\uparrow}=\theta^{y}_{\downarrow})$ as indicated in Fig. \[energyhc\](b), signaling a continuous phase transition nature with $Z_2$ symmetry. (Here $\theta_{\sigma}^{\alpha}$ is the twisted angle for spin-$\sigma$ particles in the $\alpha$-direction, which shifts the particle crystal momentum ${{\bf k}}_{\alpha}\rightarrow{{\bf k}}_{\alpha}+\theta_{\sigma}^{\alpha}/N_{\alpha}$; see the definition below). We emphasize that to fully establish the continuous ground energy evolution without level crossing we need to perform a scaling of the system size results, which is beyond our current ED limit. Instead, we will demonstrate its continuous transition for large system sizes from the DMRG calculation of ground state wavefunction fidelity and antiferromagnetic order parameters, as shown in Sec. \[dmrg\]. Alternatively, the topological index obtained in ED calculation can help us locate the phase transition boundary. The topological nature of quantum spin-Hall state is characterized by the Chern number matrix by introducing twisted boundary conditions [@Sheng2003; @Sheng2006] $\psi(\cdots,{{\bf r}}_{\sigma}^{i}+N_{\alpha},\cdots)=\psi(\cdots,{{\bf r}}_{\sigma}^{i},\cdots)\exp(i\theta_{\sigma}^{\alpha})$. The system is periodic when one flux quantum $\theta_{\sigma}^{\alpha}=0\rightarrow2\pi$ is inserted. Meanwhile, the many-body Chern number of the ground state wavefunction $\psi$ is defined as $C_{\sigma,\sigma'}=\frac{1}{2\pi}\int d\theta_{\sigma}^{x}d\theta_{\sigma'}^{y}F_{\sigma,\sigma'}^{xy}$ with the Berry curvature $$\begin{aligned}
F_{\sigma,\sigma'}^{xy}=\mathbf{Im}\left(\langle{\frac{\partial\psi}{\partial\theta_{\sigma}^x}}|{\frac{\partial\psi}{\partial\theta_{\sigma'}^y}}\rangle
-\langle{\frac{\partial\psi}{\partial\theta_{\sigma'}^y}}|{\frac{\partial\psi}{\partial\theta_{\sigma}^x}}\rangle\right).\nonumber\end{aligned}$$ Due to time-reversal symmetry, for any ground state and interaction, one has the antisymmetric properties $C_{\uparrow,\downarrow}=-C_{\downarrow,\uparrow},C_{\uparrow,\uparrow}=-C_{\downarrow,\downarrow}$ in the spanned Hilbert space, and the total Chern number related to the charge Hall conductance is equal to zero $C_q=\sum_{\sigma,\sigma'}C_{\sigma,\sigma'}=0$ for any interaction strength. Therefore we always have an antisymmetric $C$-matrix [@Sheng2006] $$\begin{aligned}
\mathbf{C}=\begin{pmatrix}
C_{\uparrow,\uparrow} & C_{\uparrow,\downarrow}\\
C_{\downarrow,\uparrow} & C_{\downarrow,\downarrow}\\
\end{pmatrix}\end{aligned}$$ For decoupled QSHE at weak interactions, we obtain $C_{\uparrow,\uparrow}=1$, and $C_{\uparrow,\downarrow}=0$. However, for strong interactions, the off-diagonal element $C_{\uparrow,\downarrow}$ related to the drag Hall conductance arising from interspecies correlation may be nonzero for two-component quantum Hall effects [@Sheng2003; @Sheng2005; @Zeng2017; @Nakagawa2017].
 Numerical ED results for two-component fermions at half-filling $N_s=2\times2\times4=16,N_{\uparrow}=N_{\downarrow}=8$ in the $\pi$-flux checkerboard lattice with $t'=0.3t,t''=0,\phi=\pi/4$. (a) The low energy spectrum as a function of onsite repulsion $U$. (b) The antiferromagnetic spin structure factors $S_{AF}^{zz},S_{AF}^{xy}$ of the lowest ground state and its many-body Chern number $C_{\uparrow,\uparrow}$ as a function of $U$. ](energy2.eps){height="1.4in" width="3.4in"}
To clarify the interaction-driven topological transition, we calculate the evolution of $C_{\uparrow,\uparrow}$ as a function of $U$. In Fig. \[energyhc\](d), $C_{\uparrow,\uparrow}$ experiences a fast drop as the interaction $U$ increases across the critical threshold $U_c$, where the distribution of Berry curvature exhibits a singular behavior, signalling the topological phase transition of a many-body system [@Carollo2005; @Zhu2006]. As a quantitative measure of the fluctuation of the Berry curvature, we take $\Delta F_{\sigma,\sigma'}=\sqrt{\int d\theta_{\sigma}^{x}d\theta_{\sigma'}^{y}[F_{\sigma,\sigma'}^{xy}-\overline{F}]^2}$ where $\overline{F}$ is the average value. Both $\Delta F_{\uparrow,\uparrow}$ and $\Delta F_{\uparrow,\downarrow}$ show a peak at the critical point where topological invariant changes, resulting from the energy level crossing at $(\theta^{x}_{\uparrow},\theta^{y}_{\uparrow})=(\pi,0)$ (see Fig. \[level\] for details). Physically, the sudden jump of $C_{\uparrow,\uparrow}$ and the singularity of $\Delta F_{\uparrow,\uparrow}$ mark the quantum phase transition from QSHE to Mott insulator, while the latter is characterized by gapless spin excitations as shown in Fig. \[dmrghc\](d).
In addition, to get a picture about the Mott insulator in the strongly large-$U$ limit, we calculate the antiferromagnetic spin structure factors $$\begin{aligned}
S_{AF}^{zz}=\sum_{\alpha,\beta}[S_{AF}^{zz}]^{\alpha,\beta}\label{zafm}\end{aligned}$$ and $$\begin{aligned}
S_{AF}^{xy}=\sum_{\alpha,\beta}[S_{AF}^{xy}]^{\alpha,\beta},\label{xafm}\end{aligned}$$ with the inner functions defined by $$\begin{aligned}
[S_{AF}^{zz}]^{\alpha,\beta}&=\frac{1}{N_s}\sum_{i,j}(-1)^{\alpha}(-1)^{\beta}\langle S_{i\alpha}^zS_{j\beta}^z\rangle,\label{szz}\\
[S_{AF}^{xy}]^{\alpha,\beta}&=\frac{1}{N_s}\sum_{i,j}(-1)^{\alpha}(-1)^{\beta}\langle S_{i\alpha}^{+}S_{j\beta}^{-}+S_{i\alpha}^{-}S_{j\beta}^{+}\rangle,\label{sxy}\end{aligned}$$ where $i,j$ denote unit cells and $\alpha,\beta\in\{A,B\}$ are sublattice indices, $(-1)^{\alpha}=1 (-1)$ for $\alpha=A (B)$. As indicated in Fig. \[energyhc\](c), both $S_{AF}^{zz}$ and $S_{AF}^{xy}$ undergo a smooth evolution, implying a continuous transition. In the Mott regime for Haldane-honeycomb lattice, the transverse $xy$-antiferromagnetism $S_{AF}^{xy}$ dominates.
 Numerical ED results for the lowest energy levels at half-filling $N_s=16,N_{\uparrow}=N_{\downarrow}=8$ as a function of Hubbard repulsion under the insertion of half flux quantum $\theta^{x}_{\uparrow}=\pi,\theta^x_{\downarrow}=0$ for (a) the Haldane-honeycomb lattice and (b) the $\pi$-flux checkerboard lattice, respectively. The parameters $t'=0.3t,t''=0$.](level.eps){height="1.4in" width="3.4in"}
Similar results have also been obtained for the $\pi$-flux CB lattice, except that the dominant antiferromagnetic order is aligned along the $z$-direction given by $S_{AF}^{zz}$ in the Mott regime. As we will show below, the different nature of antiferromagnetic orders leads to distinct finite-size scaling behaviors of CQPT, depending on the lattice details. As shown in Fig. \[energycb\] for $\pi$-flux checkerboard model, the ground energy level is continuously connected between quantum spin-hall state at weak interactions and a trivial Mott antiferromagnetism at strong interactions, as the onsite repulsion is changed. However, in the Mott regime, the dominant antiferromagnetic order is aligned along the $z$-direction given by $S_{AF}^{zz}$, instead of the transverse antiferromagnetism $S_{AF}^{xy}$ in the $xy$-plane. Similarly, the second-order transition from QSHE to Ising antiferromagnetism is also predicted in the correlated Bernevig-Hughes-Zhang model, using dynamical mean field theory [@Hohenadler2013; @Yoshida2012; @Miyakoshi2013].
However, under the insertion of flux quantum $\theta^{x}_{\uparrow}=\pi,\theta^x_{\downarrow}=0$ which breaks the $Z_2$ symmetry between spin-up and spin-down particles, the lowest two energy levels indeed cross with each other at the critical point where the Berry curvature becomes singular and topological invariant changes, as indicated in Figs. \[level\](a) and \[level\](b).
DMRG results {#dmrg}
------------
 Numerical DMRG results on a cylinder Haldane-honeycomb lattice with width $L_y=2N_y$ and length $L_x=N_x=18$ at half-filling. The evolutions of the entanglement entropy $S_{L}$, charge and spin pumpings $\Delta Q,\Delta S$, spin structure factors $S_{AF}^{zz},S_{AF}^{xy}$ as a function of $U$ on a cylinder with width (a) $L_y=6$ and (b) $L_y=8$, respectively; (c) The absolute wavefunction overlap $F(U)=|\langle\psi(U)|\psi(U+\delta U)\rangle|$, $F_i(U)=|\langle\psi(U=3t)|\psi(U)\rangle|$, and $F_f(U)=|\langle\psi(U=9t)|\psi(U)\rangle|$ with different cylinder lengths $L_x=N_x=18,15$ respectively; (d) The spin excitation gap $\Delta_s$ as a function of $U$ for different lengths. The inset shows the ground state energy derivative $\partial E/\partial U$ per site. The smooth transition is characterized by the continuous behavior of these physical quantities. There are no signs of a first-order transition. The parameters $t'=0.3t,t''=0,\phi=\pi/2$.](dmrghc.eps){height="2.6in" width="3.3in"}
To further verify the continuous interaction-driven transition, we exploit an unbiased DMRG approach for larger system sizes, using a cylindrical geometry up to a maximum width $L_y=8$ ($N_y=4$). As shown in Fig. \[dmrghc\] for HC lattice, we measure three different quantities as a function of $U$: the ground state wavefunction overlap $F(U)=|\langle\psi(U)|\psi(U+\delta U)\rangle|$ ($\delta U$ is as small as $0.1t$), the ground state entanglement entropy $S_L$, and ground state energy derivative. We also check the overlaps $F_i(U)=|\langle\psi(U=3t)|\psi(U)\rangle|$ between the ground state $\psi(U)$ and QSHE at $U=3t$, and $F_f(U)=|\langle\psi(U=9t)|\psi(U)\rangle|$ between the ground state $\psi(U)$ and AFM at $U=9t$. All the physical order parameters exhibit continuous evolutions from weak interactions to strong interactions, such that we can exclude the possibility of a first-order phase transition. The spin excitation gap $\Delta_s=E_0(S^z=1)-E_0(S^z=0)$ would tend to diminish continuously in the Mott regime.
 Numerical DMRG results of the long-range antiferromagnetic spin correlation functions $|\langle S_{i,A}^{+}S_{j,B}^{-}\rangle|$,$|\langle S_{i,A}^zS_{j,B}^z\rangle|$ as a distance $|j-i|$ in the $x$-direction between sublattice A and sublattice B on a cylinder Haldane-honeycomb lattice with finite width $L_y=2N_y=6$ and fixed length $L_x=N_x=18$ at half-filling for different Hubbard repulsions: (a) $U=3.6t<U_c$ and (b) $U=7.0t>U_c$, respectively. The blue/red dashed lines are the exponential fit to the decaying behaviors of these correlation functions. The parameters $t'=0.3t,t''=0,\phi=\pi/2$.](corre.eps){height="1.65in" width="3.4in"}
Second, we characterize the topological nature of the ground state from its topological charge pumping by inserting one flux quantum $\theta_{\uparrow}^{y}=\theta,\theta_{\downarrow}^{y}=0$ from $\theta=0$ to $\theta=2\pi$ on cylinder systems based on the newly developed adiabatic DMRG [@Gong2014] in connection to the quantized Hall conductance. The net transfer of the total charge from the right side to the left side is encoded by the expectation value $Q(\theta)=N_{\uparrow}^{L}+N_{\downarrow}^{L}=tr[\widehat{\rho}_L(\theta)\widehat{Q}]$. Here we partition the lattice system on the cylinder along the $y$-direction into two halves with equal lattice sites. $N_{\sigma}^{L}$ is the particle number of spin-$\sigma$ in the left cylinder part, and $\widehat{\rho}_L$ the reduced density matrix of the corresponding left part [@Zaletel2014]. Under the inserting of the flux $\theta_{\uparrow}^{y}=\theta,\theta_{\downarrow}^{y}=0$ in the $y$-direction, the change of $N_{\uparrow}^{L}+N_{\downarrow}^{L}$ indicates the transverse charge transfer from the right side to the left side in the $x$-direction, induced by both diagonal Hall conductance $C_{\uparrow,\uparrow}$ and drag Hall conductance $C_{\downarrow,\uparrow}$. From the Chern number matrix of two-component quantum Hall effects, in each cycle we obtain [@Zeng2017] $$\begin{aligned}
\Delta Q=Q(2\pi)-Q(0)=C_{\uparrow,\uparrow}+C_{\downarrow,\uparrow}.\end{aligned}$$ In order to quantify the spin-Hall conductance, we also calculate the spin pumping by inserting one flux quantum $\theta_{\uparrow}^{y}=\theta_{\downarrow}^{y}=\theta$ from $\theta=0$ to $\theta=2\pi$ in the $y$-direction, and define the $Z_2$ spin transfer $\Delta S$ from the right side to the left side in the $x$-direction by the physical quantity $S(\theta)=N_{\uparrow}^{L}-N_{\downarrow}^{L}=tr[\widehat{\rho}_L(\theta)\widehat{S}]$ in analogy to the charge transfer. Similarly, we obtain [@Zeng2017] $$\begin{aligned}
\Delta S=S(2\pi)-S(0)=C_{\uparrow,\uparrow}-C_{\downarrow,\uparrow}+C_{\uparrow,\downarrow}-C_{\downarrow,\downarrow}.\end{aligned}$$ For each flux cycle, we obtain both $\Delta Q\simeq1$ and $\Delta S\simeq2$ for the QSHE in the weakly interacting regime. However in the strongly interacting regime $\Delta Q\simeq0,\Delta S\simeq0$. The change of the charge pumping is shown in Fig. \[dmrghc\](a), where the critical $U_c\simeq4.8t$, while the $Z_2$ spin pumping persists a finite value deviating from the integer quantized value $2$ up to $U_{c}'>U_c$. However, with increasing $L_y=6$ to $L_y=8$, we find that the difference between $U_c$ and $U_c'$ becomes substantially reduced as shown in Figs. \[dmrghc\](a-b), which may be consistent with a direct transition from the QSHE to the Mott insulator.
 Numerical DMRG results on a cylinder $\pi$-flux checkerboard lattice with width $L_y=2N_y=6$ at half-filling with parameters $t'=0.3t,t''=0$. (a) The evolutions of the entanglement entropy $S_{L}$, charge and spin pumpings $\Delta Q,\Delta S$, spin structure factors $S_{AF}^{zz}$ as a function of $U$ on a cylinder with length $L_x=N_x=18$. (b) The absolute wavefunction overlap $F(U)=|\langle\psi(U)|\psi(U+\delta U)\rangle|$, $F_i(U)=|\langle\psi(U=3t)|\psi(U)\rangle|$, and $F_f(U)=|\langle\psi(U=9t)|\psi(U)\rangle|$ with different cylinder lengths $L_x=N_x=18,12$ respectively.](dmrgcb.eps){height="1.5in" width="3.4in"}
Third, we measure the antiferromagnetic order from spin structure factors $S_{AF}^{zz},S_{AF}^{xy}$. Both $S_{AF}^{zz},S_{AF}^{xy}$ exhibit a continuous evolution near the critical point for different system sizes as shown in Figs. \[dmrghc\](a) and \[dmrghc\](b), similar to our ED analysis. In the thermodynamic limit, $S_{AF}^{zz}$ should be vanishingly small in the strong interacting limit $U\gg t$. In Figs. \[corre\](a) and \[corre\](b), our DMRG results show that for $U<U_c$, both of the antiferromagnetic spin correlations $\langle S_{i,A}^{+}S_{j,B}^{-}\rangle,\langle S_{i,A}^zS_{j,B}^z\rangle$ decay exponentially as the distance $|j-i|$, while for $U>U_c$, only the longitudinal long-range order parameters $\langle S_{i,A}^zS_{j,B}^z\rangle$ decays exponentially as the distance $|j-i|$, but the transverse long-range order parameters $\langle S_{i,A}^{+}S_{j,B}^{-}+S_{i,A}^{-}S_{j,B}^{+}\rangle$ maintain to be a robust finite value of the order 0.01, which determines the square of transverse XY spontaneous magnetization $m_{xy}^2=\lim_{|j-i|\rightarrow\infty}|\langle S_{i,A}^{+}S_{j,B}^{-}+S_{i,A}^{-}S_{j,B}^{+}\rangle|$. For $|j-i|>6$, $\langle S_{i\alpha}^zS_{j\beta}^z\rangle$ becomes already smaller than $10^{-6}$, These are in good agreement with the physical picture proposed in Refs. [@Rachel2010; @Hohenadler2012; @Assaad2013a; @Hohenadler2014]. For our study, a very small value of spin structure factor $S_{AF}^{zz}$ in both ED and DMRG, is likely due to the finite width effects in the $y$-direction. Figures. \[dmrgcb\](a) and \[dmrgcb\](b) show the continuous phase transition through the tunable repulsion $U$ on a cylinder $\pi$-flux checkerboard lattice for large system sizes. The topological phase transition is characterized by the charge and spin pumpings when inserting one flux quantum. All the physical order parameters like spin structure factor, entanglement entropy and the wavefunction fidelity exhibit a continuous evolution from weak interactions to strong interactions.
As shown in Figs. \[dmrghc\] and \[dmrgcb\], for HC lattice, only $S_{AF}^{xy}$ shows a rapid increase, signaling an antiferromagnetic order in the transverse $xy$-plane for $U>U_c$; In contrast for $\pi$-flux CB lattice, only $S_{AF}^{zz}$ shows a rapid increase near the critical point, signaling an antiferromagnetic order in the longitudinal $z$-direction for $U>U_c$. To understand this, let us consider the antiferromagnetic long-range-ordered phase in the strongly large-$U$ limit. For $U\gg t$, similar to the usual Hubbard model, we expand the Hamiltonian in powers of $t/U$ up to the second order, and arrive at the effective spin models $J\sum_{\langle{{\bf r}},{{\bf r}}'\rangle}[S_{{{\bf r}}}^{z}S_{{{\bf r}}'}^{z}+i(S_{{{\bf r}}}^{+}S_{{{\bf r}}'}^{-}-S_{{{\bf r}}}^{-}S_{{{\bf r}}'}^{+})/2]
+J'\!\sum_{\langle\langle{{\bf r}},{{\bf r}}'\rangle\rangle}\mathbf{S}_{{{\bf r}}}\cdot\mathbf{S}_{{{\bf r}}'}
+J''\!\sum_{\langle\langle\langle{{\bf r}},{{\bf r}}'\rangle\rangle\rangle}\mathbf{S}_{{{\bf r}}}\cdot\mathbf{S}_{{{\bf r}}'}$ for $\pi$-flux CB lattice and $J\sum_{\langle{{\bf r}},{{\bf r}}'\rangle}\mathbf{S}_{{{\bf r}}}\cdot\mathbf{S}_{{{\bf r}}'}
+J'\!\sum_{\langle\langle{{\bf r}},{{\bf r}}'\rangle\rangle}[S_{{{\bf r}}}^{z}S_{{{\bf r}}'}^{z}+(e^{2i\phi}S_{{{\bf r}}}^{+}S_{{{\bf r}}'}^{-}+e^{-2i\phi}S_{{{\bf r}}}^{-}S_{{{\bf r}}'}^{+})/2]
+J''\!\sum_{\langle\langle\langle{{\bf r}},{{\bf r}}'\rangle\rangle\rangle}\mathbf{S}_{{{\bf r}}}\cdot\mathbf{S}_{{{\bf r}}'}$ for Haldane HC lattice, where $J=4t^2/U, J'=4(t')^2/U,J''=4(t'')^2/U$ (see also the related effective spin Hamiltonian for HC lattice in Refs. [@Rachel2010; @Reuther2012]).
 Numerical DMRG results for the spin structure factor on a cylinder with width $L_y=2N_y=6$ at half-filling with parameters $t''=0$. (a-b) Finite-size dependence of the structure factor $S_{AF}^{xy}$ as a function of $U$ for HC lattice with critical exponents $\beta\simeq0.31,\nu\simeq0.7$ at $t'=0.3t$. (c-d) Finite-size scaling of the structure factor $S_{AF}^{zz}$ as a function of $U$ for CB lattice with critical exponents $\beta\simeq1/8,\nu\simeq1.0$ at $t'=0.3t$.](scale.eps){height="2.4in" width="3.4in"}
When $t''=0$, for $\pi$-flux CB lattice the nearest-neighbor term is an Ising exchange, while the next-nearest-neighbor term is an isotropic antiferromagnetic Heisenberg exchange. However for Haldane HC lattice the nearest-neighbor term is an isotropic antiferromagnetic Heisenberg exchange, while the next-nearest-neighbor term is antiferromagnetic in the longitudinal direction but ferromagnetic in the transverse direction when $\phi$ is close to $\pi/2$. In our typical parameters $J'/J\lesssim0.3$ which is away from possible spin liquid regime [@Gong2014a; @Gong2013], combining all the exchange terms, we expect an antiferromagnetic order in the $z$-direction for $\pi$-flux checkerboard lattice, but in the $xy$-plane for Haldane-honeycomb lattice, due to the next-nearest-neighbor frustration term. As a result, the scaling behavior around the critical point is different for HC and CB lattices.
 Numerical DMRG results of the phase diagram for (a) Haldane-honeycomb lattice and (b) $\pi$-flux checkerboard lattice models on a cylinder with finite width $L_y=2N_y=6$ and fixed length $L_x=N_x=18$ at half-filling for $t''=0$.](phase.eps){height="1.35in" width="3.3in"}
Due to the numerical difficulty of well-controlled DMRG convergence for two-component particles on cylinder width $L_y>8 (N_y>4)$, we cannot perform a finite-size scaling in the $y$-direction, and therefore focus on the quasi-one dimensional scaling of the cylinder length $L_x$. This is different from QMC methods where the finite-size scaling is done at the same time in both $x,y$-directions. Despite its limitation, we show that it still sheds some light into the critical scaling exponents. For HC lattice, in Figs. \[scalexy\](a) and \[scalexy\](b), a finite size scaling of $S_{AF}^{xy}$ by using the scaling function $S_{AF}^{xy}/N_s\propto L_{x}^{-2\beta/\nu}f(L_{x}^{1/\nu}(U-U_c))$ gives the critical exponents $\beta=0.31,\nu=0.70$. For QMC simulations in Refs. [@Hohenadler2012; @Assaad2013a; @Assaad2013b], they extract the exponents $\beta=0.3486,\nu=0.6717$ in fully agreement with those of 3D XY model. In comparison, we can see that our DMRG results are in reasonable agreement with 3D XY universality class. While for CB lattice in Figs. \[scalexy\](c) and \[scalexy\](d), $S_{AF}^{zz}$ from different sizes can merge together by using the scaling function $S_{AF}^{zz}/N_s\propto L_{x}^{-2\beta/\nu}f(L_{x}^{1/\nu}(U-U_c))$ using the critical exponents $\beta=1/8,\nu=1$, which indicates that the phase transition falls into the 2D Ising universality class [@Pelissettoa2002; @Moukouri2012]. When $U$ approaches a critical value $U_c$, $F(U)$ shows a small bump, implying a peak of the fidelity susceptibility $\chi_F=(1-F(U))/(\delta U)^2$ which is a signature of phase transition [@Gu2010; @Saadatmand2017]. We obtain a similar picture for $\chi_F\propto L_{x}^{2/\nu}f(L_{x}^{1/\nu}(U-U_c))$. Thus we conjecture that this phase transition maybe belong to the 2D Ising universality class, which is different from that in HC lattice, although a stronger evidence of finite-size scaling in cylinder width is necessary.
 Numerical DMRG results on a cylinder lattice in the flat band limit with width $L_y=2N_y$ and length $L_x=N_x=18$ at half-filling. The evolutions of the absolute wavefunction overlap $F(U)$, spin structure factors $S_{AF}^{zz},S_{AF}^{xy}$, and entanglement entropy $S_{L}$ are shown for (a) the Haldane-honeycomb lattice with parameters $t'=0.6t,t''=0.58t,\phi=2\pi/5$ and (b) the checkerboard lattice with parameters $t'=0.3t,t''=-0.2t,\phi=\pi/4$, respectively. The smooth transition is characterized by the continuous behavior of these physical quantities.](flat.eps){height="1.5in" width="3.2in"}
Finally, we present our DMRG results of the phase diagram in the parameter plane $(U,t')$ without $t''$, as indicated in Figs. \[phase\](a) and \[phase\](b). First of all, we identify a CQPT separating the QSHE from the antiferromagnetic ground state on both HC and CB lattices, without the evidence of intermediate phase in between. The apparent non-vanishing spin pump is due to the fluctuating off-diagonal Berry curvature driven by interspecies correlation, as also identified from ED analysis in Fig. \[energyhc\](d). Here we do not consider the situation $t'\rightarrow0$ where the system is a gapless Dirac semimetal for both models, and the transition from such a Dirac semimetal to AFM has been claimed to be of Gross-Neveu universality class in several QMC simulations [@Assaad2013a; @Assaad2013b], which we leave for future study. Moreover, by including the next-next-nearest-neighbor hopping in the flat band limit, we obtain the similar physical picture of a continuous phase transition. As shown in Fig. \[flat\](a) and \[flat\](b), for both Haldane-honeycomb and $\pi$-flux checkerboard, the quantum phase transition is continuous, identified from three physical quantities including the absolute wavefunction overlap $F(U)$, spin structure factors $S_{AF}^{zz},S_{AF}^{xy}$, and entanglement entropy $S_{L}$.
Summary and Discussions {#summary}
=======================
In summary, using both ED and DMRG calculations, we have demonstrated a continuous phase transition from a quantum spin-Hall state to an antiferromagnetic Mott insulator driven by onsite Hubbard repulsion at half-filling, which is characterized by the continuous evolutions of the physical quantities, including the wave function fidelity, spin structure factors, entanglement entropy. The topological transition nature is encoded by the singular behavior of the Berry curvature driven by strong interspecies correlation, but the total charge Hall conductance remains unchanged. In close comparison, for an integer quantum Hall state with a symmetric $C$-matrix $\mathbf{C}=\begin{pmatrix}
1 & 0\\
0 & 1\\
\end{pmatrix}$ in both $\pi$-flux checkerboard and Haldane-honeycomb lattices with broken time-reversal symmetry, recent ED and DMRG studies show that a direct first-order level crossing occurs at one of high-symmetry twisted boundary conditions [@Zeng2017; @Vanhala2016] from integer quantum Hall effect (IQHE) to a trivial Mott insulator. Physically, these two classes of transitions indeed can belong to different classes as the transition between IQHE to a Mott insulator has a quantized jump of charge Chern number between two charge insulators, while the transition between QSHE and a Mott insulator only has a change of spin Chern number from a spin insulator to a gapless spin system. We believe our current work may provide a new insight into the interaction-driven topological transition nature. As one intriguing direction of our study, it is interesting and important to investigate the role of broken U(1)-spin symmetry by adding spin-orbit coupling in the transition between QSHE and the magnetic phase, which is very relevant to the transition metal oxide Na$_{2}$IrO$_{3}$ materials [@Shitade2009; @Jackeli2009], and we leave it for a future follow-up project.
W.Z. thanks Kai Sun for stimulating discussions. This work is supported by National Science Foundation Grants PREM DMR-1205734 (T.S.Z.) and DMR- 1408560 (D.N.S.) The work at Los Alamos was supported by the U.S. DOE Contract No. DE-AC52-06NA25396 through the LDRD Program (W.Z. & J.-X.Z.), and supported in part by the Center for Integrated Nanotechnologies, a U.S. DOE Office of Basic Energy Sciences user facility.
[99]{} L. D. Landau, Phys. Z. Sowjetunion [**11**]{}, 26 (1937). X. G. Wen, Int. J. Mod. Phys. B, [**04**]{}, 239 (1990). See also a recent review X. G. Wen, [arXiv:1610.03911.](https://arxiv.org/abs/1610.03911) C. N. Varney, K. Sun, M. Rigol, and V. Galitski, Phys. Rev. B [**84**]{}, 241105(R) (2011). A. Amaricci, J. C. Budich, M. Capone, B. Trauzettel, and G. Sangiovanni, Phys. Rev. Lett. [**114**]{}, 185701 (2015). B. Roy, P. Goswami, and J. D. Sau, Phys. Rev. B [**94**]{}, 041101(R) (2016). N. Samkharadze, K. A. Schreiber, G. C. Gardner, M. J. Manfra, E. Fradkin, and G. A. Csáthy, Nat. Phys. [**12**]{}, 191 (2016). L. Tsui, F. Wang, and D.-H. Lee, [arXiv:1511.07460.](https://arxiv.org/abs/1511.07460) C. Castelnovo and C. Chamon, Phys. Rev. B [**77**]{}, 054433 (2008). A. Hamma, W. Zhang, S. Haas, and D. A. Lidar, Phys. Rev. B 77, 155111 (2008). Y. Ran and X.-G. Wen, Phys. Rev. Lett. [**96**]{}, 026802 (2006). E.-G. Moon and C. Xu, Phys. Rev. B [**86**]{}, 214414 (2012). S. Whitsitt and S. Sachdev, Phys. Rev. B [**94**]{}, 085134 (2016). C. L. Kane and E. J. Mele, Phys. Rev. Lett. [**95**]{}, 146802 (2005). C. L. Kane and E. J. Mele, Phys. Rev. Lett. [**95**]{}, 226801 (2005). C. Xu, and J. E. Moore, Phys. Rev. B [**73**]{}, 045322 (2006). S. Rachel and K. Le Hur, Phys. Rev. B [**82**]{}, 075106 (2010). M. Hohenadler, T. C. Lang, and F. F. Assaad, Phys. Rev. Lett. [**106**]{}, 100403 (2011). D. Zheng, G.-M. Zhang, and C. Wu, Phys. Rev. B [**84**]{}, 205121 (2011). M. Hohenadler, Z. Y. Meng, T. C. Lang, S. Wessel, A. Muramatsu, and F. F. Assaad, Phys. Rev. B [**85**]{}, 115132 (2012). F. F. Assaad, M. Bercx, and M. Hohenadler, Phys. Rev. X [**3**]{}, 011015 (2013). F. F. Assaad and I. F. Herbut, Phys. Rev. X [**3**]{}, 031010 (2013). M. Hohenadler, F. Parisen Toldin, I. F. Herbut, and F. F. Assaad, Phys. Rev. B [**90**]{}, 085146 (2014). S.-L. Yu, X. C. Xie, and J.-X. Li, Phys. Rev. Lett. [**107**]{}, 010401 (2011). J. C. Budich, R. Thomale, G. Li, M. Laubach, and S.-C. Zhang, Phys. Rev. B [**86**]{}, 201407(R) (2012). D. Pesin and L. Balents, Nat. Phys. [**6**]{}, 376 (2010). S. Murakami, New J. Phys. [**9**]{}, 356 (2007); New. J. Phys. [**10**]{}, 029802(E) (2008). A. Vaezi, M. Mashkoori, and M. Hosseini, Phys. Rev. B [**85**]{}, 195126 (2012). Y.-H. Chen, H.-H. Hung, G. Su, G. A. Fiete, and C. S. Ting, Phys. Rev. B [**91**]{}, 045122 (2015). A. Rüegg and G. A. Fiete, Phys. Rev. Lett. [**108**]{}, 046401 (2012). J. Reuther, R. Thomale, and S. Rachel, Phys. Rev. B [**86**]{}, 155127 (2012). W. Wu, S. Rachel, W.-M. Liu, and K. Le Hur, Phys. Rev. B [**85**]{}, 205102 (2012). T. Liu, B. Douçot, and K. Le Hur, Phys. Rev. B [**88**]{}, 245119 (2013). M. Laubach, J. Reuther, R. Thomale, and S. Rachel, Phys. Rev. B [**90**]{}, 165136 (2014). T. C. Lang, A. M. Essin, V. Gurarie, and S. Wessel, Phys. Rev. B [**87**]{}, 205101 (2013). H.-H. Hung, V. Chua, L. Wang, and G. A. Fiete, Phys. Rev. B [**89**]{}, 235104 (2014). D.-H. Lee, Phys. Rev. Lett. [**107**]{}, 166806 (2011). C. Griset, and C. Xu, Phys. Rev. B [**85**]{}, 045123 (2012). M. Hohenadler and F. F. Assaad, J. Phys.: Condens. Matter [**25**]{}, 143201 (2013). Y.-F. Wang, Z.-C. Gu, C.-D. Gong, and D. N. Sheng, Phys. Rev. Lett. [**107**]{}, 146803 (2011). K. Sun, Z. Gu, H. Katsura, and S. Das Sarma, Phys. Rev. Lett. [**106**]{}, 236803 (2011). D. N. Sheng, L. Balents, and Z. Wang, Phys. Rev. Lett. [**91**]{}, 116802 (2003). D. N. Sheng, Z.-Y. Weng, L. Sheng, and F. D. M. Haldane, Phys. Rev. Lett. [**97**]{}, 036808 (2006). L. Sheng, D. N. Sheng, C. S. Ting, and F. D. M. Haldane, Phys. Rev. Lett. [**95**]{}, 136602 (2005). T.-S. Zeng, W. Zhu, and D. N. Sheng, Phys. Rev. B [**95**]{}, 125134 (2017). M. Nakagawa and S. Furukawa, Phys. Rev. B [**95**]{}, 165116 (2017). A. C. M. Carollo and J. K. Pachos, Phys. Rev. Lett. [**95**]{}, 157203 (2005). S. L. Zhu, Phys. Rev. Lett. [**96**]{}, 077206 (2006). T. Yoshida, S. Fujimoto, and N. Kawakami, Phys. Rev. B [**85**]{}, 125113 (2012). S. Miyakoshi and Y. Ohta, Phys. Rev. B [**87**]{}, 195133 (2013). S. S. Gong, W. Zhu, and D. N. Sheng, Sci. Rep. [**4**]{}, 6317 (2014). M. P. Zaletel, R. Mong, and F. Pollmann, J. Stat. Mech. P10007 (2014). S.-S. Gong, W. Zhu, D. N. Sheng, O. I. Motrunich, and M. P. A. Fisher, Phys. Rev. Lett. [**113**]{}, 027201 (2014). S.-S. Gong, D. N. Sheng, O. I. Motrunich, and M. P. A. Fisher, Phys. Rev. B [**88**]{}, 165138 (2013). A. Pelissettoa and E. Vicari, Phys. Rep. [**368**]{}, 549 (2002). S. Moukouri and E. Eidelstein, Phys. Rev. B [**86**]{}, 155112 (2012). S.-J. Gu, Int. J. Mod. Phys. B [**24**]{}, 4371 (2010). S. N. Saadatmand, and I. P. McCulloch, Phys. Rev. B [**96**]{}, 075117 (2017). T. I. Vanhala, T. Siro, L. Liang, M. Troyer, A. Harju, and P. Törmä, Phys. Rev. Lett. [**116**]{}, 225305 (2016). A. Shitade, H. Katsura, J. Kuneš, X.-L. Qi, S.-C. Zhang, and N. Nagaosa, Phys. Rev. Lett. [**102**]{}, 256403 (2009). G. Jackeli and G. Khaliullin, Phys. Rev. Lett. [**102**]{}, 017205 (2009).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The ability to create artificial intelligence (AI) capable of performing complex tasks is rapidly outpacing our ability to ensure the safe and assured operation of AI-enabled systems. Fortunately, a landscape of AI safety research is emerging in response to this asymmetry and yet there is a long way to go. In particular, recent simulation environments created to illustrate AI safety risks are relatively simple or narrowly-focused on a particular issue. Hence, we see a critical need for AI safety research environments that abstract essential aspects of complex real-world applications. In this work, we introduce the [AI Safety TanksWorld]{}as an environment for AI safety research with three essential aspects: competing performance objectives, human-machine teaming, and multi-agent competition. The [AI Safety TanksWorld]{}aims to accelerate the advancement of safe multi-agent decision-making algorithms by providing a software framework to support competitions with both system performance and safety objectives. As a work in progress, this paper introduces our research objectives and learning environment with reference code and baseline performance metrics to follow in a future work.'
author:
- |
Corban G. Rivera, Olivia Lyons, Arielle Summitt, Ayman Fatima, Ji Pak, William Shao,\
**Robert Chalmers, Aryeh Englander, Edward W. Staley, I-Jeng Wang, Ashley J. Llorens**\
*Intelligent Systems Center*\
*Johns Hopkins Applied Physics Lab*\
11100 Johns Hopkins Rd., Laurel, MD 20723\
Corresponding author: [email protected]
bibliography:
- 'references.bib'
title: 'TanksWorld: A Multi-Agent Environment for AI Safety Research'
---
Introduction
============
Emerging paradigms in machine learning (e.g., reinforcement learning) offer the potential for systems to learn complex behaviors through interaction with a learning environment. These advancements could help overcome the limitations of current autonomous systems, which largely rely on pre-determined rulesets and analytic control regimes to govern their behavior. Many open technical challenges must be overcome to realize this potential, including advancements in the safety and assurance of autonomous systems - especially goal-driven systems that learn for themselves. Recent works from the AI safety research community have identified a number of open challenges in this area. However, most existing AI testbeds and simulation environments do not explicitly address AI safety challenges. Those that do are relatively simplistic, such as gridworld environments, or narrowly-focused on isolating and illustrating a particular safety issue.
![The [AI Safety TanksWorld]{}is a competitive multi-agent environment for exploring competing performance objectives, human-machine teaming, and multi-agent competition []{data-label="fig:tanksworld"}](img/tanksworld.png){width="1.0\linewidth"}
Hence, there is a gap in environments that support AI safety research based on simulations that are sufficiently realistic to capture essential aspects of real-world AI applications while not requiring a prohibitive degree of domain expertise. To begin working towards addressing this gap, we introduce the [AI Safety TanksWorld]{}. The [AI Safety TanksWorld]{}is a multi-agent environment for exploring AI safety issues in applications that require competition and cooperation among multiple agents towards satisfying competing performance objectives. Specifically, the [AI Safety TanksWorld]{}is a team-based tanks battle where one team aims to defeat the other while avoiding unintended consequences such as losing teammates or inflicting collateral damage on the environment. We highlight three essential components of the environment: dynamic and uncertain environments, safety concerns for human and machine teams, and complex tasks with competing objectives.
*Dynamic and Uncertain Environments –* Many of the challenges surrounding AI safety arise from unknowns and uncertainty in the environment. For AI safety research, there is an important distinction between known-unknowns and unknown-unknowns. A known-unknown can often be characterized via a model that captures a degree of partial observability or structured uncertainty. Although challenging, an AI system can often be trained to mitigate the safety risks associated with known-unknowns. Real-world environments usually offer the even greater challenge of unknown-unknowns - novel situations not adequately represented during training experiences. The [AI Safety TanksWorld]{}environment includes parameters that enable the complexity, observably, and novelty of the scenario to be systematically modified. For example, each tank in our simulation can only sense a local region of the environment, making communication among allies advantageous. Our explicit modeling of the range and quality of communication enables the ability to add complexity in unexpected ways.
*AI Safety for Human-Machine Teams –* Human-machine teaming can present unique challenges for learning agents. Even when people are highly skilled at performing a given task, they may take actions that do not optimize for near-term rewards either unintentionally or intentionally due to hidden objectives. This unpredictability creates the need for human-aware adaptation in goal-driven agents. This can be particularly challenging in safety-critical applications that require safe exploration of possible actions. Further complicating the human-machine teaming relationship is that humans also need to develop accurate models of the behavior of machine teammates even after training is complete. To support research in the area of human-machine teams, the [AI Safety TanksWorld]{}includes several human surrogate policies built using behavior cloning [@dart] from human demonstrations.
*Complex Tasks with Competing Objectives –* Lastly, due to the dynamic and competitive nature of the [AI Safety TanksWorld]{}, AI agents are forced to tradeoff between the performance of a complex task (i.e., collaborate to defeat the opposing tanks) and safety objectives. This tension supports a broad exploration of methods that attempt to optimize over multiple, competing objectives. This can be particularly challenging when human developers and operators may desire to express specific preferences for certain outcomes over others. For example, a certain amount of collateral damage to property in an operating environment may be tolerable with less being preferable and more being unacceptable. Optimizing performance in these risk-sensitive regimes is an open research challenge.
Related Work
============
The AI community has a rich history of developing and adopting simulations and games to challenge the research community. Checkers, Chess [@deepblue], and Go [@alphago; @alphazero] have historically been grand challenge problems in AI. Variants of these classics have produced new AI challenges [@rbmc]. With a groundbreaking result from DeepMind [@ataridqn], a suite of Atari games became benchmarks for exploring performance of algorithms across tasks. OpenAI introduced the Gym architecture along with a suite of environments for exploring the performance of reinforcement learning algorithms in classic control, robotics, and Atari environments [@openaigym]. DeepMind released their internally developed suite of benchmark tasks for evaluating the performance of reinforcement learning algorithms [@deepmindlab]. Unity released the obstacle tower challenge environment, which provides a range of difficulty with the increasing capabilities of reinforcement learning agents [@tower]. OpenAI released Neural MMO as an environment to test AI in highly multi-agent environments [@neural_mmo].
All of these are very useful simulation environments, but they are not specifically focused on AI Safety issues. The [AI Safety TanksWorld]{}, by contrast, is specifically focused on safety concepts related to competing performance objectives, human-machine teaming, and multi-agent competition.
The AI community has also developed several test suites for AI safety scenarios. For example, a team from DeepMind released AI Safety Gridworlds [@gridworlds] as a series of environments to illustrate concrete problems in AI safety [@concrete]. However, the Gridworlds scenarios present only minimalistic examples of each AI safety concept. While the minimalism of the Gridworlds is useful for illustration, it may be beneficial to have environments that reflect more realistic challenges for AI safety. The existing AI Safety Gridworlds are also not specifically designed to explore issues that arise around human-machine teams, since each Gridworld environment consists of only a single AI-controlled entity. Other platforms like the Safety Gym [@safetygym] focus specifically on safe exploration in single agent environments.
Additionally, most of the Gridworlds scenarios present only the most difficult variations of each safety problem. This is very useful for illustrative purposes, but we believe that it would be useful to have less difficult variations of the same problems which could then be scaled up to address the more difficult varieties. For example, the “avoiding side effects" Gridworlds scenario requires the agent to avoid pushing a box into a corner in pursuit of its goal, when the reward function it is given does not include anything at all about boxes. Essentially, the agent has to be able to successfully deal with an “unknown unknown," which is an especially difficult problem. It may be useful for the community to first try to tackle easier “known unknowns" problems, with the hope of using some of the lessons to scale up to the more difficult “unknown unknowns" scenarios.
The rise of autonomous vehicles is a clear example of the AI safety hazards associated with human-machine teams. The autonomous vehicle research community has released simulation frameworks for training and evaluation of self-driving cars [@car]. While self-driving cars are an excellent example of AI safety issues surrounding human-machine teams, these simulations tend to provide a limited view of AI safety through the lens of individual cars.
The [AI Safety TanksWorld]{}Environment
=======================================
The [AI Safety TanksWorld]{}is a team-based N vs. N tanks battle (where N is a variable parameter) that motivates the design of safe multi-agent control policies that are effective, collaborative and cope with uncertainty. Figure \[fig:tanksworld\] illustrates a multi-agent scenario. The components of the environment are largely derived from assets distributed by Unity.
State and Action Representation
-------------------------------
Scenarios are composed of a closed arena, obstacles, buildings, trees, and tanks. The [AI Safety TanksWorld]{}contains multiple views including a tank-centric first person view, a top down overview, and an isometric view. The state space is a 128x128 4-channel image where each channel conveys different information. The 4-channels include: position and orientation of allies, position and orientation of visible threats, position of neutral tanks, and position of obstacles. The state space given to each tank is unique in that the 4-channel state is rotated and translated to be relative to each tank’s position and heading. The threats visible to each tank are dependent on its position relative to allies and threats as described in the Active Sensing and Teamwork section. Image \[fig:tanksworld\_state\] illustrates the unique state provided to each agent.
Three continuous actions are available to each agent: velocity (forward/reverse), turning (left/right), and shooting (yes/no). All actions are in the range $-1$ to $1$. Shots are taken for actions greater than zero, and shot frequency is rate limited.
![State is conveyed as a tank-relative 4-channel 128x128 image. The channels include position of allies, position of threats, position of neutral entities, and position of obstacles. The first three channels are shown as blue, green, and red respectively. (Top row) shows state for tanks on the red team. (Bottom row) shows states for tanks on the blue team.[]{data-label="fig:tanksworld_state"}](img/tanksworld_state.png){width="\linewidth"}
Parameterization
----------------
The [AI Safety TanksWorld]{}environment includes the following parameters, which may be selected to emphasize different research challenges:
- Communication range
- Number of neutral tanks and obstacles
- Control policies of teammates
*Communication range –* We parameterize communication range among teammates as a representation of environmental uncertainty. Since teammates may share information about the location of opposing tanks, this model controls the overall visibility of the threat environment. Opposing tanks can be seen if they are within the designated radius of a teammate and the teammate is within the radius of the current tank. This dynamic encourages teammates to stay close enough to each other to share information and far enough apart to maintain broad situational awareness of the environment.
*Number of neutral tanks and obstacles –* The number of neutral tanks and density of obstacles can be varied to control the risk of collateral damage. The neutral tanks move around the scene at random creating hazards for both teams. The density of obstacles including trees, rocks, and buildings is also controlled by a parameter. The positions of obstacles are randomized on reset.
*Control policies of teammates –* We parameterize the skill level of teammates to represent the variability that can arise in human-machine teams to the unpredictability of human decision-making. Human surrogate teammates are policies that may be obtained via behavior cloning [@dart] from human demonstrations or through explicit behavior modeling.
Rewards
-------
The environment returns separate metrics for allied, neutral, and opponent kills. The flexibility in how a reward function or constraint can be defined based on these metrics will help researchers explore AI safety across a spectrum of performance-risk trade-offs. The specific numbers described here are used for illustration.
One possible reward scheme is: the penalty for death is $1$, the penalty for destroying neutral tanks is $1$, and the penalty for allied kills is $1$. With two neutral tanks in the scene, the minimum possible score is $-7$. Given a $1$ reward for each enemy kill and 5 enemy tanks, the maximum possible reward is $5$. The components of the reward (allied, opponent, and neutral kills) are returned separately for each tank.
Discussion
==========
Active Sensing and Teamwork
---------------------------
A relatively uncommon aspect of the [AI Safety TanksWorld]{}that differentiates it from other environments is that each tank receives a unique and partial view of the world that is governed by its teamwork with allies. Within a parameterized distance, allies can communicate the threats that they observe. The communication between allies is transitive. The benefit of team coordination is better threat visibility for all allies.
![Partial observability and communications. From the perspective of Blue1, Red2 is visible because Blue1 is within a parameterized radius of Blue2 and Blue2 is within a specified radius to Red2. Red1 is not visible to either Blue1 or Blue2 because neither one is within the specified radius to Red1 or Blue3. []{data-label="fig:ppo_results"}](img/tank_comms.png){width=".5\linewidth"}
The parameter controlling communication can alter the scenario from full observability to partial observability. Distributional shifts in this parameter force policies to cope with situations that resemble unreliable communications.
Minimizing Collateral Damage
----------------------------
One of the concrete [@concrete] problems in AI safety is avoiding unintended consequences. This is made more challenging in human-machine teaming because of the difficulty in reliably modeling human decision-making in dynamic scenarios.
Neutral tanks in the [AI Safety TanksWorld]{}environment present the risk of generating collateral damage. Distributional shifts in the number of neutral tanks in the environment present additional collateral damage risk.
Human-machine Teaming
---------------------
AI that have been trained to partner with other AI teammates may be ill equipped to partner with humans. To model human teammate behavior, we can substitute control policies for teammates with human surrogate policies. We learn these human surrogate policies from demonstration using behavior cloning [@dart]. By substituting the policies of teammates with human surrogate policies, we can evaluate the challenges and safety issues that arise from an AI being partnered with human-like teammates of variable skill.
Next Steps
----------
Our ultimate goal is to host a competition for the AI research community focused on the AI safety aspects of the [AI Safety TanksWorld]{}. In the coming months, we are interested in simultaneously exploring parameterizations of the environment that elicit different AI safety challenges. We will establish expected performance using recent reinforcement learning algorithms. Once the competition rules are stable and baselines established, we will aim to host a competition track at an established AI workshop. By publishing this work in progress, we hope to attract both feedback and potential collaborators.
Acknowledgements
================
The authors would like acknowledge the APL CIRCUIT program and its organizers for the training and coordination of the interns on this project.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We carry out a systematic study of entanglement entropy in relativistic quantum field theory. This is defined as the von Neumann entropy $S_A=-{\rm Tr}\,\rho_A\log\rho_A$ corresponding to the reduced density matrix $\rho_A$ of a subsystem $A$. For the case of a 1+1-dimensional critical system, whose continuum limit is a conformal field theory with central charge $c$, we re-derive the result $S_A\sim(c/3)\log\ell$ of Holzhey et al. when $A$ is a finite interval of length $\ell$ in an infinite system, and extend it to many other cases: finite systems, finite temperatures, and when $A$ consists of an arbitrary number of disjoint intervals ([**See note added**]{}). For such a system away from its critical point, when the correlation length $\xi$ is large but finite, we show that $S_A\sim{\cal A}(c/6)\log\xi$, where $\cal A$ is the number of boundary points of $A$. These results are verified for a free massive field theory, which is also used to confirm a scaling ansatz for the case of finite-size off-critical systems, and for integrable lattice models, such as the Ising and XXZ models, which are solvable by corner transfer matrix methods. Finally the free-field results are extended to higher dimensions, and used to motivate a scaling form for the singular part of the entanglement entropy near a quantum phase transition.'
author:
- 'Pasquale Calabrese$^1$ and John Cardy$^{1,2}$'
title: Entanglement Entropy and Quantum Field Theory
---
Introduction.
=============
Recently there has been considerable interest in formulating measures of quantum entanglement and applying them to extended quantum systems with many degrees of freedom, such as quantum spin chains.
One of these measures[@Bennett] is entanglement *entropy*. Suppose the whole system is in a pure quantum state $|\Psi\rangle$, with density matrix $\rho=|\Psi\rangle\langle\Psi|$, and an observer A measures only a subset $A$ of a complete set of commuting observables, while another observer B may measure the remainder. A’s reduced density matrix is $\rho_A={\rm Tr}_B\,\rho$. The entanglement entropy is just the von Neumann entropy $S_A=-{\rm Tr}_A\,\rho_A\log\rho_A$ associated with this reduced density matrix. It is easy to see that $S_A=S_B$. For an unentangled product state, $S_A=0$. Conversely, $S_A$ should be a maximum for a maximally entangled state.**
For example, for a system with two binary (spin-$\frac12$) degrees of freedom, with $|\Psi\rangle=\cos\theta|\uparrow\downarrow\rangle+\sin\theta|\downarrow
\uparrow\rangle$, where A observes only the first spin and B the second, $S_A$ takes its maximum value of $\log 2$ when $\cos^2\theta=\frac12$, which agrees with our intuitive idea of maximal entanglement. For this system it has been shown,[@Bennett] even in the partially entangled case, that if there are $M$ copies of the state available, by making only local operations A can produce $M'<M$ states which are maximally entangled. The optimal conversion ratio $M'/M$ is given, for large $M$, by $S_A$.[^1]
Although there are other measures of entanglement,[@othermeasures] the entropy is most readily suited to analytic investigation. In several papers[@Vidal; @Korepin; @leb; @Casini], the concept has been applied to quantum spin chains. Typically, the subset $A$ consists of a commuting set of components of the spin degrees of freedom in some interval of length $\ell$, in an infinitely long chain. It is found that the entanglement entropy generally tends to a finite value as $\ell$ increases, but that this value diverges as the system approaches a quantum critical point. At such a critical point, the entropy grows proportional to $\log\ell$ for large $\ell$.
Close to a quantum critical point, where the correlation length $\xi$ is much larger than the lattice spacing $a$, the low-lying excitations and the long-distance behaviour of the correlations in the ground state of a quantum spin chain are believed to be described by a quantum field theory in 1+1 dimensions. If the dispersion relation of the low-lying excitations is linear for wave numbers $k$ such that $\xi^{-1}\ll|k|\ll a^{-1}$, the field theory is relativistic. We shall consider only those cases in this paper. At the critical point, where $\xi^{-1}=0$, the field theory is massless, and is a *conformal *field theory (CFT).**
In this case, the von Neumann entropy of subsystem corresponding to an interval of length $\ell$ was calculated some time ago by Holzhey et al.[@Holzhey], in the context of black hole physics (although that connection has been questioned), where it was termed ‘geometric’ entropy. Using methods of conformal field theory, based in part on earlier work of Cardy and Peschel[@CardyPeschel], they found $S_A\sim (c/3)\log(\ell/a)$, where $c$ is the conformal anomaly number (sometimes called the central charge) of the corresponding CFT.
This result has been verified by analytic and numerical calculations on integrable quantum spin chains corresponding to CFTs with $c=\frac12$ and $c=1$. [@Vidal; @Korepin; @leb]
In this paper, we first put the CFT arguments of Holzhey et al.[@Holzhey] on a more systematic basis, and generalise their result in a number of ways. Our methods are based on a formula for the entropy in terms of the partition function in the path integral formulation of the quantum theory as a euclidean field theory on an $n$-sheeted Riemann surface, in the limit $n\to1$.
For a 1+1-dimensional theory at a critical point, we derive analogous formulas for the entropy in the cases when the subsystem $A$ consists of an *arbitrary *number of disjoint intervals of the real line (see Eq. \[general\]), and when the whole system has itself a finite length $L$. For example, for the case when $A$ is a single interval of length $\ell$, and periodic boundary conditions are imposed on the whole system, we find $$S_A=(c/3)\log\big((L/\pi a)\sin(\pi\ell/L)\big)+c_1'\,.$$ On the other hand, for a finite system of total length $L$ with open boundaries, divided at some interior point into an interval $A$ of length $\ell$ and its complement, we find $$S_A=(c/6)\log\big((L/\pi a)\sin(\pi\ell/L)\big)+2g+c_1'\,,$$ where $g$ is the boundary entropy of Affleck and Ludwig[@AffleckLudwig]. We also treat the case when the system is infinitely long but is in a thermal mixed state at finite temperature $\beta^{-1}$: $$\label{finiteT}
S_A=(c/3)\log\big((\beta/\pi a)\sinh(\pi\ell/\beta)\big)+c_1'\,.$$ In all these cases, the constant $c_1'$ is the same, but non-universal.**
For a massive 1+1-dimensional relativistic QFT (which corresponds to an off-critical quantum spin chain where the correlation length $\xi\gg a$) the simplest results are for an infinite system divided at some point into two semi-infinite pieces. In this case we verify that the entanglement entropy is finite, and derive the universal formula $$\label{noncrit}
S_A\sim (c/6)\log(\xi/a)\,.$$ In the more general case when $A$ consists of a collection of disjoint intervals, each of length $\gg\xi$, we expect (\[noncrit\]) to be multiplied by a factor $\cal A$ which counts the number of boundary points between $A$ and $B$ (the 1d analogue of surface area.)
For ${\cal A}=1$ the entropy is exactly calculable in the case of a free field theory. We verify the above formula, and exhibit the finite-size cross-over which occurs when $\xi$ is of the size $L$ of the system. For a lattice model in this geometry, with $L$ infinite, we point out that $\rho_A$ is simply related to Baxter’s corner transfer matrix, and thus, for integrable models whose weights satisfy a Yang-Baxter relation, all its eigenvalues can be determined exactly. We treat explicitly the case of the Ising model and its anisotropic limit, the transverse Ising spin chain, and also the XXZ model, computing exactly at all values of the coupling the finite part of the entropy $S$. This agrees with our continuum result (\[noncrit\]) for $c=\frac12,1 $ when the correlation length $\xi$ is large.
The analysis for the free theory is straightforward to extend to higher dimensions, at least in suitable geometries. This leads to the well-known law, first found by Srednicki[@s-93], that the entropy should be proportional to the surface area $\cal A$ of the subsystem $A$. As pointed out by Srednicki, the coefficient of this term, for $d>1$, depends on the UV cut-off and is therefore non-universal. However, based on our calculations, we propose that there should be a non-leading piece in $S_A/{\cal A}$, proportional to $\xi^{-(d-1)}$, which depends in a singular way on the couplings near a quantum phase transition, and whose form is, moreover, *universal*.**
The layout of this paper is as follows. In the next section, we discuss the entropy in terms of the euclidean path integral on an $n$-sheeted Riemann surface. In Sec. \[sectcft\] we consider the 1+1-dimensional conformal case. We use the powerful methods of CFT to show that the partition function on the Riemann surface of interest is given, up to a constant, by a calculable correlation function of vertex operators in a $c=1$ CFT in the complex plane. Similar results apply to a system with boundaries. This general result allows us to derive all the special cases described above. In Sec. \[sectmass\] we consider the case of a massive 1+1-dimensional field theory. We derive the result $S\sim(c/6)\log(\xi/a)$ from completely general properties of the stress tensor in the relevant geometry. This is then verified for a free bosonic massive field. In the last part of this section we relate the lattice version of this problem to the corner transfer matrix, and compute the off-critical entropy for the case of the Ising and XXZ Heisenberg spin chains. In Sec. \[sectfss\] we study the off-critical case in a finite-size system, propose a scaling law, and verify it for the case of a free massive field theory. We compute the relevant scaling function in a systematic sequence of approximations. Sec. \[secthd\] is devoted to the discussion of higher dimensions.
While all this work was being carried out, some other related papers have appeared in the literature. In Ref. [@Korepin2], the result $S_A\sim(c/3)\log(\ell/a)$, first found by Holzhey et al.[@Holzhey], was obtained by the following argument: it was assumed that the entropy should be conformally invariant and therefore some function $F(x)$ of the variable $x=(\beta/\pi)\sinh(\pi\ell/\beta)$; by comparing with known case $\ell\gg\beta$ [@BCN; @Affleck] it was observed that $F(x)$ should behave as $(c/3)\log x$ as $x\to\infty$; finally it was assumed (with no justification being given) that the $\log x$ form for $F$ is valid for all values of $x$; finally the limit $\beta\gg\ell$ was taken. In the present paper, we should stress, we have derived all these statements from first principles of CFT.
Very recently Casini and Huerta[@Casini] have considered the case of two intervals $A$ and $B$, and argued that the quantity $F(A,B)\equiv S(A)+S(B)-S(A\cap B)-S(A\cup B)$ is UV finite as $a\to0$, and is given by a universal logarithmic function of the cross-ratio of the four end points. This corresponds to, and agrees with, our case $N=2$. These authors, however, assume the conformal invariance of the entropy, while, once again, we stress that in the present paper we derive this from fundamental properties of the stress tensor. These authors also give a very nice alternative derivation of Zamolodchikov’s $c$-theorem[@Zam] based on this quantity $F(A,B)$.
von Neumann Entropy and Riemann surfaces. {#sectrs}
=========================================
Consider a lattice quantum theory in one space and one time dimension, initially on the infinite line. The lattice spacing is $a$, and the lattice sites are labelled by a discrete variable $x$. The domain of $x$ can be finite, i.e. some interval of length $L$, semi-infinite, or infinite. Time is considered to be continuous. A complete set of local commuting observables will be denoted by $\{\hat\phi(x)\}$, and their eigenvalues and corresponding eigenstates by $\{\phi(x)\}$ and $\otimes_x|\{\phi(x)\}\rangle$ respectively. For a bosonic lattice field theory, these will be the fundamental bosonic fields of the theory; for a spin model some particular component of the local spin. The dynamics of the theory is described by a time-evolution operator $\hat H$. The density matrix $\rho$ in a thermal state at inverse temperature $\beta$ is $$\rho(\{\phi(x'')''\}|\{\phi(x')'\})=
Z(\beta)^{-1}\langle\{\phi(x'')''\}|e^{-\beta\hat
H}|\{\phi(x')'\}\rangle\,,$$ where $Z(\beta)={\rm Tr}\,e^{-\beta\hat H}$ is the partition function.
This may be expressed in the standard way as a (euclidean) path integral: $$\label{pathi}
\rho=Z^{-1}\int[d\phi(x,\tau)]
\prod_x\delta(\phi(x,0)-\phi(x')')\prod_x
\delta(\phi(x,\beta)-\phi(x'')'')\,e^{-S_E}\,,$$ where $S_E=\int_0^\beta L_Ed\tau$, with $L_E$ the euclidean lagrangian. (For a spin model this would be replaced by a coherent state path integral.)
The normalisation factor of the partition function ensures that ${\rm Tr}\,\rho=1$, and is found by setting $\{\phi(x)''\}=\{\phi(x)'\}$ and integrating over these variables. This has the effect of sewing together the edges along $\tau=0$ and $\tau=\beta$ to form a cylinder of circumference $\beta$.
Now let $A$ be a subsystem consisting of the points $x$ in the disjoint[^2] intervals $(u_1,v_1),\ldots,(u_N,v_N)$. An expression for the the reduced density matrix $\rho_A$ may be found from (\[pathi\]) by sewing together only those points $x$ which are not in $A$. This will have the effect of leaving open cuts, one for each interval $(u_j,v_j)$, along the the line $\tau=0$.
We may then compute ${\rm Tr}\,\rho_A^n$, for any positive integer $n$, by making $n$ copies of the above, labelled by an integer $k$ with $1\leq k\leq n$, and sewing them together cyclically along the the cuts so that $\phi(x)'_k=\phi(x)''_{k+1}$ (and $\phi(x)'_n=\phi(x)''_1$) for all $x\in A$. Let us denote the path integral on this $n$-sheeted structure by $Z_n(A)$. Then $$\label{ZoverZ}
{\rm Tr}\,\rho_A^n={Z_n(A)\over Z^n}\,.$$
Now, since ${\rm Tr}\,\rho_A^n=\sum_\lambda\lambda^n$, where $\lambda$ are the eigenvalues of $\rho_A$ (which lie in $[0,1)$,) and since ${\rm Tr}\,\rho_A=1$, it follows that the left hand side is absolutely convergent and therefore analytic for all ${\rm Re}\,n>1$. The derivative wrt $n$ therefore also exists and is analytic in the region. Moreover, if the entropy $\rho_A=-\sum_\lambda\lambda\log\lambda$ is finite, the limit as $n\to1+$ of the first derivative converges to this value.
We conclude that the right hand side of (\[ZoverZ\]) has a unique analytic continuation to ${\rm Re}\,n>1$ and that its first derivative at $n=1$ gives the required entropy: $$S_A=-\lim_{n\to1}{\partial\over\partial n}{Z_n(A)\over
Z^n}\,.$$ (Note that even before taking this limit, (\[ZoverZ\]) gives an expression for the Tsallis entropy[@Tsallis] $({\rm Tr}\,\rho_{A}^n-1)/(1-n)$.)
So far, everything has been in the discrete space domain. We now discuss the continuum limit, in which $a\to0$ keeping all other lengths fixed. The points $x$ then assume real values, and the path integral is over fields $\phi(x,\tau)$ on an $n$-sheeted Riemann surface, with branch points at $u_j$ and $v_j$. In this limit, $S_E$ is supposed to go over into the euclidean action for a quantum field theory. We shall restrict attention to the case when this is Lorentz invariant, since the full power of relativistic field theory can then be brought to bear. The behaviour of partition functions in this limit has been well studied. In two dimensions, the logarithm of a general partition function $Z$ in a domain with total area $\cal A$ and with boundaries of total length $\cal L$ behaves as $$\label{divs}
\log Z=f_1{\cal A}a^{-2}+f_2{\cal L}a^{-1}+\ldots$$ where $f_1$ and $f_2$ are the non-universal bulk and boundary free energies. Note, however, that these leading terms *cancel *in the ratio of partition functions in (\[ZoverZ\]). However, as was argued by Cardy and Peschel[@CardyPeschel] there are also *universal *terms proportional to $\log a$. These arise from points of non-zero curvature of the manifold and its boundary. In our case, these are conical singularities at the branch points. In fact, as we shall show, it is precisely these logarithmic terms which give rise to the non-trivial dependence of the final result for the entropy on the short-distance cut-off $a$. For the moment let us simply remark that, in order to achieve a finite limit as $a\to0$, the right hand side of (\[ZoverZ\]) should be multiplied by some renormalisation constant ${\cal Z}(A,n)$. Its dependence on $a$ will emerge from the later analysis.****
Entanglement entropy in 2d conformal field theory. {#sectcft}
==================================================
Now specialise the discussion of the previous section to the case when the field theory is relativistic and massless, i.e. a conformal field theory (CFT), with central charge $c$, and initially consider the case of zero temperature.
We show that in this case the ratio of partition functions in (\[ZoverZ\]) is the same as the correlation function arising from the insertion of primary scaling operators $\Phi_n(u_j)$ and $\Phi_{-n}(v_j)$, with scaling dimensions $X_n=2\Delta_n=(c/12)(1-1/n^2)$, into each of the $n$ (disconnected) sheets. Moreover, this $2N$-point correlation function is computable from the Ward identities of CFT.
In the language of string theory, the objects we consider are correlators of orbifold points in theories whose target space consists of $n$ copies of the given CFT. We expect that some of our results may therefore have appeared in the literature of the subject.
Single interval
---------------
We first consider the case $N=1$ and no boundaries, that is the case considered by Holzhey et al.[@Holzhey] of a single interval of length $\ell$ in an infinitely long 1d quantum system, at zero temperature. The conformal mapping $w\to\zeta=(w-u)/(w-v)$ maps the branch points to $(0,\infty)$. This is then uniformised by the mapping $\zeta\to z=\zeta^{1/n}=
\big((w-u)/(w-v)\big)^{1/n}$. This maps the whole of the $n$-sheeted Riemann surface ${\cal R}_n$ to the $z$-plane $\bf C$. Now consider the holomorphic component of the stress tensor $T(w)$. This is related to the transformed stress tensor $T(z)$ by[@BPZ] $$\label{schwartz}
T(w)=(dz/dw)^2\,T(z)+\ffrac c{12}\{z,w\}\,,$$ where $\{z,w\}$ is the Schwartzian derivative $(z'''z'-\frac32{z''}^2)/{z'}^2$. In particular, taking the expectation value of this, and using $\langle
T(z)\rangle_{\bf C}=0$ by translational and rotational invariance, we find $$\langle T(w)\rangle_{{\cal R}_n}=\frac c{12}\{z,w\}=
{c(1-(1/n)^2)\over 24}{(v-u)^2\over(w-u)^2(w-v)^2}\,.$$ This is to be compared with the standard form[@BPZ] of the correlator of $T$ with two primary operators $\Phi_n(u)$ and $\Phi_{-n}(v)$ which have the same complex scaling dimensions $\Delta_n=\overline\Delta_n=
(c/24)(1-(1/n)^2)$: $$\label{TPP}
\langle T(w)\Phi_n(u)\Phi_{-n}(v)\rangle_{\bf C}=
{\Delta_n\over (w-u)^2(w-v)^2(v-u)^{2\Delta_n-2}
(\bar v-\bar u)^{2\overline\Delta_n}}\,,$$ where $\Phi_{\pm n}$ are normalised so that $\langle \Phi_n(u)\Phi_{-n}(v)\rangle_{\bf C}=|v-u|^{-2\Delta_n
-2\overline\Delta_n}$. (\[TPP\]) is equivalent to the conformal Ward identity[@BPZ] $$\label{WI}
\langle T(w)\Phi_n(u)\Phi_{-n}(v)\rangle_{\bf C}=
\left({\Delta_n\over(w-u)^2}+{\Delta_n\over(w-v)^2}
+{1\over w-u}{\partial\over\partial u}+
{1\over w-v}{\partial\over\partial v}\right)
\langle\Phi_n(u)\Phi_{-n}(v)\rangle_{\bf C}\,.$$ In writing the above, we are assuming that $w$ is a complex coordinate on a single sheet $\bf C$, which is now decoupled from the others. We have therefore shown that $$\langle T(w)\rangle_{{\cal R}_n}\equiv
{\int[d\phi]T(w)e^{-S_E({\cal R}_n)}\over
\int[d\phi]e^{-S_E({\cal R}_n)}}=
{\langle T(w)\Phi_n(u)\Phi_{-n}(v)\rangle_{\bf C}
\over \langle \Phi_n(u)\Phi_{-n}(v)\rangle_{\bf C}}\,.$$ Now consider the effect of an infinitesimal conformal transformation $w\to w'=w+\alpha(w)$ of $\bf C$ which act identically on all the sheets of ${\cal R}_n$. The effect of this is to insert a factor $${1\over 2\pi i}\int_C\alpha(w)T(w)dw
-{1\over 2\pi i}\int_C\overline{\alpha(w)}\overline T(\bar w)d\bar w$$ into the path integral, where the contour $C$ encircles the points $u$ and $v$. The insertion of $T(w)$ is given by (\[TPP\]). Since this is to be inserted on each sheet, the right hand side gets multiplied by a factor $n$.
Since the Ward identity (\[WI\]) determines all the properties under conformal transformations, we conclude that the renormalised $Z_n({A})/Z^n\propto{\rm Tr}\,\rho_{A}^n$ behaves (apart from a possible overall constant) under scale and conformal transformations identically to the $n$th power of two-point function of a primary operator $\Phi_n$ with $\Delta_n=\overline\Delta_n=
(c/24)(1-(1/n)^2)$. In particular, this means that $${\rm Tr}\,\rho_{A}^n=c_n\big((v-u)/a)^{-(c/6)(n-1/n)}\,,$$ where the exponent is just $4n\Delta_n$. The power of $a$ (corresponding to the renormalisation constant $\cal Z$) has been inserted so as the make the final result dimensionless, as it should be. The constants $c_n$ are not determined by this method. However $c_1$ must be unity. Differentiating with respect to $n$ and setting $n=1$, we recover the result of Holzhey et al.
The fact that ${\rm Tr}\,\rho_{A}^n$ transforms under a general conformal transformation as a 2-point function of primary operators $\Phi_{\pm n}$ means that it is simple to compute in other geometries, obtained by a conformal mapping $z\to z'=w(z)$, using the formula[@BPZ] (z\_1,|z\_1)(z\_2,|z\_2)…=\_j|w’(z\_j)|\^[2\_n]{} (w\_1,|w\_1)(w\_2,|w\_2)….
For example, the transformation $w\to w'=(\beta/2\pi)\log w$ maps each sheet in the $w$-plane into an infinitely long cylinder of circumference $\beta$. The sheets are now sewn together along a branch cut joining the images of the points $u$ and $v$. By arranging this to lie parallel to the axis of the cylinder, we get an expression for ${\rm Tr}\,\rho_{A}^n$ in a thermal mixed state at finite temperature $\beta^{-1}$. This leads to the result for the entropy $$S_A(\beta)\sim(c/3)\log\big((\beta/\pi a)\sinh(\pi\ell/\beta)\big)+c_1'\,.$$ For $\ell\ll\beta$ we find $S_A\sim(c/3)\log(\ell/a)$ as before, while, in the opposite limit $\ell\gg\beta$, $S_A\sim(\pi c/3)(\ell/\beta)$. In this limit, the von Neumann entropy is extensive, and its density agrees with that of the Gibbs entropy of an isolated system of length $\ell$, as obtained from the standard CFT expression[@BCN; @Affleck] $\beta F\sim -(\pi c/6)(\ell/\beta)$ for its free energy.
Alternatively, we may orient the branch cut perpendicular to the axis of the cylinder, which, with the replacement $\beta\to L$, corresponds to the entropy of a subsystem of length $\ell$ in a finite 1d system of length $L$, with periodic boundary conditions, in its ground state. This gives $$S_A\sim (c/3)\log\big((L/\pi a)\sin(\pi\ell/L)\big)+c_1'\,.$$ Note that this expression is symmetric under $\ell\to L-\ell$. It is maximal when $\ell=L/2$.
Finite system with a boundary.
------------------------------
Next consider the case when the 1d system is a semi-infinite line, say $[0,\infty)$, and the subsystem $A$ is the finite interval $[0,\ell)$. The $n$-sheeted Riemann surface then consists of $n$ copies of the half-plane $x\geq0$, sewn together along $0\leq x<\ell, \tau=0$. Once again, we work initially at zero temperature. It is convenient to use the complex variable $w=\tau+ix$. The uniformising transformation is now $z=\big((w-i\ell)/(w+i\ell)\big)^{1/n}$, which maps the whole Riemann surface to the unit disc $|z|\leq1$. In this geometry, $\langle T(z)\rangle=0$ by rotational invariance, so that, using (\[schwartz\]), we find $$\label{hp}
\langle T(w)\rangle_{{\cal R}_n}=
{\Delta_n(2\ell)^2\over(w-i\ell)^2(w+i\ell)^2}\,,$$ where $\Delta_n=(c/24)(1-n^{-2})$ as before. Note that in the half-plane, $T$ and $\overline T$ are related by analytic continuation: $\overline T(\bar w)=T(w)^*$. (\[hp\]) has the same form as $\langle T(w)\Phi_n(i\ell)\rangle$, which follows from the Ward identities of boundary CFT[@JCbound], with the normalisation $\langle\Phi_n(i\ell)\rangle=(2\ell)^{-\Delta_n}$.
The analysis then proceeds in analogy with the previous case. We find $${\rm Tr}\,\rho_{A}^n\sim \tilde c_n(2\ell/a)^{(c/12)(n-1/n)}\,,$$ so that $S_A\sim(c/6)\log(2\ell/a)+{\tilde c}'_1$.
Once again, this result can be conformally transformed into a number of other cases. At finite temperature $\beta^{-1}$ we find $$S_A(\beta)\sim(c/6)\log\big((\beta/\pi a)\sinh(2\pi\ell/\beta)\big)+
{\tilde c}'_1\,.$$ By taking the limit when $\ell\gg\beta$ we find the same extensive entropy as before. However, we can now identify ${\tilde c}'_1-c'_1$ as the boundary entropy $g$, first discussed by Affleck and Ludwig[@AffleckLudwig].
For a completely finite 1d system, of length $L$, at zero temperature, divided into two pieces of lengths $\ell$ and $L-\ell$, we similarly find $$S_A=(c/6)\log\big((L/\pi a)\sin(\pi\ell/L)\big)+2g+c'_1\,.$$
General case.
-------------
For general $N$, the algebra is more complicated, but the method is the same. The uniformising transformation now has the form $z=\prod_i(w-w_i)^{\alpha_i}$, with $\sum_i\alpha_i=0$ (so there is no singularity at infinity.) Here $w_i$ can be $u_j$, $v_j$, or $\pm iu_j$ in the case of a boundary. In our case, we have $\alpha_i=\pm1/n$, but it is interesting to consider the more general transformation, and the notation is simpler. Once again we have T(w)=c[12]{}{z,w}= c[12]{}[z”’z’-32[z”]{}\^2\^2]{}. Consider $\{z,w\}/z$. As a function of $w$, this is meromorphic, has a double pole at each $w=w_i$, and is $O(w^{-2})$ as $w\to\infty$. Hence it has the form {z,w}/z=\_i[A\_i(w-w\_i)\^2]{}+\_i[B\_iw-w\_i]{}, where $\sum_iB_i=0$. In order to determine $A_i$ and $B_i$, we need to compute $z'$, etc, to second order in their singularities at $w_i$. After some algebra, we find $$\begin{aligned}
z'&=&
\left[{\alpha_i\over w-w_i}+\sum_{j\not=i}{\alpha_j\over w-w_j}\right]\,z\,;\\
z''&=&\left[{-\alpha_i+\alpha_i^2\over(w-w_i)^2}
+2{\alpha_i\over w-w_i}\sum_{j\not=i}{\alpha_j\over
w_i-w_j}+\cdots\right]\,z\,;\\
z'''&=&\left[{2\alpha_i-3\alpha_i^2+\alpha_i^3\over(w-w_i)^3}
+{-3\alpha_i+3\alpha_i^2\over(w-w_i)^2}\sum_{j\not=i}{\alpha_j\over w_i-w_j}
+\cdots\right]\,z\,.\end{aligned}$$ Let C\_i\_[j=i]{}[\_jw\_i-w\_j]{}. Then the coefficient of $(w-w_i)^{-2}$ is $$\begin{aligned}
&&\left\{\left[\alpha_i(\alpha_i-1)(\alpha_i-2)+3\alpha_i(\alpha_i-1)C_i(w-w_i)
+\cdots\right]\left[\alpha_i+C_i(w-w_i)+\cdots\right]\right.\\
&&\left.-\ffrac32\left[\alpha_i(\alpha_i-1)+2\alpha_iC_i(w-w_i)
+\cdots\right]^2\right\}/
\left[\alpha_i+C_i(w-w_i)+\cdots\right]^2\,,\end{aligned}$$ from which we find after a little more algebra that $$\begin{aligned}
A_i &=&\ffrac12(1-\alpha_i^2)\,;\\
B_i &=&C_i\,{1-\alpha_i^2\over \alpha_i}\,.\end{aligned}$$ Thus we have shown that T(w)=c[12]{} \_i. This is to be compared with the conformal Ward identity T(w)\_i\_i(w\_i)=\_i\_k\_k(w\_k). For these to be equivalent, we must have $\Delta_i=\frac12(1-\alpha_i^2)$ and \_[j=i]{}[\_jw\_i-w\_j]{} =[w\_i]{}\_k\_k(w\_k). A necessary and sufficient condition for this is that ([1-\_i\^2\_i]{} \_[j=i]{}[\_jw\_i-w\_j]{})= [w\_i]{}([1-\_k\^2\_k]{} \_[j=k]{}[\_jw\_k-w\_j]{}), for each pair $(i,k)$. This reduces to -[1-\_i\^2\_i]{}[\_k(w\_k-w\_i)\^2]{}= -[1-\_k\^2\_k]{}[\_i(w\_i-w\_k)\^2]{}, that is, $\alpha_i=\pm\alpha_k$ for each pair $(i,k)$. Since $\sum_i\alpha_i=0$ the only way to satisfy this is to have $\alpha_i=\alpha\sigma_i$, with $\sigma_i=\pm1$, and half the $\sigma_i=+1$ and the remainder $-1$. Interestingly enough, this is precisely the case we need, with $\alpha=1/n$.
If these conditions are satisfied, \_j\_j(w\_j)=c[12]{}(1-\^2)\_[k=i]{}[\_i\_k w\_i-w\_k]{}, so that \_j\_j(w\_j)=c[12]{}(1-\^2)\_[k<i]{}\_i\_k (w\_i-w\_k)+ E, where $E$ is independent of all the $w_i$. In the case with no boundary, $E$ can depend however on the $\bar w_i$. A similar calculation with $\overline T$ then gives a similar dependence. We conclude that ${\rm Tr}\,\rho_{A}^n$ behaves under conformal transformations in the same way as \_j\_j\_[j<k]{} (w\_i-w\_k)\^[c[12]{}(1-\^2)\_i\_k]{} (|w\_i-|w\_k)\^[c[12]{}(1-\^2)|\_i|\_k]{}. Taking now $\alpha=1/n$, and $\sigma=\pm1$ according as $w_i=u_j$ or $v_j$, we find $$\label{generaln}
{\rm Tr}\,\rho_{A}^n\sim
c_n^N\left({\prod_{j<k}(u_k-u_j)(v_k-v_j)\over\prod_{j\leq k}(v_k-u_j)}
\right)^{(c/6)(n-1/n)}\,.$$ The overall constant is fixed in terms of the previously defined $c_n$ by taking the intervals to be far apart from each other, in comparison to their lengths.
Differentiating with respect to $n$ and setting $n=1$, we find our main result of this section $$\label{general}
S_A=\frac c3\left(\sum_{j\leq k}\log\big((v_k-u_j)/a\big)
-\sum_{j<k}\log\big((u_k-u_j)/a\big)-\sum_{j<k}\log\big((v_k-v_j)/a\big)
\right)+Nc_1'\,.$$
A similar expression holds in the case of a boundary, with half of the $w_i$ corresponding to the image points.
Finally we comment on the recent result of Casini and Huerta[@Casini], which corresponds to $N=2$. In fact, it may be generalised to the ratio of Tsallis entropies: from (\[generaln\]) we find $${S_n(A)S_n(B)\over S_n(A\cap B)S_n(A\cup B)}=
\left({(v_1-u_1)(v_2-u_2)\over (u_2-u_1)(v_2-v_1)}\right)^{(c/6)(n-1/n)}\,,$$ where $S_n(A)\equiv{\rm Tr}\,\rho_A^n$, and the expression in parentheses is the cross-ratio $\eta$ of the four points. Notice that in this expression the dependence on the ultraviolet cut-off $a$ disappeared, so have the non-universal numbers $c_n$. Differentiating with respect to $n$ gives the result of Casini and Huerta[@Casini], who however *assumed *that the result should depend only on $\eta$.**
Entropy in non-critical 1+1-dimensional models {#sectmass}
==============================================
Massive field theory - general case
-----------------------------------
In this section we consider an infinite non-critical model in 1+1-dimensions, in the scaling limit where the lattice spacing $a\to0$ with the correlation length (inverse mass) fixed. This corresponds to a massive relativistic QFT. We first consider the case when the subset $A$ is the negative real axis, so that the appropriate Riemann surface has branch points of order $n$ at 0 and infinity. However, for the non-critical case, the branch point at infinity is unimportant: we should arrive at the same expression by considering a finite system whose length $L$ is much greater than $\xi$.
Our argument parallels that of Zamolodchikov[@Zam] for the proof of his famous $c$-theorem. Let us consider the expectation value of the stress tensor $T_{\mu\nu}$ of a massive euclidean QFT on such a Riemann surface. In complex coordinates, there are three non-zero components: $T\equiv T_{zz}$, $\overline T\equiv T_{\bar z\bar z}$, and the trace $\Theta=4T_{z\bar z}=4T_{\bar zz}$. These are related by the conservation equations $$\begin{aligned}
\label{cons}
\partial_{\bar z}T+\ffrac14\partial_z\Theta&=&0\,;\\
\partial_z\overline T+\ffrac14\partial_{\bar z}\Theta&=&0\,.\end{aligned}$$
Consider the expectation values of these components. In the single-sheeted geometry, $\langle T\rangle$ and $\langle\overline
T\rangle$ both vanish, but $\langle\Theta\rangle$ is constant and non-vanishing: it measures the explicit breaking of scale invariance in the non-critical system. In the $n$-sheeted geometry, however, they all acquire a non-trivial spatial dependence. By rotational invariance about the origin, they have the form $$\begin{aligned}
\langle T(z,\bar z)\rangle&=& F_n(z\bar z)/z^2\,;\\
\langle\Theta(z,\bar z)\rangle-\langle\Theta\rangle_1&=&
G_n(z\bar z)/(z\bar z)\,;\\
\langle\overline T(z,\bar z)\rangle&=& F_n(z\bar z)/{\bar z}^2 \,.\end{aligned}$$ From the conservation conditions (\[cons\]) we have $$(z\bar z)\left(F'_n+\ffrac14G'_n\right)=\ffrac14G_n\,.$$ Now we expect that $F_n$ and $G_n$ both approach zero exponentially fast for $|z|\gg\xi$, while in the opposite limit, on distance scales $\ll\xi$, they approach the CFT values (see previous section) $F_n\to (c/24)(1-n^{-2})$, $G_n\to0$.
This means that if we define an effective $C$-function $$C_n(R^2)\equiv \left(F(R^2)+\ffrac14G(R^2)\right)\,,$$ then $$R^2{\partial\over\partial(R^2)}C_n(R^2)=\ffrac14G_n(R^2)\,.$$
If we were able to argue that $G_n\leq0$, that is $\langle\Theta\rangle_n\leq\langle\Theta\rangle_1$, we would have found alternative formulation of the $c$-theorem. However, we can still derive an integrated form of the $c$-theorem, using the boundary conditions:[^3] $$\int_0^\infty {G_n(R^2)\over R^2}d(R^2)=-(c/6)(1-n^{-2})\,,$$ or equivalently $$\int\left(\langle\Theta\rangle_n-\langle\Theta\rangle_1\right)d^2\!R
=-\pi n(c/6)(1-n^{-2})\,,$$ where the integral is over the whole of the the $n$-sheeted surface. Now this integral (multiplied by a factor $1/2\pi$ corresponding to the conventional normalisation of the stress tensor) measures the response of the free energy $-\log Z$ to a scale transformation, i.e. to a change in the mass $m$, since this is the only dimensionful parameter of the renormalised theory. Thus the left hand side is equal to $$-(2\pi)\,m(\partial/\partial m)\left[\log Z_n-n\log Z\right]\,,$$ giving finally $${Z_n\over Z^n}= c_n(ma)^{(c/12)(n-1/n)}\,,$$ where $c_n$ is a constant (with however $c_1=1$), and we have inserted a power of $a$, corresponding to the renormalisation constant $\cal Z$ discussed earlier, to make the result dimensionless.
This shows that the $(n-1/n)$ dependence for the exponent of the Tsallis entropy is a general property of the continuum theory. Differentiating at $n=1$, we find the main result of this section $$\label{mass}
S_{A}\sim-(c/6)\log(ma)=(c/6)\log(\xi/a)\,,$$ where $\xi$ is the correlation length. We re-emphasise that this result was obtained only for the scaling limit $\xi\gg a$. However, for lattice integrable models, we shall show how it is possible to obtain the full dependence without this restriction.
So far we have considered the simplest geometry in the which set $A$ and its complement $B$ are semi-infinite intervals. The more general case, when $A$ is a union of disjoint intervals, is more difficult in the massive case. However it is still true that the entropy can be expressed in terms of the derivative at $n=1$ of correlators of operators $\Phi_n$. The above calculation can be thought of in terms of the one-point function $\langle\Phi_n\rangle$. In any quantum field theory a more general correlator $\langle\prod_{i=1}^{k}\Phi_{\pm n}(w_i)\rangle$ should obey cluster decomposition: that is, for separations $|w_i-w_j|$ all $\gg\xi$, it should approach $\langle\Phi_n\rangle^{k}$. This suggests that, in this limit, the entropy should behave as $S_A\sim{\cal A}(c/6)\log(\xi/a)$, where ${\cal A}=k$ is the number of boundary points between $A$ and its complement. This would be the 1d version of the area law[@s-93]. When the interval lengths are of the order of $\xi$, we expect to see a complicated but universal scaling form for the cross-over.
Free bosonic field theory
-------------------------
In this subsection we verify Eq. (\[mass\]) by an explicit calculation for a massive free field theory (Gaussian model.) The action =12 ((\_)\^2+m\^2\^2)d\^2r, is, as before, considered on a $n$-sheeted Riemann surface with one cut, which we arbitrarily fix on the real negative axis.
To obtain the entanglement entropy we should know the ratio $Z_n/Z^n$, where $Z_n$ is the partition function in the $n$-sheeted geometry. There are several equivalent ways to calculate such partition function. In the following, we find easier to use the identity[^4] Z\_n=-G\_n([**r**]{},[**r**]{})d\^2r, \[logZ\] where $G_n({\bf r,r'})$ is the two-point correlation function in the $n$-sheeted geometry. Thus we need the combination $G_n-nG_1$. $G_n$ obeys (-\^2\_[**r**]{}+m\^2) G\_n([**r,r’**]{}) = \^2([**r-r’**]{}) . \[Hde\] Its solution (see the Appendix) may be expressed in polar coordinates as (here $0<r,r'<\infty$ and $0\leq\theta,\theta'<2\pi n$) G\_n(r,,r’,’)= \_[k=0]{}\^d\_k \_0\^d\_k(,’), \[2pt\] where ${\cal C}_k(\theta,\theta')=
\cos(k\theta/n)\cos(k\theta'/n)+\sin(k\theta/n)\sin(k\theta'/n)$, $d_0=1$, $d_{k>0}=2$, and $J_k(x)$ are the Bessel functions of the first kind. At coincident points (i.e. $r=r'$, $\theta=\theta'$), and after integrating over $\theta$ and $\lambda$ we have G\_n(r)G\_n([**r**]{},[**r**]{})=\_[k=0]{}\^d\_k I\_[k/n]{} (m r) K\_[k/n]{} (mr), \[Gr\] where $I_k(x)$ and $K_k(x)$ are the modified Bessel functions of the first and second kind respectively [@as].
The sum over $k$ in (\[Gr\]) is UV divergent. This reflects the usual short-distance divergence which would occur even in the plane. However, if we formally exchange the order of the sum and integration we find -Z\_n= d\^2 r G\_n(r) [=]{} \_k d\_k \_0\^I\_[k/n]{} (m r) K\_[k/n]{} (mr) r dr= \_k d\_k k. \[div\] Interpreting the last sum as $2\zeta(-1)=-1/6$, we obtain the correct result, which we now derive more systematically.
Let us first regularise each sum over $k$ by inserting a function $F(k/\Lambda_n)$: $F$ is chosen so that $F(0)=1$, and all its derivatives at the origin vanish: however it goes to zero sufficiently fast at infinity. Since $k/n$ is conjugate to the angle $\theta$, we should think of this cut-off as being equivalent to a discretisation $\delta\theta$. Thus we should choose $\Lambda_n=\Lambda\cdot n$, where $\Lambda\sim(\delta\theta)^{-1}$.
To perform the sum over $k$, we use the Euler-MacLaurin (EML) sum formula [@as] \_[k=0]{}\^d\_k f(k)= \_0\^f(k) dk - f’(0)- \_[j=2]{}\^ f\^[(2j-1)]{}(0), \[eml\] where $B_{2n}$ are the Bernoulli numbers [@as]. Using standard identities of the Bessel function (namely $\partial_k K_k(x)|_{k=0}=0$ and $\partial_k I_k(x)|_{k=0}=-K_0(k)$ [@as]), we obtain G\_n(r)= 2 \_0\^dk I\_[k/n]{} (m r) K\_[k/n]{} (mr)F(k/n)+ K\_0\^2 (mr) +\_0\^r dr \_[j1]{} D\_[2j +1]{}(r), \[Greml\] where we define $D_i(x)=\partial^i(I_k(x) K_k(x))/\partial k^i|_{k=0}$. In the last term in Eq. (\[Greml\]) the order of the integral, derivative, and sum can be exchanged and each term in the sum is \_0\^x dx I\_k(x) K\_k(x)= - =0 i=2j+12, i.e. the sum in Eq. (\[Greml\]) is vanishing.
Thus, still with the regulator in place, Z\_n= -\_0\^r dr \_0\^dk I\_[k/n]{} (m r) K\_[k/n]{} (mr)F(k/n) -1[24nm\^2]{} where we have used $\int_0^\infty rK_0^2(mr)dr=1/(2m^2)$.
Now the point is that in the integral over $k$ the factor of $n$ can be scaled out by letting $k\to nk$. Thus this potentially divergent term cancels in the required combination $G_n-nG_1$. Having taken this combination, we may now remove the regulator to find the main result = (n-). The integration of the last expression wrt $m^2$ (made following the recipe for the integration limits given in the previous section) gives \^n= = (n-), \[trrhon\] and finally the entanglement entropy S=-[Tr]{} = -. \^n|\_[n=1]{}= -. (m\^2 a\^2)\^[(n-)]{} |\_[n=1]{} =- m\^2 a\^2, that agrees with the general formula we derive (\[mass\]), with $c=1$ and $m=\xi^{-1}$.
Integrable models and the corner transfer matrix
------------------------------------------------
In this subsection we verify (\[mass\]) for the transverse Ising chain and the uniaxial XXZ Heisenberg model. These results are also an independent check of the uniform convergence of the derivative wrt $n$ of $\rho^n$ when $n\rightarrow1$, since for these models we can compute the eigenvalues of $\rho_A$ exactly.
Although these systems are integrable and their ground state is known, a direct calculation of $\rho$ is difficult. The difficulties arising in a direct calculation can be avoided mapping the quantum chains onto two-dimensional classical spin systems. As firstly pointed out by Nishino et al. [@nishino] the density matrix of the quantum chain is the partition function of a two-dimensional strip with a cut perpendicular to it. In fact the ground state of a quantum chain described by an Hamiltonian $H$ is also eigenstate of the transfer matrix $T$ of a classical systems satisfying $[H,T]=0$. Therefore the reduced density matrix of a subsystem $A$ of the chain (defined in the Introduction as $\rho_A={\rm Tr}_B |\Psi\rangle\langle \Psi|$, with $B$ the complement of $A$) is the partition function of two half-infinite strips, one extending from $-\infty$ to $0$ and the other from $+\infty$ to 0, with the spins in $B$ identified.
This partition function is the product of four Baxter corner transfer matrices (CTMs)[@Baxter] $\hat A$. If the lattice is choose in a clever way (i.e. rotated by $\pi/4$ with respect to the cut) one ends in the infinite length limit with [@pkl-99] \_A=\^4=e\^[-\_[CTM]{}]{}, where $\hat{H}_{\rm CTM}$ is an effective Hamiltonian, which, for the models under consideration, may be diagonalised by means of fermionisation (see for more details about this equivalence [@pkl-99] and references therein.) Note that ${\rm Tr \hat{\rho}_A \neq 1}$, thus the usual density matrix is $\rho_A=\hat{\rho}_A/{\rm Tr}\hat{\rho}_A$.
The method just outlined is very general. However, for the integrable chains under consideration (and indeed for any model satisfying suitable Yang-Baxter equations[@Baxter]) it is possible to write $H_{\rm CTM}=\e \hat{O}$, with $\e$ the scale giving the distance between the energy levels, and $\hat{O}$ is an operator with integer eigenvalues (for the Ising and XXZ models it is expressed in terms of free fermions.) Using this property the entropy is given by[^5] S=-[Tr]{} \_A\_A= -[Tr]{} + \_A= -+ Z, \[entrint\] where we defined $Z={\rm Tr} \hat{\rho}_A={\rm Tr} e^{-\hat{H}_{\rm CTM}}$.
The Ising model in a transverse field can be described by the one-dimensional Hamiltonian H\_I=-\_[n=1]{}\^[L-1]{}\^x\_n- \_[n=1]{}\^[L-1]{}\^z\_n\^z\_[n+1]{}, \[HamI\] where $\sigma_n$ are the Pauli matrices at the site $n$, and we normalise the Hamiltonian (following [@pkl-99]) by imposing the transverse field in the $x$ direction to be one, thus the transition is driven by the parameter $\lambda$. The classical equivalent of (\[HamI\]) is two dimensional Ising model. For $\lambda=0$ the ground state of the Hamiltonian (\[HamI\]) is a quantum “paramagnet” with all the spins aligned with the magnetic field in the $x$ direction, and $\langle\sigma^z_n\rangle=0$. In the opposite limit $\lambda=\infty$ the magnetic field is negligible and the ground state is ferromagnetic with $\langle\sigma^z_n\rangle=\pm 1$. The (second-order) transition between this two regimes happens at $\lambda=1$. The exponent characterizing the divergence of the correlation length is $\nu=1$, i.e. $\xi\simeq |\lambda-1|^{-1}$.
The CTM of the Ising model may be diagonalised in terms of fermionic operators. The CTM Hamiltonian, written in terms of the fermion occupation number $\hat{n}_i$ (with eigenvalues $0$ and $1$) is [@pkl-99] \_[CTM]{}=\_[j=0]{}\^\_j \_j. The energy levels are \_j= {. = where $K(k)$ is the complete elliptic integral of the first kind [@as], and $k=\min [\lambda,\lambda^{-1}]$.
For $\lambda<1$ Z=[Tr]{} e\^[-\_[CTM]{}]{}=\_[j=0]{}\^, and the entropy from Eq. (\[entrint\]) is S= \_[j=0]{}\^+ \_[j=0]{}\^(1+e\^[-(2j+1)]{}). \[Spara\] Analogously in the quantum ferromagnetic phase ($\lambda>1$) S=\_[j=0]{}\^+ \_[j=0]{}\^(1+e\^[-2j]{}). \[Sferro\] Fig. \[Fig\] shows a plot of the entropy as function of $\lambda$, characterised by a divergence at the quantum critical point $\lambda=1$. Note that $S(0)=0$ and $S(\infty)=\log2$, in agreement with the expectation that the pure ferromagnetic ground state ($\lambda=\infty$) has two possible accessible configurations with opposite sign of magnetisation (i.e. $S(\infty)=\log2$) and the pure quantum paramagnetic ground state ($\lambda=0$) has only one configuration available with all the spins aligned in the direction of the magnetic field $x$ and the resulting entropy is zero.
Let us analyze in details the behaviour at the critical point. For $\lambda\rightarrow1$, $\e\rightarrow0$ in both the phases, thus the sums Eqs. (\[Spara\]) and (\[Sferro\]) can be approximated by the integral S \_0\^dx (+(1+e\^[-2x]{}) ) = , \[Sint\] with $\simeq$ we mean in the critical region. The $\lambda$ dependence (we use $K(0)=\pi/2$, $K(x)=-1/2 \log(1-x) +O((1-x)^0)$ [@as]) S -(1-k)=+C\_1, where in the last equality $\xi\propto |1-k|^{-1}$ has been used. This agrees with \[mass\], with $c=1/2$ for the Ising model. The constant $C_1$ is not universal.
Another model whose density matrix has been derived by using the CTM is the XXZ model H\_[XXZ]{}=\_n (\^x\_n\^x\_[n+1]{}+\^y\_n\^y\_[n+1]{}+ \^z\_n\^z\_[n+1]{}), for $\Delta>1$ (whose classical equivalent is the Baxter six-vertex model [@Baxter].) This model has an Ising-like ferromagnetic state for $\Delta>1$ and a planar XX ferromagnetic ground state at $0<\Delta<1$ (the case $\Delta<0$ describes the antiferromagnetic regime, in which we are not interested.) At $\Delta=1$ the Hamiltonian is isotropic (XXX) and approaching such a point from large $\Delta$ values the correlation length diverges as (see e.g. [@Baxter]) , \[xiXXZ\] because at $\Delta=1$ a massless excitation (Goldstone mode) is present in the spectrum.
The CTM Hamiltonian $\hat{H}_{\rm CTM}$ for this model has been obtained in Ref. [@pkl-99] \_[CTM]{}=\_[j=0]{}\^2j n\_j, with $\e={\rm arccosh} \Delta$, for $\Delta>1$. The entropy is given by Eq. (\[Sferro\]). Close to the isotropic point ($\Delta=1$) it holds $\e\simeq\sqrt2\sqrt{\Delta-1}$ and (see Eq. (\[Sint\])) S . Using Eq. (\[xiXXZ\]) to write $\Delta$ in term of the correlation length, we have S , in agreement with \[mass\], with $c=1$. Again, in the limit $\Delta\rightarrow\infty$, $S=\log2$, since the ferromagnetic ground state is Ising-like.
The method used here for the transverse Ising model and the XXZ model can in principle be applied to all those integrable model whose weights satisfy Yang-Baxter equations.
Finite-size effects {#sectfss}
===================
So far, we have studied the entropy either at the critical point, for a finite subsystem, or away from criticality in an infinite system. These two regimes may be linked by a generalisation of finite-size scaling theory. This would assert, for example, that the entropy of a subsystem $A$ which forms e.g. the left half of a finite non-critical system of length $2L$ should have the form S\_A(L,)= (L + s\_[FSS]{}(L/)), \[FSS1d\] with $s_{\rm FSS}(0)=0$ (this because we are referring to the scaling part, the constant term found before is not universal and it defines an overall additive normalisation of $S$) and $s_{\rm FSS}(x)\sim-\log x$, for large $x$, so as to recover $S=(c/6)\log\xi$ in the infinite length limit. $s_{\rm FSS}(x)$ should admit a small $x$ expansion s\_[FSS]{}(x)=\_[j1]{} s\_j x\^[2j]{}, with $s_j$ [*universal*]{} coefficients, and a large $x$ expansion s\_[FSS]{}(x)=-x+\_[j0]{} , with $s^\infty_j$ also universal (in special situations logarithmic corrections could also be generated.) In the next section we test this hypothesis for the free massive field theory.
FSS in the Gaussian model
-------------------------
Once again we consider the $n$-sheeted surface, but it now consists of $n$ discs of finite radius $L$, sewn together along the negative real axis from $-L$ to $0$. The Gaussian two-point function in a finite geometry has been calculated in the Appendix Eq. (\[2ptL\]). Setting $\theta=\theta'$, $r=r'$ and integrating over $\theta'$ the propagator at coincident point is G\_n(r)= \_[k=0]{}\^d\_k \_[i=1]{}\^ , where $\alpha_{\nu,i}$ denotes the $i-$th zero of $J_\nu(x)$. $G(r)$ is the analogous of (\[Gr\]) in a finite geometry.
As we showed in the previous section, the right order to proceed to get a sensitive $Z_n$ is first to perform the sum over $i$, then sum over $k$ and finally integrate over $r$. This is really hard, requiring a sum of Bessel functions for generic argument, over the zeroes of different Bessel functions. What one can do is inverting the order of the sums and integrations and try to understand what happens. From the previous exercise we know that this operation has to be done with care.
However, the formal result is Z\_[n]{}= \_k d\_k\_i. \[formal\] Since the large $i$ behaviour of the zeroes of the Bessel functions is $\alpha_{\nu,i}\sim\pi(i+\nu/2-1/4)$ (see e.g. [@as]), this sum diverges.
A first, qualitatively correct expression for the universal function $s_{\rm FSS}(x)$ can be obtained from Eq. (\[formal\]), truncating the EML formula Eq. (\[eml\]) in the variable $k$ at the first order in the derivative. This approximation gives a correct form of the result because it takes into account the complete infinite volume result (whose truncated expression is exact.) The integral term in the EML is divergent, but it cancels as before in the ratio $Z_n/Z^n$ if the cut-off in the angular modes is properly chosen as in the infinite volume case. The EML approximation at the first order is (we do not write the integral) Z\_[n]{}= \_i , \[Zeml1\] where we used .|\_[k=0]{}= , which can be derived using the formulas [@as] of the derivative of the Bessel functions with respect to the order. The remaining sum over $i$ (the zeroes of the Bessel function) is now finite.
This gives the entropy as S=-. \^n|\_[n=1]{}= 2 F\_1(m L,L/a), which reproduces the correct limits for $m=0$ and $L=\infty$.
From this formula we may compare with the FSS ansatz (\[FSS1d\]), with $c=1$: s\_[FSS]{}\^[(1)]{}(m L)=12 (F\_1(m L,L/a)- F\_1(0,L/a))= - \_i , \[sfs\] that, as expected, is a function only of the product $mL$ and all the dependence upon $a$ disappears after the subtraction. This agrees with $s_{\rm FSS}(0)=0$ and $s_{\rm FSS}(0)=-\log x$, for large $x$ (it can be easily shown, since in this limit the sum can be replaced by an integral.) The over-script $(1)$ is there to remind that we made a first-order approximation in EML.
Eq. (\[sfs\]) characterises completely the crossover between the mass dominated (non-critical) and the geometry dominated (critical) regime.
A plot of $s_{\rm FSS}^{(1)}$ is shown in Fig. \[figs\] (calculated as the sum of the first 1000 zeroes of $J_0$.) From this figure we see that an optimum approximation for $x\geq2$ of $s_{\rm FSS}^{(1)}(x)$ is $s_{\rm FSS}^{\rm asy}(x)= -\log x+s_0^\infty$, with $s_0^\infty=0.120912$ (this value of $s_0^\infty$ is a fit.) We can also calculate analytically all the universal coefficients of the small $mL$ expansion [@bessels]: s\_1\^[(1)]{}&=-& \_[i=1]{}\^ =-,\
s\_2\^[(1)]{}&=& \_[i=1]{}\^ =,\
s\_3\^[(1)]{}&=-& \_[i=1]{}\^ =-,\
s\_4\^[(1)]{}&& 1.15 10\^[-3]{},s\_5\^[(1)]{}-1.98 10\^[-4]{},s\_6\^[(1)]{}3.42 10\^[-5]{}. Fig. \[figs\] (see also the inset) shows a plot of $s_{\rm FSS}^{(1)}(x)$, compared with quadratic, quartic, and sextic approximation. The agreement is excellent in the region $x\leq2$.
Before starting the full calculation of the function $s_{\rm FSS}(x)$ let us summarise what we can learn from the first order approximation in the EML expansion: A rather good approximation of the full function may be obtained by matching only the first-order small argument expansion with the large $x$ behaviour $\sim\log x$ (which is exact and does not depend upon the approximation.) Thus in the following we show how to calculate the coefficients $s_j$ without any approximation.
Our starting point is again Eq. (\[formal\]), but this time we will not use EML sum formula. For this reason we have to be care since the sum is divergent.
The interesting object is $\log Z_{n}-n\log Z$. In particular we can write an FSS ansatz also for this “free-energy” and so subtract the $m=0$ part, obtaining the universal function $f(x=mL)$ f(x)=3\_k d\_k\_i, form which $s_{\rm FSS}(x)=-\left.\frac{\partial f(x)}{\partial n}\right|_{n=1}$. This formula has to be intended in a formal way: in fact it is the difference of two diverging sums. To make this difference sensible the recipe for the cut-off explained in the previous section has to be used. Anyway from the computational point of view it is simpler to make the calculation without care of the cut-off and to adjust only at the end the result. The numerical sum over the zeroes of the Bessel functions of generic order cannot be done as before, but the small $x$ expansion of $f(x)=\sum f_jx^{2j}$ with coefficients f\_j= \_k d\_k\_i, can be calculated analytically. For example s\_1\^[wrong]{}=-. f\_1 |\_[n=1]{}= 3\_k d\_k\_i = -3\_k d\_k k\_i, that using [@bessels] \_[i=1]{}\^=, leads to s\_1\^[wrong]{}=-3\_k d\_k = -\_k d\_k = -(2-1). \[app\] We use the over-script “wrong” because it is not strictly correct to make the calculation this way, since we are implicitly using the same cut-off for $Z$ and $Z_n$. In fact, applying the EML to the sum in (\[app\]), one has s\_1\^[wrong]{}=-\_k d\_k = -(2\_0\^+ 2+…), that has a first order term $s^{(1)}_1=-1/4$ (in agreement with what previously found) but the integral is not vanishing. This is a finite difference of the two divergent expressions with the wrong cut-off. The value of the integral ($2\int dk (k+1)^{-2}=2$) must be properly subtracted to have the right result s\_1=-(2-3)=-0.217401…. This is close to the first order EML result $-1/4$, signaling that such approximation, not only reproduces the qualitative physics but is also quantitatively reliable (at the level of 10%.) In the same manner one can calculate all the coefficients $s_j$, using the more complicated expressions for the sum of higher negative powers of zeroes of Bessel functions reported in the literature [@bessels].
Higher dimensions: Scaling of entropy and area law {#secthd}
==================================================
The scaling hypothesis plays a fundamental rule in understanding classical phase transitions. Crudely speaking, it asserts that the microscopic length scale $a$ does not enter explicitly into thermodynamic relations near the critical point, as long as the various thermodynamic variables are suitable normalised and allowed to scale with their non-trivial scaling dimensions, related to the various universal critical exponents.
In fact, from the scaling of the singular part of the free energy density (here $t=|T-T_c|/T_c$, $\xi\propto t^{-\nu}$ is the correlation length, $h$ the external magnetic field, $f_{\pm}(x)$ a universal function, the subscript $\pm$ refers to the two phases, and $y_h$ is the scaling dimension of $h$) f\_[sing]{}(t,h)=\^[-d]{} f\_(h\^[-y\_h]{}), the critical behaviour of all the thermodynamic observables and in particular the scaling laws may be derived. Note that $f_{\rm sing}$ is not the total free energy density: there is another non-universal piece which has an explicit $a^{-d}$ dependence on the microscopic cut-off. However, this term is analytic in the thermodynamic variables.
As argued by Srednicki[@s-93], for $d>2$ the entropy $S_A$ is proportional to the surface area $\cal A$ of the subsystem $A$. Thus we should discuss the entropy per unit area $s=S_A/{\cal A}$. In analogy with the classical case, we may conjecture a scaling form for the singular part of the entropy per unit area near a quantum phase transition s\_[sing]{}(g,h,T)=\^[-(d-1)]{}s\_(h\^[-y\_h]{},T \^[-z]{}), \[Sscal\] where $y_h>0$ describes the relevant effect of the field conjugated to the order parameter, $z>0$ the relevance of the temperature close the the quantum phase transition (in all the examples considered up to now $z=1$), and all other neglected operators are supposed to be irrelevant. $s_\pm$ should be a universal function apart from the normalisations of its arguments. The relation between $\xi$ and $g$ (the parameter driving the transition) is $\xi=|g-g_c|^{-\nu}$. As for the free energy, $s_{\rm sing}$ is not the total entropy: we also expect to find explicitly $a$-dependent pieces which are, however, analytic in $g$.
From the scaling of the entropy all the scaling laws can be obtained, e.g. the specific heat goes like C=T\~T \^[-(d-1+z)]{}. Similar identities can be derived for other observables.
This scaling can be explicitly checked in the case of the Gaussian model in all dimensions for $h=T=0$ and $z=1$. To calculate the entropy of a $d$-dimensional system close to a quantum critical point, one has to consider a $d+1$ field theory. We consider the geometry where $d-1$ dimensions are translational invariant in a domain delimited by an hypersurface of area $\cal A$, and on the remaining two dimensional plane there is a branch cut going from 0 to $\infty$ in an arbitrary direction. The choice of a different geometry is not expected to change the main results.
The two-point function is translational invariant in $d-2$ directions and coincides with Eq. (\[2pt\]) on the cut plane, i.e. G(r\_;r,,r’,’)= e\^[i k\_r\_]{} \_[k=0]{}\^d\_k \_0\^d\_k (,’), \[2pthd\] where $r_\perp$ is the vector between the two considered points in the $d-1$ dimensional subspace.
The calculation proceeds as in the one-dimensional case with the substitution $m^2\rightarrow m^2+k_\perp^2$, and the integration over $k_\perp$ in front of all the equations. The space integration is obviously over $d^{d-1}r'_\perp r dr d\theta$ (we use $r'_\perp$ to stress that it is different from that appearing in Eq. (\[2pthd\]) that instead is $r_\perp=0$, since the two-point function at coincident point has to be considered.) Thus the analogue of Eq. (\[trrhon\]) is \^n= d\^[d-1]{} r’\_ (n-). To have a finite result both the integrations are in a finite region. This means that the system must live in a finite box with $\int d^{d-1} r'_\perp={\cal A}$ and a cut-off proportional to $a^{-1}$ must be understood for the integration over $k_\perp$.
In this way the entropy is S=- . \[Sd\] In general, the integral diverges like $a^{-(d-1)}$. However the coefficient of this divergence (and other sub-leading divergences which can occur for sufficiently large $d$) are in general analytic in $m$. There is, however, a finite piece which behaves as $m^{d-1}$ in agreement with (\[Sscal\]). Note however that as $d\to1$ this singular term combines with the non-singular piece $\propto a^{-(d-1)}$ to give the previously found result $\propto\log(ma)$.
Just as for the classical free energy, one may also conjecture a finite-size scaling form for the entropy of the form s\_[sing]{}(L,g,h,T)=L\^[-(d-1)]{} s\_[FSS]{}(L\^[1/]{} |g-g\_c|, L\^[y\_h]{} h,L\^[z]{} T). For $h=T=0$ it reads s\_[sing]{}(L,g,h=T=0)= A\_s L\^[-(d-1)]{} s\_[FSS]{}(L/), with $s_{\rm FSS}(L/\xi)$ a function which satisfies $s_{\rm FSS}(0)=1$ and $s_{\rm FSS}(x)\propto 1/x$ for large $x$, in order to reproduce the infinite volume limit, and $A_s$ is a non universal constant fixing the normalisation of the entropy. In one dimension, this relation again does not make sense. The $\log L$ behaviour may be seen as resulting from a cancellation between the $L^{-(d-1)}$ behaviour of $s_{\rm sing}$ against a constant term coming from the regular part, with $A_s=O((d-1)^{-1})$ as $d\to1$. This gives the anticipated form S\_[sing]{}(L,)= (L + s\_[FSS]{}(L/)), that has been explicitly checked for the Gaussian model in Sec. \[sectfss\].
*Acknowledgments*. JC thanks D. Huse for first bringing this problem to his attention. This work was supported in part by the EPSRC under Grant GR/R83712/01 The initial phase was carried out while JC was a member of the Institute for Advanced Study. He thanks the School of Mathematics and the School of Natural Sciences for their hospitality. This stay was supported by the Bell Fund, the James D. Wolfensohn Fund, and a grant in aid from the Funds for Natural Sciences.**
Note added.
===========
We recently were made aware, some time after this paper appeared in print, that the result (3.32) for more than one interval is incorrect in general. This error can be traced to the assumption at the beginning of Sec. 3C that there exists a uniformizing transformation for the $n$-sheeted Riemann surface into the sphere. In fact the surface has in general a non-trivial genus and therefore such a transformation does not exist. For the case of two intervals and $n=2$ it can readily be seen that the surface is topologically a torus, and therefore the partition function $Z_2$ depends on the whole operator content of the theory, not only the central charge as claimed. For other values of $n$ the partition function should also depend on the structure constants (OPE coefficients). It is therefore highly unlikely that this would not remain true for the derivative at $n=1$. Indeed, a study of the computation for the 2-interval case for a compactified boson, first carried out in 1987 by Dixon et al.[@Dixon], shows that this is the case, and that our (3.32) is incorrect. This is also the result of recent numerical investigations [@Gliozzi; @Furokawa]. We are grateful to V. Pasquier and F. Gliozzi for pointing out this work.
Although this error also carries over to the case of a semi-infinite system, mentioned just below (3.32) and also to Eq. (3.33), it does not impact on our later application of these methods to studying the time-dependence of the entanglement entropy following a quantum quench[@cc3]. This is because in this case only the limiting behavior, when the various cross-ratios are either small or large, is needed, and this is insensitive to the precise form of (3.32). Likewise, as far as we are aware, all the other conclusions of the present paper remain valid.
Two-point function in a $n$-sheeted Riemann geometry: Infinite and finite volume results
========================================================================================
The Green function $G({\bf r,r'})$ of the Helmholtz differential equation $(-\nabla^2_{\bf r}+m^2)f({\bf r})=0$, with specific boundary condition, is the solution of $$(-\nabla^2_{\bf r}+m^2)G({\bf r,r'})=\delta^d({\bf r-r'})\,.$$ $G({\bf r,r'})$ admits an eigenfunction expansion $$G({\bf r,r'})=\sum_k N_k \phi_k({\bf r}) \phi_k({\bf r'})\,,$$ in terms of $\phi_k({\bf r})$, eigenfunctions of the Helmholtz differential operator $(-\nabla^2_{\bf r}+m^2)$. $N_k$ is a normalisation constant that should be derived from the orthonormality requirement of the eigenfunctions $$N_m\int \phi_m({\bf r})\phi_n({\bf r'})d^d\!r=\delta_{nm}\,.$$
In the case of a $n$-sheeted Riemann surface we are interested in, the eigenvalue problem is solved in polar coordinates ${\bf r}=(x,y)=(r\cos\theta,r\sin\theta)$. A complete set of eigenfunctions is \_[,i]{}\^a=([i]{}) J\_(\_i r),\_[,i]{}\^b=([i]{}) J\_(\_i r), with $J_\nu(x)$ Bessel functions of the first kind. Note that the Bessel functions of the second kind $Y_\nu(x)$, that are also eigenfunctions of the same differential operator, do not enter in the expansion since we require regularity at $r=0$.
Let us consider first the solution in a finite geometry. The infinite volume limit is recovered by taking the limit $L\rightarrow \infty$ in the sense of distributions. Imposing the $2\pi n$ periodicity boundary condition we have that $\nu$ is an integer multiple of $1/n$, i.e. $\nu=k/n$. Constraining the eigenfunctions to vanish at $r=L$, the eigenvalues are $\lambda_i L=\alpha_{\nu,i}$, with $\alpha_{\nu,i}$ the $i$-th zero of the Bessel function.
Using the orthogonality relation of the Bessel functions (see e.g. [@as]) $$\int_0^L r dr J_\nu(\alpha_{\nu,i} r/L)J_{\nu'}(\alpha_{\nu',i} r/L)=
\frac{L^2}{2}J^2_{\nu+1}(\alpha_{\nu,i})\delta_{\nu\nu'}\,,$$ we get from the orthonormality requirement $$N_{i,k}=\frac{d_k}{2\pi n}\frac{2/L^2}{J^2_{k/n+1}(\alpha_{k/n,i})}\,.$$ that gives the Green function G(r,,r’,’)= \_[k=0]{}\^d\_k \_[i=1]{}\^ \_k(,’). \[2ptL\] The factor $d_k$ comes from $\int_0^{2\pi} 1dx =2\pi$ and $\int_0^{2\pi}dx \cos^2 x= \int_0^{2\pi}dx \sin^2 x =\pi$.
In the limit $L\rightarrow\infty$, the index $i$ becomes continuous $\alpha_{\nu,i}/L\rightarrow\lambda$ and the $\delta_{\nu\nu'}$ in the orthonormality condition is replaced by $\delta(\lambda-\lambda')$. The limit of the normalisation factor is $N_k(\lambda)=d_k\lambda/(2\pi n)$, that leads to the two-point function reported in the text (\[2pt\]). Note that a slightly different form of this two-point function (satisfying Dirichlet boundary conditions) was used to investigate the critical behaviour at an edge [@c-83].
[99]{} C. H. Bennett, H. J. Bernstein, S. Popescu, and B. Schumacher, Concentrating Partial Entanglement by Local Operations, 1996 Phys. Rev. A [**53**]{} 2046 \[quant-ph/9511030\]. A. Osterloh, L. Amico, G. Falci and R. Fazio, Scaling of Entanglement close to a Quantum Phase Transitions, 2002 Nature [**41**]{}6 608;\
T. J. Osborne and M. A. Nielsen, Entanglement in a simple quantum phase transition, 2002 Phys. Rev. A [**66**]{} 32110 \[quant-ph/0202162\];\
H. Barnum, E. Knill, G. Ortiz, R. Somma, and L. Viola, A Subsystem-Independent Generalization of Entanglement, 2004 Phys. Rev. Lett. [**92**]{} 107902 \[quant-ph/0305023\];\
F. Verstraete, M. A. Martin-Delgado, and J. I. Cirac, Diverging Entanglement Length in Gapped Quantum Spin Systems, 2004 Phys. Rev. Lett. [**92**]{} 087201 \[quant-ph/0311087\]. G. Vidal, J. I. Latorre, E. Rico and A. Kitaev, Entanglement in quantum critical phenomena, 2003 Phys. Rev. Lett. [**90**]{} 227902;\
J. I. Latorre, E. Rico, and G. Vidal, Ground state entanglement in quantum spin chains, 2004 Quant. Inf. and Comp. [**4**]{} 048 \[quant-ph/0304098\];\
J. I. Latorre, C. A. Lutken, E. Rico, and G. Vidal, Fine-grained entanglement loss along renormalization group flows, 2004 [*Preprint*]{} quant-ph/0404120. B.-Q. Jin and V. E. Korepin, Quantum Spin Chain, Toeplitz Determinants and Fisher-Hartwig Conjecture, 2004 J. Stat. Phys. [**116**]{} 79 \[quant-ph/0304108\]. N. Lambert, C. Emary, and T. Brandes, Entanglement and the Phase Transition in Single Mode Super-radiance, 2004 Phys. Rev. Lett. [**92**]{} 073602 \[quant-ph/0309027\]. H. Casini and M. Huerta, A finite entanglement entropy and the c-theorem, 2004 [*Preprint*]{} hep-th/0405111. C. Holzhey, F. Larsen, and F. Wilczek, Geometric and Renormalized Entropy in Conformal Field Theory, 1994 Nucl. Phys. B [**424**]{} 44 \[hep-th/9403108\]. J. Cardy and I. Peschel, Finite-size dependence of the free energy in two-dimensional critical systems, 1988 Nucl. Phys. B [**300**]{} 377. I. Affleck and A. W. W. Ludwig, Universal non-integer “ground-state degeneracy” in critical quantum systems, 1991 Phys. Rev. Lett. [**67**]{}, 161. M. Srednicki, Entropy and Area, 1993 Phys. Rev. Lett. [**71**]{} 666 \[hep-th/9303048\]. V. E. Korepin, Universality of Entropy Scaling in 1D Gap-less Models, 2004 Phys. Rev. Lett. [**92**]{} 096402 \[cond-mat/0311056\]. H. W. J. Blöte, J. Cardy and M. P. Nightingale, Conformal invariance, the central charge, and universal finite-size amplitudes at criticality, 1986 Phys. Rev. Lett. [**56**]{} 742. I. Affleck, Universal term in the free energy at a critical point and the conformal anomaly, 1986 Phys. Rev. Lett. [**56**]{} 746. A. B. Zamolodchikov, Irreversibility of the flux of the renormalization group in a 2D field theory, 1986 JETP Lett. [**43**]{} 731 \[1986 Pisma Zh. Eksp. Teor. Fiz. [**43**]{} 565\]. C. Tsallis, Possible generalization of Boltzmann-Gibbs statistics, 1988 J. Stat. Phys. [**52**]{} 479. A. A. Belavin, A. M. Polyakov and A. B. Zamolodchikov, Infinite conformal symmetry in two-dimensional quantum field theory, 1984 Nucl. Phys. B [**241**]{} 333. J. Cardy, Conformal invariance and surface critical behavior, 1984 Nucl. Phys. B [**240**]{} 514. M. Abramowitz and I. A. Stegun, Handbook of mathematical functions, 1972 National Bureau of Standards, Tenth Printing. T. Nishino, Density Matrix Renormalization Group Method for 2D Classical Models, 1995 J. Phys. Soc. Jpn. [**74**]{} 3598 \[cond-mat/9508111\];\
T. Nishino and K. Okunishi, Density Matrix and Renormalization for Classical Lattice Models, in “Strongly Correlated Magnetic and Superconducting Systems”, 1997 Lect. Notes Phys. [**478**]{} 167 \[cond-mat/9610107\]. R. J. Baxter, Exactly Solved Models in Statistical Mechanics, 1982 Academic Press, San Diego. I. Peschel, M. Kaulke, and O. Legeza, Density-matrix spectra for integrable models, 1999 Ann. Physik (Leipzig) [**8**]{} 153 \[cond-mat/9810174\]. G. N. Watson, A Treatise on the Theory of Bessel Functions, 1944, Cambridge Univ. Press.\
I. N. Sneddon, On some infinite series involving the zeros of Bessel functions of the first kind, 1960 Proc. Glasgow Math. Assoc. [**4**]{} 144. J. Cardy, Critical behaviour at an edge, 1983 J. Phys. A [**16**]{} 2617. L. Dixon, D. Friedan, E. Martinec and S. Shenker, The Conformal Field Theory of Orbifolds, 1987 Nucl. Phys. B [**282**]{}, 13. M. Caraglio and F. Gliozzi, Entanglement Entropy and Twist Fields, 2008 arXiv:0808.4094. S. Furokawa, V. Pasquier and J. Shiraishi, Mutual Information and Compactification Radius in a $c=1$ Critical Phase in One Dimension, 2008 arXiv:0809.5113. P. Calabrese and J. Cardy, Evolution of Entanglement Entropy in One-Dimensional Systems, 2005, J.Stat.Mech., P04010.
[^1]: As far as we are aware, this analysis has not been extended to systems with many degrees of freedom such as we consider in this paper. Indeed, we shall see that for such systems $S_A$ can be much larger than unity, so cannot have such a simple interpretation.
[^2]: This restriction is not necessary, but the set-up is easier to picture in this case.
[^3]: We have assumed that theory is trivial in the infrared. If the RG flow is towards a non-trivial theory, $c$ should be replaced by $c_{UV}-c_{IR}$.
[^4]: This holds only for non-interacting theories: in the presence of interactions the sum of all the zero-point diagrams has to be taken into account.
[^5]: In the rest of this section, all logarithms are assumed taken to base $e$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'It is a survey of the results on stable homotopy types of polyhedra of small dimensions, mainly obtained by H.-J. Baues and the author [@bd1; @bd3; @bd4]. The proofs are based on the technique of matrix problems (bimodule categories).'
address: |
Department of Mechanics and Mathematics\
Kyiv Taras Shevchenko University\
01033 Kyiv\
Ukraine
author:
- 'Yuriy A. Drozd'
title: Matrix problems and stable homotopy types of polyhedra
---
This paper is a survey of some recent results on stable homotopy types of polyhedra. The common feature of these results is that their proofs use the technique of the so called *matrix problems*, which was mainly elaborated within framework of representation theory. I think that this technique is essential in homotopy theory too, and perhaps even in much more general setting of triangulated categories. I hope that the considerations of Section 3 are persuasive enough. Certainly, I could not cover all such results, restricting by the stable homotopy classification of polyhedra of small dimensions obtained in [@bd1; @bd3; @bd4; @bh]. I tried to present these results in a homogeneous way and also replace references to rather sophisticated topological sources by simpler ones. The latter mainly concern some basic facts about homotopy groups of spheres, which can be found in [@hu] or [@tod]. I also used the book [@sw] as a standard source of references; maybe some readers will prefer [@sp] or [@co]. Most of these references are collected in Section 1. For the matrix problems I have chosen the language of *bimodule categories* explained in Section 2, since it seems the simplest one as well as the most appropriate for applications.
Note that almost the same arguments that are used in Sections 5 and 6 can be applied to the classification of polyhedra with only $2$ non-trivial homology groups [@bd4], while the dual arguments were applied to the spaces with only $2$ non-trivial homotopy groups in [@bd2]. Rather similar are also calculations in [@hen] (see also the Appendix by Baues and Henn to [@bd1]). I hope that any diligent reader of this survey will be able to comprehend the arguments of these papers too.
I am extremely indebted to H.-J. Baues, who was my co-author and my guide to the topological problems, and to C. M. Ringel, whose wonderful organising activity had made such a pleasant and fruitful collaboration possible. Most of our joint results H.-J. Baues and I obtained during my visits to the Max-Plank-Institut für Mathematik, and I highly acknowledge its support.
Generalities on stable homotopy types {#s1}
=====================================
All considered spaces are supposed *pathwise connected* and *punctured*; we denote by $*_X$ (or by $*$ if there can be no ambiguity) the marked point of the space $X$. $B^n$ and $S^{n-1}$ denote respectively the $n$-dimensional *ball* $\setsuch{\fX\in\mR^n}{||\fX||\le1}$ and the $(n-1)$-dimensional *sphere* $\setsuch{\fX\in\mR^n}{||\fX||=1}$, both with the marked point $(1,0,\dots,0)\,$. As usually, we denote by $X\vee Y$ the *bouquet* (or one point union) of $X$ and $Y$, i.e. the factor space $X\sqcup Y$ by the relation $*_X=*_Y$, and identify it with $*_X\xx Y\cup X\xx*_Y\sb X\xx Y$; we denote by $X\wee Y$ the factor space $X\xx Y/X\vee Y$. In particular, we denote by $\Si X=S^1\wee X$ the *suspension* of $X$ and by $\Si^nX=\underbrace{\Si\dots\Si}_{n\mbox{ \scriptsize times}}X$ its $n$-th suspension. The word “*polyhedron*” is used as a synonym of “*finite* .” One can also consider bouquets of several spaces $\bigvee_{i=1}^sX_i$; if all of them are copies of a fixed space $X$, we denote such a bouquet by $sX$.
We recall several facts on stable homotopy category of . We denote by $\hot(X,Y)$ the set of homotopy classes of continuous maps $X\to Y$ and by $\cw$ the *homotopy category* of polyhedra, i.e. the category whose objects are polyhedra and morphisms are homotopy classes of continuous maps. The suspension functor defines a natural map $\hot(X,Y)\to\hot(\Si X,\Si Y)$. Moreover, the Whitehead theorem [@sw Theorem 10.28] shows that the suspension functor *reflects isomorphisms*: if $\Si f$ is an isomorphism (i.e. a homotopy equivalence), so is $f$. We set $\hos(X,Y)=\dlim_n\hot(\Si^n X,\Si^n Y)$. If $\al\in\hot(\Si^n X,\Si^n Y),\ \be\in\hot(\Si^m Y,\Si^m Z)$, one can consider the class $\Si^n\be\circ\Si^m\al\in\hot(\Si^{m+n}X,\Si^{n+m}Z)$, which is, by definition, the product $\be\al$ of the classes of $\al$ and $\be$ in $\hos(X,Z)$. Thus we obtain the *stable homotopy category* of polyhedra $\cws$. Actually, if we only deal with *finite* , we need not go too far, since the Freudenthal theorem [@sw Theorem 6.26] implies the following fact.
\[11\] If $X,Y$ are of dimensions at most $d$ and $(n-1)$-connected, where $d<2n-1$, then the map $\hot(X,Y)\to\hot(\Si X,\Si Y)$ is bijective. If $d=2n-1$, this map is surjective. In particular, the map $\hot(\Si^mX,\Si^mY)\to\hos(X,Y)$ is bijective if $m> d-2n+1$ and surjective if $m=d-2n+1$.
Here $(n-1)$*-connected* means, as usually, that $\pi_k(X)$, the $k$-th homotopy group of $X$, is trivial for $k\le n-1$. Thus for *all* polyhedra of dimension at most $d$ the map $\hot(\Si^mX,\Si^mY)\to\hos(X,Y)$ is bijective if $m\ge d$ and surjective if $m=d-1$.
Note also that the natural functor $\cw\to\cws$ reflects isomorphisms.
Since we are only interested in stable homotopy classification, we identify, in what follows, polyhedra and continuous maps with their images in $\cws$. We denote by $\cwf$ the full subcategoy of $\cws$ consisting of all spaces $X$ with torsion free homology groups $\rH_i(X)=\rH_i(X,\mZ)$ for all $i$.
Recall that any suspension $\Si^nX$ is an *H-cogroup* [@sw Chapter 2], commutative if $n\ge 2$, the category $\cws$ is an additive category. Moreover, one can deduce from the Adams’ theorem [@sw Theorem 9.21] that this category is actually *fully additive*, i.e. every idempotent $e\in\hos(X,X)$ splits. In this case it means that there is a decomposition $\Si^mX\iso Y\vee Z$ for some $m$, such that $e$ comes from the map $\eps:Y\vee Z\to Y\vee Z$ with $\eps(y)=y$ for $y\in Y$ and $\eps(z)=*_{Y\vee Z}$ for $z\in Z$. We call a polyhedron $X$ *indecomposable* if $X\iso Y\vee Z$ implies that either $Y$ or $Z$ are contractible (i.e. isomorphic in $\cw$ to the 1-point space). Actually, the category $\cws$ is a *triangulated category* [@gm]. The suspension plays the role of shift, while the *triangles* are the *cone sequences* $X\str f Y\to Cf\to \Si X$ (and isomorphic ones), where $Cf=CX\cup_f Y$ is the *cone of the map* $f$, i.e the factor space $CX\sqcup Y$ by the relation $(x,0)\sim f(x)$; $CX=X\xx I/X\xx 1$ is the cone over the space $X$. Note that cone sqeuences coincide with *cofibration sequences* in the category $\cws$ [@sw Proposition 8.30]. Recall that a cofibration sequence is a such one $$\label{e11}
X \str f Y \str g Z \str h \Si X \str{\Si f} \Si Y$$ that for every polyhedron $P$ the induced sequences $$\label{e12}
\begin{split}
& \hos(P,X) \str{f_*} \hos(P,Y) \str{g_*} \hos(P,Z) \str{h_*} \hos(P,\Si X) \str{\Si f_*} \hos(P,\Si Y), \\
& \hos(\Si Y,P) \str{\Si f^*} \hos(\Si X,P) \str{h^*} \hos(Z,P) \str{g^*} \hos(Y,P) \str{ f^*} \hos(X,P)
\end{split}$$ are exact. In particular, we have an exact sequence of *stable homotopy groups* $$\label{e13}
\pi^S_k(X) \str {f_*} \pi^S_k(Y) \str{g_*} \pi^S_k(Z) \str{h_*}
\pi^S_{k-1}(X) \str{\Si f_*} \pi^S_{k-1}(Y),$$ where $\pi^S_k(X)=\dlim_m\pi_{k+m}(\Si^mX)=\hos(S^k,X)$. Certainly, one can prolong the sequences and into infinite exact sequences just taking further suspensions.
Every is obtained by *attaching cells*. Namely, if $X^n$ is the $n$-th skeleton of $X$, then there is a bouquet of balls $B=mB^{n+1}$ and a map $f:mS^n\to X^n$ such that $X^{n+1}$ is isomorphic to the cone of $f$, i.e. to the space $X^n\cup_fB$. It gives cofibration sequences like and exact sequences like and .
We denote by $\cw^k_n$ the full subcategory of $\cw$ formed by $(n-1)$-connected $(n+k)$-dimensional polyhedra and by $\cwf^k_n$ the full subcategory of $\cw^k_n$ formed by the polyhedra $X$ with torsion free homology groups $\rH_i(X)$ for all $i$. Proposition \[11\] together with the fact that every map of is homotopic to a cell map, also implies the following result.
\[12\] The suspension functor $\Si$ induces equivalences $\cw^k_n\to\cw^k_{n+1}$ for all $n>k+1$. Moreover, if $n=k+1$, the suspension functor $\Si:\cw^k_n\to\cw^k_{n+1}$ is a *full representation equivalence*, i.e. it is full, dense and reflects isomorphisms. (*Dense* means that every object from $\cw^k_{n+1}$ is isomorphic (i.e. homotopy equivalent) to $\Si X$ for some $X\in\cw^k_n$.)
Therefore, setting $\cw^k=\cw^k_{k+2}\iso\cw^k_n$ for $n>k+1$, we can consider it as a full subcategory of $\cws$. The same is valid for $\cwf^k=\cwf^k_{k+2}$. Note also that $\cw^k_n$ naturally embeds into $\cw^k_{n+1}$. It leads to the following notion [@ba1].
\[13\] An *atom* is an indecomposable polyhedron $X\in\cw^k_{k+1}$ not belonging to the image of $\cw^k_k$. A *suspended atom* is a polyhedron $\Si^mX$, where $X$ is an atom.
Then we have an obvious corollary.
\[14\] Every object from $\cw^k_n$ with $n\ge k+1$ is isomorphic (i.e. homotopy equivalent) to a bouquet $\bigvee_{i=1}^sX_i$, where $X_i$ are suspended atoms. Moreover, any suspended atom is indecomposable (thus indecomposable objects are just suspended atoms).
Note that the decomposition in Corollary \[14\] is, in general, not unique [@fr]. That is why an important question is the structure of the *Grothendieck group* $K_0(\cw^k)$. By definition, it is the group generated by the isomorphism classes $[X]$ of polyhedra from $\cw^k$ subject to the relations $[X\vee Y]=[X]+[Y]$ for all possible $X,Y$. The following results of Freyd [@fr; @co] describe the structure of this group.
\[15\]
1. Two polyhedra $X,Y\in\cw^k$ are said to be *congruent* if there is a polyhedron $Z\in\cw^k$ such that $X\vee Z\iso Y\vee Z$ (in $\cw^k$).
2. A polyhedron $X\in\cw^k$ is said to be $p$*-primary* for some prime number $p$ if there is a bouquet of spheres $B$ such that the map $p^m1_X:X\to X$ can be factored through $B$, i.e. there is a commutative diagram $$\xymatrix{ X \ar[rr]^{p^m1_X}\ar[dr] && X \\
& B \ar[ur] }$$
\[16\] The group $K_0(\cw^k)$ (respectively $K_0(\cwf^k)\,$) is a free abelian group with a basis formed by the congruence classes of $p$-primary suspended atoms from $\cw^k$ (respectively from $\cwf^k$) for all prime numbers $p\in\mN$.
Therefore, if we know the “place” of every atom class $[X]$ in $K_0(\cw^k)$ or $K_0(\cwf^k)$, i.e. its presentation as a linear combination of classes of $p$-primary suspended atoms, we can deduce herefrom all decomosition rules for $\cw^k$ or $\cwf^k$.
Bimodule categories {#bc}
===================
We also recall main notions concerning *bimodule categories* [@d1; @d2]. Let $\bA,\bB$ be two fully additive categories. An *$\bA$-$\bB$-bimodule* is, by definition, a biadditive bifunctor $\sU:\bA^\circ\xx\bB\to\Ab$. As usually, given an element $u\in\sU(A,B)$ and morphisms $\al\in\bA(A',A),\ \be\in\bB(B,B')$, we write $\be u\al$ instead of $\sU(\al,\be)u$. Given such a functor, we define the *bimodule category $\El(\sU)$* (or the category of *elements of the bimodule* $\sU$, or the category of *matrices over* $\sU$) as follows.
- The set of *objects* of $\El(\sU)$ is the disjoint union $$\ob\El(\sU)=\bigsqcup_{\substack{A\in\ob\bA\\B\in\ob\bB}}\sU(A,B).$$
- A *morphism* from $u\in\sU(A,B)$ to $u'\in\sU(A',B')$ is a pair $(\al,\be)$ of morphisms $\al\in\bA(A,A'),\,
\be\in\bB(B,B')$ such that $u'\al=\be u$ in $\sU(A,B')$.
- The product $(\al',\be')(\al,\be)$ is defined as the pair $(\al'\al,\be'\be)$.
Obviously, $\El(\sU)$ is again a fully additive category.
Suppose that $\ob\bA\supset\set{\lst An},\ \ob\bB\supset\set{\lst Bm}$ such that every object $A\in\ob\bA$ ($B\in\ob\bB$) decomposes as $A\iso\b+_{i=1}^nk_iA_i$ (respectively, $B\iso\b+_{i=1}^ml_iB_i$). Then $\bA^{\circ}$ (respectively, $\bB$) is equivalent to the category of finitely generated projective right (left) modules over the ring of matrices $(a_{ij})_{n\xx n}$ with $a_{ij}\in \bA(A_j,A_i)$ (respectively, $(b_{ij})_{m\xx m}$ with $b_{ij}\in\bB(B_j,B_i)$). We denote these rings respectively by $|\bA|$ and $|\bB|$. We also denote by $|\sU|$ the $|\bA|\mbox{-}|\bB|$-bimodule consisting of matrices $(u_{ij})_{m\xx n}$, where $u_{ij}\in\sU(A_j,B_i)$. Then $\sU(A,B)$, where $A,B$ are, respectively, a projective right $|\bA|$-module and a projective left $|\bB|$-module, can be identified with $A\*_{|\bA|}|\sU|\*_{|\bB|}B$. Elements from this set are usually considered as block matrices $(U_{ij})_{m\xx n}$, where the block $U_{ij}$ is of size $l_i\xx k_j$ with entries from $\sU(A_j,B_i)$. To form a direct sum of such elements, one has to write direct sums of the corresponding blocks at each place. Certainly, some of these blocks can be “empty,” if $k_j=0$ or $l_i=0$. An empty block is indecomposable it is of size $0\xx 1$ (in $\sU(A_j,0)$) or $1\xx 0$ (in $\sU(0,B_i)\,$); we denote it respectively by $\0^j$ or by $\0_i$.
In many cases the rings $|\bA|$ and $|\bB|$ can be identified with *tiled subrings* of rings of integer matrices. Here a *tiled subring* in $\mat(n,\mZ)$ is given by an integer matrix $(d_{ij})_{n\xx n}$ such that $d_{ii}=1$ and $d_{ik}|d_{ij}d_{jk}$ for all $i,j,k$; the corresponding ring consists of all matrices $(a_{ij})$ such that $d_{ij}|a_{ij}$ for all $i,j$ (especially $a_{ij}=0$ if $d_{ij}=0$).
\[bc-1\] Let $\bA\subset\mat(2,\mZ)$ be the tiled ring given by the matrix $$\begin{pmatrix}
1 & 12 \\ 0 & 1
\end{pmatrix} ,$$ $\sU$ be the set of $2\xx2$-matrices $(u_{ij})$ with $u_{ij}\in \mZ/24$ if $i=1,j=2$, $u_{ij}\in\mZ/2$ otherwise. We define $\sU$ as an $\bA\mbox{-}\bA$-bimodule setting $$\begin{aligned}
\begin{pmatrix} a&12b\\ 0&c\end{pmatrix} \begin{pmatrix} u_1&u_2\\u_3&u_4\end{pmatrix}=& \begin{pmatrix} au_1+bu_3\ &
au_2+12bu_4\\cu_3&cu_4 \end{pmatrix} ;\\
\begin{pmatrix} u_1&u_2\\u_3&u_4 \end{pmatrix} \begin{pmatrix} a&12b\\ 0&c\end{pmatrix} =&
\begin{pmatrix} au_1 &\ cu_2+12bu_1\\ au_3 &cu_4+bu_3\end{pmatrix} .
\end{aligned}$$ If we need to indicate this action, we write $$\begin{pmatrix}
1 & 12^* \\ 0 & 1
\end{pmatrix}\ \text{ and }\ \begin{pmatrix}
\mZ/2 &\ \mZ/24 \ \\\ \mZ/2^* & \mZ/2
\end{pmatrix}$$ for the matrix defining the ring $\bA$ and for the bimodule $\sU$. Thus the multiplications of the elements marked by stars is given by the *$\ast$-rule*: $$\label{e21}
(12a^*)\cdot (u\mod2^*)=ac\mod2.$$
\[bc-1a\] In the classification of torsion free atoms below the following bimodule plays the crucial role. We consider the tiled rings $\bA_2\subset\mat(2,\mZ)$ and $\bB_2\subset\mat(7,\mZ)$ given respectively by the matrices $$\begin{pmatrix}
1 & 2 & 2 & 12 & 24 & 12 & 24 \\ 1 & 1 & 1 & 12 & 24 & 6 & 24 \\ 1 & 2 & 1 & 12 & 24 & 12 & 24 \\
0&0&0& 1&2&12^*&12\\ 0&0&0& 1&1&12&6\\ 0&0&0& 0&0&1&1& \\ 0&0&0& 0&0&0&1
\end{pmatrix}\ \text{ and }\ \begin{pmatrix}
1 & 12^* \\ 0 & 1
\end{pmatrix} .$$ The $\bA_2\mbox{-}\bB_2$-bimodule $\sU_2$ is defined as the set of matrices of the form $$\begin{pmatrix}
\ \mZ/24 & 0\\ \mZ/12 &0 \\ \mZ/12 &0\\ \mZ/2 &\ \mZ/24 \ \\ 0 &\mZ/12 \\ \mZ/2^* &\mZ/2 \\ 0 &\mZ/2
\end{pmatrix} .$$ The multiplication in $\sU_2$ is given by the natural matrix multiplication, but taking into account the $\ast$-rule .
We shall use the following description of indecomposable elements in $\El(\sU_2)$. Set $I_1=\set{1,2,3,4,6},\,I_2=\set{4,5,6,7}$, $V=\setsuch{v\in\mN}{1\le v\le6}$, $V_1=\setsuch{v\in\mN}{1\le v\le 12}$, $V_2=\set{1,2,3}$.
\[bc-2\] A complete list $\kL_2$ of non-isomorphic indecomposable objects from $\El(\sU_2)$ consists of
- empty objects $\0^j\ (j=1,2)$ and $\0_i\ (1\le i\le 7)$;
- objects $v^j_i\in\sU(A_j,B_i)\ (j=1,2;\,i\in I_j;\ v\in V_1$ if $i=1;\ v=1$ if $i=6,7$ or $(ij)=(14);\ v\in V$ otherwise);
- objects $v^j_{il}=\ds\binom {v^j_i}{1^j_l}\ (j=1,2;\ i=1,2,3,\,l=4,6$ if $j=1;\ i=4,5,\,l=6,7$ if $j=2$; if $(il)=(26)$ or $(57)$ then $v\in V_2$; otherwise $v\in V$);
- objects $v_{44}=(1^1_4 \ v^2_4)$ with $v\in V$;
- objects $v_{4l}=\ds \begin{pmatrix}
1^1_4 & v^2_4 \\ 0& 1^2_l
\end{pmatrix} $ with $l=6,7$ and $v\in V$;
- objects $v_iw_{44}=\ds \begin{pmatrix}
v^1_i &0 \\ 1^1_4 & w^2_4
\end{pmatrix} $ with $i=1,2,3$ and $v,w\in V$;
- objects $v_iw_{4l}=\ds \begin{pmatrix}
v^1_i & 0\\ 1^1_4 & w^2_4 \\ 0& 1^2_l
\end{pmatrix} $ with $i=1,2,3,\,l=6,7$ and $v,w\in V$.
Here the indices define the block containing the corresponding element.
Decompose $\sU$ into $2$-primary and $3$-primary parts. Since for every two matrices $M_2,M_3\in\GL(n,\mZ)$ there is a matrix $M\in\GL(n,\mZ)$ such that $M\equiv M_2\mod2$ and $M\equiv M_3\mod3$, we can consider $2$-primary part and $3$-primary part separately. Note that in the $3$-primary part the blocks $u^1_4,u^1_6,u^2_6$ and $u^2_7$ vanish, while the other non-zero blocks of $u\in\ob(\sU_2)$ are with entries from $\mZ/3$ and there are no restrictions on elementary transformation of the matrix $u$. Thus every element in the $3$-primary part is a direct sum of elements $1^j_i$ with $j=1,i=1,2,3$ or $j=2,i=4,5$.
For elements $u,u'$ of the $2$-primary part write $u<u'$ if $u'=ua$ for some non-invertible $a\in\bA_2$. Then we have the following inequalities: $$\begin{aligned}
& 1^1_1<1^1_3<1^1_2<2^1_1<2^1_3<2^1_2<4^1_1,\\
& 1^1_6<1^1_4<4^1_1 \ \text{ and }\ 1^1_6<2^1_2; \\
& 1^2_4<1^2_5<2^2_4<2^2_5<4^2_4,\\
& 1^2_7<1^2_6<4^2_4 \ \text{ and }\ 1^2_7<2^2_5.
\end{aligned}$$ Using them, one can easily decompose the parts $$\tilde u^1= \begin{pmatrix} u^1_1\\ u^1_2\\ u^1_3\end{pmatrix} \ \text{ and }\
\tilde u^2= \begin{pmatrix}u^2_4\\ u^2_5 \end{pmatrix}$$ into a direct sum of empty and $1\xx 1$ matrices. Now we obtain a column splitting of the remaining matrices, and with respect to the transformation that do not change $\tilde u^1$ and $\tilde u^2$, these columns are linearly ordered. Therefore, we can also split them into empty and $1\xx 1$ blocks. Together with $\tilde u^1$ and $\tilde u^2$, it splits the whole matrix $u$ into a direct sum of matrices of the forms from the list $\kL_2$, where $v,w$ are powers of $2$. Adding $3$-primary parts, we get the result.
\[bc-1b\] Consider the idempotents $e=\sum_{i\in I_1}e_{ii}\in\bA_2$ and $e'=e_{11}\in\bB$. Set $\bA_1=e\bA e,\,\bB_1=e'\bB_2e'
\iso\mZ$ and $\sU_1=e'\sU_2 e$. Then $\sU_1$ is an $\bA_1\mbox{-}\bB_1$-bimodule; elements from $\El(\sU_1)$ can be identified with those from $\El(\sU_2)$ having no second column and fifth row. Hence we get the following result.
\[bc-3\] A complete list $\kL_1$ of non-isomorphic indecomposable objects from $\El(\sU_1)$ consists of
- empty objects $\0_i\ (i\in I_1)$;
- objects $v_i\ (i\in I_1;\ v\in V_1$ if $i=1,\ v\in V$ if $i=2,3,\ v=1$ if $i=4,6$);
- objects $v_{il}=\ds\binom {v_i}{1_l}\ (i=1,2,3,\,l=4,6$; if $(il)=(26)$ then $v\in V_2$, otherwise $v\in V$).
Here the indices show the blocks where the corresponding elements are placed.
Bimodules and homotopy types {#bh}
============================
Bimodule categories arise in the following situation. Let $\bA$ and $\bB$ are two fully additive subcategories of the category $\Hos$. We denote by $\bA\dagg\bB$ the full subcategory of $\Hos$ consisting of all objects $X$ isomorphic (in $\Hos$) to the cones of morphisms $f:A\to B$ with $A\in\bA,\,B\in\bB$, or, the same, such that there is a cofibration sequence $$\label{bh-e1}
A \str f B \str{g} X \str{h} \Si A,$$ where $A\in\bA,\,B\in\bB$. Consider the $\bA$-$\bB$-bimodule $\sH$, which is the restriction on $\bA^\circ\xx\bB$ of the “regular” $\Hos$-$\Hos$-bimodule $\hos$. If $f\in\hos(A,B)$ is an element of $\sH$, it gives rise to an exact sequence like with $X=Cf$. Moreover, since this sequence is a cofibration one, every morphism $(\al,\be):f\to f'$, where $f'\in\Hos(A',B')$ induces a morphism $\ga:X\to X'$, where $X'=Cf'$, such that the diagram $$\label{bh-e2}
\begin{CD}
A @>f>> B @>g>> X @>h>> \Si A @>\Si f>> \Si B\\
@V\al VV @V\be VV @VV\ga V @VV\Si\al V @VV\Si\be V\\
A' @>>f'> B'@>>g'> X' @>>h'> \Si A' @>>\Si f'> \Si B'
\end{CD}$$ commutes. In what follows we suppose that the categories $\bA$ and $\bB$ satisfy the following condition: $$\label{bh-e3}
\hos(B,\Si A)=0 \quad\text{for all}\quad A\in\bA,\ B\in\bB. %\tag{\bf{\star}}$$ In this situation, given a morphism $\ga:X\to X'$, we have that $h'\ga g=0$, hence $\ga g=g'\be$ for some $\be:B\to B'$. Moreover, since the sequence $$B \str g X \str h \Si A \str{\Si f} \Si B$$ is cofibration as well, and $\Si:\Hos(A,B)\to\Hos(\Si A,\Si B)$ is a bijection, there is a morphism $\al:A\to A'$, which makes the diagram commutative.
Note that neither $\ga$ is uniquely determined by $(\al,\be)$, nor $(\al,\be)$ is uniquely restored from $\ga$. Nevertheless, we can control this non-uniqueness. Namely, if both $\ga$ and $\ga'$ fit the diagram for given $(\al,\be)$, their difference $\ol\ga=\ga-\ga'$ fits an analogous diagram with $\al=\be=0$. The equality $\ol\ga g=0$ implies that $\ol\ga=\si h$ for some $\si:\Si A\to X'$, and the equality $h'\ol\ga=0$ implies that $\ol\ga=g'\tau$ for some $\tau:X\to B$. On the contrary, if $\ol\ga=\si h=g'\tau$ for *some* morphisms $\si:X\to \Si Y\,\tau:X\to Z$, where $Y\in\bA,\,Z\in\bB$, the condition implies that $\ol\ga g=h'\ol\ga=0$, so $\ol\ga$ fits the diagram with $\al=\be=0$.
Fix now $\ga$, and let both $(\al,\be)$ and $(\al',\be')$ fit for this choice of $\ga$. Then the pair $(\ol\al,\ol\be)$, where $\ol\al=\al-\al',\,\ol\be=\be-\be'$, fits for $\ga=0$. The equality $g'\ol\be=0$ implies that $\ol\be=f'\si$ for some $\si:B\to A'$, and the equality $(\Si\al)h=0$ implies that $\Si\al=\Si\tau\Si f$, or $\al=\tau f$ for some $\tau:B\to S$. On the contrary, if $(\ol\al,\ol\be):f\to f'$ is such that $\ol\be=f'\si$ and $\ol\al=\tau f$ with $\si,\tau:B\to A'$, then $g'\be=(\Si\al) h=0$, hence this pair fits with $\ga=0$.
Summarizing these considerations, we get the following statement.
\[bh-1\] Let $\bA,\bB$ be fully additive subcategories of $\Hos$ satisfying the condition , $\bA\dagg\bB$ be the full subcategory of $\Hos$ consisting of all spaces such that there is a cofibration with $A\in\bA$, $B\in\bB$. Denote by $\sH$ the bimodule $\hos$ considered as $\bA$-$\bB$-bimodule, by $\kI$ the ideal in $\bA\dagg\bB$ consisting of all morphisms $\ga:X\to X'$ that factor both through an object from $\Si\bA$ and through an object from $\bB$, and by $\kJ$ the ideal in $\El(\sH)$ consisting of all morphisms $(\al,\be):f\to f'$ such that $\be$ factors through $f'$ and $\al$ factors through $f$. Then the factor categories $\El(\sH)/\kJ$ and $\bA\dagg\bB/\kI$ are equivalent; an equivalence is induced by the maps $f\mapsto Cf$ and $(\al,\be)\mapsto\ga$, where $\ga$ fits a commutative diagram . Moreover, $\kI^2=0$, thus the functor $\bA\dagg\bB\to\bA\dagg\bB/\kI$ reflects isomorphisms.
We only have to check the last statement. But if $\ga:X\to X'$ factors as $X\str\tau B\str{g'}X'$ and $\ga':X'\to X''$ factors as $X'\str h' \Si A\str\si X''$, where $A\in\bA,\,B\in\bB$, then $\ga'\ga=0$, since $h'g:B\to\Si A$ and $\hos(B,\Si A)=0$.
\[bh-2\] In the situation of Theorem \[bh-1\], suppose that $\hos(B,A)=0$ for each $A\in\bA,\,B\in\bB$. Then $\El(\sH)\iso\bA\dagg\bB/\kI$. Moreover, the functor $\bA\dagg\bB\to\El(\sH)$ is a *representation equivalence*, i.e. it is dense, preserves indecomposables and reflects isomorphisms.
Note also that any *isomorpism* $f:A\str\sim B$ is a zero object in $\El(\sH)/\kJ$, since its identity map $(1_A,1_B)$ can be presented as $(f^{-1}f,ff^{-1})$. Obviously, the corresponding object from $\bA\dagg\bB$ is zero (i.e. contractible) too.
Small dimensions {#sd}
================
We now use Theorem \[bh-1\] to describe s of atoms of dimensions at most 5, or, the same, indecomposable objects in the categories $\cw^1_2$ and $\cw^2_3$.
\[sd-1\] It is well known that $\pi_n(S^n)=\mZ$ (freely generated by the identity map). It allows easily to describe atoms in $\cw^1_2$. Such an atom $X$ is (stably!) of the form $Cf$ for some map $f:mS^2\to nS^2$. Since $\hos(S^n,S^{n+1})=0$, Theorem \[bh-1\] can be applied. The map $f$ is given by an integer matrix. Using automorphisms of $mS^2$ and $nS^2$, we can transform it to a diagonal form. Hence, indecomposable gluings can only be if $m=n=1$; thus $f=q1_{S^2}$. One can see that such a gluing is indecomposable $q$ is a power of a prime number. The corresponding atom $S^2\cup_qB^3$ will be denoted by $M(q)$ and called *Moore atom*. It occurs in a cofibration sequence $$S^2 \str q S^2 \xarr{g(q)} M(q) \xarr{h(q)} S^3 \str q S^3.$$
For the next section we need more information about $2$-primary Moore atoms. We denote $M_t=M(2^t)$ and write $g_t,h_t$ instead of $g(2^t),h(2^t)$. These atoms can be included into the following commutative “octahedral” diagram [@gm], where $t=r+s$: $$\label{sd-e1}
\xy<0em,2.5em>
\xymatrix@R=1em{
& S^2 \ar[rr]^{g_r} \ar[dd]^{\,2^s} && M_r \ar[dr]^{h_r} \ar[dd]_{k_{tr}} \\
S^2 \ar[ur]^{2^r} \ar[dr]_{2^t} &&&& S^3 \\
& S^2 \ar[rr]_{g_t} \ar[dr]_{g_s} && M_t \ar[ur]_{h_t}\ar[dl]^{k_{st}} \\
&& M_s
} \endxy$$ Moreover, in this diagram $h_sk_{st}=2^rh_t$.
The exact sequence is here of the form $$\pi^S_k(S^2)\str q\pi^S_k(S^2)\larr\pi^S_k(M(q))\larr \pi^S_k(S^3)\str q \pi^S_k(S^3),$$ which gives the values of stable homotopy groups of the spaces $M(q)$ shown in Table \[tab1\] below.
$$\begin{array}{|c|c|c|c|}
\hline
k & 2 & 3 & 4 \\
\hline
\ \pi^S_k(M(q)),\ q \text{ odd}\ &\ \mZ/q\ & 0 & 0 \\
\hline
\ \pi^S_k(M_t),\ t>1 \ & \mZ/q &\ \mZ/2\ &\ \mZ/2\+\mZ/2\ \\
\hline
\pi^S_k(M_1) & \mZ/2 & \mZ/2 & \mZ/4 \\
\hline
\end{array}$$
(By the way, this table implies that all Moore atoms are pairwise non-isomorphic.) Actually, the only non-trivial is the group $\pi^S_4(M_1)$. It can be obtained as $\pi_6(\Si^2M_1)$, which is isomorphic to the $2$-primary component of $\pi_6(S^3)=\mZ/12$ (cf. [@hu Lemma XI.10.2]). To prove that the sequence $$0\larr \pi^S_4(S^2)=\mZ/2 \larr \pi^S_4(M_t) \larr \pi^S_4(S^3)=\mZ/2 \larr 0$$ splits if $t>1$, it is enough to consider the commutative diagram $$\label{sd-e2}
\xy<0pt,1.5em>
\xymatrix{
0 \ar[r] & \pi^S_4(S^2) \ar[r] \ar[d]_{0\,} & \pi^S_4(M_1) \ar[r]\ar[d] & \pi^S_4(S_3) \ar@{=}[d] \ar[r]& 0 \\
0 \ar[r] & \pi^S_4(S^2) \ar[r] & \pi^S_4(M_t) \ar[r] & \pi^S_4(S_3) \ar[r] &0
}\endxy$$ arising from the diagram with $r=1$. It shows that the second row of this diagram is the pushdown of the first one along the zero map; thus it splits.
\[sd-2\] Now we are able to describe atoms in $\cw^2_3$. They are cones $Cf$ for some $f:mS^4\to Y$ with $2$-connected $Y$ of dimension $4$. Again $\Hos(Y,S^5)=0$, so Theorem \[bh-1\] can be applied. Example \[sd-1\] shows that $Y$ is a bouquet of spheres $S^3,S^4$ and suspended Moore atoms $\Si M(q)$. Note that $\pi^S_4(Y)=\pi_4(Y)$ for every $Y$; in particular $\pi_4(S^4)=\mZ,\ \pi_4(S^3)=\mZ/2$ (generated by the suspended Hopf map $\eta_1=\Si\eta$; $\eta:S^3\to S^2\iso\mC\mP^1$ is given by the rule $\eta(a,b)=(a:b)$, where $(a,b)\in\mC^2$ are such that $|a|^2+|b|^2=1$) and $$\pi_4(\Si M(q))=\pi^S_3(M(q))=\begin{cases} \mZ/2 &\mbox{if }\ q=2^r \\
0 &\mbox{otherwise}. \end{cases}$$ The Hopf map $\eta_2=\Si^2\eta:S^4\to S^3$ and the inclusion $j:S^2\to M(q)$ give rise to an epimorphism $\eta_*:\pi_4(S_4)\to \pi_4(S^3)$ and to an isomorphism $j_*:\pi_4(S^3)\to\pi_4(\Si M_r)$, where $M_r=M(2^r)$. Moreover, if $t>r$, there is a map $M(2^r)\to M(2^t)$ that induces an isomorphism $\pi_4(M_r)\to\pi_4(M_t)$. If $Y=s_4S^4\vee s_3S^3\vee(\bigvee_{r=1}^\8 m_rM_r)$, a map $f:mS^4\to Y$ can be given by a matrix of the form $$\begin{pmatrix}
F_4 & F_3 & G_1 & G_2 & \dots
\end{pmatrix} ,$$ where $F_i$ is of size $m\xx s_i$ with entries from $\pi_4(S^i)$; $G_r$ is of size $m\xx m_r$ with entries from $\pi_4(\Si M_r)$ (some of these matrices can be “empty,” containing no columns). Using automorphisms of $Y$ and $B$, one can easily transform this matrix to the shape where there is at most two non-zero elements in every row (if two, one of them necessarily in the matrix $F_4$ and even) and at most one non-zero element in every column, as shown below: $$\begin{array}{|c|c|c|}
\hline
\ F_4\ &\ F_3\ &\ G_r\ \\
\hline
q & & \\
\hline
& \eta & \\
\hline
2^t &\eta &\\
\hline
&&\eta \\
\hline
2^t & &\eta \\
\hline
\end{array}$$ Thus $X$ decomposes into a bouquet of the spaces $\Si^2M(q)$ (which are not atoms, but suspended atoms), spheres and the spaces $C(\eta),\,C(\eta 2^t)$ and $C(2^r\eta 2^t)$, which are gluings of the following forms:
$$\xymatrix@=1ex{
{5} \ar@{.} [rrrrrrrrrrr] & \cel \ar@{-} [dd] &&& \cel \ar@{-}[ddl] &&&
\cel \ar@{-}[ddl] &&& \cel \ar@{-}[ddl] \ar@{-}[d] & {} \\
{4} \ar@{.} [rrrrrrrrrrr] &&&& \cel \ar@{-}[u] && \cel \ar@{-}[d]
&&& \cel\ar@{-}[d] &\cel&{} \\
{3} \ar@{.} [rrrrrrrrrrr] & \cel && \cel &&& \cel &&& \cel&& {} \\
& {C(\eta)} && {C(\eta 2^t)} &&& {C(2^r\eta)} &&& {C(2^r\eta 2^t)} }$$
Here, following Baues, we denote the cells by bullets and the attaching maps by lines; the word in brackets shows which maps are chosen to attach bigger cells to smaller ones. We do not show the fixed point, which coincide here with $X^2$ (since $X$ is $2$-connected); thus the lowest bullets actually describe spheres, not balls. These polyhedra are called *Chang atoms*. Again one can check that all of them are pairwise non-isomorphic.
Thus we have proved the following classical result.
\[bh-5\] The atoms of dimension at most $5$ are:
- sphere $S^1$ (of dimension $1$);
- Moore atoms $M(q)$, where $q=p^r$, $p$ is a prime number (of dimension $3$);
- Chang atoms $C(\eta),\,C(\eta 2^r),\,C(2^r\eta 2^t)$ (of dimension $5$).
In what follows, we often use suspended Moore and Chang atoms. We shall denote them by the same symbols but indicating the dimension. Thus $M^d(q)=\Si^{d-3}M(q)$ and $C^d(w)=\Si^{d-5}C(w)$ for $w\in\set{\eta,2^r\eta,
\eta2^r,2^r\eta2^t}$; in particular, $M(q)=M^3(q)$ and $C(w)=C^5(w)$. The same agreement will also be used for other atoms constructed below.
Dimension 7 {#d7}
===========
We shall now consider the category $\cw^3$. Its object actually come from $\cw^3_4$, so we have to classify atoms of dimension $7$. Such an atom $X$ is $3$-connected, so we may suppose that $X^3=*$. Set $B=X^5$, then $X/B$ only has cells of dimensions $6$ and $7$. Therefore $X\in \Si^3\cw^1\dagg\Si^2\cw^1\iso\Si\cw^1\dagg\cw^1$. Consider the bifunctor $\sW(A,B)=\hos(\Si A,B)$ restricted to the category $\cw^1$. Since, obviously, $\Hos(B,\Si^2A)=0$ for $A,B\in\cw^1$, we can apply Theorem \[bh-1\]. So we first classify indecomposable elements of the bimodule category $\El(\sW)$.
Indecomposable objects of the category $\cw^1$ are spheres $S^2,S^3$ and Moore atoms $M(q)\ (q=p^r,\ r$ prime). If $q$ is odd, one easily sees that $\sW(A,M(q))=0$ for all $A$, so we may only consider the spaces $M_r=M(2^r)$.
From the cofibration sequence $$S^2\str{g_r} S^2\to M_r\str{h_r} S^3\to S^3$$ and the diagram , we get the values of the $\hos$-groups shown in Table \[tab2\].
$$\begin{array}{|c|c|c|c|c|}
\hline
& S^2 & S^3 & M_1 &\ M_r\ (r>1)\ \\
\hline
S^2 & \mZ &\ \mZ/2\ & \ \mZ/2\ & \mZ/2\\
\hline
S^3 & 0 & \mZ & \mZ/2 &\ \mZ/2^r\ \\
\hline
M_1 &\ \mZ/2\ &\ \mZ/2\ & \mZ/4& T_{1r}\\
\hline
\ M_t\ (t>1)\ &\ \mZ/2^t\ &\mZ/2 & T_{1t} & T_{tr} \\
\hline
\end{array}$$
Here $T_{tr}$ denotes the set of matrices $\ds\binom{a\ b}{0\ a}$ with $a\in2^m\mZ/2^m,\,b\in\mZ/2$, where $m=\min(r,t)$. The equality $\hos(M_1,M_1)=\mZ/4$ follows from the fact that this ring acts on $\pi^S_4(M_1)=\mZ/4$, so $2\hos(M_1,M_1)\ne0$. The diagram implies that the sequence $$\begin{gathered}
0\larr \hos(S^3,M_t) \larr \hos(M_r,M_t) \larr\\
\larr \ker\{\hos(S^2,M_t)\str{2^r}\hos(S^2,M_t)\}\larr 0
\end{gathered}$$ splits if $\min(r,t)>1$. The generator of the subgroup of diagonal matrices in $T_{tr}$ is $k_{tr}$, while the matrix $\ds\binom{0\ 1}{0\ 0}$ corresponds to the morphism $g_t\eta h_r$.
Analogous calculations, using Table \[tab1\] of the preceding section and the diagram \[sd-e1\], produce the following Table \[tab3\] for the values of the functor $\hos(\Si A,B)$.
$$\begin{array}{|c|c|c|c|c|}
\hline
& S^2 & S^3 & M_1 &\ M_r\ (r>1)\ \\
\hline
S^2 & \mZ/2 & \mZ/2 & \mZ/4 &\ \mZ/2\+\mZ/2\ \\
\hline
S^3 & \mZ &\ \mZ/2\ & \mZ/2 & \mZ/2 \\
\hline
M_1 &\ \mZ/2\ &\ \mZ/4\ & \ \mZ/2\+\mZ/2\ & \ \mZ/2\+\mZ/4\ \\
\hline
\ M_t\ (t>1)\ & \mZ/2 &\ \mZ/2\+\mZ/2\ &\ \mZ/4\+\mZ/2\ &\ \mZ/2\+\mZ/2\+\mZ/2\ \\
\hline
\end{array}$$
It is convenient to organize this result in the form of Table \[tab4\] below, as in [@bd3].
$$\begin{array}{|c|ccccccc|ccccccc|}
\hline
&& ^1\* & ^2\* & ^3\* & \dots & ^\8\* &&& *^\8 & \dots & *^3 & *^2 & *^1 &\\
\hline
\ \*_1 &&\ \mZ/2\ &\ \mZ/2\ &\ \mZ/2\ &\dots &\ \mZ/2\ &&& 0 & \dots & 0 & 0 & 0 &\\
\ \*_2 &&\ \mZ/4\ & \mZ/2 & \mZ/2 & \dots & \mZ/2 &&&\ \mZ/2\ & \dots &\ \mZ/2\ &\ \mZ/2\ & 0 &\\
\ \*_3 &&\ \mZ/4\ & \mZ/2 & \mZ/2 & \dots & \mZ/2 &&&\ \mZ/2\ & \dots &\ \mZ/2\ &\ \mZ/2\ & 0 &\\
\ \vdots && \hdotsfor5 &&& \hdotsfor5 &\\
\ \*_\8 &&\ \mZ/4\ & \mZ/2 & \mZ/2 & \dots & \mZ/2 &&&\ \mZ/2\ & \dots &\ \mZ/2\ &\ \mZ/2\ & 0 &\\
\hline
\ _\8* && 0 & 0 & 0 & \dots & \ \mZ\ &&& \mZ/2 & \dots & \mZ/2 & \mZ/2 & \mZ/2 &\\
\vdots && \hdotsfor5 &&& \hdotsfor5 &\\
\ _3* && 0 & 0 & 0 & \dots & 0 &&& \mZ/2 & \dots & \mZ/2 & \mZ/2 & \mZ/2 &\\
\ _2* && 0 & 0 & 0 & \dots & 0 &&& \mZ/2 & \dots & \mZ/2 & \mZ/2 & \mZ/2 &\\
\ _1* && 0 & 0 & 0 & \dots & 0 &&& \mZ/4 & \dots & \mZ/4 & \mZ/4 & \mZ/2 &\\
\hline
\end{array}$$
In this table the row marked by $\*_t$ (respectively, $_t*$) shows the part of the group $\hos(\Si M_r,M_t)$ that comes from $\hos(\Si M_r,S^2)$ (respectively, from $\hos(\Si M_r,S^3)$). In the same way, the column marked by $^r\*$ (respectively, $*^r$) shows the part of this group that comes from $\hos(S^3,M_t)$ (respectively, from $\hos(S^4,M_t)$). The columns $^\8\*$ and $*^\8$ correspond, respectively, to $\hos(S^4,\_\,)$ and $\hos(S^3,\_\,)$; the rows $\*_\8$ and $_\8*$ correspond, respectively, to $\hos(\_ \,,S^2)$ and $\hos(\_ \,,S^3)$.
Therefore we consider the elements from $\El(\sW)$ as block matrices $(W^x_y)$, where $x\in\set{^r\*,*^r}$, $y\in\set{\*_t,^t*}$ and the block $W^x_y$ is with entries from the corresponding cell of Table \[tab4\]. Moreover, morphisms between Moore spaces induce the following transformations of vertical stripes $W^x$ and horizontal stripes $W_y$ of such a matrix, which we call *admissible transformation*:
- replacing the stripes $M^{^r\*}$ and $M^{*^r}$ by $M^{^r\*}X$ and $M^{*^r}X$;
- replacing the stripes $M_{\*_t}$ and $M_{_t*}$ by $XM_{\*_t}$ and $XM_{_t*}$;
- replacing $M^{*^r}$ by $M^{*^r}+M^{*^{r'}}X+M^{^s\*}Y$, where $r'>r$, $s$ arbitrary;
- replacing $M_{\*_t}$ by $M_{\*_t}+XM_{\*{t'}}+YM_{_s*}$, where $t'>t$, $s$ arbitrary;
- replacing $M^{^r\*}$ and $M^{*^r}$ by $M^{^r\*}+M^{^{r'}\*}X$ and $M^{*^r}+2^{r-r'}M^{*^{r'}}X$, where $r'<r$;
- replacing $M_{_t*}$ and $M_{\*_t}$ by $M_{_t*}+XM_{_{t'}*}$ and $M_{\*_t}+2^{t-t'}XM_{\*_{t'}}$, where $t'<t$;
- replacing $M^{^1\*}$ by $M^{^1\*}+2M^{^r\*}X+2M^{*^s}Y$; $r,s$ arbitrary;
- replacing $M_{_1*}$ by $M_{_1*}+2XM_{_t*}+2YM_{\*_s}$; $r,s$ arbitrary;
- replacing $M^{*^r}_{_1*}$ by $M^{*^r}_{_1*}+2M^{^s\*}_{\*_1}X$; $s$ arbitrary;
- replacing $M^{^1\*}_{\*_t}$ by $M^{^1\*}_{\*_t}+2XM^{*^1}_{_s*}$; $s$ arbitrary.
Here $X,Y$ denote arbitrary integer matrices of the appropriate size; in the transformations of types $(\mathsf a)$ and $(\mathsf a')$ the matrix $X$ must be invertible. Two matrices $W,W'$ are isomorphic in $\El(\sW)$ $W$ can be transformed to $W'$ using admissible transformations.
It is convenient first to reduce the block $W^{^\8\*}_{_\8*}$ to a diagonal form $D=\mathrm{diag}(\lst am)$ with $a_1|a_2|\dots|a_m$. Let $a_k=2^{d_k}b_k$ with odd $b_k$. Denote by $W^{^{\8k}\*}$ and $W_{_{\8k}*}$ the parts of the stripes $W^{^\8\*}$ and $W_{_\8*}$ corresponding to the columns and rows with $d_k=d$ ($k=\8$ if $d_k=0$). Since all other matrices of these stripes are with entries from $\mZ/2$, we can make the parts $W^{^{\80}\*}$ and $W_{_{\80}*}$ zero. Moreover, using admissible transformations that do not change the block $D$, we can replace $W^{^{\8k}\*}$ by $W^{^{\8k}\*}+W^{^{\8l}\*}X$ and $W_{_{\8k}*}$ by $W_{_{\8k}*}+YW_{_{\8l}*}$ for any $l<k$. In what follows we always suppose that $W$ is already in this form.
Call two matrices of this form $W,W'$ *2-equivalent*, if there is a matrix $W''\iso W$ such that $W''\equiv W\mod2$. One can easily see that the problem of 2-equivalence of matrices from $\El(\sW)$ is actually a sort of *bunch of chains* in the sense of [@bo; @d3]. We use the paper [@d3] as the source for the further discussion. Namely, we have the chain $\kE=\set{\*_t,_t*,{_{\8k}*}}$ for the rows and the chain $\kF=\set{^r\*,*^r,{^{\8k}\*}}$ for the columns, where $$\begin{gathered}
{_1*}<{_2*}<{_3*}<\dots<{_{\8\8}}<\dots<{_{\83}*}<{_{\82}*}<{_{\81}*}< \\
<\*_\8<\dots<\*_3<\*_2<\*_1,
\end{gathered}$$ $$\begin{gathered}
{^1\*}<{^2\*}<{^3\*}<\dots<{^{\8\8}\*}<\dots{^{\83}\*}<{^{\82}\*}<{^{\81}\*}<
\\ <*^\8<\dots<*^3<*^2<*^1.
\end{gathered}$$ The equivalence relation $\sim$ on $\kX=\kE\cup\kF$ is given by the rule $$\*_t\sim {_t*}\ (t\ne\8),\ ^r\*\sim *^r\ (r\ne\8),\ {^{\8k}\*}\sim{_{\8k}*}$$ for all possible values of $t,r$ and $k\ne\8$. Thus we can get a classification of our matrices up to 2-equivalence from [@d3]. Namely, we write $x-y$ if either $x\in\kE,\,y\in\kF$ or vice versa, at least one of them belongs to $\set{\*_t}\cup\set{*^r}$, moreover, $\set{x,y}\ne\set{\*_t,*^1}$ and $\set{x,y}\ne\set{\*_1,*^r}$. We call an *$\kX$-word* a sequence $w=x_1\rho_2x_2\rho_3\dots \rho_nx_n$, where $x_i\in\kX,\,\rho_i\in\set{\sim,-}$, $\rho_i\ne \rho_{i+1}\ (i=2,\dots,n-1)$ and $x_{i-1}\rho_ix_i$ holds in $\kX$ for all $i=2,\dots,n$. Such a word is called *full* if the following conditions hold:
- either $\rho_2=\sim$ or $x_1\not\sim y$ for all $y\in\kX,\ y\ne x_1$;
- either $\rho_n=\sim$ or $x_n\not\sim y$ for all $y\in\kX,\ y\ne x_n$.
$w$ is called a *cycle* if $\rho_2=\rho_n=-$ and $x_n\sim x_1$ in $\kX$. If, moreover, $w$ cannot be written in the form $v\sim v\sim \dots \sim v$ for a shorter word $v$, it is called *aperiodic*. We call a polynomial $f(t)\in\mZ/2[t]$ *primitive* if it is a power of an irreducible polynomial with the leading coefficient $1$. We shall identify any word $w$ with its inverse and any cycle $w$ with any of its cyclic shifts. Then the set of indecomposable representations of this bunch of chains is in 1-1 correspondence with the set $\kS\cup\kB$, where $\kS$ is the set of full words (up to inversion) and $\kB$ is the set of pairs $(w,f)$, where $w$ is an aperiodic cycle (up to a cyclic shift) and $f\ne t^d$ is a primitive polynomial. We call representations corresponding to $\kS$ *strings* and those corresponding to $\kB$ *bands*.
Note that an $\kX$-word can contain at most one element $^{\8k}\*$, at most one element $*_{\8k}$ and at most one subword of the form $\*_t-*^r$ or its inverse. Replacing $w$ by its inverse, we shall suppose that there are no words of the form $*^r-\*_t$ or $^{\8k}\*\sim_{\8k}*$. It is convenient to rewrite this answer in a modified form. Namely, we replace the subword $_{\8k}*\sim ^{\8k}\*$, if it occurs, by $_k\eps^k$, also omit $x_1$ if $\rho_2=\sim$, omit $x_n$ if $\rho_n=\sim$ and omit all remaining symbols $\sim$. Then we replace every subword $^r\*-\*_t$ by $^r\*_t$, $\*_t-^r\*$ by $_t\*^r$, $_t*-*^r$ by $_t*^r$, $*^r-_t*$ by $^r*_t$ and $\*_t-*^r$ by $_t\theta^r$. Note that in the last case $r\ne1$ and $t\ne1$. We also omit all signs $\sim$, replace any double superscript $^{rr}$ by $^r$ and any double subscript $_{tt}$ by $_t$. Certainly, the original word can be easily restored from such a shortened form. Now, any full word or its inverse can be written as a subword of one of the following words: $$\begin{aligned}
& ^{r_1}\*_{t_1}*^{r_2}\*_{t_2}*\dots ^{r_n}\*_{t_n} \quad \text{ (``usual word'')}, \\
& _{t_{-m}}\*\dots ^{r_{-2}}*_{t_{-2}}\*^{r_{-1}}*_{t_{-1}}\theta^{r_1}\*_{t_1}*^{r_2}\*_{t_2}\dots *^{r_n}
\quad\text{ (``theta-word'')}, \\
& _{t_{-m}}\*\dots ^{r_{-2}}*_{t_{-2}}\*^{r_{-1}}*_k\eps^k\*_{t_1}*^{r_2}\*_{t_2}\dots *^{r_n}
\quad\text{ (``epsilon-word'')},
\end{aligned}$$ Moreover,
- $\8$ can only occur at the ends of a word, not in a theta-word or epsilon-word.
- In any theta-word $t_{-1}\ne 1$ and $r_1\ne1$.
Any cycle or its shift can be written as $$^{r_1}\*_{t_1}*{r_2}\*_{t_2}*\dots ^{r_n}\*_{t_n}*^{r_1}.$$
The description of the representations in [@d3] also implies the following properties.
\[d71\]
1. Any row (column) of a string contains at most $1$ non-zero element.
2. There are at most $2$ zero rows or columns in a string, namely, they are in the following stripes:
1. $M_{_t*}$ if $w$ has an end $\*_t$ (or $_t\*$), $t\ne\8$;
2. $M^{*^r}$ if $w$ has an end $^r\*, r\ne\8$;
3. $M_{\*_t}$ if $w$ has an end $_t*$;
4. $M^{^r\*}$ if $w$ has an end $*^r$ (or $^r*$);
5. $M_{_{\8k}*}$ if the left end of $w$ is $_k\eps^k$;
6. $M^{^{\8k}\*}$ if the right end of $w$ is $_k\eps^k$.
We call each end occurring in this list a *distinguished end*.
3. The horizontal and vertical stripes of a band can be subdivided in such a way that every new horizontal or vertical band has exactly $1$ non-zero block, which is invertible.
Recall that elements modulo 4 only occur in the stripes $W^{^1\*}$ and $W_{_1\*}$.
\[d72\] Let $W\in\El(\sW)$ (with diagonal $W^{^\8\*}_{_\8*}$). Denote by $\ol W$ its reduction modulo 2 and by $\ti W$ the matrix obtained from $W$ by replacing all invertible entries with $0$ (thus all entries of $\ti W$ are even). Suppose that $\ol W=\b+_{i=1}^m\ol W_i$, where all $\ol W_i$ are strings or bands. Then $W\iso W'$, where $\ol W'=\ol W$ and the only non-zero rows and columns of $\ti W'$ can be those corresponding to the distinguished ends of types 2(a-d) of Proposition \[d71\]. In particular, if some of $\ol W_i$ is a band, a theta-string or an epsilon-string, $W'$ has a direct summand $W_i$ such that $\ol W_i\equiv W_i\mod2$ and $\ti W_i=0$.
Thus we only have now to consider the case, when $W=W'$ and every $\ol W_i$ is a usual string. Suppose that $W_i$ corresponds to a string $w_i$. It is easy to verify that if $w_i$ and $w_j$ have a common distinguished end, there is a sequence of distinguished transformations, which does not change $\ol W$ and adds the row (or column) corresponding to this end in $\ti W_i$ to the row (or column) corresponding to this end in $\ti W_j$ or vice versa. Hence, such rows (columns) are in some sense linearly ordered. As a consequence, we can transform $\ti W$ to a matrix having at most one non-zero element in every row and every column (without changing $\ol W$). It gives us the following description of indecomposable matrices from $\El(\sW)$ with $\ti W\ne0$.
\[d73\] Suppose that $W$ is an indecomposable matrix from $\El(\sW)$, such that $\ti W'\ne0$ for every matrix $W'\iso W$. Let $\ol W=\b+_{i=1}^m\ol W_i$, where each $\ol W_i$ is a usual string. There are, up to isomorphism, the following possibilities:
1. $m=1$, $\ol W$ corresponds to a word $w$ and $\ti W$ has a unique non-zero element in the block $W^a_b$ for the following choices: $$\begin{aligned}
& w={_{t_1}*}^{r_2}\*_{t_2}\dots,\ a={^1\*},\ b=\*_{t_1}\ (t_1=\ne 1); \tag{a}\\
& w={^{r_1}\*}_{t_1}*^{r_2}\*\dots, \ a=*^{r_1},\ b={_1*}\ (r_1\ne1); \tag{b},\\
& w={_1*}^{r_1}\*_{t_1}*^{r_2}\dots,\ a=*^r,\ b={_1*}; \tag{c},\\
& w={^1\*}_{t_1}*^{r_2}\*_{t_2}\dots,\ a={^1\*},\ b=\*_t; \tag{d}.
\end{aligned}$$
2. $m=2$, $\ol W_i\ (i=1,2)$ correspond to the words $w_i$ and $\ti W$ has a unique non-zero element in the block $W^a_b$, where $$\begin{aligned}
& w_1={_1*}^{r_{-1}}\*_{t_{-1}}*^{r_{-2}}\*\dots,\ w_2={^{r_1}\*}_{t_1}*^{r_2}\*_{t_2}\dots,\
a=*^{r_1},\ b={_1*}, \tag{e}\\
& w_1={^1\*}_{t_1}*^{r_1}\*_{t_2}*\dots,\ w_2={_{t_{-1}}}*^{r_{-1}}\*_{t_{-2}}*^{r_{-2}}\*\dots,\
a={^1\*},\ b=\*_{t_{-1}}. \tag{f}
\end{aligned}$$
We encode these matrices by the following words $w$: $$\begin{aligned}
& w=\dots \*^{r_2}*_{t_2}\*^{r_1}*_{t_1}\theta^1 \quad\text{in case (a)},\\
& w={_1\theta}^{r_1}\*_{t_1}*^{r_2}\*_{t_2}*\dots \quad\text{in case (b)},\\
& w=\dots *_{t_2}\*^{r_2}*_{t_1}\*^{r_1}*_1\theta^r \quad\text{in case (c)},\\
& w={_t\theta}^1\*_{t_1}*^{r_2}*_{t_2}\*^{r_3}*\dots \quad\text{in case (d)},\\
& w=\dots *^{r_{-2}}\*_{t_{-2}}*^{r_{-1}}*_1\theta^{r_1}\*_{t_1}*^{r_2}\*\dots \quad\text{in case (e)},\\
& w=\dots *{t_{-2}}\*^{r_{-1}}*_{t_{-1}}\theta^1\*_{t_1}*^{r_2}\*_{t_2}*\dots \quad\text{in case (f)},\\
\end{aligned}$$ We call these words “theta-words” as well.
Obviously, cases (a-d) always give indecomposable matrices. On the other hand, one can check that in case (e) $W$ is indecomposable $(r_{-1}+1,t_1,r_{-2},t_2,\dots)<(r_1,t_{-2},r_2,t_{-3},\dots)$ with respect to the lexicographical order [@bd3]. In case (f) $W$ is indecomposable $(t_1+1,r_{-1},t_2,r_{-2},\dots)<(t_{-1},r_2,t_{-2},r_3,\dots)$ lexicographically. Thus we obtain a complete list of non-isomorphic indecomposable matrices from $\El(\sW)$. Moreover, it is easy to verify that they remain pairwise non-isomorphic and indecomposable in $\El(\sW)\kJ$ as well. Thus, using Theorem \[bh-1\], we get the following result.
\[d74\] Indecomposable polyhedra from $\cw^3_4$ are in 1-1 correspondence with usual words, theta-words, epsilon-words and bands defined above, with the only restriction that in a theta-word $w=\dots ^{r_{-2}}*_{t_{-2}}\*^{r_{-1}}*_{t_{-1}}\theta^{r_1}\*_{t_1}*^{r_2}\*_{t_2}\dots$ the following conditions hold: *$$\begin{aligned}
\text{ if $\,t_{-1}=1$, then }\ &(r_{-1}+1,t_1,r_{-2},t_2,\dots)<(r_1,t_{-2},r_2,t_{-3},\dots), \\
\text{ if $\,r_1=1$, then }\ &(t_1+1,r_{-1},t_2,r_{-2},\dots)<(t_{-1},r_2,t_{-2},r_3,\dots)
\end{aligned}$$ (lexicographically in both cases).*
The gluings of spheres corresponding to these words can be described as follows: $$\xymatrix@R=.1ex@C=2ex@!R{
{7} \ar@{.} [rrrrrrrrrrrrrr] &&&& \cel \ar@{-}[dd] &&&& \cel \ar@{-}[dd]
&&&& \cel \ar@{-}[dd] && {} \\
{}\\
{6} \ar@{.} [rrrrrrrrrrrrrr] &&&& \cel \ar@{-}[ddddrr] &&&& \cel \ar@{-}[ddddrr]
&&&& \cel \ar@{-}[ddr] && {} \\
& {\cdots\quad } \\
{5} \ar@{.} [rrrrrrrrrrrrrr] & *=0{}& \cel \ar@{-}[uuuurr] \ar@{-}[dd] &&&&
\cel \ar@{-}[uuuurr] \ar@{-}[dd] &&&& \cel \ar@{-}[uuuurr] \ar@{-}[dd] &&& *=0{} & {} \\
&&&&&&&&&&&&& {\quad \cdots} \\
{4} \ar@{.} [rrrrrrrrrrrrrr] && \cel@\ar@{-}[uul] &&&& \cel &&&& \cel &&&& {}
}$$
**for a usual word**
$$\xymatrix@R=.1ex@C=2ex@!R{
{7} \ar@{.} [rrrrrrrrrrrrrrrr] && \cel \ar@{-}[dd] \ar@{-}[ddddrr] &&&&
\cel \ar@{-}[dd] \ar@{-}[ddddrr] &&&& \cel \ar@{-}[dd] &&&& \cel \ar@{-}[dd] && {} \\
{} \\
{6} \ar@{.} [rrrrrrrrrrrrrrrr] && \cel \ar@{-}[ddl] &&&& \cel \ar@{-}[ddddll]
&&&& \cel \ar@{-}[ddddrr] &&&& \cel \ar@{-}[ddr] && {} \\
{} \\
{5} \ar@{.} [rrrrrrrrrrrrrrrr] & *=0{}&&& \cel \ar@{-}[dd] &&&&
\cel \ar@{-}[dd] &&&& \cel \ar@{-}[uuuurr] \ar@{-}[dd] &&& *=0{} & {} \\
& {\cdots\quad} &&&&&&&&&&&&&& {\quad \cdots} \\
{4} \ar@{.} [rrrrrrrrrrrrrrrr] &&&& \cel &&&& \cel \ar@{-}[uuuuuurr] &&&& \cel &&&& {}
}$$
**for a theta-word**
$$\xymatrix@R=.1ex@C=2ex@!R{
{7} \ar@{.} [rrrrrrrrrrrrrrrr] &&&& \cel \ar@{-}[dd] \ar@{-}[ddddrr] &&&&&&
\cel \ar@{-}[dd] \ar@{-}[ddddll] &&&& \cel \ar@{-}[dd] \ar@{-}[ddddll] & & {} \\
&{\cdots\quad} \\
{6} \ar@{.} [rrrrrrrrrrrrrrrr] &&&& \cel \ar@{-}[ddddll] &&&&
\cel \ar@{-}[ddddll] \ar@{-}[dd] && \cel \ar@{-}[ddddrr] &&&& \cel \ar@{-}[ddr] && {} \\
{} \\
{5} \ar@{.} [rrrrrrrrrrrrrrrr] && \cel \ar@{-}[uul] \ar@{-}[dd] &&&& \cel \ar@{-}[dd]
&& \cel &&&& \cel \ar@{-}[dd] &&& *=0{}& {} \\
& &&&&&&&&&&&&&& {\quad \cdots} \\
{4} \ar@{.} [rrrrrrrrrrrrrrrr] && \cel &&&& \cel &&&&&& \cel &&&& {}
}$$
**for an epsilon-word**
In these diagrams vertical segments present the suspended atoms $M_r$, slanted lines correspond to the gluings arising from Hopf maps $S^{d+1}\to S^d$, while the long slanted line in a theta-word shows the gluing arising from the doubled Hopf map $S^6\to S^4$.
Note that all atoms from $\cw^3_4$ are $p$-primary ($2$-primary, except $M(q)$ with odd $q$). Therefore, we have the uniqueness of decomposition of spaces from $\cw^3$ into bouquets of suspended atoms.
Bigger dimensions. Wildness {#bdw}
===========================
Unfortunately, if we pass to bigger dimensions, the calculations as above become extremely complicated. In the representations theory the arising problems are usually called “*wild*.” Non-formally it means that the classification problem for a given category contains the classification of representations of arbitrary (finitely generated) algebras over a field. It is well-known, since at least 1969 [@gp], that it is enough to show that this problem contains the classification of pairs of linear mappings (up to simultaneous conjugacy), or, equivalently, the classification of triples of linear mappings $$\label{bde1}
\xymatrix{{V_1\,} \ar[r] \ar@<1ex>[r] \ar@<-1ex>[r] & {\ V_2}}$$ On the other hand, the problems like one considered in the preceding section, where indecomposable objects can be parameterised by several “discrete,” or combinatorial parameters (as $\kX$-words above) and at most one “continuous” parameter (as a primitive polynomial in the description of bands), are called “*tame*.” The problems, where the answer is purely combinatorial, like the classification of atoms of dimensions $d\le 5$, are called “finite.” I shall not precise these notions formally. The reader can consult, for instance, the survey [@d2], where it is done within the framework of representation theory. An important question in the representation theory is to distinguish finite, tame and wild cases. The following result accomplishes such an investigation for stable homotopy types.
\[bdw1\] The classification problem for the category $\cw^4$ is wild.
Let $\bB$ be the category of bouquets of Moore atoms $M=M_1$, $\bA=\Si^2\bB$. Then $\cw^4$ contains the subcategory $\Si^3(\bA\dagg\bB)\iso\bA\dagg\bB$. Corollary \[bh-2\] shows that the category $\bA\dagg\bB$ is representation equivalent to $\El(\sH)$, where $\sH$ is the restriction of $\hos$ onto $\bA^\circ\xx\bB$. We know that $\hos(M,M)=\mZ/4$. Therefore, we only have to show that $\hos(\Si^2M,M)\iso\mZ/2\mZ/2\xx\mZ/2$. Indeed, it implies the category $\El(\sH)$ is representation equivalent to the category of diagrams of the shape .
The cofibration sequence $S^2\str2S^2\to M\to S^3\str2 S^3$ and the Hopf map $\eta:S^5\to S^4$ produce the following commutative diagram: $$\begin{CD}
0 @>>> \mZ/2 @>>> \pi^S_4(M) @>>> \mZ/2 @>>> 0 \\
&& @V\eta^*VV @VVV @VV\wr V \\
0 @>>> \mZ/2 @>>> \pi^S_5(M) @>>> \mZ/2 @>>> 0 ,
\end{CD}$$ Since $\eta^3=4\nu$, where $\nu$ is the element of order $8$ in $\pi^S_5(S^2)=\mZ/24$ [@tod], actually $\eta^*=0$, so the lower row splits and $\pi^S_5(M)=\mZ/2\+\mZ/2$. Just in the same way we show that $\hos(\Si^2M,S^2)=\mZ/2\+\mZ/2$. Now, applying the functors $\hos(\_\,,S^2)$ and $\hos(\_\,,M)$ to the same cofibration sequence, we get the commutative diagram $$\begin{CD}
0 @>>> \mZ/2 @>>> \hos(\Si^2M,S^2) @>>> \mZ/2 @>>> 0 \\
&& @VVV @VVV @VV\wr V \\
0 @>>> \mZ/2\+\mZ/2 @>>> \hos(\Si^2M,M) @>>> \mZ/2 @>>> 0 .
\end{CD}$$ Since the upper row of this diagram splits, the lower one splits as well, hence $\hos(\Si^2M,M)=\mZ/2\+\mZ/2\+\mZ/2$. It accomplishes the proof.
We can summarize the obtained results in the following theorem.
The category $\cw^k$ is of finite type for $k\le 2$, tame for $k=3$ and wild for $k\ge4$.
Torsion free atoms. Dimension 9. {#tf}
================================
Nevertheless, if we consider *torsion free* atoms, the situation becomes much simpler. Namely, in this case neither sphere of dimension $d$ can be attached to the spheres of dimension $d-1$, thus in the picture describing the gluing of spheres there is no fragments of the sort $$\xymatrix@=1ex{ d &{} \ar@{.}[rr] & \cel\ar@{-}[d] &{}\\ {d-1} &{}\ar@{.}[rr] &\cel &{} }$$ Therefore, a calculation of atoms from $\cwf^k_{k+1}$ can be organized as follows. Denote by $\bB_k$ the full subcategory of $\cw$ consisting of bouquets of torsion free suspended atoms of dimension $2k$ and by $\bS_k$ the category of bouquets of spheres $S^{2k}$. Let $\Ga_m(X)$ denote the subgroup $\im\{\pi_m^S(X^{m-1})\to\pi_m(X)\}$ of $\pi_m^S(X)$. When $X$ runs through $\bB_k$, $\Ga_{2k}$ can be considered as an $\bS_k$-$\bB_k$-bimodule; we denote this bimodule by $\sG_k$. Then the following analogue of Theorem \[bh-1\] holds (with essentially the same proof).
\[tf-1\] Denote by $\kI$ the ideal of the category $\cwf^k_{k+1}$ consisting of all morphisms $X\to X'$ that factor through an object from $\bB_k$, and by $\kJ$ the ideal of the category $\El(\sG_k)$ consisting of such morphisms $(\al,\be):f\to f'$ that $\al$ factors through $f$ and $\be$ factors through $f'$. Then $\cwf^k_{k+1}/kI\iso \El(\Ga_k)/kJ$. Moreover, both $\kI^2=0$ and $\kJ^2=0$, hence the categories $\cwf^k_{k+1}$ and $\El(\sG_k)$ are representation equivalent.
The only new claim here is that $\kJ^2=0$. But this equality immediately follows from the fact that if a morphism $X\to S^m$ factors through $X^{m-1}$, it is zero.
Thus a torsion free atom of dimension 7 can be obtained as a cone of a map $f:mS^6\to Y$, where $Y$ is a bouquet of spheres $S^4,S^5$ and suspended Chang atoms $C^6(\eta)$, while $f\in\Ga_6(Y)$. Easy calculations, like above, give the following values of $\Ga_6$:
$$\begin{array}{|c|c|c|c|}
\hline
X & S^4 & S^5 & \ C^6(\eta)\ \\
\hline
\ \Ga_6\ &\ \mZ/2\ &\ \mZ/2\ & 0 \\
\hline
\end{array}$$
(The last $0$ is due to the fact that the map $\eta_*:\pi_6(S^5)\to\pi_6(S^4)$ is an epimorphism [@tod]). The Hopf map $\eta:S^5\to S^4$ induces an isomorphism $\Ga_6(S^5)\to\Ga_6(S_4)$. Therefore, the only indecomposable torsion free atom of dimension 7 is the gluing $C(\eta^2)=C^7(\eta^2)=S^4\cup_{\eta^2}B^7$. (Note that such an atom must contain at least one $4$-dimensional cell.) Moreover, all torsion free atoms of dimensions $d\le7$ are $2$-primary.
A torsion free atom of dimension 9 is a cone of some map $f:mS^8\to Y$, where $Y$ is a bouquet of spheres $S^i\ (5\le i\le 7)$, suspended Chang atoms $C^7(\eta),C^8(\eta)$ and suspended atoms $C^8(\eta^2)$. One can calculate the following table of the groups $\Ga_8$ for these spaces:
$$\begin{array}{|c|c|c|c|c|c|c|}
\hline
X & S^5 & S^6 & S^7 &\ C^7(\eta)\ &\ C^8(\eta)\ &\ C^8(\eta^2) \ \\
\hline
\ \Ga_8\ &\ \mZ/24\ &\ \mZ/2\ &\ \mZ/2\ & \mZ/12 & 0 & \mZ/12 \\
\hline
\end{array}$$
Morphisms between these spaces induce epimorphims $\Ga_8(S^5)\to\Ga_8(C^7(\eta^2))\to\Ga_8(C^7(\eta))$, $\Ga_8(S^7)\to\Ga_8(S_6)$ and monomorphisms $\Ga_8(S^7)\to\Ga_8(C^7(\eta)\to\Ga_8(S^5)$, $\Ga_8(S^6)\to\Ga_8(S^5)$. It can be deduced either from [@tod] or, perhaps easier, from the results of [@un], cf. [@bd1]. (The only non-trivial one is the monomorphism $\Ga_8(S^7)\to\Ga_8(\Si^2A(\eta))\,$.) Again we consider the map $f$ as a block matrix $$F= \begin{pmatrix}
F_1 & F_2 &\ F_3 &\ F_4 &\ F_6
\end{pmatrix}^\top .$$ Here $F_i$ is of size $m_i\xx m$ with entries from $\Ga_8(Y_i)$, where $$Y_i= \begin{cases}
\ S^5 &\text{if } i=1,\\ \ C^7(\eta) &\text{if } i=2, \\ \ C^8(\eta^2)&\text{if } i=3,\\
\ S^6 &\text{if } i=4,\\ \ S^5 &\text{if } i=6.
\end{cases}$$ We have written $F_6$, not $F_5$, in order to match the notations of the Example \[bc-1a\]; so we set $I_1=\set{1,2,3,4,6}$. Using the automorphisms of $mS^7$ and of $Y$, one can replace the matrix $F$ by $PFQ$, where $P\in\GL(m,\mZ)$ and $Q=(Q_{ij})_{i,j\in I_1}$ is an invertible integer block matrix, where the block $Q_{ij}$ is of size $m_i\xx m_j$ with the following restrictions for the entries $a\in Q_{ij}$: $$\begin{aligned}
a=&0 \quad\text{ for } i\in\set{4,6},\ j<i, \\
a\equiv& 0 \mod2\, \text{ for } (ij)\in\set{(12),(13),(23)},\\
a\equiv& 0 \mod6\, \text{ for } (ij)=(26),\\
a\equiv& 0 \mod{12}\, \text{ for } j\in{4,6},\ i\in\set{1,2,3},\ (ij)\ne(26).
\end{aligned}$$ Thus we have come to the *bimodule category* $\El(\sU_1)$ considered in Example \[bc-1b\], so we can use Corollary \[bc-3\], which describes all indecomposable objects of this category. Certainly, we are not interested in the “empty” objects $\0_i$, since they correspond to the spaces with no 9-dimensional cells. Note also that the matrices $(1_4),(1_6)$ correspond not to atoms, but to suspended atoms $C^9(\eta^2)$ and $C^9(\eta)$. We use the following notation for the atoms corresponding to other indecomposable matrices $F$: $$\def\for{&& \text{for} &&}
\begin{array}{ccccc}
A(v) \for (v_1),\\
A(\eta v) \for (v_2),\\ \vspace*{.5ex}
A(\eta^2 v) \for (v_3),\\ \vspace*{1ex}
A(v\eta) \for \ds\binom{v_1}{1_6},\\ \vspace*{1ex}
A(v\eta^2) \for \ds\binom{v_1}{1_4},\\ \vspace*{1ex}
A(\eta v\eta) \for \ds\binom{v_2}{1_6},\\ \vspace*{1ex}
A(\eta v\eta^2) \for \ds\binom{v_2}{1_4},\\ \vspace*{1ex}
A(\eta^2 v\eta) \for \ds\binom{v_3}{1_6},\\
A(\eta^2 v\eta^2) \for \ds\binom{v_3}{1_4}.
\end{array}$$
So we have proved
\[tf-2\] Every torsion free atom of dimension $9$ is isomorphic to one of the atoms $A(w)$ with $w\in\set{v,\eta v,\eta^2 v,v\eta,v\eta^2,\eta v\eta,\eta v\eta^2,\eta^2 v\eta,\eta^2 v\eta^2}$.
Using the gluing diagrams, these atoms can be described as in Table \[tab9\] below.
$$\xymatrix@R=.7ex@C=2ex{
9 \ar@{.}[rrrrrrrrrr] & \cel \car[dddd] && \cel \car[ddddl] && \cel \car[ddddl] &&
\cel \car[dd] \car[ddddl] && \cel \car[ddd] \car[ddddl] & {} \\
8 \ar@{.}[rrrrrrrrrr] & &&& \cel \car[ddd] &&&&&& {} \\
7 \ar@{.}[rrrrrrrrrr] & & \cel \car[dd] &&&&& \cel &&& {} \\
6 \ar@{.}[rrrrrrrrrr] & &&&&&&&& \cel & {} \\
5 \ar@{.}[rrrrrrrrrr] & \cel & \cel && \cel && \cel && \cel && {} \\
& A(v) & A(\eta v) && A(\eta^2 v) && A(v\eta) && A(v\eta^2) }$$
$$\xymatrix@R=.7ex@C=2ex{
9 \ar@{.}[rrrrrrrrrrrr] && \cel \car[ddddl]\car[dd] &&& \cel \car[ddddl]\car[ddd] &&& \cel \car[ddddl]\car[dd]
&&&\cel \car[ddd] \car[ddddl] & {} \\
8 \ar@{.}[rrrrrrrrrrrr] & \cel \car[ddd] &&& \cel \car[ddd] &&&&&&&& {} \\
7 \ar@{.}[rrrrrrrrrrrr] & & \cel &&&&& \cel \car[dd] &\cel && \cel\car[dd] &&{} \\
6 \ar@{.}[rrrrrrrrrrrr] & &&&& \cel &&&&&& \cel & {} \\
5 \ar@{.}[rrrrrrrrrrrr] &\cel &&& \cel &&& \cel &&& \cel && {} \\
& A(\eta^2 v\eta) &&& A(\eta^2 v\eta^2) &&& A(\eta v\eta) &&& A(\eta v\eta^2) }$$
One can also check that the $2$-primary atoms in this list are those with $v$ divisible by $3$, while the only $3$-primary one is $A(8)$. Thus there are altogether $29$ primary suspended atoms of dimension at most 9. The congruent ones are only $A(3)$ and $A(9)$. Indeed, $A(3)\vee S^5$ corresponds to the matrix $\ds \binom 30 \mod{24}$. But the latter can be easily transformed to $\ds\binom 90 \mod{24}$, which corresponds to $A(9)\vee S^5$: $$\binom 30 \to \binom 3{12} \to \binom {-9}{12} \to \binom 9{12} \to \binom 90.$$ (At the last step we add the first row multiplied by $4$ to the second one; all other transformations are obvious.) One can verify that all other $2$-primary atoms are pairwise non-congruent.
\[tf-3\] The Grothendieck group $K_0(\cwf^4)$ is a free abelian group of rank 29.
Note that the matrix presentations allows easily to find the images in $K_0(\cwf^4)$ of all atoms. For instance, the equivalence of matrices $$\begin{pmatrix} 8_1 &0 \\ 0 & 3_1 \end{pmatrix} \sim
\begin{pmatrix} 1_1 &0 \\ 0 & 0_1 \end{pmatrix}$$ implies that $A(8)\vee A(3)\iso A(1)\vee S^5\vee S^9$, thus in $K_0(\cwf^4)$ we have $$[A(1)]=[A(8)]+[A(3)]-[S^5]-[S^9].$$ The reader can easily make analogous calculations for all atoms of Table \[tab9\].
Torsion free atoms. Dimension 11. {#tf1}
=================================
For torsion free atoms of dimension 11 analogous calculations have been done in [@bd4]. Nevertheless, they are a bit cumbersome, so we propose here another, though rather similar, approach. Namely, denote by $\bS'_k$ the category of bouquets of spheres $S^{2k-1}$ and $S^{2k}$, by $\bB'_k$ the category of bouquets of suspended atoms of dimension $2k-1$ and by $\sG'_k$ the $\bS'_k$-$\bB'_k$-bimodule such that $$\sG'_k(S^{2k-1},B)=\Ga_{2k-1}(B)\quad\text{and}\quad
\sG'_k(S^{2k},B)=\Ga_{2k}(B).$$
\[tf1-1\] Denote by $\kI'$ the ideal of the category $\cwf^k_{k+1}$ consisting of all morphisms $X\to X'$ that factors through an object from $\bB'_k$, and by $\kJ'$ the ideal of the category $\El(\sG'_k)$ consisting of such morphisms $(\al,\be):f\to f'$ that $\al$ factors through $f$ and $\be$ factors through $f'$. Then $\cwf^k_{k+1}/\kI'\iso \El(\Ga'_k)/\kJ'$. Moreover, both $(\kI')^2=0$ and $(\kJ')^2=0$, hence the categories $\cwf^k_{k+1}$ and $\El(\Ga'_k)$ are representation equivalent.
Thus we obtain torsion free atoms if dimension $11$ as cones of maps $S\to Y$, where $S$ is a bouquet of spheres of dimensions $9$ and $10$, while $Y$ is a bouquet of $5$-connected suspended atoms of dimensions $6\le d\le 9$. Note that at least one of these atoms must have a cell of dimension $6$ in order that such a cone be an atom.
Just as above, we have the following values of $\Ga_9$ and $\Ga_{10}$ for such atoms:
$$\begin{array}{|c|c|c|c|c|c|c|c|}
\hline
X & S^6 &\ C^8(\eta)\ &\ C^9(\eta^2) \ & S^7 &\ C^9(\eta)\ & S^8 & S^9 \\
\hline
\ \Ga_9\ &\ \mZ/24\ &\ \mZ/12\ &\ \mZ/12\ &\ \mZ/2\ & 0&\ \mZ/2\ & 0 \\
\hline
\ \Ga_{10} & 0 &0&0 & \mZ/24 &\ \mZ/12\ &\ \mZ/2\ & \ \mZ/2\ \\
\hline
\end{array}$$
(We have arranged this table taking into account the known maps between these groups, as above.) The Hopf map $S^{10}\to S^9$ induces monomorphisms in the 4th and the 6th columns of this table, while the maps between suspended atoms induce homomorphisms analogous to those of the preceding section. Thus a morphism $f:S\to Y$ can be described by a matrix $$F= \begin{pmatrix}
F_1 & F_2 & F_3 & F_4 & 0 & F_6 & 0 \\ 0 & 0 & 0 & G_4 & G_5&G_6& G_7
\end{pmatrix}^\top ,$$ where the matrix $F_i$ ($G_i)$ has entries from the first row (respectively, second row) and the $i$-th column of the table above. Two matrices, $F$ and $F'$, define homotopic polyhedra if $F'=PFQ$, where $P,Q$ are matrices over the tiled orders, respectively, $$\begin{pmatrix}
\ 1 & 2 & 2 & 12 & 24 & 12 & 24 \,\\ \ 1 & 1 & 1 & 12 & 24 & 6 & 24 \,\\ \ 1 & 2 & 1 & 12 & 24 & 12 & 24\, \\
\ 0&0&0& 1 & 2 & 12^* & 12\, \\ \ 0&0&0& 1 & 1 & 12 & 6\, \\\ 0&0&0&0&0& 1&1\\ \ 0&0&0&0&0&0&1 \,
\end{pmatrix}
\quad\text{and}\quad
\begin{pmatrix} \ 1 & 12^*\, \\ \ 0&1\, \end{pmatrix}. \\$$ Here $12^*$ shows that the corresponding element obeys the $\ast$-rule , i.e. induces a *non-zero* map $\mZ/2\to\mZ/2$ and acts as usual multiplication by $12$ in all other cases.
Thus we have obtained the *bimodule category* $\El(\sU_2)$ from Example \[bc-1a\], so can use the list of indecomposable objects from Theorem \[bc-2\]. Moreover, we only have to consider the matrices having non-empty $G$-column and one of the parts $F_1,F_2,F_3$ (otherwise we have no $11$-dimensional or no $6$-dimensional cells). Therefore, a complete list of atoms arises from the following matrices: $$\begin{pmatrix} v_i & 0 \\ 1_4 & w_4 \end{pmatrix} ,\quad
\begin{pmatrix} v_i & 0 \\ 1_4 & w_4 \\ 0 & 1_6 \end{pmatrix} ,\quad
\begin{pmatrix} v_i & 0 \\ 1_4 & w_4 \\ 0 & 1_7 \end{pmatrix} ,$$ where $i\in\set{1,2,3}$, $v,w\in\set{1,2,3,4,5,6}$. We omit the upper indices of Theorem \[bc-2\], since here they coincide with the column number; the lower indices show to which horizontal stripe of the matrix $F$ the corresponding elements belong. It gives the following list of $11$-dimensional torsion free atoms.
\[tf1-2\] Every torsion free atom of dimension 11 is isomorphic to one of the atoms of Table \[tab11\] below.
$$\xymatrix@R=.7ex@C=2.2ex{
11 \ar@{.}[rrrrrrrrrrrrrrrrrrrr] &&&\cel\ar@{-}[ddddl] &&&&\cel\ar@{-}[ddddl]\ar@{-}[dd] &&&&
\cel\ar@{-}[ddd] \ar@{-}[ddddl]&&&& \cel\ar@{-}[ddddl]\ar@{-}[dd] &&&& \cel\ar@{-}[ddddl] &{} \\
10 \ar@{.}[rrrrrrrrrrrrrrrrrrrr] &&\cel\ar@{-}[ddddl]\ar@{-}[ddd] &&&&\cel\ar@{-}[ddddl]\ar@{-}[ddd]
&&&&\cel\ar@{-}[ddddl] \ar@{-}[ddd] &&&& \cel\ar@{-}[ddddl]\ar@{-}[ddd] &&&& \cel\ar@{-}[ddd] \ar@{-}[ddddl]&&{} \\
9 \ar@{.}[rrrrrrrrrrrrrrrrrrrr] &&&&&&&\cel &&\cel\ar@{-}[ddd] &&&&&&\cel &&&& &{} \\
8 \ar@{.}[rrrrrrrrrrrrrrrrrrrr] &&&&&\cel\ar@{-}[dd] &&&&&&\cel &&&&&& \cel\ar@{-}[dd] &&& {} \\
7 \ar@{.}[rrrrrrrrrrrrrrrrrrrr] &&\cel &&&&\cel &&&&\cel &&&& \cel &&&& \cel &&{} \\
6 \ar@{.}[rrrrrrrrrrrrrrrrrrrr] & \cel &&&&\cel &&&&\cel &&&& \cel &&&& \cel &&&{} \\
&& *=0{A(v\eta^2w))} &&&& *=0{A(\eta v\eta^2w\eta)} &&&& *=0{A(\eta^2v\eta^2w\eta^2)}
&&&& *=0{A(v\eta^2w\eta)} &&&& *=0{A(\eta v\eta^2w)} &&
}$$
$$\xymatrix@R=.7ex@C=2.2ex{
11 \ar@{.}[rrrrrrrrrrrrrrrr] &&&\cel\ar@{-}[ddddl]\ar@{-}[ddd] &&&&\cel\ar@{-}[ddddl] &&&&
\cel\ar@{-}[dd] \ar@{-}[ddddl]&&&& \cel\ar@{-}[ddd] \ar@{-}[ddddl]&{} \\
10 \ar@{.}[rrrrrrrrrrrrrrrr] &&\cel\ar@{-}[ddddl]\ar@{-}[ddd] &&&&\cel\ar@{-}[ddddl]\ar@{-}[ddd]
&&&&\cel\ar@{-}[ddddl] \ar@{-}[ddd] &&&& \cel\ar@{-}[ddddl]\ar@{-}[ddd] &&{} \\
9 \ar@{.}[rrrrrrrrrrrrrrrr] &&&&&\cel\ar@{-}[ddd] &&&&\cel\ar@{-}[ddd] &&\cel &&&& &{} \\
8 \ar@{.}[rrrrrrrrrrrrrrrr] &&&\cel &&&&&&&&&&\cel\ar@{-}[dd] && \cel & {} \\
7 \ar@{.}[rrrrrrrrrrrrrrrr] &&\cel &&&&\cel &&&&\cel &&&& \cel &&{} \\
6 \ar@{.}[rrrrrrrrrrrrrrrr] & \cel &&&&\cel &&&&\cel &&&& \cel &&&{} \\
&& *=0{A(v\eta^2w\eta^2)} &&&& *=0{A(\eta^2 v\eta^2w)} &&&& *=0{A(\eta^2v\eta^2w\eta)}
&&&& *=0{A(\eta v\eta^2w\eta^2)}&&
}$$
Again $2$-primary atoms are those with $v,w\in\set{3,6}$ and there are no $3$-primary spaces in this table. Moreover, the new $2$-primary atoms are pairwise non-congruent, wherefrom we obtain the following result.
\[tf1-3\] The Grothendieck group $K_0(\cwf^5)$ is a free abelian group of rank 85.
We end up with the following statements about the higher dimensional torsion free spaces.
\[tf1-4\]
1. There are infinitely many non-isomorphic (even non-congruent) $2$-primary atoms of dimension $13$. Hence the Grothendieck group $K_0(\cwf^k)$ is of infinite rank for $k\ge6$.
2. If $k\ge11$, the classification problem for the category $\cwf^k$ is wild.
We shall show first that $\pi_{12}^S(A^{11}(\eta^2v)$, or, the same, $\pi_{10}^S(A(\eta^2v)$ equals $\mZ/2\+\mZ/2$. We consider the cofibration sequences $$\begin{aligned}
\tag{a}& S^8 \str f \Si C \str g A \str h S^9 \str{\Si f} \Si^2C,\\
\tag{b}& S^6 \larr S^4 \larr C \larr S^7 \larr S^5,
\end{aligned}$$ where $A=A(\eta^2v),\,C=C(\eta^2)$. Note that the map $f$ factors through $S^5$. From the sequence (b) we get $\pi_9^S(C)\iso\pi_9^S(S^7)\iso\mZ/2$ and $\pi^S_9(\Si C)\iso\pi^S_8(C)=0$. The second equality follows from the fact that the induced map $\pi^S_8(S^7)\to\pi^S_8(S^5)$ is known to be injective [@tod]. Since $\pi^S_{10}(S^5)=\pi^S_{10}(S^6)=0$, the sequence (a) gives then an exact sequence $$0 \larr \pi_{10}^S(\Si C)\iso\mZ/2 \larr \pi_{10}^S(A) \larr \pi^S_{10}(S^9)\iso\mZ/2 \larr 0.$$ To show that this sequence splits, we have to check that $2\al=0$ for every $\al\in\pi^S_{10}(A)$. In any case, $2\al$ factors through $\Si C$, which gives rise to a commutative diagram $$\begin{CD}
M^{10}(2) @>\phi>> S^{10} @>2>> S^{10} \\
@V\ga VV @VV\be V @VV\al V \\
S^8 @>>f> \Si C @>>g> A
\end{CD}$$ for some $\be,\ga$ (we have used the cofibration sequence for $M(2)$). Since $\pi^S_9(S^5)=\pi^S_{10}(S^5)=0$, also $\hos(M^{10}(q),S^5)=0$. But the map $\be \phi=f\ga$ factors through $S^5$, so $\be\phi=0$ and $\be=2\si$ for some $\si\in\pi^S_{10}(\Si C)\iso\mZ/2$. Hence $\be=0$ and $2\al=0$.
Analogous calculations show that any endomorphism of $A$ acts as a homothety on $\pi_{10}^S(A)$. Since, obviously, $\hos(A,S^{10})=0$, Corollary \[bh-2\] shows that the category of spaces arising as cones of mappings $mS^{12}\to nA^11(\eta^2v)$ is equivalent to the category of representations of the Kronecker quiver $\ti A_1$, or, the same, of diagrams of $\mZ/2$-vector spaces of the shape $V_1\rightrightarrows V_2$. But it is well-known that this quiver is of infinite type, i.e. has infinitely many non-isomorphic indecomposable representations. Obviously, all corresponding spaces are $2$-primary and non-congruent, which proves the claim (1).
The claim (2) follows from the equality $\pi^S_{20}(S^{11})\iso(\mZ/2)^3$. It implies that the category of spaces, which are cones of mappings $mS^{20}\to mS^{11}$ is equivalent to that of diagrams $\ds\xymatrix{{V_1\,} \ar[r] \ar@<1ex>[r] \ar@<-1ex>[r] & {\ V_2}}$. The latter is well-known to be wild.
Perhaps, the estimate $11$ in the claim (2) of Proposition \[tf1-4\] is too big, but at the moment I do not know a better one. On the other hand, we can hope that the classification problem for $\cws^6$ is still tame.
[99]{} Baues, H.-J. *Homotopy Type and Homology*. Oxford University Press, 1996. Baues, H.-J. Atoms of Topology. *Jahresber. Dtsch. Math.-Ver.* 104, No.4 (2002) 147-164 . Baues, H.-J. and Drozd, Yu. A. The homotopy classification of $(n-1)$-connected $(n+4)$-dimensional polyhedra with torsion free homology. *Expo. Math.* 17 (1999), 161–179. Baues, H.-J. and Drozd, Yu. A. Representation theory of homotopy types with at most two non-trivial homotopy groups. *Math. Proc. Cambridge Phil. Soc.* 128 (2000), 283–300. Baues, H.-J. and Drozd, Yu. A. Indecomposable homotopy types with at most two non-trivial homology groups, in: Groups of Homotopy Self-Equivalences and Related Topics. *Contemporary Mathematics* 274 (2001) 39–56. Baues, H.-J. and Drozd, Yu. A. Classification of stable homotopy types with torsion-free homology. *Topology* 40 (2001) 789–821. Baues, H.-J. and Hennes. The homotopy classification of $(n-1)$-connected $(n+3)$-dimensional polyhedra, $n\ge4$. *Topology* 30 (1991) 373–408. Bondarenko, V. M. Representations of bundles of semichained sets and their applications. *St. Petersburg Math. J.* 3 (1992) 973–996. Cohen, J. M. *Stable Homotopy*. Lecture Notes in Math. Springer-Verlag, 1970. Drozd, Yu. A. Matrix problems and categories of matrices. *Zapiski Nauch. Semin. LOMI* 28 (1972) 144–153. Drozd, Yu. A. Finitely generated quadratic modules. *Manus. Math.* 104 (2001) 239–256. Drozd, Yu. A. Reduction algorithm and representations of boxes and algebras. *Comptes Rendues Math. Acad. Sci. Canada* 23 (2001) 97–125. Freyd, P. Stable homotopy II. *Applications of Categorical Algebra*. Proc. Symp. Pure Math. 17 (1970), 161–191. Gelfand, I. M. and Ponomarev, V. A.. Remarks on the classification of a pair of commuting linear transformations in a finite-dimensional space. *Funk. Anal. Prilozh.* 3:4 (1969) 81-82. Gelfand, S. I. and Manin, Yu. I. *Methods of Homological Algebra*. Springer–Verlag, 1996. Henn, H.-W. Classification of $p$-local low dimensiona; spectra. *J. Pure and Appl. Algebra* 45 (1987) 45–71. Hu Sze-Tsen. *Homotopy Theory*. Academic Press, 1959. Spenier, E. *Algebraic Topology*. McGraw-Hill, 1966. Switzer, R. M. *Algebraic Topology — Homotopy and Homology*. Springer-Verlag, 1975. Toda, H. *Composition Methods in the Homotopy Groups of Spheres*. Ann. Math. Studies 49. Princeton, 1962. Unsöld. H. M. $A^4_n$-Polyhedra with free homology. *Manus. Math.* 65 (1989) 123–145. Whitehead, J. H. C. The homotopy type of a special kind of polyhedron. Ann. Soc. Polon. Math. 21 (1948), 176–186.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Two hypotheses are presented, outlining a new cause for global warming. We propose that the crucial factor in global warming is the amount and position of water vapour through the atmosphere. The purpose of this report is to open the debate and to encourage discussion among scientists.'
author:
- 'Vladimir Shaidurov[^1]'
title: 'Atmospheric hypotheses’ of Earth’s global warming'
---
When analyzing the mean-year trend of the Earth’s surface temperature for the past 140 years (see for Fig. 1(a)) one can discern two sections of monotone linear increase of temperature during two last industrial centuries.[^2] The first one begins somewhere in the period 1906-1909. The previous segment demonstrates a weak decrease in the temperature trend, not increase (see for Figs. 1(b), 4), though this includes the beginning of industrial period, with a subsequent increase in production of greenhouse gases. If we suppose that it is not only greenhouse gases that launched global warming, then what could be an additional cause of the comparatively sudden change of mean-year trend of the Earth’s surface temperature? We should look for a phenomenon of cosmic scale during this time which could have given rise to beginning of global warming with a significant probability. On the 30th June 1908 Tungus meteorite crossed almost all of the atmosphere and exploded. Instrumental measures with numerical modelling reconstruct an explosion of the power of approximately by 15 Mt TNT at an altitude approximately of 10 km. Such an explosion could cause considerable stirring of the high layers of atmosphere and change its structure.
[**Hypothesis 1.**]{} The Tungus meteorite considerably changed the thermo-protective properties of the Earth’s atmosphere and turned out to be one of the agencies which launched global warming.
From our point of view, water vapor plays a more important role in Earth’s thermal regime than other gases. The point is that additionally to the usual accumulation of heat because of increasing temperature, water vapour (as molecules of ${\rm H}_2{\rm O}$) possesses three more significant properties in the real range of pressure and temperature in the lower layers of atmosphere: liquid-gas transition, solid-liquid transition, and dissociation into ions (${\rm H}_2{\rm O} \rightarrow {\rm H}^+ + {\rm OH}^-$). At constant temperature each of these transitions consumes heat, and the opposite transitions give the same heat that is more often greater than energy due to moderate changes of temperature. Moreover, the considerable heat capacity of water in comparison with the atmospheric gases makes the World Ocean a very important accumulator of heat and source of vapour. Together, the complex of atmosphere, land surface, and ocean, assimilates 25% of the solar radiation incident on the Earth. After all stages of reradiation 97% of this energy goes out of Earth. Of the components of this complex, changes in amount of water vapor form the primary influence on assimilation of solar radiation.
For a rough estimate for the influence of increase or decrease of primary assimilation of solar energy, let as look at the minimal temperature reached at the Earth’s surface: -89in the Antarctic. Thermal processes inside the Earth do not lead to significant temperature differences at the surface in the absence of solar heating. Other parts of the Earth’s surface are at greater distance from center and liquid core owing to the Earth’s shape. Having an average temperature of land surface and ocean of about +15$^{\rm o}$, we see that the increase of assimilation of solar energy by 1% would entail an increase of temperature by more than 4in stationary regime.
But the real dynamic regime intensifies this effect. The point is that the average increase of land and ocean temperature produces higher average absolute humidity. In its turn, this raises the assimilation ability of the atmosphere even at constant content of carbon dioxide. But increasing the average ocean temperature is responsible for lower water solubility of carbon dioxide, which then arrives in the atmosphere. Moreover, increase in land temperature is responsible for growth of bogs, at least in Northern Russia, due to the removal of permafrost deep down. The rise in area and activity of bogs leads to more active production of methane. Thus, a self-stimulated process was launched for the increase of average temperature of the Earth’s surface. Therefore the rise of greenhouse gas concentration is more a consequence of warming but not a main reason.
[**Hypothesis 2:**]{} The above mentioned variant of self-stimulated process (with a permanent rise of average absolute humidity, and resulting concentration of carbon dioxide and other greenhouse gases) was launched in 1908 after the atmosphere reconstruction due to the Tungus meteorite.
Following a further examination of the curve in Figure 1(a), one can see oscillatory behavior in the period 1945 - 1976. What was happened in this time? The 16th July 1945 saw the first explosion of a nuclear bomb, and this heralded the period of nuclear bomb tests. From the standpoint of atmosphere hypothesis the nuclear tests in the atmosphere are opposite to consequences of the Tungus meteorite. When a nuclear charge explodes at the Earth’s surface or in the atmosphere, the shock wave vents water vapor from the troposphere to the stratosphere through tropopause. For some period ($\approx$ 3 years) water vapor in the stratosphere and aerosol, and dust in the troposphere and stratosphere suffice for the defense of the Earth from solar radiation. But then all gradually settled, and global warming continued. All nuclear explosions above the ground and the sea together gave rise to tendency for decreasing the global temperature of Earth’s surface. The last nuclear test in open atmosphere was on the 16th October 1980. All subsequent nuclear tests were under the ground or the sea, which does not generates the shock wave reaching the stratosphere. But the tendency of global warming recommenced earlier, approximately in 1977. Discordance between beginning of second period of warming and the finish of nuclear tests in the open atmosphere may be attributed to superposition with cyclic variations of temperature, which were apparent before 1908 too.
Now let us raise a question about some regulation of the protective properties of the atmosphere. In the first instance we have the previous consideration of nuclear explosions in the open atmosphere. But this gave a comparatively short-term effect because of the gradual subsidence of ice in pearl clouds through the tropopause, due to their greater density when compared with air. However, clouds reflect the significant part of solar radiation. But in the mesosphere there is one more type of clouds, the so called silver clouds. They persist much longer. The distinction is the following. Almost the whole of the stratosphere and mesosphere consists of molecular oxygen O$_2$ and molecular nitrogen N$_2$. Also ozone O$_3$ is formed in comparatively small quantities with the help of solar radiation. The first distinction is the temperature gradient: temperature grows with altitude in the stratosphere, approximately from -55to 0 and diminishes in the mesosphere, from about 0to -95(see Fig. 2). The second distinction consists in the different pressure and density, which are several times less in the mesosphere than in the stratosphere. Therefore, water vapour in the troposphere (such as is formed during atmospheric nuclear tests), comes to a temperature below freezing point almost everywhere except its upper border. Thus it forms crystals having greater density than the ambient gas. In rapidly moving flows the crystals migrate down through the tropopause into the troposphere. High speeds and agitation do not end this process quickly; it may continue for months depending upon the tropospheric humidity in the test region.
But water vapour in the mesosphere is another matter. At a pressure hundreds of times less than at atmospheric pressure at sea level, the freezing point of water vapour shifts to a vastly negative temperature without the intermediate liquid state (see Fig. 3). Therefore there exists a sizable layer spanning the higher part of the stratosphere and lower part of the mesosphere where water is in the gas state. The mesospheric composition is slightly distinct from the stratospheric one at significant, with less density of gases. The gaseous state of water vapor has lower density than the ambient gases (atomic masses of H$_2$O, O$_2$, and N$_2$ equal 18, 32, and 28, respectively). Therefore it has some tendency to move up in rapidly moving flows with some stirring against the background of diffusion. When it migrates, gas climbs to a temperature below freezing point, crystallizes and migrates down. There it evaporates missing the liquid state, and the process repeats. Thus, mesopause with a strongly negative temperature of around -95prevents water vapor leaving beyond the upper bound of the mesosphere.
From these discussions the following idea results about some deceleration and possible ceasing of global warming. With this purpose it is possible to start reconstruction of the protective layer in middle part of mesosphere, which consists of water vapor. For this it is enough to transport and combust molecular hydrogen, H$_2$ in the appropriate part of the mesosphere. The ambient quantity of molecular oxygen and ozone is enough to generate water vapor, which will be 9 times greater by weight than the transported hydrogen. Along with this, the density of molecular hydrogen is small enough in comparison with that of the ambient gases over the region of transport that it induces lift. Therefore hydrogen may serve as a means for transport itself.
Modern technical tools seem to be sufficient to realize such a transport in small parts and to observe the consequences. In addition, modern mathematical and computational modeling are at level when more detailed quantitative estimates are possible both for the immediate effects of such intervention as well as its influence on global climate.
Note the obvious environmental safety of this suggestion since the combustion of hydrogen in higher level of atmosphere can generate water vapor only. Combustion of small amounts gives the possibility of observing any change before it becomes too dramatic.
[**Acknowledgements:**]{} In conclusion I express my gratitude to Professor J. Levesley and Professor A. Gorban for the invitation to visit the University of Leicester and for fruitful work and discussions.
[9]{} . The Regents of the University of Michigan. University Corporation for Atmospheric Research (UCAR). http://www.windows.ucar.edu. University of Waterloo. http://www.science.uwaterloo.ca. U.S. Government Printing Office, Washington, D.C.
![(a) The Earth’s surface temperature is shown year by year (red bars) and approximately decade by decade (black line, a filtered annual curve suppressing fluctuations below near decadal time-scales). There are uncertainties in the annual data (thin black whisker bars represent the 95% confidence range) due to data gaps, random instrumental errors and uncertainties, uncertainties in bias corrections in the ocean surface temperature data and also in adjustments for urbanization over the land. (b) Additionally, the year by year (blue curve) and 50 year average (black curve) variations of the average surface temperature of the Northern Hemisphere for the past 1000 years have been reconstructed from “proxy" data calibrated against thermometer data (see list of the main proxy data in the diagram). The 95% confidence range in the annual data is represented by the grey region. These uncertainties increase in more distant times and are always much larger than in the instrumental record due to the use of relatively sparse proxy data. Nevertheless the rate and duration of warming of the 20th century has been much greater than in any of the previous nine centuries. Similarly, it is likely that the 1990s have been the warmest decade and 1998 the warmest year of the millennium. (The figure is reprinted from \[1\].)](FigSh1.eps){width="11cm" height="13cm"}
![The average temperature profile through Earth’s atmosphere. (The figure is reprinted from \[2\].) ](FigSh2.eps){width="11cm" height="16cm"}
![The pressure, the freezing point of water vapor, and the temperature of average atmosphere depending on altitude. (The data are taken from \[2\], \[3\], \[4\].) The abscissa is the altitude in km. The pressure is plotted at left ordinate as logarithm of mercury column in mm. The temperature is plotted at right ordinate in Celsius degree. The red line gives an estimate of the pressure at corresponding altitude. The blue line gives the freezing point of water vapor depending on pressure (at corresponding point). The black line demonstrates the temperature of average atmosphere at corresponding altitude and copies the analogous graph from Figure 2 in some different form. The node of intersection 1 indicates the point of formation for pearl clouds; the node 2 does for silver ones. Between them the water vapor is in gas state. Near above and near below it is in solid state. ](FigSh3.eps){width="12cm" height="12cm"}
{width="15cm" height="19cm"}
[^1]: Institute of Computational Modeling of Siberian Branch of Russian Academy of Sciences, shidurov$@$icm.krasn.ru. Visited Leicester in April 2005.
[^2]: The best estimate for this trend for the last 100 years is that the global average surface temperature has increased by $0.6 \pm 0.2$ C.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'I discuss recent progress in low-energy tests of symmetries and conservation laws, including parity nonconservation in atoms and nuclei, electric dipole moment tests of time-reversal invariance, $\beta$-decay correlation studies, and decays violating separate (family) and total lepton number.'
address: 'Inst. for Nuclear Theory and Dept. of Physics, University of Washington, Seattle, WA 98195 USA'
author:
- 'W. C. Haxton'
title: Fundamental Symmetries and Conservation Laws
---
parity nonconservation ,electric dipole moments ,correlations in $\beta$ decay ,lepton number 11.30.Er ,11.30.Fs ,11.30.Hv ,14.60.Pq ,23.40.-s
Parity Nonconservation (PNC)
============================
The use of parity in atomic spectroscopy dates from the 1920s, when it was introduced as a wave function label, prompting Wigner to demonstrate that such labeling is a consequence of the mirror symmetry of the electromagnetic interaction. Today parity and its violation are tools for probing aspects of the standard model (SM) (e.g., to isolate the strangeness-conserving hadronic weak interaction) and new physics beyond the SM (e.g., the contributions of a new boson $Z_0^\prime$ to the running of weak couplings).
The weak interaction between atomic electrons and the nucleus is dominated by the coherent $A(e)-V(N)$ contribution. As the SM tree-level coupling to protons, $c_V(p) = 1 - 4\sin^2 \theta_W \sim 0.1$, is suppressed while $c_V(n) = -1$, the weak charge of the nucleus $Q_{weak} \sim Z c_V(p)+ N c_V(n)$ is approximately $-N$. Consequently $$H^{\mathrm{Atomic~PNC}}_{A(e)-V(N)} \sim {G_F \over 2 \sqrt{2}}~ Q_{weak}~ \rho_N(\vec{r}) ~\gamma_5,$$ where $\rho_N(\vec{r})$ is the neutron density. The effects of this short-range interaction grow $\sim~Z^3$ and thus are most easily detected in heavy atoms.
Atomic PNC was first observed in 1978, while the best current limit comes from the 1997 JILA experiment, $Q_{weak}$($^{133}$Cs) = $-73.16 \pm 0.29 \mathrm{(exp)} \pm 0.20 \mathrm{(theor)}$ [@JILA; @Derevianko]. The $\sim 0.3$% precision poses a challenge for theoreticians attempting to calculate the associated atomic mixing. Advances in the atomic structure calculations include improved evaluations of relativistic (Breit) and radiative corrections, vacuum polarization, the neutron distribution, strong-field self-energies, and weak vertex corrections. The corrections have been at the sub-1% level for Cs, and in total have brought SM calculations into agreement with experiment at $\lsim 1\sigma$. Consequently beyond-the-SM contributions are constrained, yielding, e.g., a bound on the mass of an extra neutral boson, $M(Z_0^\prime) \gsim 1.3$ TeV [@Derevianko].
The atomic PNC constraint on $c_V(p)$ and $c_V(n)$, or equivalently on the underlying quark couplings, is essentially orthogonal to that from PNC electron scattering experiments, as shown in the right frame of Fig. \[fig:PVES\] [@Young]. In the left frame experimental values for $\sin^2 \theta_W$ are superimposed on SM predictions for its running [@Erler]. There is good agreement. Included on this graph are the error bars experimentalist expect to achieve in future intermediate-energy measurements at JLab.
![In the left frame [@Erler] experimental determinations of $\sin^2 \theta_W$ are compared to SM predictions, normalized to the $Z$ pole. The right frame [@Young] shows the complementary constraints from atomic PNC ($Q_{weak} \sim N$) and PNC electron scattering (where protons have been the favorite target).[]{data-label="fig:PVES"}](PVES_final.pdf){width="13.7cm"}
The Cs experimenters also studied the hyperfine dependence of the signal, from which a small nuclear-spin-dependent contribution to PNC, $V(e)-A(N)$, was extracted, $$H^{\mathrm{Atomic~PNC}}_{V(e)-A(N)} = {G_F \over \sqrt{2}} \kappa \vec{\alpha} \cdot \vec{I}\rho(\vec{r}) ~\Rightarrow~\kappa = \kappa_{Z_0} + \kappa_{\mathrm{HF}} + \kappa_{\mathrm{A}} = 0.112 \pm 0.016.$$ The measured $\kappa$, obtained from 7000 hours of data, constrains the sum of three terms, tree-level $Z_0$ exchange ($\kappa_{Z_0} \sim 0.014$, suppressed because the nuclear coupling is no longer coherent and the vector $Z_0$ coupling to the electron $\sim (4 \sin^2 \theta_W-1)/2 \sim -0.05$ is small), spin-dependent effects arising from the combination of hyperfine and $Q_{weak}$ interactions ($\kappa_{\mathrm{HF}} \sim 0.0078$), and the nuclear anapole moment ($\kappa_{\mathrm{A}} \sim 0.090 \pm 0.016$), which dominates the signal. The nuclear anapole moment is a PNC coupling of a photon to the nucleus (see Fig. \[fig:anapole\]), part of a set of weak radiative corrections. It arises from a PNC torroidal current winding within the nucleus, growing with atomic number as A$^{2/3}$ (proportional to the torroid’s cross section). This growth leads to a the weak radiative correction that exceeds the tree-level axial $Z_0$ exchange in heavy nuclei.
![The diagram on the left, a contribution to the anapole moment (a weak radiative correction arising from single-photon exchange), is part of a larger class of such corrections, including the box diagram at the center. The anapole constraints from atomic PNC studies of Cs and Tl are included in the diagram on the right, which summarizes the current status of hadronic PNC studies.[]{data-label="fig:anapole"}](anapole.pdf){width="13.7cm"}
Because the largest contribution to $\kappa_{\mathrm{A}}$ for a heavy nucleus comes from opposite-parity admixtures in the ground-state nuclear wave function, $\kappa_{\mathrm{A}}$ provides a constraint on hadronic PNC. As discussed in the next section, the Cs anapole moment seems to be somewhat larger than one would expect based on other tests of hadronic PNC.
The Nucleon-Nucleon Parity-Nonconserving Interaction
====================================================
The NN PNC interaction at low energies is characterized by the five S-P Danilov amplitudes of Table \[tab:PNC\], which in turn are often parameterized in terms of a potential derived from $\rho$, $\omega$, and $\pi^{\pm}$ exchange [@DDH]. This parameterization can be viewed as a phenomenological effective theory, with the heavy meson exchanges playing the role of short-range interactions $\vec{\nabla}_{12} \delta(\vec{r}_{12})$ in each of the five S-P channels, and with the pion separately determining the long-range behavior of the potential (see Fig. \[fig:meson\]). In fact, systematic effective field theory formulations exist for pionless theory and with explicit pions [@Dan].
Transition I $\leftrightarrow$ I$^\prime$ $\Delta$I n-n n-p p-p meson exchanges
----------------------------------- --------------------------------------- ------------------- ------------- ------------- ------------ -------------------------
${}^3S_1 \leftrightarrow {}^1P_1$ 0 $\leftrightarrow$ 0 0 x $\rho,\omega$
${}^1S_0 \leftrightarrow {}^3P_0$ 1 $\leftrightarrow$ 1 0 x x x $\rho,\omega$
1 x x $\rho,\omega$
2 x x x $\rho$
${}^3S_1 \leftrightarrow {}^3P_1$ 0 $\leftrightarrow$ 1 1 x $\pi^\pm,\rho,\omega$
: S-P weak PNC amplitudes and the corresponding meson-exchanges [@Adelberger]
\[tab:PNC\]
One of the goals of the field has been to isolate the neutral current contribution to hadronic PNC. While the weak interaction can be observed in flavor-changing hadronic decays, the neutral current contribution to such decays is suppressed by the GIM mechanism and thus unobservable. The NN and nuclear systems are thus the only practical laboratories for studying the hadronic weak interaction in all of its aspects [@Adelberger].
As the weak contribution to the NN interaction is much smaller than the strong and electromagnetic contributions, PNC is exploited to isolate weak effects. The most common observables are pseudoscalars arising from the interference of weak and strong amplitudes, e.g., the circular polarization of $\gamma$ rays emitted from an unpolarized excited nuclear state, or the $\gamma$ ray asymmetry if the nuclear state can be polarized. As the observable depends on a product of parity-conserving and PNC amplitudes, the weak interaction appears linearly. Alternatively, there are processes, such as the $\alpha$ decay of an unnatural-parity state to a 0$^+$ final state, where the observable depends on the square of a weak amplitude, and consequently is not a pseudoscalar.
The isospin of meson-nucleon couplings has an interesting relation to the underlying bare charged and neutral currents. The hadronic weak interaction is $$L^{eff} = {G \over \sqrt{2}} \left[ J^\dagger_W J_W + J_Z^\dagger J_Z \right] + h.c.~~~~~~~~
J_W = \cos{\theta_C} J_W^{\Delta S=0} + \sin{\theta_C} J_W^{\Delta S =-1},$$ where the charge-changing current is the sum of $\Delta$I=1 $\Delta$S=0 and $\Delta$I=1/2 $\Delta$S=-1 terms. Consequently the $\Delta$S=0 interaction has the form $$L^{eff}_{\Delta S=0} = {G \over \sqrt{2}} \left[ \cos^2{\theta_C}J_W^{0 \dagger} J_W^0 +
\sin^2{\theta_C} J_W^{1 \dagger} J_W^1+ J_Z^\dagger J_Z \right]$$ where the first term, a symmetric product of $\Delta$I=1 currents, has $\Delta$I=0,2, while the second term, a symmetric product of $\Delta$I=1/2 currents, is $\Delta$I=1 but Cabibbo suppressed. Consequently a $\Delta$I=1 PNC meson-nucleon vertex should be dominated by the neutral current term – a term not accessible in strangeness-changing processes. This is the $\pi^\pm$ exchange channel. One could isolate this term by an isospin analysis of a complete set of PNC NN observables or, alternatively, by finding a case in which the isospins of the admixed nuclear states select only the $\Delta$I=1 contribution
![A single-boson-exchange contribution to $V_{PNC}$ contains one weak vertex (left) and one strong one (right). DDH [@DDH] related the weak vertex to SM quark currents, using factorization, the quark model, and sum rules.[]{data-label="fig:meson"}](meson.pdf){width="12cm"}
Unfortunately, experimental progress in this field has been slow. Ideally one would like to avoid constraints from complex nuclei, as calculated nuclear mixing matrix elements are generally needed in the analysis. But there too few constraints available from $NN$ and few-body experiments. The latest effort to improve this situation, the LANSCE measurement of the analyzing power $A_\gamma(\vec{n}+ p \rightarrow d + \gamma)$, fell short of its goal and thus will need to be revisited when more intense neutron sources become available at the SNS. The best available data are [@Adelberger]
$$A_L(\vec{p}+p, 45~\mathrm{MeV}) = (-1.57 \pm 0.23) \times 10^{-7}~~~~~P_\gamma(^{18}\mathrm{F}) = (12 \pm 38) \times 10^{-5} ~~~$$ $$A_L(\vec{p}+\alpha, 46~\mathrm{MeV}) = (-3.34 \pm 0.93) \times 10^{-7} ~~~~~A_\gamma(^{19}\mathrm{F} )= (-7.4 \pm 1.9) \times 10^{-5}$$
The mixing of nearly degenerate opposite parity doublets accounts for the large nuclear PNC signals. $P_\gamma(^{18}$F) is important because 1) the mixing is isovector and 2) the nuclear matrix element can be extracted from ancillary measurements (related axial-charge $\beta$ decay) with very little uncertainty. The $^{19}$F case is also partly constrained by similar data. Finally, there is the Cs anapole moment where, despite the apparent complexity of the nuclear structure physics, the existing theoretical analyses are in good accord.
The analyses of these data can be displayed as constraints on effective isoscalar and isovector weak couplings, as shown in the right panel of Fig. \[fig:anapole\]. If the Cs anapole moment is excluded, there is a region of overlap, corresponding to a small isovector coupling (compared to the “best value" of [@DDH]) and an isoscalar coupling $\gsim$ the DDH “best value." The isovector coupling is constrained by the $^{18}$F measurement and consistent with zero: this component is the test for neutral currents. But the conclusion of a suppressed isovector NN PNC interaction rests entire on this one measurement and associated analysis. When the $^{133}$Cs anapole moment is added, no solution is found: the anapole moment appears to be larger than one would have expected, based on direct measurements of hadronic PNC. A small corner of the upper bound on the anapole moment of $^{205}$Tl is also shown: while the error bar on the Tl measurement is quite large, the result favors a coupling opposite to that of $^{133}$Cs, contradicting theory expectations.
Clearly the field needs a new generation of higher precision experiments, including neutron observables (e.g., in the $n+p$ system or $\vec{n}$ + $^4$He), to make progress.
Electric Dipole Moments and CP Violation
========================================
A permanent electric dipole moment $d$ of an elementary particle or of a composite system (such as an atom) requires both time-reversal and parity violation: $H_{edm} = d~ \vec{E} \cdot \vec{s}$ reverses sign under $t \rightarrow -t$ and under $\vec{r} \rightarrow -\vec{r}$. The signature of an edm is precession of the particle’s spin around the direction of the applied field, with a frequency proportional to $d$ and to the strength of the applied electric field. By the CPT theorem, a nonzero T-violating edm implies CP nonconservation (CPNC). Searches for edms
- test the SM’s two sources of CPNC, the CKM phase measured in neutral kaon decays and the unmeasured QCD $\theta$ parameter (edm searches require $| \theta | \lsim 10^{-10}$); and
- probe CPNC beyond the SM (baryogenesis appears to require new sources).
The sensitivity of measurements is remarkable. The dipole moment limit for $^{199}$Hg corresponds to a strain of about 10$^{-19}$, if interpreted in terms of a linear displacement. That is, were one to expand the atom to the dimensions of the earth, such an edm corresponds to a displacement of initially overlapping uniformly charged spheres (+ and -) by 10$^{-4}$ angstroms. The Hg precession sensitivity, given typical electric fields of $\sim 10^5$ v/m, corresponds to shifts in energy level splittings of $\sim 10^{-26}$ eV.
The connection between experimental limits and fundamental Lagrangians is generally not simple. Particularly in the case of the edms of diamagnetic atoms – where the spin is carried by the nucleus – a theorist must relate the fundamental CPNC phases to effective low-energy couplings, determine the nuclear interactions these couplings induce, calculate the resulting CPNC mixing of nuclear states and thus the nuclear edm, and finally evaluate the atomic screening effects that determine the residual atomic edm, the quantity experimentalists measure.
The general electromagnetic current for a spin-1/2 fermion $\langle p^\prime | j_\mu^{em} | p \rangle$ is $$\bar{N}(p^\prime) \left( F_1(q^2) \gamma_\mu
+ F_2(q^2) \sigma_{\mu \nu} q^\nu + {a(q^2) \over m^2} (\not q q_\mu - q^2 \gamma_\mu) \gamma_5
+ d(q^2) \sigma_{\mu \nu} q^\nu \gamma_5 \right) N(p),$$ where $F_1$ and $F_2$ are the ordinary charge and magnetic couplings, $a$ is the anapole coupling, and $d$ the edm. In an atom or nucleus, this last term will generate odd static charge multipoles (C1 (the edm), C3, ...) and even static magnetic multipoles (M2, ...) that are CPNC and PNC (provide the spin allows M2 and C3 moments).
Edm experiments have been done on various neutral systems, including free neutrons, paramagnetic atoms or molecules (the unpaired electron provides sensitivity to the electron edm), and diamagnetic atoms (paired electrons but a nonzero nuclear spin, so that the valence nucleon’s edm and CPNC nuclear state mixing can be probed). Table \[tab:edm\] lists some of the bounds. The field is quite active, with new or proposed efforts including ultracold neutrons (Ill, PSI, Munich, SNS), $^{199}$Hg (Seattle), liquid $^{129}$Xe (Princeton), trapped $^{225}$Ra (Argonne, KVI) and $^{213}$Ra (KVI), trapped $^{223}$Rn, and the deuteron (BNL ring experiment). Future cold-neutron efforts, for example, may improve the current upper bound, 2.9 $\times 10^{-26}$ e cm, by a factor $\sim$ 60 in the next decade [@CERN].
Particle edm limit System SM prediction (CKM phase)
------------- ------------------------------- ---------------------- -----------------------------
e 1.9$ \times 10^{-27}$ e cm atomic $^{205}$Tl $10^{-38}$ e cm
p 6.5$ \times 10^{-23}$ e cm molecular TlF $10^{-31}$ e cm
n 2.9$ \times 10^{-26}$ e cm ultracold n $10^{-31}$ e cm
$^{199}$Hg 3.1$ \times 10^{-29}$ e cm atom vapor cell $10^{-33}$ e cm
: Edm limits, direct or derived, and expected SM level, based on the CKM phase [@CERN; @Hg].
\[tab:edm\]
A new result for $^{199}$Hg, anticipated when this talk was delivered, has been announced. The experiment uses $\sim 10^{14}$ neutral atoms in a vapor cell designed to extend the spin relaxation time for Hg (which has atomic spin 0) to 100-200 s. The edm resides on the nucleus, which is shielded from an applied field by the polarization of the atomic cloud. Consequently a net interaction energy is generated only through nuclear finite-size effects, resulting in a reduced sensitivity to the nuclear edm, $$d_{\mathrm{atomic}} \sim 10 Z^2 \left( {R_N \over R_A} \right)^2 d_{\mathrm{nucleus}},$$ where $R_N$ and $R_A$ are the nuclear and atomic sizes. Such Schiff-shielding effects are less severe in heavy atoms, because of the stronger Coulomb field at the nucleus and larger $R_N$. The new result, $|d(^{199}\mathrm{Hg})| \lsim 3.1 \times 10^{-29}$ e cm [@Hg], provides the most stringent bounds on a variety of possible sources of hadronic CPNC.
There are plans for a new generation of experiments employing trapped stable or radioactive isotopes. As traps allow more flexibility in the choice of nuclear and atomic spins, this technique may open up opportunities to exploit certain isotopes with enhanced polarizabilities (though higher spins may also increase sensitivity to field inhomogeneities). For example, the CPNC mixing of the 160 eV 5/2$^-$-5/2$^+$ ground-state parity doublet in $^{229}$Pa is expected to enhance the nuclear edm by a factor of $\sim 10^3-10^4$ [@Henley]. Theoretical studies of the collective enhancements of dipole moments of octupole-deformed nuclei, where parity doublets arise naturally [@Doba], helped motivate Argonne and KVI proposals for $^{225}$Ra.
Precise Measurements of Weak Decays
===================================
As approximately a dozen contributions to this conference discuss the use of precise decay measurements to probe the weak interaction, I wish there were more time available to discuss this field. There are several motivations for these difficult experiments:
- probing general properties of weak rates, such as universality, mixing angles (e.g., the extraction of $V_{ud}$ from Fermi $\beta$ decay or from the neutron), and coupling strengths (e.g., the pseudoscalar coupling $F_P$ or the second-class tensor coupling $F_T$);
- constraining symmetry-breaking new interactions by their effects on muon or neutron decay, such as the exotic P-even, pseudo-T-odd neutron-decay $D$ coefficient $$\begin{aligned}
{d \omega \over dE_e d \Omega_e d \Omega_\nu} \propto p_e E_e (E_0-E_e)^2
&&\left[1 + a\vec{\beta}_e \cdot \hat{p}_\nu + A \vec{\sigma}_n \cdot \vec{\beta}_e \right. \nonumber \\
&&\left. + B \vec{\sigma}_n \cdot \hat{p}_\nu + b {m_e \over E_e} + D \vec{\sigma}_n \cdot
(\vec{\beta}_e \times \hat{p}_\nu) \right] \mathrm{;~and}\end{aligned}$$
- verifying SM relations among decay parameters, e.g., $$a = {g_V^2-g_A^2 \over g_V^2 -3 g_A^2}~~~~~~~~~~A=-2 {g_V^2+g_Ag_V \over g_V^2-2g_A^2}~~~~~~~~~~B=2 {g_V^2-g_A g_V \over g_V^2-3 g_A^2}.$$
Nuclei can be useful in such tests: selection rules can simplify constraints. For example, in a high-Q $0^+ \rightarrow 0^+$ $\beta$ decay, the back-to-back emission of the $e^+$ and $\nu_e$ is forbidden for a $V-A$ interaction because of unbalanced angular momentum associated with the handedness of the leptons. The addition of scalar ($S,S^\prime$) interactions $$H_\beta = \bar{\psi}_n \gamma_\mu \psi_p~\bar{\psi}_\nu C_V \gamma_\mu (1-\gamma_5) \psi_e~
+ ~\bar{\psi}_n \psi_p~\bar{\psi}_\nu (C_S + C_{S^\prime} \gamma_5) \psi_e$$ produces leptons with identical chiralities, so that emission in the same direction is forbidden, and back-to-back leptons preferred. Thus the daughter nucleus recoil momentum distribution (the observable) can be a sensitive test for nonzero $C_S$ and $C_{S^\prime}$. Among the interesting cases are $^{32}$Ar [@Ar] (where improved momentum resolution was achieved by measuring final-state delayed protons) and $^{38}$K$^m$ [@K] (where a magneto-optical trap allowed the low-energy recoiling nucleus to freely escape to a detector).
Flavor and Total Lepton Number
==============================
The discovery of neutrino oscillations (e.g., $\nu_e \rightarrow \nu_\mu$) demonstrates that flavor is not conserved among the leptons and motivates further tests of lepton flavor number and total lepton number nonconservation, $$\sum_{in} l_e \neq \sum_{out} l_e~~~~~~~~~\mathrm{or}~~~~~~~~~\sum_{in} l_e+l_\mu+l_\tau \neq
\sum_{out} l_e+l_\mu+l_\tau.$$ Total lepton number plays a special role in descriptions of neutrino mass.
As the sources of lepton flavor violation (LFV) are varied, the relative sensitivities to new physics are highly model dependent. Classic low-energy tests include $\mu \rightarrow e + \gamma$, $\mu + (N,Z) \rightarrow e + (N,Z)$, and $\mu \rightarrow e+e+e$, as well as corresponding $\tau$ decays. Fig. \[fig:muon\] shows representative sensitivities for two LFV mechanisms.
![Illustrations from [@CERN; @Paradisi] of the expected branching ratios for Higgs-mediated and gaugino-mediated LFV, vs. the respective Higgs boson and SUSY masses.[]{data-label="fig:muon"}](muon.pdf){width="13.5cm"}
Table \[tab:muon\] lists some of the LFV branching ratio bounds that have been obtained in the past two decades. The generally favored $\mu^+ \rightarrow e^+ \gamma$ mode can be mimicked by final-state accidentals, limiting the advantages of high-intensity muon beams. In contrast, as beam intensity is often a limiting factor in searches for $\mu^- \rightarrow e^-$ conversion in nuclei, improved beams can lead to large increases in sensitivity. The conversion process requires good energy resolution, $\lsim$ 1 MeV, to exclude the background from ordinary $\mu$ decay.
Mode Bound (90% c.l.) Year Experiment/Laboratory
------------------------------------------- -------------------------- -------- -------------------------
$\mu^+ \rightarrow e^+ \gamma$ 1.2$ \times 10^{-11}$ 2002 MEGA/LAMPF
$\mu^+ \rightarrow e^+ e^+ e^- $ 1.0$ \times 10^{-12}$ 1988 SINDRUM I/PSI
$\mu^+ e^- \leftrightarrow \mu^- e^+$ 8.3$ \times 10^{-11}$ 1999 PSI
$\mu^-$Ti $\leftrightarrow e^-$ Ti 6.1$ \times 10^{-13}$ 1998 SINDRUM II/PSI
$\mu^-$Ti $\leftrightarrow e^+$ Ca$^*$ 3.6$ \times 10^{-11}$ 1998 SINDRUM II/PSI
$\mu^-$Pb $\leftrightarrow e^-$ Pb 4.6$ \times 10^{-11}$ 1996 SINDRUM II/PSI
$\mu^-$Au $\leftrightarrow e^-$ Au 7.0$ \times 10^{-13}$ 2006 SINDRUM II/PSI
: Limits on LFV from experiments with muons [@CERN]
\[tab:muon\]
Both J-PARC and FermiLab have plans for next generation $\mu \rightarrow e$ conversion experiments that will substantially improve limits on LFV. These experiments will use pulsed proton beams to remove pion backgrounds by timing, large-acceptance capture solenoids to increase the $\mu$ flux, and bent solenoids to transport the muons, removing neutrals and separating charge. The FermiLab experiment will use 8 GeV protons from a new driver and has a branching ratio goal of $4 \times 10^{-17}$, while the J-PARC experiment will use a 40 GeV proton beam and has a goal of $5 \times 10^{-19}$ [@CERN]. This program will push LFV sensitivities for scalar exchanges from the current level of $\sim$ 1 TeV to $\sim$ 10 TeV.
Tests of total lepton number are important to the description of massive neutrinos. Neutrinos are unique among SM fermions in lacking an obvious charge or other additively conserved quantum number that would reverse sign under particle-antiparticle conjugation. Consequently, particle and antiparticle could be identical, $\nu=\bar{\nu}$. That is, the neutrino might be a Majorana particle, rather than Dirac ($\nu \perp \bar{\nu}$). Prior to 1957 the absence of neutrinoless $\beta \beta$ decay appeared to rule out a Majorana neutrino, since the process illustrated in Fig. \[fig:bb\]a) would lead to relatively rapid decay. This result seemed to require a “charge" – lepton number – to distinguishing $\nu$ and $\bar{\nu}$, and the assumption of conservation of that charge to account for the absence of $\beta \beta$ decay, $$l_e(e^-) = l_e(\nu_e)=+1~~~~~l_e(e^+)=l_e(\bar(\nu)_e)=-1~~~~~~~~~~\sum_{in} l_e = \sum_{out} l_e.$$ The process in Fig. \[fig:bb\]b) would then be forbidden, as there are no leptons in the initial state, while the final state carries a net lepton number of two.
But this conclusion ignores neutrino helicity: a massless right-handed neutrino has the wrong handedness to be reabsorbed in the second $\beta$ decay of Fig. \[fig:bb\]a). That is, absence of neutrinoless $\beta \beta$ tells us nothing about the Dirac or Majorana nature of the neutrino if the process is independently forbidden by maximal PNC.
![Neutrinoless $\beta \beta$ decay scenarios: a) the pre-1957 Majorana case, which appeared to conflict with experiment; b) the Dirac case, where the process is forbidden by lepton number conservation; and c) the Majorana case, where the handedness mismatch is not total, due to the neutrino mass.[]{data-label="fig:bb"}](doublebeta.pdf){width="10cm"}
The discovery of neutrino mass, however, changes this argument. Handedness is no longer exact and thus does not forbid $\beta \beta$ decay. Instead, the rate is suppressed by a factor $(m_\nu/E_\nu)^2$, where $E_\nu$ is the typical energy of the exchanged neutrino, as illustrated in Fig. \[fig:bb\]c). If one can overcome this suppression factor by doing a very sensitive experiment, $\beta \beta$ decay would be observed, provided Majorana neutrinos exist. Furthermore, the rate might tell us something about the scale of neutrino mass – important because oscillation experiments only probe $m^2$ differences.
One might expect the neutrino to have both Dirac $\bar{\Psi}_R M_D \Psi_L$ and Majorana $\bar{\Psi}_R^c M_R \Psi_R$ components. The latter breaks the global gauge invariance $\Psi \rightarrow e^{i \alpha} \Psi$ associated with a conserved lepton number $l_e$, and thus can contribute to manifestly $l_e$-forbidden processes like neutrinoless $\beta \beta$ decay. As was discussed in Concha Gonzalez-Garcia’s talk, there is a prejudice for such masses because they also explain, via the seesaw mechanism, why neutrinos are light. In the seesaw mechanism the diagonalization of the mass matrix $$\left( \begin{array}{cc} 0 & M_D \\ M_D & M_R \end{array} \right) ~\longrightarrow~
m_\nu^{\mathrm{light}} ~\sim~ M_D \left( {M_D \over M_R} \right)$$ provides the needed “small parameter" $M_D/M_R$ that explains why neutrinos are so much lighter than other SM fermions (which can only have Dirac masses). Thus light neutrinos are a reflection of the greater freedom available in building neutrino masses. Current oscillation results suggest $M_R \sim 3 \times 10^{15}$ GeV, a value near the GUT scale.
The possibility of discovering total-lepton-number violation is high because
- Nature likely makes use of Majorana masses;
- atmospheric neutrino experiments suggest that the mass scale is not unreasonably small, $m_\nu \gsim 0.05$ eV; and
- extraordinary efforts are underway to mount massive new $\beta \beta$ decay experiments that will extend current sensitivities by an additional factor $\sim$ 1000.
Summary
=======
Precise tests of symmetries and conservation laws, using low-energy techniques, remain one of our best windows on physics of and beyond the SM, complementing the experiments performed at the energy frontier. Their utility derives from
- the many opportunities to isolate interesting interactions in both elementary and composite systems, using angular momentum, parity, and kinematics;
- exquisite experimental sensitivities (atomic shifts of $\sim 10^{-26}$ eV, $\beta \beta$ decay lifetimes of $10^{26}$ years);
- level degeneracies and collective responses enhancing interesting interactions in atoms and nuclei;
- unique sensitivities, such as the GUT-scale reach of $\beta \beta$ decay;
- the improving intensities of muon and cold neutron beams; and
- the capacity of theory to connect what is learned at low-energies to both astrophysics (e.g., neutrino mass) and accelerator physics (e.g., supersymmetry at the LHC).
Acknowledgment
==============
This work was supported in part by the Office of Nuclear Physics, US Department of Energy. I thank the organizers of PANIC08 for a most enjoyable meeting.
[00]{}
C. S. Wood et al., *Science* **275**, 1759 (1997); S. C. Bennett and C. E. Wieman, *Phys. Rev. Lett.* **82**, 2484 (1999)
S. G. Porsev, K. Beloy, and A. Derevianko, arXiv: 0902.03351
R. D. Young, R. D. Carlini, A. W. Thomas, and J. Roche, *Phys. Rev. Lett.* **99**, 122003 (2007)
J. Erler and M. J. Musolf, *Phys. Rev. D* **72**, 073003 (2005)
B. Desplanques, J. F. Donoghue, and B. R. Holstein, *Annals Phys.* **124**, 449 (1980)
D. R. Phillips, M. R. Schindler, and R. P. Springer, arXiv:0812:2073; S. L. Zhu [*et al.*]{}, *Nucl. Phys. A* **748**, 435 (2005)
E. G. Adelberger and W. C. Haxton, *Ann. Rev. Nucl. Part. Sci* **35**, 501 (1985)
M. Raidal [et al.]{}, arXiv:0801:1826
W. C. Griffith, M. D. Swallows, T. H. Loftus, M. V. Romalis, B. R. Heckel, and E. N. Fortson, arXiv:0901.2328
W. C. Haxton and E. M. Henley, *Phys. Rev. Lett.* **51**, 1937 (1983); O. P. Sushkov, V. V. Flambaum, and I. B. Khriplovich, *Zh. Exp. Teor. Fiz.* **87**, 1521 (1984)
N. Auerbach, V. V. Flambaum, and V. Spevak, *Phys. Rev. Lett.* **76**, 4316 (1996); J. Dobaczewski and J. Engel, *Phys. Rev. Lett.* **94**, 232502 (2005)
E. G. Adelberger [*et al.*]{}, *Phys. Rev. Lett.* **83**, 1299 (1999)
A. Gorelov [*et al.*]{}, *Phys. Rev. Lett.* **94**, 142501 (2005)
P. Paradisi, *JHEP* **0608**, 047 (2006)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Developing effective coarse grained (CG) approach is a promising way for studying dynamics on large size networks. In the present work, we have proposed a strength-based CG ([$s$-CG ]{}) method to study critical phenomena of the Potts model on weighted complex networks. By merging nodes with close strength together, the original network is reduced to a CG-network with much smaller size, on which the CG-Hamiltonian can be well-defined. In particular, we make error analysis and show that our strength-based CG approach satisfies the condition of statistical consistency, which demands that the equilibrium probability distribution of the CG-model matches that of the microscopic counterpart. Extensive numerical simulations are performed on scale-free networks, without or with strength-correlation, showing that this [$s$-CG ]{}approach works very well in reproducing the phase diagrams, fluctuations, and finite size effects of the microscopic model, while the [$d$-CG ]{}approach proposed in our recent work \[Phys. Rev. E 82, 011107(2010)\] does not.'
author:
- 'Chuansheng Shen$^{1,2}$'
- 'Hanshuang Chen$^{1}$'
- 'Zhonghuai Hou$^{1}$'
- 'Houwen Xin$^{1}$'
title: 'Coarse-grained Monte Carlo simulations of the phase transition of Potts model on weighted networks'
---
Introduction
============
In the last two decades, we have witnessed dramatic advances in complex networks research, which has been one of the most active topics in statistical physics and closely related disciplines [@RMP02000047; @AIP02001079; @SIR03000167; @PRP06000175; @PRP08000093]. The central issue in this field is to study how the topology of networks influences dynamics, such as phase transition, self-organized criticality and epidemic spreading, etc. Usually, Monte Carlo (MC) simulations [@Lan2000] have been widely used to study such dynamics. However, the sizes of many real-world networks are very large, such as human brain composed of about $10^{11}$ neurons and $10^{14}$ synapses [@RMP06001213], and thereby brute-force simulations are quite expensive and sometimes even become impossible. Phenomenological models, such as mean-field description, may capture certain properties of the system, but often ignore microscopic details and fluctuation effects which may be important near some critical points. Therefore, a promising way to bridge the gap between the microscopic details and system level behaviors is to develop coarse-grained (CG) approaches, aiming at significantly reducing the degree of freedom while properly preserving the microscopic information of interest.
Recently, several CG approaches have been proposed in the literature. Renormalization transformation has been used to reduce the size of self-similar networks while preserving the most relevant topological properties of the original ones [@PRL04016701; @NTR05000392; @PRL06018701; @PRL08148701]. Gfeller and Rios proposed spectral decomposition technique to obtain a CG-network which can reproduce the random walk and synchronization dynamics of the original network [@PRL08174104]. Kevrekidis et al. developed equation-free multiscale computational methods to accelerate simulation using a coarse time-stepper [@CMS03000715], which has been successfully applied to study the CG dynamics of oscillator networks [@PRL06144101], gene regulatory networks [@JCP06084106], and adaptive epidemic networks [@EPL08038004]. Nevertheless, none of the works mentioned above has considered critical phenomena in complex networks, which has been a frontier topic in the context of network science [@RMP08001275].
Very recently, we have proposed a degree-based CG ([$d$-CG ]{}) approach to study the critical phenomena of the Ising model and the SIS-epidemic model in unweighted networks [@PRE10011107]. A local mean field (LMF) scheme was introduced to generate the CG network from the microscopic one. Specifically, we have proposed a so-called condition of statistical consistency (CSC) that the CG-model should satisfy to guarantee the validity of the CG-approach. We showed that the CSC can be exactly fulfilled if we merge nodes with the same degree together. Extensive numerical simulations showed that our [$d$-CG ]{}approach does work very well to reproduce the phase transition behaviors of the original network, including the critical point and the fluctuation properties, but with much less computational efforts. Our method also makes it feasible to investigate the finite size effects of both models, which should be much more expensive and even forbidden if we use brute-force methods. However, this [$d$-CG ]{}approach can only apply to binary networks, i.e., each of the link in the network either exists or not, but with no weight. As we know, many real-world networks are intrinsically weighted, with their links having diverse strengths. Examples include the collaboration networks [@PRE01016131; @PRE01016132; @PHA02000590], airport networks [@PNA04003747; @PRE05036124], metabolic networks [@NAT04000839] predator-prey relationship networks [@Pin2002], and so on. Therefore, a straightforward question is: Can we use CG approaches to study the critical phenomena in weighted networks?
To answer this question, in the present work, we have considered the critical phenomena of the Potts model in weighted complex networks. The Potts model is related to a number of important topics in statistical and mathematical physics [@RMP82000235; @Gar1979] and was successfully applied to neural networks, multiclass classification problems, graph coloring problem, and so on. It contains a system of coupled nodes, each of which has *p* possible states. Only when two nodes are in the same state, they have pairwise interactions. With the increment of temperature, the Potts model undergoes an order-disorder phase transition at some critical temperature. For $p = 2$, Potts model is equivalent to the well-known Ising model. Instead of the [$d$-CG ]{}scheme, we have proposed a strength-based CG ([$s$-CG ]{}) approach, where those nodes with *similar* strength are merged together to form a CG-node. Note that in weighted networks, it is unpractical to merge nodes with exactly the same strength together. By detailed analysis of the discrepancy between the Hamiltonian of a CG configuration and that of its corresponding microscopic configurations, we show that the [$s$-CG ]{}approach can approximately satisfy the CSC defined on weighted networks. Extensive numerical simulations are performed on scale-free(SF) networks, without or with strength-correlation, showing that our [$s$-CG ]{}approach works very well in reproducing the phase diagrams, fluctuations, and finite size effects of the microscopic model, while the simple [$d$-CG ]{}does not. Compared to our previous work [@PRE10011107], the present study step forward several important steps. First of all, we should note that [$s$-CG ]{}is a brand new method compared to [$d$-CG ]{}and the latter cannot apply to weighted networks, although they share some similar ideas. Secondly, weighted networks are of more ubiquitous importance than binary unweighted ones, thus the [$s$-CG ]{}approach should find more applications. What is more, we have extended the study from the simple two-state Ising model to a more general one, the multi-state Potts model. In addition, we have performed error analysis in the present study, which clearly demonstrates the robustness of our approaches.
Coarse Graining Procedure {#sec2}
=========================
CG Potts Model
--------------
In this paper, we consider the $p$-states Potts model on a weighted network consisted of $N$ nodes, whose Hamiltonian is given by $$H = - \sum\limits_{i<j} {w_{ij} \delta _{\alpha _i ,\alpha _j } } ,
\label{eq1}$$ where $w_{ij}$ is the weight on the edge connecting a pair of nodes $i$ and $j$ ($w_{ij}=0$ if the nodes $i$ and $j$ are not connected). $\alpha _i(= 1,\cdots,p)$ denotes the state of node $i$, $\delta
_{\alpha_i,\alpha_j } = 1$ if $\alpha_i=\alpha_j$ and $0$ otherwise.
To setup the CG-Potts model, one needs to obtain the CG-Hamiltonian defined on the CG-network, followed by CG-MC simulations to study the dynamic behaviors. The CG-network is simply obtained by node-merging, i.e., $q_\mu$ nodes within the original micro-network are merged into a single CG-node $C_\mu$, where $\mu=1,...,N^c$ labels the CG-node and $N^c$ is size of the CG-network. Following the LMF scheme used in Ref. [@PRE10011107], the weight of link between two CG nodes $\mu$ and $\nu$ reads, $$\label{eqCGA}
\bar w_{\mu\nu} =
\begin{cases}
\frac{2}{q_\mu(q_\mu-1)}\sum\limits_{i,j \in C_\mu; i<j} w_{ij}
& \text{for $\mu = \nu$}, \\
\frac{1}{q_\mu q_\nu}\sum\limits_{i \in C_\mu,j \in C_\nu} w_{ij}
& \text{for $\mu \ne \nu$}.
\end{cases}$$
The CG-Hamiltonian $\bar H$ can be readily obtained, $$\bar H = \bar H_1 + \bar H_2$$ where
\[eqCGH\] $$\begin{aligned}
\bar H_1 &=& - \sum\limits_\mu {\bar w_{\mu\mu} } \sum\limits_\alpha {\frac{\eta _{\mu,\alpha} (\eta_{\mu,\alpha} - 1)}{2}}
\label{eq4a} \\
\bar H_2 &=& - \sum\limits_{\mu,\nu( > \mu)} {\bar w_{\mu\nu} } \sum\limits_\alpha {\eta _{\mu,\alpha} \eta _{\nu,\alpha} }
\label{eq4b}
\end{aligned}$$
Herein, $\bar H_1$($\bar H_2$) denote CG interactions inside(among) the CG-nodes, respectively. $\eta_{\mu,\alpha}$ stands for the number of $\alpha$-state micro-nodes inside $C_\mu$. Since there are $\frac{{\eta _{\mu,\alpha} (\eta _{\mu,\alpha} - 1)}}{2}$ possible distinct pairs of $\alpha$-state micro-nodes inside $C_\mu$, and each pair has a weighted coupling $w_{\mu\mu}$, the CG-interactions among all the $\alpha$-state nodes inside $C_\mu$ is given by $$\bar H_{\mu,1}^{(\alpha)} = - \bar w_{\mu\mu} \frac{{\eta _{\mu,\alpha} (\eta _{\mu,\alpha} - 1)}}{2}.$$ Summation this over all CG-nodes $\mu$ and states $\alpha$ gives the result in Eq.(3a). Eq.(3b) can be interpreted in a similar way. Note that Eq.(3) are closed at the CG level, i.e., as long as one has constructed the CG-network, $\bar {\bm w}$ and $\bar H$ are then both well defined, based on which one can perform CG-MC simulations without going back to the micro-level.
CSC: Condition of statistical consistency
-----------------------------------------
The above procedure tells us how to calculate the CG-Hamiltonian if we already have the CG-network. However, which $q_\mu$ nodes are merged together to form a CG-node $C_\mu$ is yet not determined. Generally speaking, one may construct the CG-network deliberately, for instance, one may simply generate $N^c$ values, $q_\mu$ obeying $\sum_{\mu=1}^{N^c} q_\mu=N$ and then just randomly merge $q_\mu$ micro-nodes to form $C_\mu$. Therefore, an important question arises: How to guarantee that the CG-model can reproduce the dynamics of the corresponding microscopic model correctly?
We address this problem by extending the so-called CSC as proposed in [@PRE10011107]. We introduce $\vec
\eta_\mu=\{\eta_{\mu,\alpha}\}_{\alpha=1,..., p}$ to denote the state of $C_\mu$ and $\vec {\bm \eta}= \{\vec
\eta_\mu\}_{\mu=1,...,N^c}$ to denote the configuration of the CG-network. Note that a given CG configuration $\vec {\bm \eta}$ corresponds to many microscopic configurations, which defines the degeneracy factor $g(\vec {\bm \eta})$. In the equilibrium state of the CG-model, the probability of finding a given CG-configuration $\vec {\bm \eta}$ is given by the canonical distribution, i.e., $$p_{\rm CG}(\vec {\bm \eta})=g(\vec {\bm \eta})e^{-\bar H/k_BT}/\bar Z,$$ where $\bar Z=\sum_{\vec {\bm \eta}} p_{\rm CG}(\vec {\bm \eta}) $ is the CG partition function. It is important to note, however, that $p(\vec {\bm \eta})$ can be calculated exactly from the equilibrium distribution of the micro-model, $$p_{\rm micro}(\vec {\bm \eta}) = \sum\nolimits_{}' e^{ - H/k_B T}/Z,$$ where $Z$ is the partition function of the micro-model, and the prime means summation over all the microscopic configurations that contribute to $\vec {\bm \eta}$. Since we are interested in the equilibrium phase transition behavior of the Potts model, we thus assert that for the CG-model to be statistically consistent with the micro-model, $p_{\rm CG}(\vec {\bm \eta})$ and $ p_{\rm micro}(\vec
{\bm \eta})$ must be equal, i.e., the CSC reads $$\label{eqCSC}
g(\vec {\bm \eta})e^{-\bar H/k_BT}/\bar Z =\sum\nolimits_{}' e^{ -
H/k_B T}/Z.$$
[$s$-CG ]{}Scheme and error analysis
------------------------------------
In the present work, we propose a [$s$-CG ]{}scheme to construct the CG-network, i.e., nodes with same or similar strengths are merged together to form a CG-node, where the strength $s_i$ of node $i$ is defined as $s_i = \sum\nolimits_j {w_{ij} } $ [@PRL01005835; @PNA04003747]. In the following, we will show that if nodes inside each CG-node have same strengths, the CSC will hold exactly within the ANA. In addition, if the strengths within $C_\mu$ are nearly the same, the CSC can also hold approximately.
In the literature, ANA [@RMP08001275; @PRE03036112; @PRL02258702; @PRE02035108] has been widely used to study the ensemble averaged dynamics of complex networks and proved to be successful. ANA assumes that one can replace the dynamics on a given network by that on a weighted fully connected graph with connectivity $ A_{ij}=d_i d_j/(D N)$, where $d_i$ ($d_j$) denotes the degree of node $i$ ($j$) and $D$ is the mean degree of the network. Analogously, in weighted networks link weight can be expressed as $$w_{ij}=s_i
s_j/(S N)
\label{eqANA}$$ where $S$ is the mean strength of the network. Substituting Eq.(\[eqANA\]) into Eq.(\[eqCGA\]), the adjacency matrix of the CG-network now reads,
\[eqCGA\_ANA\] $$\begin{aligned}
\bar w_{\mu\mu} &=& {\frac{2}{{q_\mu (q_\mu - 1)}}} \sum\limits_{i
< j \in C_\mu } {\frac{{(S_\mu + \delta s_i )(S_\mu + \delta s_j
)}}{{S N}}} \nonumber \\
&=& \frac{S_\mu^2}{{S N}} \left( {1 - \Omega_\mu } \right)
\\
\bar w_{\mu\nu} &=& {\frac{1}{{q_\mu q_\nu}}}\sum\limits_{i \in C_\mu ,j \in C_\nu } {\frac{{(S_\mu + \delta s_i )(S_\nu +
\delta s_j )}}{{S N}}} \nonumber \\
& =& \frac{1}{{S N}}\sum\limits_{i \in C_\mu ,j \in C_\nu } {S_\mu
S_\nu }
= \frac{S_\mu S_\nu}{{S N}}
\end{aligned}$$
Herein, we have written $s_i=S_\mu+\delta s_i$, with $S_\mu=\frac{1}{q_\mu}\sum_{i\in C_\mu}s_i$ being the mean strength within $C_\mu$. $\Omega_\mu = \frac{{ \left< \delta s^2 \right
>_\mu }}{S_\mu^2 (q_\mu - 1)}$ where $\left< \delta s^2 \right
>_\mu=\frac{1}{q_\mu}\sum_{i\in C_\mu} (\delta s_i)^2$ is the variance of strength within $C_\mu$. In the first equation, we have used the fact that $ ( \sum_{i \in C_\mu} \delta s_i )^2 = 2\sum_{i < j \in
C_\mu } \delta s_i \delta s_j + \sum_{i \in C_\mu } (\delta s_i) ^2=
0.$ The second equation holds simply because $\sum_{i \in C_\mu, j
\in C_\nu }\delta s_i \delta s_j = (\sum_{i \in C_\mu} \delta s_i)(
\sum_{j \in C_\nu} \delta s_j)=0.$ Substituting Eq.(\[eqCGA\_ANA\]) into Eq.(\[eqCGH\]), we can get
\[eqCGH\_ANA\] $$\begin{aligned}
\bar H_1 &=& - \frac{1}{{S N}}\sum\limits_\mu { {S_\mu^2 \left( {1 - \Omega_\mu } \right)\sum\limits_\alpha
{\frac{{\eta _{\mu,\alpha} (\eta _{\mu,\alpha} - 1)}}{2}} } } \\
\bar H_2 &=& \frac{1}{{S N}}\sum\limits_{\mu,\nu( > \mu)} {S_\mu S_\nu } \sum\limits_\alpha {\eta _{\mu,\alpha} \eta
_{\nu,\alpha} } ,
\label{eq12}
\end{aligned}$$
To compare the CG-Hamiltonian with the microscopic one, we now group the micro-nodes with same state $\alpha$ inside $C_\mu$ as $C_{\mu,\alpha}$. Clearly, the size of $C_{\mu,\alpha}$ is $\eta_{\mu,\alpha}$. As in Eq.(\[eqCGH\]), we can also split the micro-Hamiltonian $H$ into two parts , $$H = H_1 + H_2$$ where $H_1$ and $H_2$ denote energy contributions from intra and inter the CG-nodes respectively. With ANA, and noting the fact only nodes with same states have interactions at the micro-level, one has
\[eqMicH\] $$\begin{aligned}
H_1 &=& - \sum\limits_\mu \sum\limits_\alpha {\sum\limits_{i < j \in C_{\mu,\alpha} }} \frac{s_i s_j}{S N} \frac{\eta
_{\mu,\alpha} (\eta _{\mu,\alpha} - 1)}{2} \\
H_2 &=& - \sum\limits_{\mu,\nu( > \mu)} \sum\limits_\alpha {\sum\limits_{ i \in C_{\mu,\alpha},j \in C_{\nu,\alpha} } }
\frac{s_i s_j}{S
N} \eta _{\mu,\alpha} \eta _{\nu,\alpha}
\end{aligned}$$
Following similar steps to obtain Eq.(\[eqCGH\_ANA\]), we may also write $s_i=S_{\mu,\alpha}+\delta s_i$ (here node $i$ belongs to the group $C_{\mu,\alpha}$) and Eqs.(\[eqMicH\]) change to
\[eqMicH\_ANA\] $$\begin{aligned}
H_1 &=& - \sum\limits_\mu {\sum\limits_\alpha {\sum\limits_{i < j \in C_{\mu,\alpha} } {\frac{{(S_{\mu,\alpha}
+ \delta s_i )(S_{\mu,\alpha} + \delta s_j )}}{{S N}} } } } \times \nonumber \\ & & \frac{{\eta _{\mu,\alpha} (\eta _{\mu,\alpha} -
1)}}{2} \nonumber \\
&=& - \frac{1}{{S N}}\sum\limits_\mu {\sum\limits_\alpha { {S_{\mu,\alpha}^2
(1 - \Omega _{\mu,\alpha} )\frac{{\eta _{\mu,\alpha} (\eta _{\mu,\alpha} - 1)}}{2}} } } \\
H_2 &=& - \sum\limits_{\mu,\nu( > \mu)} {\sum\limits_\alpha {\sum\limits_{i \in C_{\mu,\alpha} ,j \in C_{\nu,\alpha} }
{\frac{{(S_{\mu,\alpha} + \delta s_i )(S_{\nu,\alpha} + \delta s_j )}}{{S N}} } } }\times \nonumber \\ & & \eta _{\mu,\alpha} \eta
_{\nu,\alpha} \nonumber \\
&=& - \frac{1}{{S N}}\sum\limits_{\mu,\nu( > \mu)} { \sum\limits_\alpha
{\eta _{\mu,\alpha} \eta _{\nu,\alpha} S_{\mu,\alpha} S_{\nu,\alpha}} }
\end{aligned}$$
Here $\Omega_{\mu,\alpha} = \frac{{ \left< \delta s^2 \right
>_{\mu,\alpha} }}{S_{\mu,\alpha}^2 (\eta _{\mu,\alpha} - 1)}$ where $\left< \delta s^2 \right
>_{\mu,\alpha}=\frac{1}{\eta _{\mu,\alpha} }\sum_{i\in C_{\mu,\alpha}} (\delta s_i)^2$ is the variance of strength within the group of nodes $C_{\mu,\alpha}.$
Comparing Eq.(\[eqCGH\_ANA\]) with Eq.(\[eqMicH\_ANA\]), the discrepancy between the CG-Hamiltonian and the micro-Hamiltonian, is given by
\[eqCG\_Mic\] $$\begin{aligned}
\bar H_1 - H_1 &=& - \frac{1}{{S N}}\sum\limits_\mu
{\sum\limits_\alpha { \frac{{\eta _{\mu,\alpha} (\eta _{\mu,\alpha}
- 1)}}{2} } } \times \nonumber \\ & & {[ {S_\mu^2 ( {1 - \Omega _\mu
} ) - S_{\mu,\alpha }^2 (1 - \Omega _{\mu,\alpha } )} ]}
\\
\bar H_2 - H_2 & = &- \sum\limits_{\mu,\nu} {\sum\limits_\alpha
{\frac{\eta _{\mu,\alpha} \eta _{\nu,\alpha} ( S_\mu S_\nu -
S_{\mu,\alpha }S_{\nu,\alpha }) }{{S N}} }}\end{aligned}$$
Obviously, for the exact [$s$-CG ]{}algorithm where all the nodes inside a given CG-node have same strengths, $\Omega_\mu =
\Omega_{\mu,\alpha}=0 \ \forall (\mu,\alpha) $ and $S_\mu=S_{\mu,\alpha}$, hence $\bar H_1=H_1$ and $\bar H_2=H_2$. In this case, all those microscopic configurations contributing to a CG-configuration $\vec {\bm \eta}$ have exactly the same Hamiltonian $H$, which also equals to the CG-Hamiltonian $\bar H$. Since the constrained summation $\sum\nolimits '$ contains exactly $g(\vec
{\bm \eta})$ items, the numerators on both sides of Eq.(\[eqCSC\]) are exactly equal, i.e., $g(\vec {\bm \eta}) e^{-\bar H/k_B
T}=\sum\nolimits 'e^{- H/k_B T} $. Since we can also write the microscopic partition function as $Z=\sum_{\vec {\bm \eta}} (
\sum\nolimits ' e^{-H/k_BT})$, it is readily to show that the two partition functions equal, $\bar Z = Z$. Therefore, the CSC, Eq.(\[eqCSC\]), exactly holds.
However, we should note that for a weighted network, the exact [$s$-CG ]{}method is not practical, since the strength of a given node is generally not an integer. Therefore, usually one can only merge nodes with close strength together. Let us analyze Eq.(\[eqCG\_Mic\]) again. The factor $\Omega_\mu$ scales as $\frac{ { \left< \delta s^2
\right >_\mu }} {S_\mu^2 q_\mu}$, hence if we merge many nodes with similar strengths together, $\Omega_\mu \ll 1$ is expected to be true. One may also expect that $\Omega_{\mu,\alpha}\ll 1$ for the same reason. Therefore, the discrepancy between $\bar H$ and $H$ mainly depends on the difference between $S_\mu$ and $S_{\mu,\alpha}$. Here, we note that the nodes with $\alpha$-state flip with time during the simulation. In the equilibrium state, one expects that $C_{\mu,\alpha}$ may scan throughout $C_\mu$ for many times, such that $S_{\mu,\alpha}$ averaged over time is close to $S_\mu$. Hence $(\bar H-H)/H$ averaged over long time could be small. Note that if we merge nodes randomly, $\Omega_\mu \ll 1$ and $\Omega_{\mu,\alpha} \ll 1$ will be violated and the above reasoning should fail. We thus conclude that the pratical [$s$-CG ]{}approach, by merging nodes with similar strength together, can satisfy the CSC approximately.
Numerical Results {#sec3}
=================
To show the validity of our [$s$-CG ]{}approach, we perform extensive simulations on weighted SF networks. SF networks are much heterogeneous and serve as better candidates to test our method than other homogeneous networks, such as small-world or random networks (other types of complex networks have also been investigated, the qualitative results are the same and not shown here). We first generate a regular (unweighted) SF network by using the Barabási–Albert (BA) model [@SCI99000509] with power-law degree distribution $P(k)\sim k^{-3}$. To convert this unweighted SF network into a weighted one, we use the algorithm as proposed in Ref. [@PRE04026109]: The weight of a link between node $i$ and $j$ ($1\le i,j\le N$) is given by $ w_{ij} =(
\frac{i}{N}+\frac{j}{N} )^\theta / 2$, where $\theta$ is a tunable parameter. Note that $\theta = 0$ corresponds to an unweighed network.
The MC simulation at the microscopic level follows standard Metropolis dynamics: At each step, a micro-node is randomly selected and its state is randomly updated with an acceptance probability $\min(1, e^{-\Delta H/k_B T})$, where $\Delta H$ is the associated change of the micro-Hamiltonian, $k_B$ is the Boltzmann constant, $T$ is the temperature. In the present work, we set $k_B=1$. Similarly, during each CG-MC step, a CG-node $C_\mu$ is randomly chosen with probability proportional to its size $q_\mu$. The probability for the process that an $\alpha$-node changes to a $\beta$-node, with correspondingly $\eta_{\mu,\alpha}\rightarrow
\eta_{\mu,\alpha}-1$ and $\eta_{\mu,\beta} \rightarrow
\eta_{\mu,\beta}+1$, is given by $\eta_{\mu,\alpha} \min(1, e^{ -
\Delta \bar H/k_B T} )$, where $\Delta \bar H $ is the change of CG-Hamiltonian during this process. Since $N^c$ can be much smaller than $N$, the CG-MC is expected to be much faster and memory-saving than the micro-level MC simulation.
The collective state of the system is described by the total magnetic moment $ M = \frac{1}{{2N}}\sum\limits_{\mu,\alpha }
{\left| {M_{\mu,\alpha } } \right|}$, where $ M_{\mu,\alpha } =
\frac{{p\eta _{\mu,\alpha } - 1}} {{p - 1}}{\kern 8pt} ( \mu= 1,
\cdots ,N^c)$ denotes the $\alpha$-component of the magnetic moment within $C_\mu$. With increasing temperature $T$, the Potts model undergoes a phase transition at some critial temperature $T_c$ from an ordered state, where $M \sim O(1) $ is strictly nonzero, to a disordered state with $M \simeq 0$. We use the similar [$s$-CG ]{}approach to construct the CG-network with different $N^c$ and compare the results obtained from CG-MC simulations with those of micro-MC simulations.
To begin, we show the results in Fig.1 for $\theta = 0$, where the network are essentially unweighted and the [$s$-CG ]{}approach is identical to the [$d$-CG ]{}. Fig.1(a) and 1(b) show the moment $M$ and susceptibility $\chi=\beta N(\langle M^2 \rangle - \langle M\rangle ^2 )$ as functions of $T$, respectively. The susceptibility is related to the variance of the total magnetization according to the fluctuation-dissipation theorem. Apparently, our results (empty squares and solid circles) are in excellent agreements with the micro-level counterparts (solid lines). As comparisons, we have also shown the results obtained by a random-merging (RM) CG-model (dotted lines) and the heterogeneous mean field theories (HMFT)[@EPJ04000177] (empty triangles). Here, the RM model means that one simply merge $N/N^c$ randomly selected nodes to form a CG-node. Evidently this random scheme fails to reproduce the microscopic behaviors at all. The results of the HMFT are obtained by numerically solving the self-consistent equations of order parameter[@EPJ04000177]. We find that the HMFT can predict the curve of $M \sim T$ quite well, however, it fails to predict the curve of $\chi \sim T$. Strikingly, even when the original network is reduced to one with only $16$ CG-nodes, the CG model still faithfully reproduces the phase transition curves and fluctuation properties. Since $N^c$ is largely reduced compared to $N$, a considerable speed-up of CPU time can be achieved which makes it feasible to study system size effects. Fig.1(c) plots $T_c$ as a function of $\ln N$, obtained by our CG method with $N^c=64$. $T_c$ is determined as the location of the peak in the $\chi \sim T $ curve, see Fig.(1b). The dependence is linear with a slope $\simeq
1.68$, which agrees rather well with a theoretical prediction ${T_c}/\ln N=\frac{S}{4p}\simeq1.67$ [@EPJ04000177], where $S$ is the average node strength in the network.
![(color online). (a)-(b) $M$ and $\chi$ as functions of $T$ (in unit of $J/k_B$) for the Potts model on unweighted SF networks ($\theta =0$), obtained from brute-force MC simulation (solid line), HMFT(triangle), random-merging CG(dotted line), and the [$s$-CG ]{}(square and circle). $N=16384$, $p=3$. (c) Dependence of $T_c$ on the network size $N$ obtained by the [$s$-CG ]{}approach with fixed $N^c=64$. All the networks have fixed mean degree $D=20$. The error bars (not shown) are smaller than the symbol sizes. \[fig1\]](fig1.eps){width="1.0\columnwidth"}
For $\theta \ne 0$, the networks are weighted. Here we take $\theta
=2.4$ as an example to ensure the heterogeneity of the link weights. Figure(2a) and (2b) show $M$ and $ \chi $ as functions of $T$ respectively. As in Fig.1(b), the peak in $\chi$ locates the critical point $T_c$. Clearly, the [$s$-CG ]{}results (solid circle) are still in excellent agreements with the MC results (solid lines), however, the [$d$-CG ]{}(solid squares) [@PRE10011107] and RM-CG (dotted lines) both fails. For such weighted networks, the dynamic equations of HMFT is not available either. Thus, for such weighted networks, our [$s$-CG ]{}approach is the only promising CG approach so far. In Fig.2(c), we have also shown the dependence of $T_c$ on the network size. Apparently, there is also a linear dependence between $T_c$ and $\ln N$ with the slope being about $1.288$. As mentioned in the last paragraph, this slope depends on the average strength $S$. For a weighted network, one may estimate $S$ by $\langle w_{ij}
\rangle D$, where $\langle w_{ij} \rangle \simeq \int_0^2 x^\theta
/4dx = \frac{1}{{4(\theta + 1)}}2^{\theta + 1} $. Substituting $ D
= 20$, $\theta = 2.4$ and $p = 3$ to these formula, we obtain ${T_c}/\ln N=\frac{S}{4p}\simeq1.293$, which is consistent with the simulation value.
In real-world networks, correlation is an ubiquitous feature. For instance, social networks show that nodes with large degrees tend to connect together, a property referred to as assortative mixing[@PRL02208701]. In contrast, many technological and biological networks show disassortative mixing, i.e., connections between high-degree and low-degree nodes are more probable [@PRL01258701; @SCI02000910]. Previous studies showed that correlations may play important roles in network dynamics [@PRL02208701; @PRL01258701; @SCI02000910; @PRE08051105; @PRE02047104]. In the present work, we have used our [$s$-CG ]{}method to study the phase transition of Potts model on correlated networks, which can not be studied by the HMFT which assumes no degree correlation. To characterize the assortative property of the weighted network, a strength correlation coefficient $r$, an extension of the degree correlation [@PRL02208701], can be defined as $$r = (\langle s_i s_j \rangle - \langle s_i \rangle \langle s_j
\rangle )/(\langle s_i^2 \rangle - \langle s_i \rangle ^2 ) .
\label{eq13}$$ Here $s_i$ and $s_j$ are the strengths of the two end-nodes of an edge. $r$ is zero for networks with no strength-correlation, such as BA-SF networks, and positive or negative for assortative or disassortative mixing networks, respectively.
![(color online). Phase transition behaviors of the Potts model on unweighted correlated networks. (a) $T_c$ plotted as a function of the network correlation coefficient $r$, obtained via CG-MC and MC simulations. $N=1024$ and $N^c=64$. (b) Dependence of $T_c$ on the network size $N$. All the networks have fixed mean degree $D=20$. \[fig3\]](fig3.eps){width="0.9\columnwidth"}
Figure 3(a) shows $T_c$ as a function of $r$, obtained from our [$s$-CG ]{}approach and micro-MC simulations for $\theta = 0$. Again, the fits between CG-MC and MC are good. Figure 3(b) shows the effects of correlated network size on $T_c$. Interestingly, we find that the linear dependence between $T_c$ and $\ln N$ is lost for correlated networks. For assortative(disassortative) networks $T_c$ grows monotonically much faster(slower) than $\ln N$, respectively. In other words, the ordered state in an assortative(disassortative) network is harder(easier) to be destroyed with increasing temperature than in an un-correlated network. This is understandable since a hub-node in the network is more difficult to change its state than a leaf-node due to larger energy barrier. In an assortative network, hub-nodes are connected together, such that they tend to freeze into a local ordered state which is stable to thermal fluctuations. For a disassortative network, a hub-node is usually connected to many leaf-nodes. Since leaf-nodes can change state easily, the alone${\kern 1pt}$ hub-node is more likely to change state with the help of their boiling${\kern 1pt}$ neighbors. Therefore, assortative correlations tend to increase $T_c$ as observed here.
![ Magnetization $M$ and susceptibility $\chi$ as functions of temperature $T$ for the Potts model on weighted correlated networks. Symbols and lines correspond to the CG-MC and micro-MC simulation results, respectively. Other parameters are same as in Fig.2. \[fig4\]](fig4.eps){width="0.9\columnwidth"}
In Fig.(4), the magnetization $M$ and susceptibility $\chi$ of the ferromagnetic Potts model on weighted networks are plotted as functions of temperature $T$ at different correlation coefficient $r$, obtained from our [$s$-CG ]{}approach and micro-MC simulations. Again, the agreements between CG-MC and MC are excellent, further demonstrating the validity of our method.
Conclusions {#sec4}
===========
In summary, we have developed a stength-based [$s$-CG ]{}approach for coarse-graining study of the phase transition of the Potts model on weighted networks. We have utilized a mean-field scheme to generate the connectivity of the CG-network and derived the CG-Hamiltonian. To address the problem how to guarantee the validity of the CG-model, we have proposed the so-called CSC, which requires that the probability to find a given CG-configuration in the equilibrium state, calculated from the CG-model, should be the same as that calculated from the original microscopic model. We show, by performing error analysis, that our [$s$-CG ]{}approach, by merging nodes with close strengths together, holds the CSC approximately with ANA. Detailed numerical simulations demonstrate clearly that our [$s$-CG ]{}approach can reproduce the microscopic MC simulation results very well, not only for the onset of phase transition, but also for the fluctuations and system size effects.
This work was supported by the National Science Foundation of China under Grants No.20933006 and No.20873130.
[39]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, , , , , ****, ().
, ** (, , ).
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
, , , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , , , ****, ().
, ** (, , ).
, ****, ().
, ** (, , ).
, , , ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We comment on a recent paper [@Deng2009] by Deng et al. in which the Eddington-Robertson parameters for our modified gravity theory (MOG) are derived. We show by explicit calculation that the role of the vector field $\phi_\mu$ cannot be ignored in this derivation.'
author:
- |
J. W. Moffat$^{\dag *}$ and V. T. Toth$^{\dag}$\
\
bibliography:
- 'refs.bib'
title: 'Comment on “Modified scalar-tensor-vector gravity theory and the constraint on its parameters” by Deng, et al.'
---
In [@Deng2009], Deng et al. present a generalization of STVG [@Moffat2006a; @Moffat2007e], a scalar-tensor-vector formulation of our modified gravity theory (MOG). Their argument is based on a Parameterized Post-Newtonian (PPN, [@Chandra1965; @Will1993]) representation of STVG, which they derive in the appendices of their paper. In this derivation, the authors explicitly ignore the MOG vector field: “Then, we rewrite Eqs. (35) and (42) while abandoning the vector field \[...\]”. Through this approach, the authors reduce MOG to a specific case of Brans-Dicke theory [@BD1962], depriving MOG of the essential feature of a repulsive massive vector field.
Yet the vector field cannot be ignored. Considering Eq. (\[eq:12\]) in [@Deng2009], we find that the term $\omega m^2k^2\phi_\mu\phi_\nu/c^2$ appears in sums also containing the trace of the matter stress-energy tensor. Clearly, this term can be omitted only if the condition $\omega m^2k^2\phi_\mu\phi^\mu/c^2\ll T$ is satisfied. Otherwise, this term must be retained. Considering our spherically symmetric, static vacuum solutions, this is precisely the case: for instance, in the near vacuum of the outer solar system, $T$ is vanishingly small, yet $\phi_0$ is nonvanishing due to the “fifth force” influence of the Sun. Even near the Sun, for instance along the path of a radio signal received from distant spacecraft at opposition, the energy density associated with the fifth force vector field is comparable to the energy density of the local medium.
To demonstrate this through explicit calculation, we begin with the following equations from [@Deng2009], ignoring terms of order ${\cal O}(\epsilon^4)$ and higher: $$\begin{aligned}
g_{00}&=-1+\epsilon^2N\tag{3}\label{eq:3},\\
g_{ij}&=\delta_{ij}+\epsilon^2H_{ij}\tag{5}\label{eq:5},\\
G&=G_0(1+\epsilon^2\overset{(2)}G)\tag{25}\label{eq:25},\\
m&=m_0(1+\epsilon^2\overset{(2)}m)\tag{29}\label{eq:29},\\
\omega&=\omega_0(1+\epsilon^2\overset{(2)}\omega)\tag{30}\label{eq:30},\\
R_{00}&=-\frac{1}{2}\epsilon^2\Box N\tag{A13}\label{eq:A13},\\
T_{00}&=\sigma-\epsilon^2(2N\sigma+\sigma_{kk})\tag{A14}\label{eq:A14},\\
T_{ij}&=\epsilon^2\sigma_{ij}\tag{A16}\label{eq:A16},\\
T&=-\sigma+\epsilon^2(N\sigma+2\sigma_{kk})\tag{A17}\label{eq:A17}\end{aligned}$$ from which it follows that $\partial_\mu G\partial^\mu G\propto\epsilon^4$, $\partial_\mu m\partial^\mu m\propto\epsilon^4$, and $\partial_\mu\omega\partial^\mu\omega\propto\epsilon^4$ can all be ignored at the $\epsilon^2$ level. Terms quadratic in $B_{\mu\nu}$ can also be ignored, leading us to rewrite Eq. (\[eq:12\]) from [@Deng2009] as follows: $$\begin{aligned}
R_{\mu\nu}=&\frac{8\pi G}{c^2}\left[T_{\mu\nu}-\frac{1}{2}g_{\mu\nu}T+\omega\frac{m^2k^2}{c^2}\phi_\mu\phi_\nu\right]\nonumber\\
&-\frac{G_{;\mu\nu}}{G}-\frac{1}{2}g_{\mu\nu}\frac{G_{;\kappa}^{;\kappa}}{G}.\tag{12}\label{eq:12}\end{aligned}$$ The $00$-component of (\[eq:12\]) can be written, using (\[eq:3\]), (\[eq:A14\]), and (\[eq:A17\]), as $$\begin{aligned}
R_{00}=&\frac{8\pi G}{c^2}\bigg[\sigma-\epsilon^2(2N\sigma+\sigma_{kk})\nonumber\\
&\hskip 24pt+\frac{1}{2}(1-\epsilon^2N)\left\{-\sigma+\epsilon^2(N\sigma+2\sigma_{kk})\right\}\nonumber\\
&\hskip 24pt+\omega\frac{m^2k^2}{c^2}\phi_0\phi_0\bigg]+\frac{1}{2}(1-\epsilon^2N)\frac{\Box G}{G},\tag{12a}\end{aligned}$$ or, after simplification, while making use of (\[eq:25\]) and (\[eq:A13\]), we obtain a modified version of (\[eq:A18\]) of [@Deng2009]: $$-\frac{1}{2}\epsilon^2\Box N=\frac{8\pi G}{c^2}\left[\frac{1}{2}\sigma-\epsilon^2N\sigma+\omega\frac{m^2k^2}{c^2}\phi_0\phi_0\right]+\frac{1}{2}\epsilon^2\Box\overset{(2)}G.\tag{A18}\label{eq:A18}$$
In a similar vein, we obtain the $ij$-component of (\[eq:12\]). First, we make use of the gauge condition (32) in [@Deng2009] and rewrite (\[eq:A11\]) from [@Deng2009] as $$R_{ij}=\epsilon^2\left(-\frac{1}{2}H_{ij,kk}-\overset{(2)}G_{,ij}\right).\tag{A11}\label{eq:A11}$$ Using (\[eq:A11\]), we then write the $ij$-component of (\[eq:12\]) as $$\begin{aligned}
&\epsilon^2\left(-\frac{1}{2}H_{ij,kk}-\overset{(2)}G_{,ij}\right)=\frac{8\pi G}{c^2}\bigg[\epsilon^2\sigma_{ij}\nonumber\\
&-\frac{1}{2}(\delta_{ij}+\epsilon^2H_{ij})\{-\sigma+\epsilon^2(N\sigma+2\sigma_{kk})\}+\omega\frac{m^2k^2}{c^2}\phi_i\phi_j\bigg]\nonumber\\
&-\frac{G_{;ij}}{G}-\frac{1}{2}(\delta_{ij}+\epsilon^2H_{ij})\frac{G^{;\kappa}_{;\kappa}}{G}.\tag{12b}\end{aligned}$$ After simplification, we obtain (\[eq:A20\]) from [@Deng2009], again with slight modifications: $$\begin{aligned}
-\frac{1}{2}\epsilon^2H_{ij,kk}&=\frac{8\pi G}{c^2}\bigg[\frac{1}{2}\delta_{ij}\sigma+\epsilon^2\bigg(\sigma_{ij}-\frac{1}{2}\delta_{ij}N\sigma-\delta_{ij}\sigma_{kk}\nonumber\\
-\frac{1}{2}&\sigma H_{ij}\bigg)+\omega\frac{m^2k^2}{c^2}\phi_i\phi_j\bigg]-\frac{1}{2}\epsilon^2\delta_{ij}\Box\overset{(2)}G.\tag{A20}\label{eq:A20}\end{aligned}$$
Next, we rewrite Eq. (\[eq:20\]) from [@Deng2009] as follows: $$\begin{aligned}
&(\theta+3)\frac{G_{;\kappa}^{;\kappa}}{G}=\nonumber\\
&-\frac{8\pi G}{c^2}\bigg(-\sigma+\epsilon^2(N\sigma+2\sigma_{kk})-\omega\frac{m^2k^2}{c^2}\phi_\mu\phi^\mu\bigg),\tag{20}\label{eq:20}\end{aligned}$$ from which $$\epsilon^2\Box\overset{(2)}G=\frac{1}{\theta+3}\frac{8\pi G}{c^2}\bigg(\sigma-\epsilon^2(N\sigma+2\sigma_{kk})+\omega\frac{m^2k^2}{c^2}\phi_\mu\phi^\mu\bigg).\tag{20a}\label{eq:20a}$$
Using (\[eq:20a\]), we can rewrite (\[eq:A18\]) as $$\begin{aligned}
-\frac{1}{2}\epsilon^2\Box N=&\frac{8\pi G}{c^2}\bigg[\left(\frac{1}{2}+\frac{1}{2(\theta+3)}\right)\sigma+\omega\frac{m^2k^2}{c^2}\phi_0\phi_0\nonumber\\
&+\frac{1}{2(\theta+3)}\omega\frac{m^2k^2}{c^2}\phi_\mu\phi^\mu+{\cal O}(\epsilon^2)\bigg],\tag{A18a}\end{aligned}$$ and (\[eq:A20\]) as $$\begin{aligned}
-\frac{1}{2}\epsilon^2H_{ij,kk}=&\frac{8\pi G}{c^2}\bigg[\left(\frac{1}{2}-\frac{1}{2(\theta+3)}\right)\delta_{ij}\sigma+\omega\frac{m^2k^2}{c^2}\phi_i\phi_j\nonumber\\
-\frac{1}{2(\theta+3)}&\delta_{ij}\omega\frac{m^2k^2}{c^2}\phi_\mu\phi^\mu+{\cal O}(\epsilon^2)\bigg].\tag{A20a}\end{aligned}$$
As in [@Deng2009], we define $H_{ij}=\delta_{ij}V$. If we ignore terms containing $\phi_i$, we get $$\epsilon^2\Box N=-\frac{8\pi G}{c^2}\left[\frac{\theta+4}{\theta+3}\sigma+\frac{2\theta+5}{\theta+3}\omega\frac{m^2k^2}{c^2}\phi_0^2+{\cal O}(\epsilon^2)\right],\tag{A22}$$ and $$\epsilon^2\Box V=-\frac{8\pi G}{c^2}\left[\frac{\theta+2}{\theta+3}\sigma+\frac{1}{\theta+3}\omega\frac{m^2k^2}{c^2}\phi_0^2+{\cal O}(\epsilon^2)\right].\tag{A24}$$ If $\sigma\gg\omega m^2k^2\phi_0^2/c^2$, we get back the result of [@Deng2009] for the Eddington–Robertson parameter $\gamma$: $$\gamma=\frac{V}{N}\simeq\frac{\Box V}{\Box N}\simeq\frac{\theta+2}{\theta+4}.\tag{A25}\label{eq:A25}$$ However, if $\phi_0^2$ is the dominant term, we obtain $$\begin{aligned}
\gamma\simeq\frac{1}{2\theta+5}.\tag{A25a}\label{eq:A25a}\end{aligned}$$ The quantity $\phi_0$ is nonvanishing [@Moffat2007e]: $$\phi_0\simeq -\sqrt{\frac{G_N}{\omega}}M\frac{e^{-mr}}{r}\nonumber$$ is determined by the magnitude of the source mass $M$ (in this case, the mass of the Sun) and the inverse of the distance from the source. This expression allows us to calculate numerically the mass density associated with the vector field at a specific distance from the Sun, for instance at $r=0.1$ A.U.: $$\rho_\phi{}_{[r=0.1~\mathrm{A.U.}]}=\frac{m^2k^2G_NM^2e^{-2mr}}{c^2r^2}=5.4\times 10^{-19}~\mathrm{kg}/\mathrm{m}^3.\nonumber$$
In contrast, a typical proton density of 5 protons per cubic centimeter at a distance of 1 A.U. from the Sun (and assuming that the density of the solar wind falls off as $1/r^2$) we get, at $r=0.1$ A.U., given the proton mass at $m_p=1.67\times 10^{-27}$ kg, a proton mass density of $$\rho_p{}_{[r=0.1~\mathrm{A.U.}]}=100m_pN_p{}_{[r=1~\mathrm{A.U.}]}=8.35\times 10^{-19}~\mathrm{kg}/\mathrm{m}^3.\nonumber$$
Given that these two values, $\rho_\phi$ and $\rho_p$, are comparable in magnitude, neither (\[eq:A25\]) nor (\[eq:A25a\]) properly represent the observed value of the Eddington-parameter $\gamma$. Further, one must consider significant fluctuations (up to an order of magnitude) in the solar wind mass density, making it likely that no constant value of $\theta$ can satisfy the stringent limits on $\gamma$ from precision solar system observations.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The quantum de Finetti theorem asserts that the $k$-body density matrices of a $N$-body bosonic state approach a convex combination of Hartree states (pure tensor powers) when $N$ is large and $k$ fixed. In this note we review a construction due to Christandl, Mitchison, König and Renner [@ChrKonMitRen-07] valid for finite dimensional Hilbert spaces, which gives a quantitative version of the theorem. We first propose a variant of their proof that leads to a slightly improved estimate. Next we provide an alternative proof of an explicit formula due to Chiribella [@Chiribella-11], which gives the density matrices of the constructed state as a function of those of the original state.'
address:
- 'CNRS & Laboratoire de Mathématiques (UMR 8088), Université de Cergy-Pontoise, F-95000 Cergy-Pontoise, France.'
- 'CNRS & Laboratoire de Mathématiques (UMR 8088), Université de Cergy-Pontoise, F-95000 Cergy-Pontoise, France.'
- 'Université Grenoble 1 & CNRS, LPMMC (UMR 5493), B.P. 166, F-38 042 Grenoble, France'
author:
- Mathieu LEWIN
- Phan Thành NAM
- Nicolas ROUGERIE
date: 'August 21, 2014'
title: Remarks on the quantum de Finetti theorem for bosonic systems
---
Introduction {#sec:intro}
============
Consider a system of $N$ bosons with one-particle state space $\gH$, a separable Hilbert space. Pure states of this system are rank-one projectors $|\Psi\rangle \langle \Psi|$ onto normalized vectors $\Psi \in {\gH ^{N} }$, the symmetric tensor product of $N$ copies of $\gH$. Mixed states are then convex combinations of pure states, that is, self-adjoint positive trace-class operators with trace $1$: $$\label{eq:states}
\cS ({\gH ^{N} }):= \left\{ \Gamma \in \gS ^1 ({\gH ^{N} }),\: \Gamma = \Gamma ^*, \: \Gamma {\geqslant}0, \: \operatorname{Tr}\Gamma = 1 \right\}.$$ It is a well-known fact that Hartree states, i.e. projections onto tensor powers $\Psi=u ^{\otimes N}\in \gH^N$ (for $u$ a normalized vector in $\gH$), play a very special role in the physics of bosonic systems. Indeed, since bosons, contrarily to fermions, do not satisfy the Pauli exclusion principle, there is a possibility for many particles to occupy the same quantum state, which is the meaning of the ansatz $u ^{\otimes N}$. It is in fact the case for non interacting particles in the ground state of a one-body Hamiltonian, or in the thermal state at low enough temperature. This is the famous *Bose-Einstein condensation* phenomenon.
The case of interacting bosonic particles is much more subtle. It is no longer true that the bosonic ground state of an interacting Hamiltonian (say with two-body interactions) is given exactly by a Hartree state. In fact, one may expect that this is almost the case for large systems with properly scaled interactions, that is in the limit $N\to \infty$ with a $N$-dependent two-body coupling ensuring that the total energy of the system stays of order $N$. Proofs of this expectation for various models are available in the literature (see [@PetRagVer-89; @RagWer-89; @Werner-92; @LieSeiSolYng-05; @LewNamRou-13] and references therein), emphasizing the relevance of mean-field theories obtained from the ansatz $\Psi = u ^{\otimes N}$ to the study of large bosonic systems.
One possible strategy to investigate these mean-field limits was employed in [@PetRagVer-89; @RagWer-89; @Werner-92; @LewNamRou-13] and is based on a remarkable structure property of the set of bosonic states in the limit of infinitely many particles. Roughly speaking, *any $N$-body bosonic state is almost a convex combination of Hartree states for large $N$*. Since the quantum mechanical energy is linear in the state[^1], it is then immediate to guess that the infimum over all states should coincide with the infimum over Hartree states in the large $N$ limit. Hence the validity of the mean-field approximation follows using very little of the particular form of the Hamiltonian. Of course, to obtain a rigorous proof, one must justify the preceding heuristic statement in a sufficiently strong sense, and this is precisely the purpose of the *quantum de Finetti theorem*.
Let us start with a sequence of (mixed) states $\Gamma_N \in \cS ({\gH ^{N} })$ and construct the corresponding reduced $k$-body density matrices $$\label{eq:reduced matrices}
\Gamma_N ^{(k)} := \operatorname{Tr}_{k+1\to N} \left[ \Gamma_N \right] \in \cS ({\gH ^{k} }).$$ Here $\operatorname{Tr}_{k+1\to N}$ is a notation for the partial trace with respect to all the variables but $k$. Modulo a diagonal extraction one can always assume that there is a subsequence (not relabeled) along which $$\label{eq:weak conv}
\Gamma_N ^{(k)} \wto_* \Gamma ^{(k)}$$ weakly-$\ast$ in the trace-class, when $N\to \infty$. If in addition the convergence in is strong (that is, holds in trace-class norm), then we have the consistency relations $$\label{eq:consistent}
\Gamma ^{(k)} = \operatorname{Tr}_{k+1} \Gamma ^{(k+1)}$$ for any $k=0,1,2,\ldots $. The hierarchy $(\Gamma ^{(k)})_{k\in \NN}$ can then be thought of as describing a system with infinitely many particles and the quantum de Finetti theorem of Størmer [@Stormer-69] and Hudson-Moody [@HudMoo-75] states that there exists a unique Borel probability measure $\mu$ on the sphere $S\gH$ of the one-particle Hilbert space, invariant under the group action of $S^1$, such that $$\label{eq:deF strong}
\boxed{\Gamma ^{(k)} = \int_{S\gH} |u ^{\otimes k}\rangle \langle u ^{\otimes k} | d\mu (u).}$$ The above result is a generalization of the famous classical de Finetti theorem of Hewitt and Savage [@DeFinetti-31; @DeFinetti-37; @HewSav-55]. It is one way to give a rigorous meaning to the previous heuristic statement, already extremely useful in applications to mean-field systems: typically it is safe to assume that particles interact only pairwise and then the energy of a state depends only on the reduced $2$-body density matrix.
A second step is to wonder whether one could obtain a quantitative version of the de Finetti theorem, much like Diaconis and Freedman obtained a quantitative version of the classical de Finetti theorem [@DiaFre-80]. Namely, one may ask the following question: given a state $\Gamma_N \in \cS ({\gH ^{N} })$, does there exist another state ${\tilde{\Gamma}}_N$ of the form $${\tilde{\Gamma}}_N = \int_{S\gH} |u ^{\otimes N}\rangle \langle u ^{\otimes N} | d\mu_N (u)$$ with $\mu_N$ a probability measure, such that $$\label{eq:quant deF}
\boxed{ \operatorname{Tr}_{{\gH ^{k} }} \left| \Gamma_N ^{(k)} - {\tilde{\Gamma}}_N ^{(k)} \right| {\leqslant}C(N,k)}$$ with $C(N,k)\to 0$ when $N\to \infty$ and $k$ is fixed ? The answer to this question is still open for general Hilbert spaces, but important progress has been made in recent years in the case of finite dimensional spaces [@KonRen-05; @FanVan-06; @ChrKonMitRen-07; @Chiribella-11; @Harrow-13]. Motivated by applications in quantum information theory, variants of the above question have also been investigated [@ChrKonMitRen-07; @ChrTon-09; @RenCir-09; @Renner-07; @BraHar-12] by either adding assumptions on the state or weakening the conclusion.
As concerns , the best estimate available, proved by Christandl, König, Mitchison and Renner in [@ChrKonMitRen-07], is $C(N,k)=4dk/N$ where[^2] $d$ is the dimension of $\gH$. It has the desired property that for fixed $d$ and $k$, $C(N,k) \to 0$ when $N\to \infty$. We emphasize that here we discuss only bosonic states, but in [@ChrKonMitRen-07] it is shown that one may generalize to general symmetric states (boltzons), with $C(N,k) = 4kd ^2 /N$. A result of this form had appeared before in [@FanVan-06] for the case $k=2$.
In this note we do three things: first we review the construction of [@ChrKonMitRen-07] and discuss some related heuristics in Section \[sec:motiv\]. We then propose a variant of their proof which gives a slightly better bound $\sim 2 kd /N$ when $N\gg dk$, see Theorem \[thm:CKMR\]. Next we express the construction, which is in fact given by an anti-Wick quantization, in terms of creation/annihilation operators. This yields our Theorem \[thm:CKMR-identity\], an explicit expression of the density matrices of the state ${\tilde{\Gamma}}_N$ as a function of the density matrices of the original state $\Gamma_N$. A variant of the bound follows immediately, see Remark \[rm:non opt bound\].
The explicit formula we obtain for the density matrices of the state ${\tilde{\Gamma}}_N$ was derived before by Chiribella in [@Chiribella-11] (see also [@Harrow-13]), where it is interpreted as a relation between “universal measure and prepare quantum channels” and “optimal cloning maps”. Our proof is new, and proceeds by expressing the problem at hand in terms of the second quantization formalism and the CCR algebra. This method, along with the relationship between de Finetti measures and upper/lower symbols in a coherent state representation that we discuss in Section \[sec:motiv\], provides an alternative interpretation of the fundamental principles at work.
We end this introduction by noting that, for finite dimensional spaces, the quantum de Finetti theorem in the form can be obtained from the aforementioned bounds simply by passing to the limit $N\to \infty$ at fixed $d$. Then, as we proved in [@LewNamRou-13] (see Section 2 therein), a stronger theorem for general spaces can be obtained by combining the localization method of [@Lewin-11] and in finite dimensional spaces. Namely, under the sole assumption , the conclusion in still holds provided the measure $\mu$ is allowed to live on the ball of the Hilbert space rather than on the sphere. This is connected to recent results of Ammari and Nier [@AmmNie-08; @AmmNie-09; @AmmNie-11]. It is then simple to deduce the usual statement when there is strong convergence. The method in [@LewNamRou-13] thus shows in particular that the de Finetti theorem for general spaces can be deduced from the version in finite dimensional spaces, which in turn follows from the quantitative bounds we discuss in the sequel. This has the advantage of providing a more constructive (but longer) proof of the result than that given in the original references [@Stormer-69; @HudMoo-75] and in [@CavFucSch-02].
**Acknowledgments.** NR thanks Denis Basko, Thierry Champel and Markus Holzmann for a helpful discussion. We also benefited from interesting exchanges with Matthias Christandl, Isaac Kim and Jan Philip Solovej and financial support from the European Research Council (FP7/2007-2013 Grant Agreement MNIQS 258023) and the ANR (Mathostaq project, ANR-13-JS01-0005-01). NR and PTN acknowledge the hospitality of the Institute for Mathematical Science of the National University of Singapore.
An explicit construction and the associated estimates {#sec:CKMR}
=====================================================
Motivation and heuristics {#sec:motiv}
-------------------------
When the one-body Hilbert space $\gH$ is finite dimensional, one can base an explicit construction on Schur’s lemma: by rotational invariance of the (normalized) Haar (uniform) measure $du$ on the unit sphere $S\gH$ we have $$\label{eq:Schur}
\dim \gH^N \int_{S\gH} | u^{\otimes N} \rangle \langle u^{\otimes N} | \, du = {{\ensuremath {\mathds 1} }}_{\gH^N}$$ where $$\dim \gH^N = {N+d-1 \choose {d-1}}\quad\text{and}\quad \dim\gH=d.$$ The decomposition is a coherent state representation [@KlaSka-85; @ZhaFenGil-90], and in this respect, one may rephrase the problem we are looking at as the quest for an *upper symbol* associated to a given (mixed) state $\Gamma$ on $\gH^N$. That is we may look for a positive measure $d{\mu ^{\rm up}}$, absolutely continuous with respect to $du$, such that the identity $$\label{eq:up symb}
\Gamma = \int_{S\gH} d{\mu ^{\rm up}}(u) | u^{\otimes N} \rangle \langle u^{\otimes N} |,$$ or at least an approximation of it, holds. Of course there need not be such a *positive* upper symbol[^3] for every state (but there is always one if $d{\mu ^{\rm up}}$ is allowed to be a signed measure [@Simon-80]). This point of view is helpful nevertheless because there is always a (positive) *lower symbol* (or Husimi function) associated to the state in the representation : $$\label{eq:low symb}
d{\mu ^{\rm low}}(u) :=\dim \gH^N {{\ensuremath{\left\langle u^{\otimes N},\Gamma u^{\otimes N} \right\rangle}}} du.$$ Now, the limit $N\to \infty$ is reminiscent of a semi-classical limit, in which one may expect that upper and lower symbols will approximately coincide [@Lieb-73b; @Simon-80]. A good guess is then to define the candidate de Finetti measure as being the lower symbol of $\Gamma$ in the representation .
We can even motivate this choice more explicitly: assume that the state in question does have a positive upper symbol to begin with, i.e. that holds with ${\mu ^{\rm up}}$ a positive measure. We are then looking for an approximate expression of the upper symbol in terms of the state itself. Let us compute the lower symbol[^4] of : $$\begin{aligned}
d{\mu ^{\rm low}}(u) &=\dim \gH^N {{\ensuremath{\left\langle u^{\otimes N},\Gamma u^{\otimes N} \right\rangle}}} du \\
&= \dim \gH^N \left( \int_{S\gH} |\langle u, v \rangle| ^{2N} d{\mu ^{\rm up}}(v) \right) du.\end{aligned}$$ It is clear that for large $N$, the contribution to the above integral of vectors $v$ that are not exactly colinear to $u$ will be negligible, and therefore that the integral will simply converge to $d{\mu ^{\rm up}}(u)$, which means that $$d{\mu ^{\rm low}}\to d{\mu ^{\rm up}}$$ as measures when $N\to \infty$. If the upper symbol was not positive to begin with, this still shows that it can be approximated by a positive measure (the lower symbol) when $N$ is large. This should not be surprising because (as we used above) $\langle u ^{\otimes N}, v ^{\otimes N}\rangle_{{\gH ^{N} }} \to 0$ if $u$ and $v$ are not colinear and $N\to \infty$. The coherent state basis $(u ^{\otimes N})_{u\in \S \gH}$ thus tends to be “less and less overcomplete” in this limit, i.e. closer and closer to a true basis. Then the upper symbol should become positive simply by positivity of the state.
All this suggests that the lower symbol should be a good approximation to the upper symbol. Of course the lower symbol also has the advantage of being always well-defined as an explicit, positive, function of the state. It is thus natural to take $$\label{eq:Husimi deF}
\boxed{d\mu_N (u) :=\dim \gH^N {{\ensuremath{\left\langle u^{\otimes N},\Gamma u^{\otimes N} \right\rangle}}} du}$$ as the candidate de Finetti measure of a given $N$-body state $\Gamma$, and this is the choice of [@ChrKonMitRen-07]. In the following we state an estimate of the form , which confirms its sensibility.
A quantitative de Finetti theorem and an explicit formula {#sec:results}
---------------------------------------------------------
We are interested in relating the density matrices of an $N$-body state $\Gamma_N$ $$\gamma_N^{(k)}:=\operatorname{Tr}_{k+1\to N} \Gamma_N$$ to those of the state defined by taking as an upper symbol in the representation $$\label{eq:def representation}
\widetilde \gamma _N^{(k)} :=\int_{S\gH} d\mu_N(u) |u^{\otimes k} \rangle \langle u^{\otimes k}|,\quad\text{with}\quad d\mu_N (u) :=\dim \gH^N {{\ensuremath{\left\langle u^{\otimes N},\Gamma_N u^{\otimes N} \right\rangle}}}du.$$ The main estimate is the following.
\[thm:CKMR\]\
Let $\gH$ be a Hilbert space of dimension $d$. For every state $\Gamma_N$ on $\gH^N$, let $\widetilde \gamma _N^{(k)}$ be defined as in . Then for every $k=1,2,...,N$ we have $$\label{eq:error-CKMR-improved}
\boxed{\operatorname{Tr}_{\gH^k} \Big| \gamma_N ^{(k)} - \widetilde \gamma _N^{(k)} \Big| \le
\begin{cases}
2 &\text{if $N{\leqslant}2kd$},\\
\displaystyle\frac{2kd}{N-kd}&\text{if $N>2kd$}.\\
\end{cases}}$$
\[rm:Stormer\] As claimed before, such a statement implies the Størmer-Hudson-Moody theorem recalled in , for finite dimensional spaces. Indeed, starting from one can always extract convergent subsequences from reduced density matrices as in , and the convergence is strong since $\gH ^k$ is finite dimensional. The measure $\mu_N$ on the other hand is a probability over a compact set, so along a further subsequence $\mu_N \to \mu$, a probability, weakly as measures. Then, for any $V_k\in \gH ^k$ $$\langle V_k , \widetilde \gamma _N^{(k)} V_k \rangle = \int_{S\gH} \left| \langle u ^{\otimes k}, V_k \rangle \right| ^2 d\mu_N (u) \to \int_{S\gH} \left| \langle u ^{\otimes k}, V_k \rangle \right| ^2 d\mu (u)$$ since $u\mapsto \left| \langle u ^{\otimes k}, V_k \rangle \right| ^2$ is clearly a continuous function of $u$. In finite dimension this implies strong convergence of the reduced density matrices of ${\tilde{\Gamma}}_N$: $$\widetilde \gamma _N^{(k)} \to \int_{S \gH} |u ^{\otimes k} \rangle \langle u ^{\otimes k} |d\mu(u)$$ and thus we deduce from the estimate .
Recall that $\operatorname{Tr}\gamma_N ^{(k)} = \operatorname{Tr}\widetilde \gamma _N^{(k)} =1$, and therefore the bound $$\operatorname{Tr}_{\gH^k} \Big| \gamma_N ^{(k)} - \widetilde \gamma _N^{(k)} \Big| \le 2$$ is an obvious consequence of the triangle inequality for the trace-norm. On the other hand, we have $$\frac{\delta}{1-\delta}{\leqslant}2\delta, \quad \forall \: 0{\leqslant}\delta{\leqslant}\frac12,$$ and this yields the simpler bound $$\label{eq:error-CKMR}
\operatorname{Tr}_{\gH^k} \Big| \gamma_N ^{(k)} - \widetilde \gamma _N^{(k)} \Big| \le
\frac{4kd}{N}.$$ The slightly weaker inequality is the one that was proved by Christandl, König, Mitchison and Renner in [@ChrKonMitRen-07] (with a different convention for the trace norm, which is divided by $2$). In Section \[sec:proof CKMR\], we provide a proof of the estimate , which is very similar but not identical to that of in [@ChrKonMitRen-07]. While the proof in [@ChrKonMitRen-07] uses a “lifting up" argument (namely going from the $N$ particle space to the $(N+k)$ particle space), we rather use a “lifting down" argument. Related arguments and estimates may also be found in [@Chiribella-11], where the bound is improved to $2kd/N$.
\[rm:infinite dimension\] It is clear from the discussion is Subsection \[sec:motiv\] that the construction used to prove Theorem \[thm:CKMR\] works only for finite dimensional spaces. It is an open problem to obtain an estimate with a better $d$-dependence, in particular one that would apply to infinite dimensional systems.
As we mentioned at the end of Section \[sec:intro\], the construction used here is useful, when combined with localization methods, to provide a constructive proof of the infinite dimensional quantum de Finetti theorem of Størmer-Hudson-Moody mentioned in the Introduction, see [@LewNamRou-13 Section 2]. This proof, based on the finite dimensional constructions discussed here, has applications to infinite dimensional settings [@LewNamRou-14; @LewNamRou-14b].
We next state the explicit formula relating $\tilde\gamma^{(k)}_N$ to $(\gamma^{(\ell)}_N)_{\ell= 0 \ldots k}$. It was first derived by Chiribella [@Chiribella-11] (along with the associated Remark \[rm:non opt bound\]) and we will provide a different proof.
\[thm:CKMR-identity\]\
For the density matrices $\gamma_N^k$ and $\widetilde \gamma _N^{(k)}$ given above, we have $$\label{eq:CKMR exact}
\widetilde \gamma _N^{(k)} = {{N+k+d-1}\choose k}^{-1}\sum_{\ell=0}^{k} {N \choose \ell} \gamma_N^{(\ell)} \otimes _s {{\ensuremath {\mathds 1} }}_{\gH^{k-\ell}}$$ with the convention that $$\gamma_N^{(\ell)} \otimes _s {{\ensuremath {\mathds 1} }}_{\gH^{k-\ell}}= \frac{1}{\ell!\,(k-\ell)!}\sum_{\sigma\in S_k} (\gamma_N^\ell)_{\sigma(1),...,\sigma(\ell)} \otimes ({{\ensuremath {\mathds 1} }}_{\gH^{k-\ell}})_{{\sigma(\ell+1)},...,{\sigma(k)}}.$$
\[rm:non opt bound\]
A bound of the form can also be deduced easily from the formula . Indeed, from the representation in Theorem \[thm:CKMR-identity\] we may write $$\widetilde \gamma _N^{(k)} - \gamma_N ^{(k)} = ( C(d,k,N) - 1) \gamma_N ^{(k)} + B = -A + B \label{eq:estim CKMR}$$ where $$C(d,k,N) = \frac{(N+d-1)!}{(N+k+d-1)!} \frac{N!}{(N-k)!} < 1,$$ and $A,B$ are non-negative operators. Since from it is clear that $\operatorname{Tr}(-A+B)=0$, the triangle inequality gives $$\operatorname{Tr}\Big|\widetilde \gamma _N^{(k)} - \gamma_N ^{(k)} \Big| {\leqslant}\operatorname{Tr}A + \operatorname{Tr}B = 2 \operatorname{Tr}A = 2 (1- C(d,k,N)).$$ By the elementary inequality $$\begin{aligned}
C(d,k,N) &= \prod_{j=0} ^{k-1} \frac{N-j}{N+j+d}\ge \left( 1 - \frac{2k + d -2}{N + d + k - 1}\right) ^k {\geqslant}1 -k \frac{2k + d -2}{N + d + k - 1}\end{aligned}$$ we find that $$\label{eq:error-NLR}
\operatorname{Tr}\Big| \gamma_N ^{(k)} - \widetilde \gamma _N^{(k)} \Big| \le \frac{2 k(d+2k)}{N}.$$ The $k$-dependence in is not as good as that of , but at least for fixed $k$ and $d\ll N$ we recover the same dependence on $d/N$. Recall that fixed $k$ and large $N$ is the relevant limit for studying mean-field approximations of many-body systems. Only $k=2$ is needed for systems comprising two-body interactions.
The main idea behind our proof of Theorem \[thm:CKMR-identity\] is that the density matrices of $\widetilde \gamma _N^{(k)}$ turn out to be defined via an anti-Wick (anti-normal order of creation and annihilation operators) quantization, whereas the original density matrices $\gamma _N^{(k)}$ are of course defined by a Wick (normal order) quantization. Once this has been observed, the proof of , given in Section \[sec:Wick\], consists in using the Canonical Commutation Relation repeatedly, with the upshot that, since there are many particles but few available degrees of freedom ($d\ll N$), annihilation and creation almost commute: their commutators are of order $1$ whereas the operators themselves should roughly be of order $\sqrt{N/d}$. The connection between quantum de Finetti theorems for bosonic states and the Wick versus anti-Wick quantization issue was inspired to us by the approach of Ammari and Nier [@Ammari-HDR; @AmmNie-08; @AmmNie-09; @AmmNie-11]. We also remark that, independently of our work, Lieb and Solovej [@LieSol-13] use a formula very similar to in their investigation of the classical entropy of quantum states.
Proof of the main estimate, Theorem \[thm:CKMR\] {#sec:proof CKMR}
================================================
In this section we give the proof of the bound , following ideas from [@ChrKonMitRen-07]. For simplicity of writing, we only deal with pure states $\Gamma_N=|\Psi_N\rangle\langle\Psi_N|$, as it is clear that the general case follows from the triangle inequality.
Denote $P_u:=|u \rangle \langle u|$ for every $u\in S\gH$. Note that $P_u^{\otimes k}=|u^{\otimes k} \rangle \langle u^{\otimes k}|$ for every $k\in \mathbb N$. Thus for every bounded operator $A$ on $\gH^k$, using Schur’s formula (\[eq:Schur\]) we find that $$\operatorname{Tr}[A \gamma_N^{(k)}]= \langle \Psi_N, (A \otimes {{\ensuremath {\mathds 1} }}_{\gH ^{N-k}}) \Psi_N\rangle = \dim \gH^N \int_{S\gH}\langle \Psi_N, P_u ^{\otimes N} (A \otimes {{\ensuremath {\mathds 1} }}_{\gH ^{N-k}}) \Psi_N\rangle du .$$ On the other hand, by the definition of $\widetilde\gamma_N^{(k)}$, $$\begin{aligned}
\operatorname{Tr}[A \widetilde\gamma_N^{(k)}] &= \dim \gH^N \int_{S\gH} \langle Au^{\otimes k}, u^{\otimes k} \rangle . |\langle \Psi_N, u ^{\otimes N} \rangle|^2 du \\
&=\dim \gH^N \int_{S\gH} \big\langle (A\otimes {{\ensuremath {\mathds 1} }}_{\gH ^{N-k}}) u^{\otimes N}, u^{\otimes N} \big\rangle . \big|\langle \Psi_N, u ^{\otimes N} \rangle\big|^2 du \\
&=\dim \gH^N \int_{S\gH} \big\langle \Psi_N, P_u^{\otimes N} (A\otimes {{\ensuremath {\mathds 1} }}_{\gH ^{N-k}}) P_u^{\otimes N} \Psi_N \big\rangle du .\end{aligned}$$ Thus $$\begin{aligned}
\label{eq:diff-gammaN}
\operatorname{Tr}[A (\gamma_N^{(k)}-\widetilde\gamma_N^{(k)})] =
\dim \gH^N \int_{S\gH} \big\langle \Psi_N, P_u^{\otimes N} (A \otimes {{\ensuremath {\mathds 1} }}_{\gH ^{N-k}}) ({{\ensuremath {\mathds 1} }}_{\gH ^{N}}-P_u^{\otimes N}) \Psi_N \big\rangle du .\end{aligned}$$ Using $$P_u^{\otimes N} (A \otimes {{\ensuremath {\mathds 1} }}_{\gH ^{N-k}}) ({{\ensuremath {\mathds 1} }}_{\gH ^{N}}-P_u^{\otimes N}) = P_u^{\otimes k} A({{\ensuremath {\mathds 1} }}_{\gH ^{k}}-P_u^{\otimes k}) \otimes P_u^{\otimes N-k}
$$ we find that $$\begin{gathered}
\int_{S \gH } P_u^{\otimes N} (A \otimes {{\ensuremath {\mathds 1} }}_{\gH ^{N-k}}) ({{\ensuremath {\mathds 1} }}_{\gH ^{N}}-P_u^{\otimes N}) du
\\ = A\otimes {{\ensuremath {\mathds 1} }}_{\gH ^{N-k}} \int_{S\gH} \left( {{\ensuremath {\mathds 1} }}_{\gH ^{k}}- P_u ^{\otimes k}\right) \otimes P_u ^{\otimes N-k}du
- \int_{S\gH} \left( {{\ensuremath {\mathds 1} }}_{\gH ^{k}}- P_u ^{\otimes k}\right) A \left( {{\ensuremath {\mathds 1} }}_{\gH ^{k}}- P_u ^{\otimes k}\right) \otimes P_u ^{\otimes N-k} du.\end{gathered}$$ Then, using Schur’s formula in $\gH ^N$ and $\gH ^{N-k}$ we have $$\label{eq:rewrite dif}
\int_{S\gH} \left( {{\ensuremath {\mathds 1} }}_{\gH ^{k}}- P_u ^{\otimes k}\right) \otimes P_u ^{\otimes N-k} du = \left( \left(\dim \gH ^{N-k} \right) ^{-1} - \left(\dim \gH ^N \right) ^{-1}\right){{\ensuremath {\mathds 1} }}_{\gH ^{N}}$$ and since $$\left( {{\ensuremath {\mathds 1} }}_{\gH ^{k}}- P_u ^{\otimes k}\right) A \left( {{\ensuremath {\mathds 1} }}_{\gH ^{k}}- P_u ^{\otimes k}\right) {\leqslant}\Vert A \Vert \left( {{\ensuremath {\mathds 1} }}_{\gH ^{k}}- P_u ^{\otimes k}\right)$$ we conclude from and that $$\begin{aligned}
\Big| \operatorname{Tr}[A (\gamma_N^{(k)}-\widetilde\gamma_N^{(k)})] \Big| \le 2 \|A\| \Big( \frac{\dim \gH^N}{\dim \gH^{N-k}} -1\Big)\end{aligned}$$ for every bounded operator $A$ on $\gH^k$. This implies the upper bound $$\begin{aligned}
\label{eq:trace-bound-cN}
\operatorname{Tr}\Big|\gamma_N^{(k)}-\widetilde\gamma_N^{(k)} \Big| \le 2 \Big( \frac{\dim \gH^N}{\dim \gH^{N-k}} -1\Big).\end{aligned}$$ Finally, due to Bernoulli’s inequality we have $$\begin{aligned}
\frac{\dim \gH^{N-k}}{\dim \gH^{N}} &= \frac{{{N+d-k-1} \choose {d-1}}}{{ {N+d-1} \choose {d-1}}}= \frac{(N-k+1)...(N-k+d-1)}{(N+1)...(N+d-1)} \\
& = \Big( 1- \frac{k}{N+1} \Big) ... \Big( 1- \frac{k}{N+d-1}\Big) \ge \Big(1- \frac{k}{N}\Big)^d \ge 1- \frac{dk}{N},\end{aligned}$$ which implies that, in the case $dk < N$ $$\begin{aligned}
\frac{\dim \gH^N}{\dim \gH^{N-k}} -1 \le \Big( 1- \frac{dk}{N} \Big)^{-1}-1 = \frac{dk}{N-dk}.\end{aligned}$$ The desired estimate (\[eq:error-CKMR-improved\]) then follows immediately from (\[eq:trace-bound-cN\]).
Proof of the explicit formula, Theorem \[thm:CKMR-identity\] {#sec:Wick}
============================================================
Our proof of Theorem \[thm:CKMR-identity\] is based on the fact that $\widetilde \gamma _N^{(k)}$ turns out to be linked to an anti-Wick representation, while $\gamma _N^{(k)}$ is defined via a standard Wick representation. The difference between $\gamma _N^{(k)}$ and $ \widetilde \gamma _N^{(k)}$ can then be computed by comparing Wick and anti-Wick representation, that is, by looking at the difference between normal ordered and anti-normal ordered polynomials in annihilation and creation operators.
Recall that for every $f_{k}\in \gH$, we can define the creation operator $a^*(f_{k}): \gH^{k-1} \to \gH^{k}$ by $${a^*}({f_{k}})\left( {\sum\limits_{\sigma \in {S_{k-1}}} {{f_{\sigma (1)}} } \otimes ... \otimes {f_{\sigma (k-1)}}} \right) = (k) ^{-1/2} \sum\limits_{\sigma \in {S_{k}}} {{f_{\sigma (1)}} } \otimes ... \otimes {f_{\sigma (k)}}$$ The annihilation operator $a(f): \gH^{k+1} \to \gH^{k}$ is the adjoint of $a^*(f)$, given by $$a(f) \left( {\sum\limits_{\sigma \in {S_{k+1}}} {{f_{\sigma (1)}} } \otimes ... \otimes {f_{\sigma (k+1)}}} \right) = (k+1) ^{1/2} \sum\limits_{\sigma \in {S_{k+1}}} \left\langle f,f_{\sigma(1)} \right\rangle {{f_{\sigma (2)}} } \otimes ... \otimes {f_{\sigma (k)}}$$ for all $f,f_1,...,f_{k}$ in $\gH$. These operators satisfy the [*canonical commutation relations*]{} $$\label{eq:CCR}
[a(f),a(g)]=0,\quad[a^*(f),a^*(g)]=0,\quad [a(f),a^*(g)]= \langle f,g \rangle_{\gH}.$$
We shall need two lemmas. The first one says that any bosonic $k$-body density matrix can be completely determined by its expectation against Hartree states $u^{\otimes k}$.
\[le:uk-g-uk=0\]\
If a trace class self-adjoint operator $\gamma^{(k)}$ on $\gH^k$ satisfies \[eq:uk-g-uk=0\] u\^[k]{}, \^[(k)]{} u\^[k]{} =0 u, then $\gamma^{(k)} \equiv 0$.
In connection with the discussion in Section \[sec:motiv\], this result says that the state is uniquely determined by its lower symbol. This is a well-known fact even in more abstract settings [@Klauder-64; @Simon-80]. For the reader’s convenience, a standard proof is given in Appendix \[app:Hartree-determine\].
In the second lemma, we compare normal and anti-normal ordering of creation and annihilation operators. Thanks to Lemma \[le:uk-g-uk=0\] we need only do this for a single mode $v\in \gH$.
\[le:Wick A Wick\]\
Let $v\in S\gH$ with associated creation and annihilation operators $a ^* (v)$ and $a(v)$. Then $$\label{eq:Wick A Wick}
a(v) ^n a ^*(v) ^n = \sum_{k=0} ^n \binom{n}{k} \frac{n!}{k!} a ^*(v) ^k a (v) ^k \mbox{ for any } n\in \NN.$$
Recall that the $n$-th Laguerre polynomial is given by $$L_n (x) = \sum_{k=0} ^n \binom{n}{k} \frac{(-1) ^k}{k!} x ^k$$ and these polynomials satisfy the relation $$(n+1) L_{n+1} (x) = (2n +1) L_n (x) - x L_n (x) - n L_{n-1} (x).$$ The identity is equivalent to $$\label{eq:Wick A Wick proof}
a(v) ^n a ^*(v) ^n = \sum_{k=0} ^n c_{n,k} \, a^*(v) ^k a (v) ^k$$ where the $c_{n,k}$’s are the coefficients of the polynomial $$\tilde{L}_n (x) := n! \, L_n (-x).$$
We prove by induction on $n$. Note first that the CCR immediately gives for $n=1$, while it is easy to see that $$\label{eq:Wick A Wick 2}
a(v) ^2 a ^*(v) ^2 = a ^*(v) ^2 a (v) ^2 + 4 a ^* (v) a (v) + 2$$ by a repeated use of the CCR. This is for $n=2$, so we simply need an induction formula giving $a(v) ^{n+1} a ^*(v) ^{n+1}$ as a function of $a(v) ^n a ^*(v) ^n$ and $a(v) ^{n-1} a ^*(v) ^{n-1}$. We claim that $$\label{eq:induction}
a(v) ^{n+1} a ^*(v) ^{n+1} = a^* (v) a(v) ^n a ^*(v) ^n a(v) + (2n+1) a(v) ^n a ^*(v) ^n - n ^2 a(v) ^{n-1} a ^*(v) ^{n-1}.$$ Note the order of creation and annihilation operators in the first term: knowing a normal ordered representation of $a(v) ^n a ^*(v) ^n$ we can deduce a normal ordered representation for this term. Since the modified Laguerre polynomials satisfy $$\tilde{L}_{n+1} (x) = (2n +1) \tilde{L}_n (x) + x \tilde{L}_n (x) - n ^2 \tilde{L}_{n-1} (x)$$ it is then clear that follows from .
We now prove our claim . We use the relations $$\begin{aligned}
\label{eq:CCR n}
a(v) a ^* (v) ^n &= a ^* (v) ^n a(v) + n a ^* (v) ^{n-1} \nonumber \\
a(v) ^n a ^* (v) &= a ^* (v) a(v) ^n + n a (v) ^{n-1}\end{aligned}$$ that follow from the CCR. Then $$\begin{aligned}
a^* (v) a(v) ^n a ^*(v) ^n a(v) &= a(v) ^n a ^* (v) ^{n+1} a (v) - n a (v) ^{n-1} a ^* (v) ^n a(v)\\
& = a(v) ^{n+1} a ^*(v) ^{n+1} - (n + 1) a(v) ^n a ^*(v) ^n \\
&- n a(v) ^n a ^*(v) ^n + n ^2 a(v) ^{n-1} a ^*(v) ^{n-1},\end{aligned}$$ which is .
Now we are able to give the
Clearly it is sufficient to consider only the case of a pure state $|\Psi_N \rangle \langle \Psi_N |$. Using Lemma \[le:uk-g-uk=0\], the $k$-particle density matrix of $\Psi_N$ is uniquely defined by v\^[k]{}, \_N\^[(k)]{} v\^[k]{} = \_N, a\^\*(v)\^k a(v)\^k \_N for all $v\in \gH$ such that ${\left\lVert v \right\rVert}=1$. In contrast, the $ \widetilde \gamma _N^{(k)}$ satisfies a similar formula but with the order of the creation and annihilation operators reversed: $$\begin{aligned}
\langle v^{\otimes k}, \widetilde \gamma _N^{(k)} v^{\otimes k} \rangle &= \dim \gH^N \int_{S\gH} du |\langle u^{\otimes N}, \Psi_N \rangle|^2 | \langle u^{\otimes k},v^{\otimes k} \rangle|^2 \\
&= \dim \gH^N \int_{S\gH} du |\langle u^{\otimes (N +k)}, v^{\otimes k} \otimes \Psi_N \rangle|^2 \\
&= \frac{N!}{(N+k)!} \dim \gH^N \int_{S\gH} du |\langle u^{\otimes (N +k)}, a^*(v)^{k} \Psi_N \rangle|^2 \\
&= \frac{N!}{(N+k)!}\frac{\dim \gH^N}{\dim \gH^{N+k}} \| a^*(v)^{k} \Psi_N \|^2 \\
&= \frac{(N+d-1)!}{(N+k+d-1)!} \langle \Psi_N, a(v)^k a^*(v)^{k} \Psi_N \rangle\end{aligned}$$ where we used Schur’s formula for the fourth equality. There only remains to use Lemma \[le:Wick A Wick\]: $$\begin{aligned}
\frac{(N+k+d-1)!}{(N+d-1)!} \langle v^{\otimes k}, \widetilde \gamma _N^{(k)} v^{\otimes k} \rangle &= \langle \Psi_N, a(v)^k a^*(v)^{k} \Psi_N \rangle \\
&= \sum_{\ell=0}^{k} \binom{k}{\ell} \frac{k!}{\ell!} \langle \Psi_N, a^*(v)^\ell a(v)^\ell \Psi_N \rangle\\
& = \sum_{\ell=0}^{k} {N \choose \ell} \binom{k}{\ell} k! \langle v^{\otimes \ell}, \gamma_N^{(\ell)} v^{\otimes \ell}\rangle
\end{aligned}$$ and then follows from Lemma \[le:uk-g-uk=0\].
Expectations in Hartree vectors determine the state {#app:Hartree-determine}
===================================================
In the following we use the symmetric tensor product $$\Psi_k\otimes_s\Psi_\ell(x_1,...,x_{k})=\frac{1}{\sqrt{\ell!(k-\ell)!k!}}\sum_{\sigma\in {S}_{k}}\Psi_\ell(x_{\sigma(1)},...,x_{\sigma(\ell)})\Psi_{k-\ell}(x_{\sigma(\ell+1)},...,x_{\sigma(k)})$$ of two functions $\Psi_\ell \in \gH^{\ell}$ and $\Psi_{k-\ell}\in\gH^{k-\ell}$. Note that for every $f\in \gH$, $$f\otimes_s \Psi_\ell = a^*(f) \Psi_\ell.$$
By replacing $u$ by $u+tv$ in (\[eq:uk-g-uk=0\]) and taking the derivative with respect to $t$, we obtain v \_s u\^[(k-1)]{}, \^[(k)]{} v \_s u\^[(k-1)]{} =0 for all $u,v\in \gH$. Taking $v$ in the form $v=v_1\pm \widetilde v_1$ and then $v = v_1 \pm i \widetilde v_1$ we deduce v\_1 \_s u\^[(k-1)]{}, \^[(k)]{} v\_1 \_s u\^[(k-1)]{} =0 for all $u,v_1,\widetilde v_1\in \gH$. We may then again replace $u$ by $u+tv$ in the above, and take the derivative with respect to $t$. Repeating this process $k$ times we conclude that v\_1 \_s v\_2 \_s …\_s v\_k, \^[(k)]{} v\_1 \_s v\_2 \_s …\_s v\_k =0 for all $v_j, \widetilde v_j \in \gH$. Since vectors of the forms $v_1 \otimes_s v_2 \otimes_s \ldots \otimes_s v_k$ form a complete basis for $\gH^k$, we conclude that $\gamma^{(k)}\equiv 0$.
[10]{}
, [*Systmes hamiltoniens en théorie quantique des champs : dynamique asymptotique et limite classique.*]{} Habilitation [à]{} Diriger des Recherches, University of Rennes I, February 2013.
, [*Mean field limit for bosons and infinite dimensional phase-space analysis*]{}, Annales Henri Poincaré, 9 (2008), pp. 1503–1574. 10.1007/s00023-008-0393-5.
, [*Mean field limit for bosons and propagation of [W]{}igner measures*]{}, J. Math. Phys., 50 (2009).
, [*[Mean field propagation of Wigner measures and BBGKY hierarchies for general bosonic states]{}*]{}, J. Math. Pures Appl., 95 (2011), pp. 585–626.
, [*Quantum de [F]{}inetti theorems under local measurements with applications*]{}, Proc. of the 45th ACM Symposium on theory of computing (STOC 2013), pp. 861-870, (2013), pp. 861–870.
, [*Unknown quantum states: the quantum de [F]{}inetti representation*]{}, J. Math. Phys., 43 (2002), p. 4535.
, [*On quantum estimation, quantum cloning and finite quantum de [F]{}inetti theorems*]{}, in Theory of Quantum Computation, Communication, and Cryptography, vol. 6519 of Lecture Notes in Computer Science, Springer, 2011.
, [ *One-and-a-half quantum de [F]{}inetti theorems*]{}, Comm. Math. Phys., 273 (2007), pp. 473–498.
, [*Finite de finetti theorem for conditional probability distributions describing physical theories*]{}, J. Math. Phys., 50 (2009), p. 042104.
, [*[de Finetti Representation Theorem for Infinite-Dimensional Quantum Systems and Applications to Quantum Cryptography]{}*]{}, Phys. Rev. Lett., 102 (2009), p. 110504.
, [*Funzione caratteristica di un fenomeno aleatorio*]{}. Atti della R. Accademia Nazionale dei Lincei, 1931. Ser. 6, Memorie, Classe di Scienze Fisiche, Matematiche e Naturali.
height 2pt depth -1.6pt width 23pt, [*La prévision, ses lois logiques, ses sources subjectives*]{}, Annales de l’IHP, 7 (1937), pp. 1–68.
, [*Finite exchangeable sequences*]{}, Ann. Probab., 8 (1980), pp. 745–764.
, [*Finite size mean-field models*]{}, J. Phys. A, 39 (2006), pp. 13843–13860.
, [*The church of the symmetric subspace*]{}, preprint arXiv, (2013).
, [*Symmetric measures on [C]{}artesian products*]{}, Trans. Amer. Math. Soc., 80 (1955), pp. 470–501.
, [*Locally normal symmetric states and an analogue of de [F]{}inetti’s theorem*]{}, Z. Wahrscheinlichkeitstheorie und Verw. Gebiete, 33 (1975/76), pp. 343–351.
, [*Continuous-representation theory. iii. on functional quantization of classical systems*]{}, J. Math. Phys., 5 (1964), pp. 177–187.
, [*Coherent States, Applications in Physics and Mathematical Physics*]{}, World Scientific, Singapore, 1985.
, [*A de finetti representation for finite symmetric quantum states*]{}, J. Math. Phys., 46 (2005), p. 122108.
, [*Geometric methods for nonlinear many-body quantum systems*]{}, J. Funct. Anal., 260 (2011), pp. 3535–3595.
, [*Derivation of [H]{}artree’s theory for generic mean-field [B]{}ose systems*]{}, Advances in Mathematics, 254 (2014).
height 2pt depth -1.6pt width 23pt, [*Derivation of nonlinear [G]{}ibbs measures from many-body quantum mechanics*]{}, in preparation, (2014).
height 2pt depth -1.6pt width 23pt, [*The mean-field approximation and the non-linear [S]{}chrödinger functional for trapped bose gases*]{}, preprint arXiv, (2014).
, [*The classical limit of quantum spin systems*]{}, Comm. Math. Phys., 31 (1973), pp. 327–340.
, [*The mathematics of the [B]{}ose gas and its condensation*]{}, Oberwolfach [S]{}eminars, Birkh[ä]{}user, 2005.
, [*in preparation*]{}.
, [*Asymptotics of [V]{}aradhan-type and the [G]{}ibbs variational principle*]{}, Comm. Math. Phys., 121 (1989), pp. 271–282.
, [*Quantum statistical mechanics of general mean field systems*]{}, Helv. Phys. Acta, 62 (1989), pp. 980–1003.
, [*Symmetry of large physical systems implies independence of subsystems*]{}, Nature Physics, 3 (2007), pp. 645–649.
, [*The classical limit of quantum partition functions*]{}, Comm. Math. Phys., 71 (1980), pp. 247–276.
, [*Symmetric states of infinite tensor products of [$C^{\ast} $]{}-algebras*]{}, J. Functional Analysis, 3 (1969), pp. 48–68.
, [*Large deviations and mean-field quantum systems*]{}, in Quantum probability & related topics, QP-PQ, VII, World Sci. Publ., River Edge, NJ, 1992, pp. 349–381.
, [*Coherent states : theory and some applications*]{}, Rev. Mod. Phys., 62 (1990), p. 867.
[^1]: That is, in the density matrix $|\Psi\rangle\langle \Psi|$. It is of course quadratic in the wave-function $\Psi$.
[^2]: Note the convention in [@ChrKonMitRen-07] where the trace norm is divided by $2$.
[^3]: Strictly speaking the upper symbol designates the density of $d{\mu ^{\rm up}}$ with respect to $\dim \gH^N\: du$. We shall abuse language throughout this discussion.
[^4]: The “lower symbol of the upper symbol”.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The inelastic scattering of electrons on weakly-bound nuclei is studied with a simple model based on the long range behavior of the bound state wavefunction and on the effective-range expansion for the continuum wavefunctions. Three mechanisms have been considered: (a) dissociation of halo nuclei by high energy electrons, (b) dissociation by electrons present in a fixed target, and (c) Coulomb dissociation. It is shown that the properties of halo nuclei can be studied in electron-radioactive beam colliders using the electro-disintegration process. A comparison with fixed-target experiments is also performed.'
author:
- 'C.A. Bertulani'
title: Electron scattering on halo nuclei
---
**Introduction**
================
A study of properties of weakly-bound neutron-rich, or halo nuclei, has been carried out intensively worldwide during the last decades [@CS00]. Because of their short beta-decay lifetimes, halo nuclei are often studied in fragmentation facilities, where they are produced in-flight. The probes are hadronic, usually stable nuclear targets. Typically, one uses Coulomb dissociation, stripping, elastic scattering, etc. [@BCH93] as nuclear structure probes. Such studies are complicated because the reaction mechanisms are not as well understood as with stable nuclear projectiles. The use of electromagnetic probes, e.g. electron scattering, is thus highly desirable. In fact, new experimental facilities for electron-scattering on unstable nuclear beams are under construction [@GSI]. An accurate determination of charge distributions in exotic nuclei can be obtained with electrons using inverse kinematics in a electron-nucleus collider mode [@GSI]. Electronic excitation, or dissociation, of nuclear beams can also be exploited for a deeper understanding of their structure.
It is the aim of this work to explore basic results of electron scattering on the simplest of all nuclear halo structures, namely, a one-neutron halo system. The physics mechanisms and the conditions for the realization of electron scattering experiments are assessed. Such study has also an impact in nuclear astrophysics as it allows to deduce what are the lowest binding energies of halo nuclei possible in stellar environments, where free electrons are available.
A high energy beam of weakly-bound neutron-rich nuclei dissociates as it penetrates a target due to the interaction with the atomic electrons. Since a heavy element target, e.g., $^{208}$Pb, contains almost 100 electrons, the dissociation cross sections are large, assuming that each electron in the atom scatters independently on the projectile. Moreover, due to the atomic orbital motion, the innermost electrons have large relative energy with the incoming nucleus, increasing the dissociation probability. This process is of crucial importance in designing experiments aiming at studying properties of halo nuclei with the Coulomb dissociation method.
The dissociation of neutron-rich nuclei, with small neutron separation energies, in stars can impose stringent limits on the stellar scenario where these nuclei play a role. For example, if the r-process proceeds partially out of equilibrium, the neutron radiative capture cross sections would have to be large enough to match the electron dissociation cross sections, with the appropriate neutron and electron density weights.
Electron scattering on neutron halo nuclei
==========================================
I will consider the process $e+a\longrightarrow e^{\prime}+b+c$ at small momentum transfers, $\mathbf{q=}\left( \mathbf{p}^{\prime}\mathbf{-p}\right)
/\hbar$, such that $qR\ll1$ ($R$ is the nuclear size). For simplicity, particle $b$ is taken as a neutron and $c$ as a core (inert) nucleus. The results obtained here are general and can be easily extended to the case of two-neutron halos.
The differential cross section for this process is given by [@EG88]$$\frac{d\sigma_{e}}{d\Omega}=\frac{2e^{2}}{\left( \hbar c\right) ^{4}}\left(
\frac{p^{\prime}}{p}\right) \frac{2J_{f}+1}{2J_{i}+1}\frac{EE^{\prime}+c^{2}\mathbf{p\cdot p}^{\prime}+m_{e}^{2}c^{4}}{q^{4}}\left\vert \rho
_{fi}\left( \mathbf{q}\right) \right\vert ^{2}N_{f}\ , \label{elect_scat}$$ where $E(E^{\prime})$ and $\mathbf{p}(\mathbf{p}^{\prime})$ are the initial (final) energy and momentum of the electron, respectively. $J_{i}(J_{f})$ is the initial (final) nuclear spin, and $N_{f}$ is the density of final states of the nucleus. The nuclear form factor $\rho_{fi}\left( q\right) $ is given by$$\rho_{fi}\left( \mathbf{q}\right) =\int\rho_{fi}\left( \mathbf{r}\right)
\ e^{i\mathbf{q\cdot r}}\ d^{3}r\ , \label{form_fact}$$ where $\rho_{fi}\left( \mathbf{r}\right) =\psi_{f}^{\ast}\
\psi_{i}$ is the nuclear charge transition density, with $\psi_{i}$ $(\psi_{f})$ equal to the initial (final) nuclear wavefunction. The cross section given by eq. \[elect\_scat\] only includes longitudinal (also called Coulomb) excitations, dominant at low energy transfers [@Sch54; @Wei64; @Sch66].
Eq. \[elect\_scat\] is based on the first Born approximation. It gives good results for light nuclei (e.g. $^{12}$C) and high-energy electrons. For large-Z nuclei the agreement with experiments is only of a qualitative nature. The effects of the distortion of the electron waves have been studied by many authors (see, e.g. ref. [@MF48; @Yenn53]). For a rough estimate of this effect, I follow ref. [@Cut67]. For transition densities peaked at the nuclear surface with radius $R_{0}$, the correction due to Coulomb distortion is approximately given by$$Q=\frac{d\sigma_{\mathrm{Born}}/d\Omega}{d\sigma_{\mathrm{Corrected}}/d\Omega
}\simeq\frac{1}{1+\beta Ze^{2}/\hbar c},\label{Coulcorr}$$ with$$\begin{aligned}
\beta & =\frac{120}{x^{2}}\left\{ -\frac{1}{160x^{3}}\left[ 1+\frac{3}{2}\cos\left( 2x\right) +3x\sin\left( 2x\right) +\frac{x^{2}}{3}\left(
4+5\cos\left( 2x\right) \right) +\frac{10}{3}x^{3}\sin\left( 2x\right)
\right] \right. \nonumber\\
& \left. +\frac{x}{60}\left[ \frac{9}{4}-\cos\left( 2x\right) \right]
+\frac{x^{2}}{60}\left[ \pi-2\operatorname{Si}\left( 2x\right) \right]
+n_{1}\left( 2x\right) \left[ \frac{1}{16x}+\frac{x}{40}\right] \right\}
,\label{Coulfac}$$ where $x=pR_{0}/\hbar$, $\operatorname{Si}$ is the sine integral, $\operatorname{Si}(x)=\int_{0}^{\infty}dt\sin t/t$, and $n_{i}\left(
x\right) $ is the spherical Bessel function of the second kind. The above result is valid for monopole ($l=0$) transitions. Corresponding expressions for higher order transitions are found in ref. [@Cut67].
Table \[tab:Coulcorr\] shows the correction due to Coulomb distortion, eq. \[Coulcorr\], for $^{11}$Li and $^{19}$C targets and several electron kinetic energies $K_{e}$. One sees that below $K_{e}=100$ MeV it is important to account for Coulomb distortion of the electronic waves. A radius $R_{0}=3.5$ fm was assumed for both nuclei. The Coulomb distortion correction decreases approximately linearly with $R_{0}$.
An additional correction, due to nuclear recoil [@EG88], changes eq. \[elect\_scat\] by a factor $f_{\mathrm{rec}}\simeq1+(2E_{x}/Mc^{2})\sin(\theta/2)$, where $E_{x}$ is the excitation energy, $M$ is the nuclear mass, and $\theta$ is the electron scattering angle. For the dissociation of weakly-bound nuclei, $E_{x}\ll Mc^{2}$ and this correction is much less relevant than the distortion of the electronic waves. I will neglect the Coulomb distortion and recoil effects from here on, bearing in mind that they should be taken into account in a more precise calculation.
\[c\][|l|l|l|]{}$K_{e}\ $\[MeV\] & Q$\left( ^{11}\mathrm{Li}\right) $ & Q$\left(
^{19}\mathrm{C}\right) $\
0.1 & 0.879 & 0.784\
1 & 0.880 & 0.786\
10 & 0.887 & 0.797\
10$^{2}$ & 0.949 & 0.903\
10$^{3}$ & 0.994 & 0.989\
In a simplified model for the halo nucleus the radial parts of the initial and final wavefunctions are represented by single-particle states of the form$$u_{i}(r)=A_{i}\ h_{l_{i}}(i\eta r)\ ,\ \ \ \ \ \ \ \ \ \ \ \ u_{f}(r)=\cos(\delta_{l_{f}})\ j_{l_{f}}(kr)-\sin(\delta_{l_{f}})\ n_{l_{f}}(kr)\label{wav_func}$$ where $\eta$ is related to the neutron separation energy $S_{n}=\hbar^{2}\eta^{2}/2\mu$.$\ h_{l_{i}}(i\eta r)$ represents the large distance behavior of the bound state wavefunction, $\mu$ is the reduced mass of the neutron + core system and $\hbar k$ their relative momentum in the final state. $h_{l_{i}}$, $j_{l_{f}}$, and $n_{l_{f}}$ are the spherical Hankel, Bessel, and Neumann functions, respectively. $A_{i}\ $ is the ground state asymptotic normalization coefficient, which includes the normalization of the neutron single-particle wavefunction, and a spectroscopic factor which accounts for the many-body aspects. This single-particle picture has been used previously to study Coulomb excitation of halo nuclei with success [@BB86; @BS92; @Ots94; @MOI95; @KB96; @TB04].
The constant $A_{i}$ (with spectroscopic factor equal to the unity) is used to normalize the bound state wavefunction, and corrects for the nuclear interaction range, $r_{0}$. In the case of an s-wave ground state ($l_{i}=0$), one has [@HJ87] $A_{i}=\exp\left( \eta
r_{0}\right) \sqrt{\eta /2\pi\left( 1+\eta r_{0}\right) }$. For weakly-bound nuclei, $1/\eta\gg r_{0}$ and $A_{i}\simeq\sqrt{\eta/2\pi}$. The ground state wavefunction entering the transition density integral, eq. \[form\_fact\], is well represented by the Hankel function $h_{l_{i}}(i\eta r)$. Note that the wavefunctions in eq. \[wav\_func\] are not orthonormal. However, the transition density matrix elements of relevance for Coulomb excitation (and similarly for electron scattering) are dominated by the outside region ($r>R$) [@Ots94; @MOI95]. Far from a resonance, the continuum wavefunction $u_{f}(r)$ is small inside the nuclear radius. Its asymptotic dependence is well described by eq. \[wav\_func\]. Using eq. \[wav\_func\] the form factor in eq. \[form\_fact\] can be calculated analytically by expanding $\ e^{i\mathbf{q\cdot r}}$ into multipoles. The results will depend on the parameters $S_{n}$, $R$, and $\delta_{l_{f}}$. To eliminate the dependence on $R$, the lower limit of the radial integral in eq. \[form\_fact\] is extended to $r=0$. The results for an s-wave ground state and the lowest order continuum angular momenta ($l_{f}=0,1,2$) are particularly simple. They are:$$\begin{aligned}
\rho_{fi}^{(0)}\left( q\right) & =\frac{e_{\mathrm{eff}}^{(0)}\ \pi A_{i}}{qk}\left\{ L-\frac{2k}{-1/a_{0}+r_{0}k^{2}/2}M\right\} \nonumber\\
\rho_{fi}^{(1)}\left( q\right) & =\frac{e_{\mathrm{eff}}^{(1)}\ i\pi
A_{i}}{q^{2}k^{2}}\left\{ \frac{\eta^{2}+k^{2}+q^{2}}{2}L-2qk-\frac{k^{3}}{-1/a_{1}+r_{1}k^{2}/2}\left[ 2\eta q+\left( \eta^{2}+k^{2}+q^{2}\right)
M\right] \right\} \nonumber\\
\rho_{fi}^{(2)}\left( q\right) & =\frac{e_{\mathrm{eff}}^{(2)}\ \pi A_{i}}{4k^{3}q^{3}}\left\{ 8kq\left( \eta^{2}+k^{2}+q^{2}\right) -\frac
{3k^{4}+3\left( \eta^{2}+q^{2}\right) ^{2}+2k^{2}\left( 3\eta^{2}+q^{2}\right) }{2}\ L\right. \nonumber\\
& +\left. \frac{k^{5}}{-1/a_{2}+r_{2}k^{2}/2}\left[ 6\eta\left( k^{2}+\eta^{2}\right) q+10\eta q^{3}+\left( 3k^{4}+3\left( \eta^{2}+q^{2}\right) ^{2}+2k^{2}\left( 3\eta^{2}+q^{2}\right) \right) M\right]
\right\} {}{}{},\label{effr}$$ where$$L=\ln\left( \frac{\eta^{2}+\left( k+q\right) ^{2}}{\eta^{2}+\left(
k-q\right) ^{2}}\right) ,\ \ \ \ \ \ \ \ \text{and}\ \ \ M=\tan^{-1}\left(
\frac{k-q}{\eta}\right) -\tan^{-1}\left( \frac{k+q}{\eta}\right) \ .$$ In these equations $e_{\mathrm{eff}}^{(\lambda)}$ $=eZ\left( -1/A\right)
^{\lambda}$ is the neutron-core effective charge which depends on the transition multipolarity $\lambda$ ($\lambda=l_{f}$ for $l_{i}=0$). The effective range approximation $k^{2l+1}\cot\delta_{l}=-1/a_{l}+r_{l}k^{2}/2$ has been used, where the parameters $a_{l}$ and $r_{l}$ are the scattering length and the effective range, respectively. Notice that only for $l=0$ the scattering length and effective range have dimensions of length.
The $l=0$ form factor has a large sensitivity to the orthogonality of the wavefunctions. If one assumes a zero-range potential for the neutron-core interaction, the scattering wavefunction, orthogonal to the bound-state wavefunction, is given by $\psi_{\mathbf{k}}^{(+)}(\mathbf{r})=\exp (i\mathbf{k\cdot
r})-\exp(ikr)/[(\eta+ik)r]$. The s-wave scattering length is then just $a_{0}=1/\eta$. Using this value, together with $r_{0}=0$, in the equation for $\rho_{fi}^{(0)}\left( q\right) $ leads to a large cancelation between the first and second terms. The $l=1,2$ form factors are also very sensitive to the scattering lengths and effective ranges. For example using $a_{1}\simeq5$ fm$^{3}$ and $r_{1}=0$ fm$^{-1}$ reduces the magnitude of $\rho_{fi}^{(1)}\left(
q\right) $ by 10% for scattering at forward angles. These results show that it is very important to include the correct energy dependence of the phase-shifts to obtain an accurate description of electron scattering off halo nuclei [@TB04].
In what follows, I will use $e_{\mathrm{eff}}^{(\lambda)}=e,$ $R=0$, $\mu=m_{N}$ (nucleon mass), and neglect the terms containing the effective-range expansion parameters in eq. \[effr\]. These approximations are not necessary but, with these choices, the numerical results will not depend on the charges and mass parameters of a particular nucleus; only on its neutron separation energy $S_{n}$.
The total electron-disintegration cross section is obtained from eq. \[elect\_scat\], with the density of states given by $N_{f}=d^{3}k/\left( 2\pi\right) ^{3}$, and integrating over $\mathbf{k}$ and $\Omega.$ Figure \[sn\] shows the electro-dissociation cross sections obtained by a numerical integration of eq. \[elect\_scat\], as a function of the separation energy, $S_{n}$, for electron bombarding energies equal to 0.1 and 0.5 MeV, respectively. One observes that only for very low neutron separation energies ($S_{n}\lesssim50$ keV) the electro-disintegration cross section becomes larger than 1 mb. If a more realistic model for $l=1$ transitions is used, the cross section will be further reduced by: (a) a factor $\left( Z_{a}/A_{a}\right) ^{2}\lesssim1/4$ due to the effective charge, (b) by properly orthogonalized wavefunctions, and (c) by the energy dependence of the phase-shifts.
Figure \[sn\] also shows that the electron dissociation cross section increases appreciably with the electron energy. It is thus instructive to study the dependence of the cross section on the electron energy at high energies. The electron energy will be considered to be much larger than the energy transfer in the dissociation, i.e. $E\gg\Delta E=E_{x}$ ($E_{x}$ denotes the excitation energy). The scattering is peaked at forward angles and, from kinematics, $q=k^{\prime}\cos\theta-k\simeq\Delta k\simeq
E_{x}/\hbar c$. For energy transfers $E_{x}$ of the order of a few MeV, one also has $q\ll p\mathbf{,\ }\eta$. Using eqs. \[effr\] one obtains for the leading multipolarity ($l=1$)$$\rho_{fi}^{(1)}\left( q\right) \cong\frac{4\pi i\ e\ A\ qk}{\left(
k^{2}+\eta^{2}\right) ^{2}}\ .\label{rhoapp}$$
Using eqs. \[rhoapp\] and \[elect\_scat\] one obtains$$\frac{d\sigma_{e}}{d\Omega dE_{x}}=\frac{48\sqrt{2}}{\pi}\frac{e^{2}\left[
e_{\mathrm{eff}}^{(1)}\right] ^{2}p^{2}}{\hbar^{2}\mu c^{2}}\frac{1}{q^{2}}\frac{\sqrt{S_{n}}\left( E_{x}-S_{n}\right) ^{3/2}}{E_{x}^{4}}.\label{sigeapp}$$ The solid scattering angle can be related to the momentum transfer by means of $d\Omega\cong2\pi\hbar^{2}qdq/p^{2}.$ The minimum momentum transfer for an excitation energy $E_{x}$ is given by $q_{\min}=\Delta k\cong
E_{x}/\hbar c$, so that the integral over the scattering angle yields$$\frac{d\sigma_{e}}{dE_{x}}=96\sqrt{2}\ \frac{e^{2}\left[ e_{\mathrm{eff}}^{(1)}\right] ^{2}}{\mu c^{2}}\ \frac{\sqrt{S_{n}}\left( E_{x}-S_{n}\right) ^{3/2}}{E_{x}^{4}}\ \ln\left( \frac{pc}{E_{x}}\right)
.\label{sigex}$$ Eq. \[sigex\] shows that, for large $p$, the energy spectrum in electro-disintegration depends weakly on the electron energy through the logarithm function. This means that there is no great advantage (in terms of number of events) in increasing the electron energy when $E_{e}\gg m_{e}c^{2}$. From eq. \[sigex\] one also sees that the energy spectrum increases sharply starting at $E_{x}=S_{n}$, peaks at $E_{x}=8S_{n}/5$, and decreases with $E_{x}^{-5/2}$ at large energies. This is the same characteristic spectrum as found in Coulomb dissociation of halo nuclei [@BS92].
The integral over the excitation energy gives, to leading order,$$\sigma_{e}\left( p\right) =6\pi\sqrt{2}e^{2}\left[ e_{\mathrm{eff}}^{(1)}\right] ^{2}\frac{1}{\mu c^{2}S_{n}}\ln\left( \frac{pc}{S_{n}}\right)
. \label{sigef}$$
For stable nuclei, with $S_{n}\simeq$ few MeV, the electron-disintegration cross section is small. The dependence of eq. \[sigef\] on the inverse of the separation energy is most important for loosely bound nuclei. Using $S_{n}=100$ keV, $E_{e}\cong pc=10$ MeV, $e_{\mathrm{eff}}^{(1)}=e,$ and $\mu
c^{2}=10^{3}$ MeV, equation \[sigef\] yields 25 mb for the dissociation cross section by high energy electrons. Note that the above equations are valid only if $E_{e}\gg m_{e}c^{2}$. They show that the electro-disintegration cross section increases very slowly with the electron energy. In contrast, as shown in figure \[sn\], at low electron energies the cross sections increase much faster with $E_{e}$.
The arguments used here are only valid for dissociation (breakup) experiments. In the case of electron excitation of bound states, the matrix elements can become large for small excitation energies and cases where there is a large overlap of the wavefunctions. Consequently, the cross section can be much higher when these conditions are met.
Dissociation of halo nuclei beams on a fixed target
===================================================
Dissociation by atomic electrons in the target
----------------------------------------------
I use the Thomas-Fermi model to describe the electronic distribution in an atom. This approximation is well known, being described in many textbooks (see, e.g., ref. [@Fri90]). In this model, the electron density as a function of the distance from the atomic nucleus with charge $Ze$ is given by$$\rho\left( r\right) =\frac{1}{3\pi^{2}}\left[ 2\frac{m_{e}}{\hbar^{2}}Ze^{2}\frac{\Phi\left( x\right) }{r}\right] ^{3/2},\ \ \mathrm{where}\ \ \ x=br,\ \ \ \mathrm{and}\ \ \ b=2\left( \frac{4}{3\pi}\right)
^{2/3}\frac{m_{e}}{\hbar^{2}}e^{2}Z^{1/3}.$$ The function $\Phi\left( x\right) $ is the solution of the Thomas-Fermi equation$$\frac{d^{2}\Phi}{dx^{2}}=\frac{\Phi^{3/2}}{x^{1/2}}\ .$$ Numerical solutions of this equation date back to refs. [@Bak30; @BC31]. An excellent approximation was found by Tietz [@Tie56]:$$\Phi\left( x\right) =\frac{1}{\left( 1+ax\right) ^{2}},\ \ \ \mathrm{where}\ \ \ \ \ \ \ \ a=0.53625.$$
The probability density (normalized to $Z$) to find an electron with momentum $p$ is given by$$\mathcal{P}(p)=\left\vert D\left( \mathbf{p}\right) \right\vert
^{2},\ \ \ \ \mathrm{where}\ \ \ D\left( \mathbf{p}\right) =\frac{1}{\left(
2\pi\right) ^{3/2}}\int d^{3}re^{i\mathbf{p\cdot r}}\ \sqrt{\rho\left(
r\right) }.\label{FT1}$$ The electronic density $\rho\left( r\right) $ has to be Lorentz transformed to the frame of reference of the projectile nucleus. Assuming a straight-line projectile motion with impact parameter $b$ from the atomic center, the transformed density is$$\rho^{\prime}\left( r\right) =\gamma\rho\left( \sqrt{b^{2}+\gamma^{2}z^{2}}\right) ,\label{Dboost}$$ where $\gamma=\left( 1-v^{2}/c^{2}\right) ^{-1/2}$ is the Lorentz factor, and $v$ is the projectile velocity.
The Fourier transform in eq. \[FT1\] becomes$$D^{\prime}\left( \mathbf{p}\right) =\frac{1}{\left( 2\pi\right) ^{3/2}}\int d^{3}r\ e^{i\mathbf{p\cdot r}}\ \sqrt{\rho^{\prime}\left( r\right)
}=\frac{1}{\left( 2\pi\right) ^{3/2}}\frac{1}{\sqrt{\gamma}}\int
d^{3}r^{\prime}\ e^{i\mathbf{P\cdot r}^{\prime}}\ \sqrt{\rho\left( r^{\prime
}\right) }\ ,\label{FT2}$$ where $$\mathbf{P=}\left( \mathbf{p}_{t},p_{z}/\gamma\right)
,\ \ \ \ \ \ \ r^{\prime}=\left( \mathbf{b},\gamma z\right) \ ,$$ with $\mathbf{p}_{t}\left( p_{z}\right) $ being the transverse (longitudinal) momentum.
Since $\rho\left( r\right) $ is spherically symmetric, eq. \[FT2\] can be rewritten as$$D^{\prime}\left( \mathbf{p}\right) =\sqrt{\frac{2}{\pi\gamma}}\frac{1}{P}\int dr\ r\ \sin\left( Pr\right) \ \sqrt{\rho\left( r\right) }\ .$$
For an atom at rest, very few electrons have orbital kinetic energies larger than 100 keV. In the case of $^{92}$U only 3% of the electrons (2 electrons!) have kinetic energies larger than that. But in the reference frame of a 100 MeV/nucleon projectile, 50% of the electrons have energies greater than 100 keV.
Assuming that each electron scatters independently, the total dissociation cross section by the target atomic electrons is given by$$\sigma_{e}^{\left( TAE\right) }\left( p\right) =\int d^{3}p\ \mathcal{P}^{\prime}(\mathbf{p})\ \sigma_{e}\left( p\right) =2\pi\int_{0}^{\infty
}dp_{t}\ p_{t}\int_{-\infty}^{\infty}dp_{z}\ \mathcal{P}^{\prime}(p_{z},p_{t})\ \sigma_{e}\left( p\right) \ .$$ The separation of the above integral into longitudinal and transverse momenta is convenient because only the longitudinal momentum component of the electrons is relevant for the dissociation of the projectile.
Figure \[ate\] shows the dissociation cross section for a halo nucleus, with separation energy $S_{n}=100$ keV, incident on a Pb target as a function of the bombarding energy. Although the cross sections are small for incident energies equal to a few hundred MeV/nucleon, they increase drastically as the bombarding energy becomes close to 1 GeV/nucleon. At 10 GeV/nucleon the dissociation cross section is of the order of 1 barn.
Comparing the above results with those obtained in section 2, we notice that there are different energy scales for electro-disintegration on fixed targets (by atomic electrons) and on a collider-beam mode. This is due to the Lorentz transformation and to the large density (compared to an electron beam) of electrons in a heavy atom. Thus, with beams of halo nuclei with a few GeV/nucleon one could, in principle, perform similar studies as with electron-radioactive beam colliders. The disadvantage is that the Coulomb dissociation cross sections of loosely-bound nuclei are much larger, as shown in the next section.
Coulomb dissociation
--------------------
Coulomb dissociation of halo nuclei has been considered long time ago [@BB86]. For the leading electric dipole transitions from an s- to a p-wave, the Coulomb dissociation cross section is given by [@BB86; @BS92]$$\frac{d\sigma_{C}}{dE_{x}}=\frac{32}{3}\frac{Z^{2}e^{2}\left[ e_{\mathrm{eff}}^{(1)}\right] ^{2}}{\hbar^{2}c^{2}}\ \frac{\sqrt{S_{n}}\left( E_{x}-S_{n}\right) ^{3/2}}{E_{x}^{4}}\ \ln\left( \frac{\gamma\hbar c}{\delta
E_{x}R}\right) ,\label{sigcoul}$$ where $Z$ is the nuclear target charge, $\delta=0.681...$, and $R$ is the strong interaction radius ($R\simeq R_{P}+R_{T}$). Note the similarity with eq. \[sigex\] in the dependence on the excitation energy $E_{x}$, because the dipole operator is the same in both cases. However, the argument of the logarithm is different because of the small electron mass. Moreover, the coherent electric field of the projectile yields a factor $Z^{2}$ which substantially increases the Coulomb dissociation cross section for large-$Z$ targets.
The total cross section for Coulomb dissociation as a function of the bombarding energy (i.e., as a function of $\gamma$) is given by$$\sigma_{C}=\frac{2\pi}{3}\frac{Z^{2}e^{2}\left[ e_{\mathrm{eff}}^{(1)}\right] ^{2}}{\hbar^{2}c^{2}}\ \frac{\hbar^{2}}{\mu S_{n}}\ \ln\left(
\frac{\gamma\hbar c}{\delta S_{n}R}\right) .$$
Using the same values listed after eq. \[sigef\], for 10 GeV/nucleon projectiles impinging on Pb targets, yields cross sections of approximately 24 barns. This is much larger than that due to the dissociation by electrons in the target. But the contribution of the later process comprises 5% of the total disintegration cross section, and should be considered in experimental analysis.
Conclusions
===========
In this article the inelastic scattering of electrons off halo nuclei was studied, with emphasis on the energy dependence of the dissociation cross sections. It is shown that the cross sections for electro-dissociation of weakly-bound nuclei reach ten milibarns for 10 MeV electrons and increase logarithmically at higher energies. This means that extracting information about the continuum structure of weakly-bound nuclei (e.g. scattering lengths and effective ranges, as in eq. \[effr\]) can only be done if the intensity of the radioactive beam is very large, or if the collider allows for a large number of sequential interactions between the electrons and the nuclei at different crossing points. This conclusion can be drawn from figure \[sn\], where a steep decrease of the dissociation cross section with $S_{n}$ is seen. Halo breakup experiments (common in fixed-target radioactive beam facilities) are difficult to carry out in electron-radioactive beam colliders, but not impossible if $S_{n}$ is small.
A new facility is under construction at the GSI/Darmstadt, Germany. Experiments in a collider mode are planned so that electron beams will cross radioactive beams with center-of-mass energies of 1.5 GeV, i.e. 0.5 GeV electrons impinging on a 740 MeV/A counter propagating ion [@GSI; @Haik]. For light, neutron-rich, nuclei luminosities of $10^{29}$/cm$^{2}$.s are expected. The approximate eq. \[sigef\] yields cross sections of the order of 1 mb for $S_{n}\simeq1$ MeV, what means an estimated 100 events/second.
I have also shown that electrons present in a fixed nuclear target access similar scattering conditions as in an electron-radioactive beam collider. However, Coulomb excitation cross sections are much larger in the case of a heavy nuclear target. In view of the scientific impact of an electron-radioactive beam facility these results are useful for guidance in planning future experiments. The role of electron (and photon) scattering on exotic nuclei in stellar environments is also of interest for stellar modeling and work in this direction is in progress.
Acknowledgements
================
I thank useful discussions with P.G. Hansen, H. Schatz, H. Simon, U. van Kolck and V. Zelevinsky. This research was supported in part by the Department of Energy under Grant No. DE-FG02-04ER41338.
[99]{}
R.F. Casten and B.M. Sherrill, *Prog. Part. Nucl. Phys.* **45** (2000) S171.
C.A. Bertulani, L.F. Canto and M.S. Hussein, Phys. Rep. **226**, 281 (1993).
FAIR: Facility for Antiproton and Ion Research, Conceptual Design Report, GSI, 2002, p. 162.
J.M. Eisenberg and W. Greiner, Excitation Mechanisms of the Nucleus", (North-Holland, Amsterdam, 1988).
L.I. Schiff, Phys. Rev. **96**, 765 (1954).
L.J. Weigert and J.M. Eisenberg, Nucl. Phys. **53**, 508 (1964).
F. Scheck, Nucl. Phys. **77**, 577 (1966).
W.A. McKinley and H. Feshbach, Phys. rev. **74**, 1759 (1948).
D.R. Yennie, D.G. Ravenhall, and R.R. Wilson, Phys. Rev. **92**, 1325 (1953); Phys. Rev. **95**, 500 (1954).
L.S. Cutler, Phys. Rev. **157**, 885 (1967).
C.A. Bertulani and G. Baur, Nucl. Phys. **A480**, 615 (1988). [Note that a factor 1/3 is missing in eqs. 3.2b and 4.3b of this reference]{}.
C.A. Bertulani and A. Sustich, Phys. Rev. **C 46**, 2340 (1992).
T. Otsuka et al., Phys. Rev. **C 49**, R2289 (1994).
A. Mengoni, T. Otsuka and M. Ishihara, Phys. Rev. C **52**, R2334 (1995).
D.M. Kalassa and G. Baur, J. Phys. **G 22**, 115 (1996).
S. Typel and G. Baur, Phys. Rev. Lett. **93**, 142502 (2004).
P.G. Hansen and B. Jonson, Europhys. Lett. **4**, 409 (1987).
H. Friedrich, Theoretical Atomic Physics", (Springer-Verlag, Heidelberg, 1990).
E.B. Baker, Phys. Rev. **36**, 630 (1930).
V. Bush and S.H. Caldwell, Phys. Rev. C **38**, 1898 (1931).
T. Tietz, J. Chem. Physics **25**, 787 (1956); Z. Naturforsch. **23a**, 191 (1968).
R. Hofstadter, Rev. Mod. Phys. **28**, 214 (1956).
D. G. Ravenhall, Rev. Mod. Phys. **30**, 430 (1958).
Haik Simon, private communication.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Statistical analysis of evolutionary-related protein sequences provides insights about their structure, function, and history. We show that Restricted Boltzmann Machines (RBM), designed to learn complex high-dimensional data and their statistical features, can efficiently model protein families from sequence information. We here apply RBM to twenty protein families, and present detailed results for two short protein domains, Kunitz and WW, one long chaperone protein, Hsp70, and synthetic lattice proteins for benchmarking. The features inferred by the RBM are biologically interpretable: they are related to structure (such as residue-residue tertiary contacts, extended secondary motifs ($\alpha$-helix and $\beta$-sheet) and intrinsically disordered regions), to function (such as activity and ligand specificity), or to phylogenetic identity. In addition, we use RBM to design new protein sequences with putative properties by composing and turning up or down the different modes at will. Our work therefore shows that RBM are a versatile and practical tool to unveil and exploit the genotype-phenotype relationship for protein families.'
author:
- Jérôme Tubiana
- Simona Cocco
- Rémi Monasson
bibliography:
- 'bibliography.bib'
title: Learning protein constitutive motifs from sequence data
---
Sequencing of many organism genomes has led over the recent years to the collection of a huge number of protein sequences, gathered in databases such as UniProt or PFAM [@finn2013pfam]. Sequences with a common ancestral origin, defining a family (Fig. 1A), are likely to code for proteins with similar functions and structures, hence providing a unique window into the relationship between genotype (sequence content) and phenotype (biological features). In this context, various approaches have been introduced to infer protein properties from sequence data statistics, in particular amino-acid conservation and coevolution (correlation) [@teppa2012disentangling; @de2013emerging].
A major objective of these approaches is to identify positions carrying amino acids with critical impact on the protein function, such as catalytic or binding sites, or specificity-determining sites controlling ligand specificity. Principal Component Analysis (PCA) of the sequence data can be used to unveil groups of coevolving sites with specific functional role [@russ2005natural; @rausell2010protein; @halabi2009protein]. Other methods rely on phylogeny [@rojas2012ras], entropy (variability in amino-acid content) [@reva2011predicting; @reva2007determinants], or hybrid combination of both [@mihalek2004family; @ashkenazy2016consurf].
Another objective is to extract structural information, such as the contact map of the three-dimensional fold. Considerable progress was brought by maximum-entropy methods, which rely on the computation of direct couplings between sites reproducing the pairwise coevolution statistics in the sequence data [@lapedes1999; @weigt2009mp; @jones2011psicov; @cocco2018inverse]. Direct couplings provide very good estimators of contacts [@morcos2011direct; @hopf2012three; @kamisetty2013assessing; @ekeberg2014fast], and capture pairwise epistasis effects necessary to model fitness changes resulting from mutations [@chakraborty2014hiv; @figliuzzi2016; @hopf2017mutation].
Despite these successes, a unique, accurate framework capable of extracting the structural and functional features common to a protein family, as well as the phylogenetic variations specific to sub-families is still missing. Hereafter, we consider Restricted Boltzmann Machines (RBM) for this purpose. RBM are a powerful concept coming from machine learning [@ackley1987learning; @hinton2012practical]; they are unsupervised (sequence data need not be annotated) and generative (able to generate new data). Informally speaking, RBM learn complex data distributions through their statistical features (Fig. 1B).
In the present work, we have developed a method to learn efficiently RBM from protein sequence data. To illustrate the power and versatility of RBM, we have applied our approach to the sequence alignments of twenty different protein families. We report the results of our approach, with special emphasis on four families: the Kunitz domain, a protease inhibitor, historically important for protein structure determination [@ascenzi2003bovine], the WW domain, a short module binding different classes of ligands [@sudol1995characterization], Hsp70, a large chaperone protein [@bukau1998hsp70], and lattice-protein *in silico* data [@shakhnovich1990enumeration; @mirny2001] to benchmark our approach on exactly solvable models [@jacquin2016benchmarking]. Our study shows that RBM are able to capture (1) structure-related features, either local, such as tertiary contacts, or extended, such as secondary structure motifs ($\alpha$-helix and $\beta$-sheet) or characteristic of intrinsically disordered regions; (2) functional features, i.e. groups of amino acids controlling specificity or activity; (3) phylogenetic features, related to sub-families sharing evolutionary determinants. Some of these features involves two residues only (as direct pairwise couplings do), others extend over large and not necessarily contiguous portions of the sequence (as in collective modes extracted with PCA). The pattern of similarities of each sequence with the inferred features defines a multi-dimensional representation of this sequence, highly informative about the biological properties of the corresponding protein (Fig. 1C). Focusing on representations of interest allows us, in turn, to design new sequences with putative functional properties. In summary, our work shows that RBM offer an effective computational tool to characterize and exploit quantitatively the genotype-phenotype relationship specific to a protein family.
![ [**Reverse and forward modeling of proteins.**]{} [**A.**]{} Example of Multiple-Sequence Alignment (MSA), here of the WW domain (PF00397). Each column $i=1,...,N$ corresponds to a site on the protein, and each line to a different sequence in the family. Color code for amino acids: red = negative charge (E,D), blue = positive charge (H, K, R), purple = non-charged polar (hydrophilic) (N, T, S, Q), yellow = aromatic (F, W, Y), black = aliphatic hydrophobic (I, L, M, V), green = cysteine (C), grey = other, small (A, G, P). [**B.**]{} In a Restricted Boltzmann Machine (RBM), weights $w_{i\mu}$ connect the visible layer (carrying protein sequences $\bf v$) to the hidden layer (carrying representations $\bf h$). Biases on the visible and hidden units are introduced by the local potentials $g_i(v_i)$ and ${\cal U}_\mu(h_\mu)$. Due to the bipartite nature of the weight graph, hidden units are independent conditioned to a visible configuration, and vice versa. [**C.**]{} Sequences $\bf v$ in the MSA (dots in sequence space, left) code for proteins with different phenotypes (dot colors). RBM define a probabilistic mapping from sequences $\bf v$ onto the representation space $\bf h$ (right), indicative of the phenotype of the corresponding protein, and encoded in the conditional distribution $P({\bf h}|{\bf v})$, Eqn. (3) (black arrow). The reverse mapping from representations to sequences is $P({\bf v}|{\bf h})$, Eqn. (4) (black arrow). Sampling a subspace in the representation space (colored domains) defines in turn a complex subset of the sequence space, and allows one to design sequences with putative phenotypic properties, either found in the MSA (green circled dots) or not encountered in Nature (arrow out of blue domain). [**D.**]{} Three examples of potentials ${\cal U}$ defining the hidden-unit type in RBM, see Eqn. (1) and Panel B: quadratic (black, $\gamma=0.2$,$\theta=0$) and double Rectified Linear Unit (dReLU) (Green: $\gamma_+=\gamma_-=0.1$, $\theta_+=-\theta_-=1$; Purple: $\gamma_+=1$, $\gamma_-=20$, $\theta_+=-6$, $\theta_-=25$) potentials. In practice, the parameters of the hidden unit potentials are fixed through learning of the sequence data. [**E.**]{} Average activity of hidden unit $h$, calculated from Eqn. (3), as a function of the input $I$ defined in Eqn. (2). The three curves correspond to the three choices of potentials in panel A. For the quadratic potential (Black), the average activity is a linear function of $I$. For dReLU1 (Green), small inputs $I$ barely activate the hidden unit, whereas dReLU2 (Purple) essentially binarizes the inputs $I$. ](figure1.pdf){width="17.8cm"}
\[fig1\]
Results {#results .unnumbered}
=======
Restricted Boltzmann Machines {#restricted-boltzmann-machines .unnumbered}
-----------------------------
### Definition
A Restricted Boltzmann Machine (RBM) is a joint probabilistic model for sequences and representations, see Fig. 1C. It is formally defined on a bipartite, two-layer graph (Fig. 1B). Protein sequences ${\bf v} = (v_1,v_2,...,v_N)$ are displayed on the Visible layer, and representations ${\bf h} = (h_1,h_2,...,h_M)$ on the Hidden layer. Each visible unit takes one out of $21$ values (20 amino acids + 1 alignment gap). Hidden-layer unit values $h_\mu$ are real. The joint probability distribution of ${\bf v}, {\bf h}$ is $$\label{Energy}
P({\bf v},{\bf h}) \propto \exp \bigg( \sum_{i=1}^N g_i(v_i) - \sum_{\mu =1}^M \mathcal{U}_\mu(h_\mu) + \sum_{i,\mu} h_\mu\, w_{i\mu} (v_i) \bigg) \ ,$$ up to a normalization constant. Here, the weight matrix $w_{i\mu}$ couples the visible and the hidden layers, and $g_i(v_i)$ and $\mathcal{U}_\mu(h_\mu)$ are local potentials biasing the values of, respectively, the visible and the hidden variables (Figs. 1B,D).
### From sequence to representation, and back
Given a sequence $\bf v$ on the visible layer, the hidden unit $\mu$ receives the input $$\label{inputmu}
I_\mu ({\bf v}) = \sum_{i} w_{i\mu} (v_i) \ .$$ This expression is analogous to the score of a sequence with a position-specific weight matrix. [ ]{} $I_\mu$ [ ]{} $\mu$, [ ]{} $|I_\mu|$ [ ]{}
The input $I_\mu$ determines, in turn, the conditional probability of the activity $h_\mu$ of the hidden unit, $$\label{condmu}
P( h_\mu | {\bf v}) \propto \exp \big(-\mathcal{U}_\mu(h_\mu) + h_\mu\, I_\mu({\bf v}) \big) \ ,$$ up to a normalization constant. The nature of the potential ${\cal U}$ is crucial to determine how the average activity $h$ varies with the input $I$, see Fig. 1E and below.
In turn, given a representation (set of activities) $\bf h$ on the hidden layer, the residues on site $i$ are distributed according to $$\label{condv}
P( v_i | {\bf h}) \propto \exp \bigg( g_i(v_i) + \sum _\mu h_\mu\, w_{i\mu}(v_i) \bigg)\ .$$ Hidden units with large activities $h_\mu$ strongly bias this probability, and favor values of $v_i$ corresponding to large weights $w_{i\mu}(v_i)$.
Use of Eqn. (3) allows us to sample the representation space given a sequence, while Eqn. (4) defines the sampling of sequences given a representation, see both directions in Fig. 1C. Iterating this process generates high-probability representations, which, in turn produce very likely sequences, and so on.
### Probability of a sequence
[ ]{}, $P({\bf v})$, [ ]{} $P({\bf v},{\bf h})$ [ ]{} $\bf h$. $$\label{marginal}
P({\bf v}) = \int \prod_{\mu=1}^M dh_\mu P({\bf v}, {\bf h}) \propto \exp \bigg[ \sum_{i=1}^N g_i(v_i) + \sum_{\mu=1}^M \Gamma_\mu\big (I_\mu ({\bf v})\big) \bigg] \ ,$$ [ ]{} $\Gamma_\mu(I) = \log \int dh \, e^{-U_\mu(h) + h \,I} $ [ ]{} ${\cal U}_\mu$ [ ]{} $\mu$, see Eqn. (\[inputmu\]).
[ ]{} ${\cal U}_\mu(h)=\frac {\gamma_\mu}{ 2} h^2 + \theta_\mu h$ (Fig. 1E), [ ]{}$\Gamma_\mu(I) = \frac{1}{2 \gamma_\mu}(I-\theta_\mu)^2$ [ ]{}Eqn. (\[marginal\]) [ ]{} $J_{ij} (v_i,v_j) =\sum _\mu \frac{1}{\gamma_\mu} w_{i\mu} (v_i) w_{j\mu}(v_j)$. [ ]{} [@cocco2013principal], [ ]{} $M$ [ ]{}
[ ]{}${\cal U}_\mu$, [ ]{} $\Gamma(I)$, [ ]{} *all orders* [ ]{} $w_{i\mu}$. [ ]{}$M$[ ]{} $$\mathcal{U}_\mu(h) = \frac12 \gamma_{\mu,+} h_+^2 + \frac12 \gamma_{\mu,-} h_-^2 + \theta_{\mu,+} h_+ + \theta_{\mu,-} h_-\ ,\quad \text{where} \quad h_+ = \max(h,0)\ , \quad h_- = \min(h,0) \ ,$$ [ ]{}$h_\mu$[ ]{}$h_\mu$ [ ]{}
[ ]{}
### Learning
The weights $w_{i\mu}$ and the defining parameters of the potentials $g_i$ and ${\cal U}_\mu$ are learned by maximizing the average log-probability ${\left\langle \log P(\bf v) \right\rangle}_{MSA}$ of the sequences $\bf v$ in the Multiple Sequence Alignment (MSA). In practice, estimating the gradients of the average log-probability with respect to these parameters requires to sample from the model distribution $P({\bf v})$, which is done through Monte Carlo simulation of the RBM, see Methods.
We also introduce penalty terms over the weights $w_{i\mu}(v)$ (and the local potentials $g_i(v)$ on visible units) to avoid overfitting, and to promote sparse weights. Sparsity facilitates the biological interpretation of weights and thus, emphasizes the correspondence between representation and phenotypic spaces (Fig. 1C). Crucially, imposing sparsity also forces the RBM to learn a so-called compositional representation, in which each sequence is characterized by a subset of strongly activated hidden units, of size large compared to 1 but small compared to $M$ [@tubiana2017emergence]. All technical details about the learning procedure are reported in Methods.
In the next sections, we present results for selected values of the number of hidden units and of the regularization penalty. The values of these (hyper-)parameters are justified afterwards.
Kunitz domain {#kunitz-domain .unnumbered}
-------------
### Description {#description .unnumbered}
The majority of natural proteins are obtained by concatenating functional building blocks, called protein domains. The Kunitz domain, with a length of about 50-60 residues (protein family PF00014 [@finn2013pfam]) is present in several genes and its main function is to inhibit serine protease such as trypsin. Kunitz domains play a key role in the regulation of many important processes in the body such as tissue growth and remodeling, inflammation, body coagulation and fibrinolysis. They are implicated in several diseases such as tumor growth, Alzheimer, cardiovascular and inflammatory diseases and, therefore, have been largely studied and shown to have a large potential in drug design [@shigetomi2010anti; @bajaj2001structure].
Some examples of Kunitz domain-containing proteins include the Basic Pancreatic Trypsin Inhibitor (BPTI, 1 Kunitz domain), the Bikunin (2 domains) [@fries2000bikunin], Hepatocyte growth factor activator inhibitor (HAI, 2 domains) and tissue factor pathway inhibitor (TFPI, 3 domains) [@shigetomi2010anti; @bajaj2001structure].
Figure 2A shows the MSA sequence logo and the secondary structure of the Kunitz domain. It is characterized by two $\alpha$ helices and two $\beta$ strands; Cystein-Cystein disulfide bridges largely contribute to the thermodynamic stability of the domain, as frequently observed in small proteins. BPTI structure was the first one ever resolved [@ascenzi2003bovine], and is often used to benchmark folding predictions based on simulations [@levitt1975computer] and coevolutionary approaches [@morcos2011direct; @hopf2012three; @kamisetty2013assessing; @cocco2013principal; @haldane2018coevolutionary]. We train a RBM with $M=100$ dReLU on the MSA of PF00014, constituted by $B=8062$ sequences with $N=53$ consensus sites.
### Inferred weights and interpretations. {#inferred-weights-and-interpretations. .unnumbered}
Figure 2B shows the weights $w_{i\mu}(v)$ attached to 5 selected hidden units. Each logo identifies the amino-acid motifs in the sequences $\bf v$ giving rise to large (positive or negative) inputs $I$ onto the associated hidden unit, see Eqn. (2).
Weight 1 in Fig. 2B has large components on sites 45 and 49, in contact in the final $\alpha_2$ helix (Figs. 2A&D). The distribution of the inputs $I_1$ partitions the MSA in three subfamilies (Fig. 2C, top panel, dark blue histogram). The two peaks in $I_1\simeq -2.5$ and $I_1\simeq 1.5$ identify sequences in which the contact is due to an electrostatic interaction with, respectively, $(+,-)$ and $(-,+)$ charged amino acid on sites 45 and 49; the other peak in $I_1\simeq 0$ identify sequences realizing the contact differently, e.g. with an aromatic amino acid on site 45. Weight 1 shows also a weaker electrostatic component on site 53 in Fig. 2B; the 4-site separation between sites 45–49–53 fits well with the average helix turn of 3.6 amino acids (Fig. 2D).
Weight 2 focuses on the contact between residues 11-35, realized in most sequences by a C-C disulfide bridge (Fig. 2B and negative $I_2$ peak in Fig. 2C, top). A minority of sequences in the MSA, corresponding to $I_2>0$ and mostly coming from nematode organisms (Appendix 1, Fig. 19), do not show the C-C bridge. A subset of these sequences strongly and positively activate hidden unit 3 (Appendix 1, Fig. 19 and $I_3>0$ peak in Fig. 2C). Positive components in weight 3 logo suggest that these proteins stabilize their structure through electrostatic interactions between sites 10 ($-$ charge) and 33-36 ($+$ charges both), see Figs. 2B&D, to compensate the absence of C-C bridge on the neighbouring sites 11-35.
Weight 4 describes a feature mostly localized on the loop preceding the $\beta_1$-$\beta_2$ strands (sites 7 to 16), see Figs. 2B&D. Structural studies of the trypsin-trypsin inhibitor complex have shown that this loop binds to proteases [@marquart1983geometry]; site 12 is in contact with the active site of the protease and is therefore key to the inhibitory activity of the Kunitz domain. The two amino acids (R, K) having a large positive contribution to weight 4 in position 12 are basic and bind to negatively charged residues (D, E) on the active site of trypsin-like serine proteases. While several Kunitz domains with known trypsin inhibitory activity, such as BPTI, TFPI, TPPI-2,... give rise to large and positive inputs $I_4$, Kunitz domains with no trypsin/chymotrypsin inhibition activity, e.g. associated to COL7A1 and COL6A3 genes [@chen2001carboxyl; @kohfeldt1996conversion], correspond to negative or vanishing values of $I_4$. Hence, hidden unit 4 possibly separates the Kunitz domains having trypsin-like protease inhibitory activity from the others.
This interpretation is also in agreement with mutagenesis experiments carried out on sites 7 to 16 to test the inhibitory effects of Kunitz domains BPT1, HAI-1, and TFP1 against trypsine-like proteases [@bajaj2001structure; @kirchhofer2003tissue; @shigetomi2010anti; @grzesiak2000inhibition; @chand2004structure]. In [@kirchhofer2003tissue] it was shown that mutation R12A on the first domain (out of two) of HAI-1 destroyed its inhibitory activity; a similar effect was observed with R12X, with X a non basic residue, in the first two domains (out of three) of TFP1 as discussed in [@bajaj2001structure]. The affinity between human serine proteases and the mutants G9F, G9S, G9P of bovine BPTI was shown to decrease in [@grzesiak2000inhibition]. Conversely, in [@kohfeldt1996conversion] it was shown that the set of mutations P10R, D13A, F14R could convert the COL6A3 domain into a Trypsin inhibitor. All these results are in agreement with the above interpretation and the logo of weight 4. Note that, though several sequences have large $I_4$ (top histogram in Fig. 2C), many correspond to small or negative values. This may be explained by the facts that (i) many of the Kunitz domains analyzed are present in two or more copies, and as such, are not all required to strongly bind trypsin [@bajaj2001structure] and (ii) Kunitz domain may have other specificities encoded by other hidden units. In particular, weight 34 in Supporting Information, displays on site 12 large components associated to medium to large size hydrophobic residues (L, M, Y), and is possibly related to other serine protease specificity classes such as chymotrypsin [@appel1986chymotrypsin].
Weight 5 codes for a complex extended mode. To interpret this feature, we display in Fig. 2C (bottom histogram) the distributions of Hamming distances between all pairs of sequences in the MSA (gray histograms) and between the 100 sequences $\bf v$ in the MSA with largest inputs $|I_\mu({\bf v})|$ to the corresponding hidden unit (light blue histograms). For hidden unit 5, the distances between those top-input sequences are smaller than between random sequences in the MSA, suggesting that weight 5 is characteristic of a cluster of closely related sequences. Here, these sequences correspond to the protein Bikunin present in most mammals and some other vertebrates [@shigetomi2010anti]. Conversely, for other hidden units (e.g. 1,2), both histograms are quite similar, showing that the corresponding weight motifs are found in evolutionary distant sequences.
[ ]{} The remaining 95 inferred weights are shown in Supporting Information. We find a variety of structural features, *e.g.* pairwise contacts as in weights 1 and 2, [ ]{}[@cocco2013principal], and phylogenetic features (activated by evolutionary related sequences as hidden unit $5$); the latter include in particular stretches of gaps, mostly located at the extremities of the sequence [@cocco2013principal]. Several weights have strong components on the same sites as weight 4, showing the complex pattern of amino acids controlling binding affinity.
![ [**Modeling Kunitz Domain with RBM.**]{} [**A.**]{} Sequence logo and secondary structure of the Kunitz domain (PF00014), showing two $\alpha$-helices and two $\beta$-strands. Note the presence of the three C-C disulfide bridges between 11-35, 2-52, 27-48. [**B.**]{} Weight logos for five hidden units, see text. Positive and negative weights are shown by letters located, respectively, above and below the zero axis. Values of the norms $\|W_\mu\|_2 = \sqrt{\sum _{i,v} w_{i\mu}(v)^2}$ are given. Same color code for the amino acids as in Fig. 1A. [**C.**]{} Top: Distribution of inputs $I_\mu({\bf v})$ over the sequences $\bf v$ in the MSA (dark blue), and average activity vs. input function (full line, left scale); red points correspond to activity levels used for design in Fig. 5. Bottom: Histograms of Hamming distances between sequences in the MSA (grey) and between the 20 sequences (light blue) with largest (for unit 2,3,4) or smallest (1,5) $I_\mu$. [**D.**]{} 3D visualization of the weights, shown on PDB structure 2knt [@merigeau19981] using VMD [@humphrey1996vmd]. White spheres denote the positions of the 3 disulfide bridges in the wild type sequence. Green spheres locate residues $i$ such that $\sum_v |w_{i\mu}(v) |>S$, with $S= 1.5$ for hidden units $\mu=1,2,3$, $S=1.25$ for $\mu=4$, and $S=0.5$ for $\mu=5$. []{data-label="fig2"}](figure2.pdf)
WW domain {#ww-domain .unnumbered}
---------
### Description
WW is a protein-protein interaction domain found in many eukaryotes and human signalling proteins, involved in essential cellular processes such as transcription, RNA processing, protein trafficking, receptor signalling. WW is a short domain of length 30-40 amino-acids (Fig. 3A, PFAM PF00397, $B=7503$ sequences, $N=31$ consensus sites), which folds into a three-stranded antiparallel $\beta$-sheet. The domain name stems from the two conserved tryptophans (W) at positions 5-28 (Fig. 3A), which serve as anchoring sites for the ligands. WW domains bind to a variety of proline (P)-rich peptide ligands, and can be divided into four groups, based on their preferential binding affinity [@sudol2000new]. Group I binds specifically to PPXY motif - where X is any amino acid; Group II to PPLP motifs; Group III to proline-arginine containing sequences (PR); Group IV to phosphorylated serine/threonine-proline sites \[p(S/T)P\]. Modulation of binding properties allow hundreds of WW domain to specifically interact with hundreds of putative ligands in mammalian proteomes.
### Inferred weights and interpretation
Four weight logos of the inferred RBM are shown in Fig. 3B; the remaining 96 weights are given in Supporting Information. Weight 1 codes for a contact between sites 4-22 realized either by two amino acids with oppositive charges ($I_1<0$), or by one [ ]{} and one negatively charged amino acid ($I_1>0$). Weight 2 shows a $\beta$-sheet–related feature, with large entries defining a set of mostly hydrophobic ($I_2>0$) or hydrophilic ($I_2<0$) residues localized on the $\beta_1$ and $\beta_2$ strands (Fig. 3B) and in contact on the 3D fold, see Fig. 3D. The activation histogram in FIg. 3C, with a large peak on negative $I_2$, suggest that this part of the WW domain is exposed to the solvent in most, but not all, natural sequences.
Weights 3 and 4 are supported by sites on the $\beta_2$-$\beta_3$ binding pocket and on the $\beta_1$-$\beta_2$ loop of the WW domain. The distributions of activities in Fig. 3C highlight different groups of sequences in the MSA that strongly correlate with experimental ligand-type identification, see Fig. 3E. We find that (i) Type I domains are characterized by $I_3<0$ and $I_4>0$; (ii) Type II/III domains are characterized by $I_3 >0 $ and $I_4>0$; (iii) There is no clear distinction between Type II and Type III domains; (iv) Type IV domains are characterized by $I_3>0$ and $I_4<0$. These findings are in good agreement with various studies:
\(i) Mutagenesis experiment have shown the importance of sites 19, 21, 24, 26 for binding specificity [@espanel1999single; @fowler2010high]. For the YAP1 WW domain, as confirmed by various studies (see Table 2 in [@fowler2010high]), the mutations H21X and T26X reduce the binding affinity to Type I ligands, while Q24R increases it and S12X has no effect. This is in agreement with the negative components of weight 3 (Fig. 3B): $I_3$ increases upon mutations H21X and T26X, decreases upon Q24R and is unaffected by S12X. Moreover the mutation L19W alone, or combined with H21\[D/G/K/R/S\] could switch the specificity from Type I to Type II/III [@espanel1999single]. These results are consistent with Fig. 3E: YAP1 (blue cross) is of Type I but one or two mutations move it to the right side, closer to the other cluster (orange crosses). Espanel and Sudol [@espanel1999single] also proposed that Type II/III specifity required the presence of an aromatic amino acid (W/F/Y) on site 19, in good agreement with weight 3.
\(ii) The distinction between Types II and III is unclear in the literature, because WW domains often have high affinity with both ligand types.
\(iii) Several studies [@russ2005natural; @kato2002determinants; @jager2006structure] have demonstrated the importance of the $\beta_1$-$\beta_2$ loop for achieving Type IV specificity, which requires a longer, more flexible loop, as opposed to short rigid loop for other types. The length of the loop is encoded in weight 4 through the gap symbol on site 13: short and long loops correspond to, respectively, positive and negative $I_4$. The importance of residues R11 and R13 was shown in [@kato2002determinants] and [@russ2005natural], where removing R13 of Type IV hPin1 WW domain reduced its binding affinity to \[p(S/T)P\] ligands. These observations agree with the logo of weight 4, whoch authorizes substitutions between K and R on sites 11 and 13.
\(iv) A specificity-related sector of eight sites was identified in [@russ2005natural], five of which carry the top entries of weight 3 (green balls in Fig. 3D). Our approach not only provides another specificity-related feature (weight 4) but also the motifs of amino acids affecting Type I & IV specificity, in good agreement with the experimental findings of [@russ2005natural].
![[**Modeling WW Domain with RBM.**]{} [**A.**]{} Sequence logo and secondary structure of the WW domain (PF00397), with three $\beta$-strands. Note the two conserved W in positions 5 and 28. [**B.**]{} Weight logos for four representative hidden units, same as Fig. 3B. [**C.**]{} Corresponding inputs, average activities and distances between top-20 feature activating sequences, same as Fig. 3C. [**D.**]{} 3D visualization of the features, shown on the PDB structure 1e0m [@macias2000structural]. White spheres locate the two W. Green spheres locate residues $i$ such that $\sum_v |w_{i\mu}(v)|>0.7$ for each hidden unit $\mu$. [**E.**]{} Scatter plot of inputs $I_3$ vs. $I_4$. Gray dots represent the sequences in the MSA; they cluster into three main groups. Colored dots show artificial or natural sequences whose specificities, given in the legend, were tested experimentally. Upper triangle: natural, from [@russ2005natural]. Lower triangle: artificial, from [@russ2005natural]. Diamond: natural, from [@otte2003ww]. Crosses: YAP1 (0) and variants (1 and 2 mutations from YAP1), from [@espanel1999single]. The three clusters match the standard ligand type classification. []{data-label="fig3"}](figure3.pdf)
Hsp70 Protein {#hsp70-protein .unnumbered}
-------------
### Description
70-kDa heat shock proteins (Hsp70) form a highly-conserved family represented in essentially all organisms. Hsp70, together with other chaperone proteins, perform a variety of essential functions in the cell: they can assist folding and assembly of newly synthetized proteins, trigger refolding cycles of misfolded proteins, transport unfolded proteins through organelle membranes, and when necessary, deliver non-functional proteins to the proteasome, endosome or lysosome for recycling [@bukau1998hsp70; @young2004pathways; @zuiderweg2017remarkable]. There are 13 HSP70s protein-encoding genes in humans, differing by where (nucleus/cytoplasm, mitochondria, endoplasmic reticulum) and when they are expressed. Some, such as HSPA8 (Hsc70) are constitutively expressed whereas others such as HSPA1 and HSPA5 are stress-induced (respectively by heat shock and glucose deprivation). Notably, Hsc70 can make up to 3% of the total total mass of proteins within the cell, and is thus one of its most important housekeeping genes. Structurally, Hsp70 are multi-domain proteins of length of 600-670 sites (631 for E-Coli DNaK gene). They consist of
- A Nucleotide Binding Domain (NBD, 400 sites) that can bind and hydrolyse ATP.
- A Substrate Binding Domain (SBD sites), folded in a beta-sandwich structure, which binds to the target peptide or protein.
- A flexible, hydrophobic interdomain-linker linking the NBD and the SBD.
- A LID domain, constituted by several (up to 5) $\alpha$ helices, which encapsulates the target protein and blocks its release.
- An unstructured C-terminal tail of variable length, important for detection and interaction with other co-chaperones, such as Hop proteins [@scheufler2000structure].
Hsp70 functions by adopting two different conformations, see Figs. 4A&B. When the NBD is bound to ATP, the NBD and the SBD are held together and the LID is open, such that the protein has low binding affinity to substrate peptides. After hydrolysis of ATP to ADP, the NBD and the SBD detach from one another, and the LID is closed, yielding high binding affinity and effectively trapping the peptides between the SBD and the LID. By cycling between both conformations, Hsp70 can bind to misfolded proteins, unfold them by stretching (e.g. with two Hsp70 bound at two ends of the protein) and release them for refold cycles. Since Hsp70 alone have low ATPase activity, this cycle requires another type of co-chaperone, J-protein, which simultaneously binds to the target protein and the Hsp70 to stimulate its ATPase activity, as well as a Nucleotide Exchange Factor (NEF) that favors swaps of the ADP back to ATP and hence release of the target protein, see Fig. 1 in [@zuiderweg2017remarkable].
We have constructed a multiple sequence alignment for HSP70 with $N=675$ consensus sites and $B=32,170$ sequences, starting from the seeds of [@malinverni2015large], and queried SwissProt and Trembl UniprotKB databases using HMMER3 [@eddy2011accelerated]. Annotated sequences were grouped based on their phylogenetic origin and functional role. Prokaryotes mainly express two Hsp70 proteins: DnaK ($B=17,118$ sequences in the alignment), which are the prototype Hsp70, and HscA ($B=3,897$), which are specialized in chaperoning of Iron-Sulfur cluster containing proteins. Eukaryotes Hsp70 were grouped by location of expression (Mitochondria: $B=851$, Chloroplaste: $B=416$, Endoplasmic reticulum: $B=433$, Nucleus/Cytoplasm and others: $B=1,452$). We also singled out Hsp110 sequences, which, despite the high homology with Hsp70, correspond to non-allosteric proteins ($B=294$). We have then trained a dReLU RBM over the full MSA with $M=200$ hidden units. We show below the weight logos, structures and input distributions for ten selected hidden units, see Fig. 4 and Appendix 1, Figs. 21-26.
### Inferred weights and interpretation
Weight 1 encodes a variability of the length of the loop within the IIB subdomain of the NBD, see stretch of gaps from sites 301 to 306. As shown in Fig. 4D (projection along x axis), it separates prokaryotic DNaK proteins - for which the loop is 4-5 sites longer - from most Eukaryotic Hsp70 proteins and prokaryotic HscA. An additional hidden unit (Weight 6 in Appendix 1, Fig. 21) further separates Eukaryotic Hsp70 from HscA proteins, whose loops are 4-5 sites shorter (distribution of inputs $I_6$ in Appendix 1, Fig. 26). This structural difference between the three families was previously reported and is of high functional importance to the NBD [@buchberger1994conserved; @brehmer2001tuning]. Shorter loops increase the nucleotide exchange rates (and thus the release of target protein) in the absence of NEF, and the loop size controls interactions with NEF proteins [@brehmer2001tuning; @briknarova2001structural; @sondermann2001structure]. Hsp70 proteins having long and intermediate loop size interact specifically with respectively GrpE and Bag-1 NEF proteins, whereas short, HscA-like loops do not interact with any of them. This cochaperone specificity allows for functional diversification within the cell; for instance, Eukaryotic Hsp70 proteins expressed within mitochondria and chloroplasta, such as the human gene HSPA9 and the Chlamydomonas reinhardtii HSP70B share the long loop with prokaryotic DNaK proteins, and therefore do not interact with Bag proteins. Within the DNaK subfamily, two main variants of the loop can be isolated as well (Weight 7 in Appendix 1, Fig. 22), hinting at more NEF-protein specificities.
Weight 2 encodes a small collective mode localized on $\beta_4-\beta_5$ strands, at the edge of the $\beta$ sandwich within the SBD. Weight are quite large ($w\sim2$), and the input distribution is bimodal, separating notably HscA and chloroplastal Hsp70 ($I_2>0$) from mitochondrial Hsp70 and the other Eukaryotic Hsp70 ($I_2<0$). We note also a similarity in structural location and amino-acid content with weight 3 of the WW–domain, which controls binding specificity (Fig. 3B). Though we have not found trace of this motif in the literature, this suggests that it could be important for binding substrate specificity. Endoplasmic reticulum-specific Hsp70 proteins can also be separated from the other Eukaryotic proteins by looking at appropriate hidden units, see Weight 8 in Appendix 1, Fig. 22 and distribution of input $I_8$ in Appendix 1, Fig. 26.
RBM can also extract collective modes of coevolution spanning multiple domains, as shown by Weight 3 (Appendix 1, Fig. 21). The residues supporting Weight 3 (green spheres in Figs. 4A&B) are physically contiguous in the ADP conformation, but not in the ATP conformation. Hence, Weight 3 captures inter-domain coevolution between the SBD and the LID domains.
Weight 4 (sequence logo in Appendix 1, Fig. 21) also codes for a wide, inter–domain collective mode, localized at the interface between the SBD and the NBD domains. When the Hsp70 protein is in the ATP conformation, the sites carrying weight 4 are physically contiguous, whereas in the ADP state they are far apart, see yellow spheres in Fig. 4A&B. Moreover, its input distribution, shown in Fig. 4E, separates the non-allosteric Hsp110 subfamily ($I_4 \sim 0$) from the other subfamilies ($I_4 \sim 40$), suggesting that this motif is important for allostery. Several mutational studies have highlighted 21 important sites for allostery within E-Coli DNaK [@smock2010interdomain]; 7 of these positions carries the top entries of Weight 3, 4 appear in another Hsp110-specific hidden unit (Weight 9 in Appendix 1, Fig. 22), and several others are highly conserved and do not coevolve at all.
Lastly, Weight 5 (Fig. 4C) codes for a collective mode located mainly on the unstructured C-terminal tail, with few sites on the LID domain. Its amino-acid content is strikingly similar across all sites: positive weights for hydrophilic residues (in particular, lysine), and negative weights for tiny, hydrophobic residues. Hydrophobic-rich or hydrophilic-rich sequences are found in the MSA, see Appendix 1, Fig. 28. This motif is consistent with the role of the tail for cochaperone interaction: hydrophobic residues are important for formation of Hsp70-Hsp110 complexes via the Hop protein [@scheufler2000structure]. High-charge content is also frequently encountered and at the basis of recognition mechanism in intrinsically disordered protein regions [@oldfield2014intrinsically], which could suggest the existence of different protein partners.
Some of the results presented here were previously obtained with others coevolutionary methods. In [@malinverni2015large], the authors showed that Direct Coupling Analysis could detect conformation-specific contacts; this is similar to hidden units, respectively, 3 and 4 presented here, located on contiguous sites in the, respectively, ADP-bound and ATP-bound conformations. In [@smock2010interdomain], an inter-domain sector of sites discriminating between allosteric and non-allosteric sequences was found. This sector share many sites with our weight 4, and is also localized at the SBD/NBD edge. However, only a sector could be retrieved with sector analysis, whereas many other meaningful collective modes could be extracted using RBM.
![[**Modeling HSP70 with RBM.**]{} [**A, B.**]{} 3D structures of the DNaK E-coli HSP70 protein in the ADP-bound ([**A**]{}: PDB: 2kho [@bertelsen2009solution]) and ATP-bound ([**B**]{}: PDB: 4jne [@qi2013allosteric]) conformations. The colored spheres show the sites carrying the largest entries in the weights in panel C. [**C.**]{} Weight logos for hidden units $\mu=1$, 2 and 5; see Appendix 1, Fig. 21 for the other hidden units. Due to the large protein length, we show only weights for positions $i$ with large weights ($\sum_v |w_{i\mu}(v)| > 0.4\times \max_i \sum_v |w_{i\mu}(v)|$), with surrounding positions up to $\pm 5$ sites away; dashed lines vertical locate the left edges of the intervals. Protein backbone colors: Blue=NBD, Cyan=Linker, Red=SBD, Gray=LID. Colors: Orange=Unit 1 (NBD loop), black = Unit 2 (SBD $\beta$ strand), green= Unit 3 (SBD/LID), yellow = Unit 4 (Allosteric). [**D.**]{} Scatter plot of inputs $I_1$ vs. $I_2$. Gray dots represent the sequences in the MSA, and cluster into four main groups. Colored dots represent the main sequence categories based on gene phylogeny, function and expression. [**E.**]{} Histogram of input $I_4$, showing separation between allosteric and non-allosteric protein sequences in the MSA.[]{data-label="fig4"}](figure4.pdf)
-4cm
Sequence Design {#sequence-design .unnumbered}
---------------
The biological interpretation of the features inferred by the RBM guides us to sample new sequences $\bf v$ with putative functionalities. In practice, we sample from the conditional distribution $P( {\bf v} | {\bf h} )$, Eqn. (3), where a few hidden-unit activities in the representation $\bf h$ are fixed to desired values, while the others are sampled from Eqn. (4). For WW domains, we condition on the activities of hidden units 3 and 4, related to binding specificity. Fixing $h_3$ and $h_4$ to levels corresponding to the peaks in the histograms of inputs in Fig. 3C allows us to generate sequences belonging specifically to each one of the three ligand-specificity clusters, see Fig. 5A.
In addition, sequences with combinations of activities that are not encountered in the natural MSA can be engineered. As an illustration, we generate by conditional sampling hybrid WW-domain sequences with strongly negative values of $h_3$ and $h_4$, corresponding to a Type I-like $\beta_2$-$\beta_3$ binding pocket and a long, Type IV-like $\beta_1$-$\beta_2$ loop, see Fig. 5A&B.
For Kunitz domains, the property ‘no 11-35 disulfide bond’ holds only for some sequences of nematode organisms, whereas the Bikunin-AMBP gene is present only in vertebrates; they are thus never observed simultaneously in natural sequences. Sampling our RBM conditioned to appropriate levels of $h_2$ and $h_5$ allows us to generate sequences with both features activated, see Figs. 5C&D.
The sequences designed by RBM are far away from all natural sequences in the MSA, but have comparable probabilities, see Figs. 5E (WW) and 5F (Kunitz). Their probabilities estimated with pairwise direct-coupling models (trained on the same data), whose ability to identify functional and artificial sequences has already been tested [@balakrishnan2011learning; @cocco2018inverse], are also large, see Appendix 1, Fig. 7.
Our RBM framework can also be modified to design sequences with very high probabilities, even larger than in the MSA, by appropriate duplication of the hidden units (Methods). This trick can be combined with conditional sampling, see Fig. 5E&F.
Contact Predictions {#contact-predictions .unnumbered}
-------------------
As illustrated above, co-occurrence of large weight components in highly sparse features often corresponds to nearby sites on the 3D fold. To extract structural information in a systematic way, we use our RBM to derive effective pairwise interactions between sites, which can then serve as estimators for contacts as in direct-coupling based approaches [@cocco2018inverse]. The derivation is sketched in Fig. 6A. We consider a sequence ${\bf v}^{a,b}$ with residues $a$ and $b$ on, respectively, sites $i$ and $j$. Single mutations $a\to a'$ or $b\to b'$ on, respectively, site $i$ or $j$ are accompanied by changes in the log probability of the sequence indicated by the full arrows in Fig. 6A. Comparing the change resulting from the double mutation with the sum of the changes resulting from the two single mutations provides our RBM-based estimate of the epistatic interaction, see Eqns. (10,11) in Methods. These interactions are well correlated with the outcomes of the Direct-Coupling Analysis, see Appendix 1, Fig. 9.
Figure 6 shows that the quality of prediction of the contact maps of the Kunitz (panel B) and the WW (panel C) domains with RBM is comparable to state-of-the-art methods based on direct couplings [@morcos2011direct]; predictions for long-range contacts are reported in Appendix 1, Fig. 10. The quality of contact prediction with RBM
- does not seem to depend much on the choice of the hidden-unit potential, compare the Gaussian and dReLU PPV performances in Figs. 6B,C&D, though the latter have better performance in terms of sequence scoring than the former, see Appendix 1, Figures 1, 2&5.
- strongly increases with the number of hidden units, see Appendix 1, Fig. 11,12. This dependence is not surprising, as the number $M$ of hidden units acts in practice as a regularizor over the effective coupling matrix between residues. In the case of Gaussian RBM, the value of $M$ fixes the maximal rank of the matrix $J_{ij}(v_i,v_j)$, see Methods. The value $M=100$ of the number of hidden units is small compared to the maximal ranks $R=20\times N$ of the couplings matrices of the Kunitz ($R=1060$) and WW ($R=620$) domains, and explains why Direct-Coupling Analysis gives slightly better performance than RBM in the contact predictions of Figs. 6B&C.
- worsens with stronger weight-sparsifying regularizations, see Appendix 1, Fig. 12, as expected.
[ ]{} [@ekeberg2014fast][ ]{} $M = 0.3\, R$, *i.e.* [ ]{} $\lambda_1^2=0.1$. [ ]{}
![[**Contact predictions using RBM.** ]{} [**A.**]{} Sketch of the derivation of effective epistatic interactions between residues with RBM. The change in log probability resulting from a double mutation (purple arrow) is compared to the sum of the changes accompanying the single mutations (blue and red arrows), see text and Methods, Eqns. (10,11). [**B**]{}. Positive Predictive Value (PPV) vs. pairs $(i,j)$ of residues, ranked according to their scores for the Kunitz domain. RBM predictions with quadratic (Gaussian RBM) and dReLU potentials are compared to direct coupling-based methods – Pseudo-Likelihood Method (plmDCA) [@ekeberg2014fast], and Boltzmann Machine (BM) learning [@Sutto13567]. [**C.**]{} Same as panel [**B**]{} for the WW domain. [**D.**]{} [ ]{} []{data-label="fig5"}](figure6.pdf)
Benchmarking on Lattice Proteins {#benchmarking-on-lattice-proteins .unnumbered}
--------------------------------
Lattice protein (LP) models were introduced in the $90's$ to study protein folding and design [@mirny2001]. In one of those models [@shakhnovich1990enumeration], a ‘protein’ of $N=27$ amino acids may fold into $\sim 10^5$ distinct structures on a $3\times 3\times 3$ cubic lattice, with probabilities depending on its sequence (Methods and Figs. 7A&B). LP sequence data were used to benchmark the Direct-Coupling Analysis in [@jacquin2016benchmarking], and we here follow the same approach to assess the performances of RBM in a case where the ground truth is known. We first generate a MSA containing sequences having large probabilities ($p_{nat}>0.99$) of folding into one structure shown in Fig. 7A [@jacquin2016benchmarking]. A RBM with $M=100$ dReLU hidden units is then learned, see Appendix 1 for details about regularization and cross-validation.
Various structural LP features are encoded by the weights as in real proteins, including complex negative-design related modes, see Figs. 7C&D and the remaining weights in Supporting Information. Performances in terms of contact predictions are comparable to state-of-the art methods on LP, see Appendix 1, Fig. 11.
The capability of RBM to design new sequences with desired features and high values of fitness, exactly computable in LP as the probability of folding into the native structure in Fig. 7A, can be quantitatively assessed. Conditional sampling allows us to design sequences with specific hidden-unit activity levels, or combinations of features not found in the MSA (Fig. 7E). These designed sequences are diverse and have large fitnesses, comparable to the MSA sequences and even higher when generated by duplicated RBM (Fig. 7F), and well correlated with the RBM probabilities $P({\bf v})$ (Appendix 1, Fig. 6).
![[**Benchmarking RBM with lattice proteins.**]{} [**A.**]{} $S_A$, one of the $103,406$ distinct structures that a 27-mer can adopt on the cubic lattice [@shakhnovich1990enumeration]. Circled sites are related to the features shown in panel 6C. [**B.**]{} $S_G$, another fold with a contact map (set of neighbouring sites) close to $S_A$ [@jacquin2016benchmarking]. [**C.**]{} Four weight logos for a RBM inferred from sequences folding into $S_A$, see Supporting Information for the remaining 96 weights. Weight 1 corresponds to the contact between sites 3 and 26, see black dashed contour in panel A; the contact can be realized by amino acids of opposite (-+) charges ($I_1>0$), or by hydrophobic residues ($I_1<0$). Weights 2 and 3 are related to, respectively, the triplets of amino acids 8-15-27 and 2-16-25, each realizing two overlapping contacts on $S_A$ (blue dashed contours). Weight 4 codes for electrostatic contacts between 3-26, 1-18 and 1-20, and imposes that the charges of amino acids 1 and 26 have same sign. The latter constraint is not due to the native fold (1 and 26 are ‘far away’ on $S_A$) but impedes folding in the ‘competing’ structure, $S_G$ (Fig. 7B and Methods), in which sites 1 and 26 are neighbours [@jacquin2016benchmarking]. [**D.**]{} Distributions of inputs $I$ and average activities (full line, left scale). All features are activated across the entire sequence space (not shown). [**E.**]{} Conditional sampling with activities $(h_2,h_3)$ fixed to $(h_2^\pm, h_3^\pm)$, see red dots in panel D. Designed sequences occupy specific clusters in the sequence space, corresponding to different realizations of the overlapping contacts encoded by weights 2 and 3 (panel 6C). Conditioning to $(h_2^-, h_3^+)$ makes possible to generate sequences combining features not found together in the MSA, see bottom left corner, even with very high probabilities (Methods). [**F.**]{} Scatter plot of the number of mutations to the closest natural sequence vs. the probability $p_{nat}$ of folding into structure $S_A$ (see [@jacquin2016benchmarking] for a precise definition) for natural (gray) and artificial (colored) sequences. Note the large diversity and the existence of sequences with higher $p_{nat}$ than in the training sample. []{data-label="fig7"}](figure7.pdf)
Cross-validation of the model and interpretability of the representations
-------------------------------------------------------------------------
Each RBM was trained on a randomly chosen subset of 80% of the sequences in the MSA, while the remaining 20% (called test set) were used for validation of its predictive power. In practice, we compute the average log-probability of the test set to assess the performances of the RBM for various values of the number $M$ of hidden units, of the regularization strength $\lambda_1^2$ and for different hidden-unit potentials. Results for the WW and Kunitz domains and for Lattice Proteins are reported in Fig. 8 and in Appendix 2 (Model Selection). The dReLU potential, which include quadratic and Bernoulli (another popular choice for RBM) potentials as special cases, is consistently better than the latters. As expected, increasing $M$ allows RBM to capture more features in the data distribution and, therefore, improves performances up to some level, after which overfitting starts to occur.
The impact of the regularization strength $\lambda_1^2$ favoring weight sparsity (see definition in Methods Eqn. (8)) is twofold, see Fig. 8A for the WW domain. In the absence of regularization ($\lambda_1^2=0$) weights have components on all sites and residues, and the RBM overfit the data, as found from the large difference between the log-probabilities of the training and test sets. [ ]{}. Imposing mild regularization allows the RBM to avoid overfitting and maximize the log-probability of the test set ($\lambda_1^2=0.03$ in Fig. 8A), but most sites and residues carry non-zero weights. Interestingly, imposing stronger regularizations has low impact on the generalization abilities of RBM (weak decrease of test set log-probability), while making weights much sparser ($\lambda_1^2=0.25$ in Fig. 3). For too large regularizations, too few non-zero weights remain available and the RBM is not powerful enough to adequately model the data (drop in log-probability of the test set).
Favoring sparser weights in exchange for a small loss in log-probability has a deep impact on the nature of the representation of the sequence space by the RBM, see Fig. 8B. Good representations are expected to capture invariant properties of sequences across evolutionary divergent organisms, rather than idiosyncratic features attached to a limited set of sequences (mixture model in Fig. 8C). For sparse enough weights, the RBM is driven into the compositional representation regime (see [@tubiana2017emergence]) of Fig. 8E, in which each hidden unit encodes a limited portion of a sequence and the representation of a sequence is defined by the set of hidden units with strong inputs. Hence, the same hidden unit (e.g. weights 1 and 2 coding for the realizations of contacts in the Kunitz domain in Fig. 2B) can be recruited in many parts of the sequence space corresponding to very diverse organisms (see bottom histograms attached to weights 1 and 2 in Fig. 2C, showing that the sequences corresponding to strong inputs are scattered all over the sequence space). In addition, silencing or activating one hidden unit affects only a limited number of residues (contrary to the entangled regime of Fig. 8D), and a large diversity of sequences can be generated through combinatorial choices of the activity states of the hidden units, which guarantees efficient sequence design.
[ ]{}[@ekeberg2014fast] [ ]{}
- [ ]{}
- [ ]{} $\alpha$[ ]{} $\sim 4$[ ]{}
- [ ]{} $\beta$[ ]{}
[ ]{}
- [ ]{}
- [ ]{}
- [ ]{}$\beta-$[ ]{}
- [ ]{}
[ ]{}
![ [**Nature of representations built by RBM and interpretability of weights.**]{} [**A.**]{} *Effect of sparsifying regularization.* Left: log-probability (Methods, Eqn. (8)) as a function of the regularization strength $\lambda_1^2$ (square root scale) for RBM with $M=100$ hidden units trained on WW domain sequence data. Right: Weights attached to three representative hidden units are shown for $\lambda_1^2=0$ (no regularization) and 0.03 (optimal log-likelihood for the test set, see left panel); weights shown in Fig. 3 were obtained at higher regularization $\lambda_1^2=0.25$. For larger regularization, too many weights vanish, and the log-likelihood diminishes. [**B.**]{} Sequences (purple dots) in the MSA attached to a protein family define a highly sparse subset of the sequence space (symbolized by the blue square), from which a RBM model is inferred. The RBM then defines a distribution over the entire sequence space, with high scores for natural sequences and over many more other sequences putatively belonging to the protein family. The representations of the sequence space by RBM can be of different types, three examples of which are sketched in the following panels. [**C.**]{} *Mixture model:* each hidden unit focuses on a specific region in sequence space (color ellipses, different colors correspond to different units), and the attached weights form a template for this region. The representation of a sequence thus involves one (or a few) strongly activated hidden units, while all remaining units are inactive. [**D.**]{} *Entangled model:* all hidden units are moderatly active across the sequence space. The pattern of activities vary from one sequence to another in a complex manner. [**E.**]{} *Compositional model:* a moderate number of hidden units are activated for each protein sequence, each recognizing one of the motifs (shown by colors) in the sequence and controlling one of the protein biological properties. Composing the different motifs in various way (right circled compositions) generates a large diversity of sequences. []{data-label="fig8"}](figure8.pdf)
![ [**Representative weights of the protein families selected in**]{} [@ekeberg2014fast] \[coding for structural properties\]. [ ]{} []{data-label="fig9"}](figure9.png)
![ [**Representative weights of the protein families selected in**]{} [@ekeberg2014fast] \[coding for non-structural properties\]. [ ]{} []{data-label="fig10"}](figure10.png)
Discussion {#discussion .unnumbered}
==========
In summary, we have shown that RBM are a promising, versatile, and unifying method for modeling and generating protein sequences. RBM, when trained on protein sequence data, reveal a wealth of structural, functional and evolutionary features. To our knowledge, no other method has been able to extract such detailed information in a unique framework so far. In addition, RBM can be used to design new sequences: hidden units can be seen as representation-controlling knobs, tunable at will to sample specific portions of the sequence space corresponding to desired functionalities. A major and appealing advantage of RBM is that the two-layer architecture of the model embodies the very concept of genotype-phenotype mapping (Fig. 1C). Codes for learning and visualizing RBM are attached to this publication (Methods).
From a machine-learning point of view, the values of RBM defining parameters (class of potentials and number $M$ of hidden units, regularization penalties) were selected based on the log-probability of a test set of natural sequences not used for training and on the interpretability of the model. The dReLU potentials we have introduced in this work (Eqn. (6)) consistently outperform other potentials for generative purposes. As expected, increasing $M$ improves likelihood up to some level, after which overfitting starts to occur. Adding sparsifying regularization not only prevents overfitting but also facilitates the biological interpretation of weights (Fig. 8A). It is thus an effective way to enhance the correspondence between representation and phenotypic spaces (Fig. 1C). It also allows us to drive the RBM operation point in which most features can be activated across many regions of the sequence space (Fig. 8E); examples are provided by hidden units 1 and 2 for the Kunitz domain in Figs. 2B&C and hidden unit 3 for the WW domain in Figs. 3B&C. Combining these features allows us to generate a variety of new sequences with high probabilities, such as those shown in Fig. 5. Note that some inferred features, such as hidden unit 5 in Figs. 2C&D and, to a lesser extent, hidden unit 2 in Figs. 3B&C, are, on the contrary, activated by evolutionary close sequences. Our inferred RBMs thus share some partial similarity with the mixture models of Fig. 8C. Interestingly, the identification of specific sequence motifs with structural, functional or evolutionary meaning does not seem to be restricted to a few protein domains or proteins, but could be a generic property as suggested by our study of 16 more families (Figs. 9&10).
Despite the algorithmic improvements developed in the present work (Methods), training RBM is challenging as it requires intensive sampling. Generative models alternative to RBM and not requiring Markov Chain sampling exist in machine learning, such as Generative Adversarial Networks [@goodfellow2014generative] and Variational Auto–encoders (VAE) [@kingma2013auto]. VAE were recently applied to protein sequence data for fitness prediction [@novak; @marks]. Our work differs in several important points: our RBM is an extension of direct-based coupling approaches, requires much less hidden units (about 10 to 50 times less than [@novak] and [@marks]), has a simple architecture with two layers carrying sequences and representations, infers interpretable weights with biological relevance, and can be easily tweaked to design sequences with desired statistical properties. We have shown that RBM can successfully model small domains (with a few tens of amino acids) as well as much longer proteins (with several hundreds of residues). The reason is that, even for very large proteins, the computational effort can be controlled through the number $M$ of hidden units, see Methods for discussion about the running time of our learning algorithm. Choosing moderate values of $M$ makes the number of parameters to be learned reasonable and avoids overfitting, yet allowing for the discovery of important functional and structural features. It is, however, unclear how $M$ should scale with $N$ to unveil ‘all’ the functional features of very complex and rich proteins (such as Hsp70).
From a computational biology point of view, RBM unifies and extends previous approaches in the context of protein coevolutionary analysis. From the one hand, the features extracted by RBM identify ‘collective modes’ controlling the biological functionalities of the protein, in a similar way to the so-called sectors extracted by statistical coupling analysis [@halabi2009protein]. However, contrary to sectors, the collective modes are not disjoint: a site may participate to different features, depending on the value of the residue it carries. On the other hand, RBM coincide with direct-coupling analysis [@morcos2011direct] when the potential ${\cal U}(h)$ is quadratic in $h$. For non-quadratic potentials ${\cal U}$, couplings to all orders between the visible units are present. The presence of high-order interactions allows for a significantly better description of gap modes [@feinauer2014improving], of multiple long-range couplings due to ligand binding, and of outliers sequences (Appendix 1, Fig. 5). Our dReLU RBM model offers an efficient way to go beyond pairwise coupling models, without an explosion in the number of interaction parameters to be inferred, as all high-order interactions (whose number, $~q^N$, is exponentially large in $N$) are effectively generated from the same $M\times N\times q$ weights $w_{i\mu}(v)$. [ ]{} [@cocco2013principal], [ ]{}
[ ]{} and could be unambiguously interpreted and related to existing literature. However, the biological significance of some of the inferred features remains unclear, and would require experimental investigation. Similarly, the capability of RBM to design new functional sequences need experimental validation besides the comparison with past design experiments (Fig. 5E) and the benchmarking on *in silico* proteins (Fig. 7). While recombining different parts of natural proteins sequences from different organisms is a well recognized procedure for protein design [@stemmer1994rapid; @khersonsky2016reinvent], RBM innovates in a crucial aspect. Traditional approaches cut sequences into fragments at fixed positions based on secondary structure considerations, but such parts are learnt and need not be contiguous along the primary sequence in RBM models. We believe protein design with detailed computational modeling methods, such as Rosetta [@simons1997assembly; @khersonsky2016reinvent], could be efficiently guided by our RBM-based approach, in much the same way protein folding greatly benefited from the inclusion of long-range contacts found by direct-coupling analysis [@marks2011protein; @hopf2012three].
Future projects include developing systematic methods for identifying function-determining sites, and analyzing more protein families. [ ]{}. In addition, it would be very interesting to use RBM to determine evolutionary paths between two, or more, protein sequences in the same family, but with distinct phenotypes. In principle, RBM could reveal how functionalities continuously change along the paths, and provide a measure of viability of intermediary sequences.
Materials and Methods
=====================
Data preprocessing
------------------
We use the PFAM sequence alignments of the V31.0 release (March 2017) for both Kunitz (PF00014) and WW (PF00397) domains. All columns with insertions are discarded, after which duplicate sequences are removed. We are left with, respectively, $N=53$ sites and $B= 8062$ unique sequences for Kunitz, and $N=31$ and $B=7503$ for WW; each site can carry $q=21$ different symbols. To correct for the heterogeneous sampling of the sequence space, a reweighting procedure is applied: each sequence ${\bf v}^\ell$ with $ \ell=1,...,B$ is assigned a weight $w_\ell$ equal to the inverse of the number of sequences with more than $90\%$ amino-acid identity (including itself). In all that follows, the average over the sequence data of a function $f$ is defined as $$\langle f({\bf v}) \rangle_{MSA}=\left(\sum_{\ell=1}^B w_\ell \; f({\bf v^\ell})\right)\bigg/\left(\sum_{\ell=1}^B w_\ell \right)\ .$$
Learning procedure
------------------
### Objective function and gradients
Training is performed by maximizing, through stochastic gradient ascent, the difference between the log-probability of the sequences in the MSA and the regularization costs, $$\label{cost_total}
\langle \log P({\bf v})\rangle_{MSA} - \frac{\lambda_f}{2} \sum_{i,v} g_i(v)^2 - \frac{\lambda_1^2}{2 q N} \sum_\mu \left( \sum_{i,v} | w_{i\mu}(v)| \right)^2\ ,$$ Regularization termes include a standard $L_2$ penalty for the potentials acting on the visible units, and a custom $L_2/L_1$ penalty for the weights. The latter penalty corresponds to an effective $L_1$ regularization with an adaptive strength increasing with the weights, thus promoting homogeneity among hidden units [^1]. Besides, it prevents hidden units from ending up entirely disconnected ($w_{i\mu}(v) = 0 \; \forall i, v$), and makes the determination of the penalty strength $\lambda_1^2$ more robust, see Appendix 1, Fig. 2.
According to Eqn. (5), the probability of a sequence $\bf v$ can be written as, $$\label{pdev}
P ({\bf v}) = e^{-E_\text{eff}({\bf v})} \bigg/ \bigg(\sum _{{\bf v}'}e^{-E_\text{eff}({\bf v}')}\bigg)\ , \quad \text{where}\quad E_\text{eff}({\bf v}) = - \sum_{i=1}^N g_i(v_i) - \sum _{\mu=1}^M \Gamma \big( I_\mu({\bf v})\big)$$ is the effective ‘energy’ of the sequence, which depends on all the model parameters. The gradient of $\langle \log P({\bf v})\rangle_{MSA}$ over one of these parameters, denoted generically by $\psi$, is therefore $$\label{grad56}
\frac{\partial }{\partial\psi} \langle \log P({\bf v})\rangle_{MSA} =\sum_{\bf v} P({\bf v}) \frac{\partial E_\text{eff}}{\partial\psi} ({\bf v})- \bigg\langle \frac{\partial E_\text{eff}}{\partial\psi} ({\bf v})\bigg\rangle_{MSA} \ .$$ Hence, the gradient is the difference between the average values of the derivative of $E_{eff}$ with respect to $\psi$ over the model and the data distributions.
### Moment evaluation
[ ]{}(\[grad56\]) [@fischer2012introduction]. [ ]{}[@ackley1987learning]. [ ]{} [@tieleman2008training][ ]{}($N_{MC} \sim 10$) [ ]{}[@parallel; @tempering; @future_algo].
### Stochastic Gradient Ascent
[ ]{}($=100$) [ ]{}
- Weights $w_{i\mu}(v)$, are randomly and independently drawn from a Gaussian distribution with zero mean and variance equal to $ \frac{0.1}{N}$. The scaling factor $\frac{1}{N}$ ensures that the initial input distribution has variance of the order of $1$.
- The potentials $g_i(v)$ are given their values in the independent-site model: $g_i(v) = \log {\left\langle \delta_{v_i,v}\right\rangle}_{\text{MSA}}$, where $\delta$ denotes the Kronecker function.
- For all hidden-unit potentials, we set $\gamma_+=\gamma_-=1$, $\theta_+=\theta_-=0$.
[ ]{}$0.1$[ , and decays exponentially after a fraction of the total training time (e.g. 50%) until it reaches a final, small value, e.g. ]{}$10^{-4}$.
### Dynamic reparametrization
[ ]{}$\gamma_\mu \rightarrow \lambda^2 \gamma_\mu$, $w_{i\mu} \rightarrow \lambda w_{i\mu}$, $\theta_\mu \rightarrow \lambda \theta_\mu$ [ ]{}$\theta_\mu \rightarrow \theta_\mu + K_\mu$, $g_i \rightarrow g_i - \sum_\mu w_{i\mu} \frac{K_\mu}{\gamma_\mu}$. [ ]{}$\gamma_\mu =1, \; \theta_\mu=0 \; \forall \mu$ [ ]{}${\left\langle h_\mu\right\rangle} \sim 0$ and $\text{Var}(h_\mu) \sim 1$ [ ]{} [@future_algo].
### Gauge choice
[ ]{}$g_i(v) \rightarrow g_i(v) + \lambda_i$ and $w_{i\mu}(v) \rightarrow w_{i\mu}(v) + K_{i\mu}$ [ ]{}$\sum_v g_i(v) = 0$, $\sum_v w_{i\mu}(v) = 0$. [ ]{}$g_i(v) \rightarrow g_i(v) - \frac{1}{q} \sum_{v'} g_i(v')$ [ ]{}$w_{i\mu}(v) - \frac{1}{q} \sum_{v'} w_{i\mu}(v')$ [ ]{}
### Evaluating the partition function {#evaluate_Z}
[ ]{}$P({\bf v})$ [ ]{} $Z = \displaystyle{\sum_{\bf v} \exp \left( -E_\text{eff}({\bf v}) \right)}$, [ ]{}Eqn. (\[pdev\]). [ ]{}$q^N$ [ ]{} $Z$ [ ]{} [@neal2001annealed; @salakhutdinov2008quantitative]. [ ]{}$P_1({\bf v}) = \frac{P_1^*({\bf v})}{Z_1}$, $P_0 = \frac{P_0^*({\bf v})}{Z_0}$ [ ]{}$Z_1$, $Z_0$[ ]{} $${\left\langle \frac{P_1^*({\bf v})}{P_0^*({\bf v})} \right\rangle}_{ {\bf v} \sim P_0 } = \sum_{ {\bf v}} \frac{P_1^*({\bf v})}{P_0^*({\bf v})} \frac{P_0^*({\bf v})}{Z_0} = \frac{1}{Z_0} \sum_{ {\bf v}} P_1^*({\bf v}) = \frac{Z_1}{Z_0}$$ [ ]{}$Z_0$ [ ]{}$P_0$[ ]{}$Z_1$ [ ]{}$P_1$, $P_0$ [ ]{} $P_1$ [ ]{} $P_0$[ ]{}${\left\langle .\right\rangle}$[ ]{} $P_\beta ({\bf v})= P_1({\bf v})^\beta\; P_0({\bf v})^{1-\beta}$, [ ]{} $Z_1$ [ ]{} $$Z_1 = \frac{Z_1}{Z_{\beta_{l_{max}}}} \frac{Z_{\beta_{l_{max}-1}}}{Z_{\beta_{l_{max}-2}}} ... \frac{Z_{\beta_1}}{Z_0} \times Z_0\ ,$$ [ ]{}$\beta_l = \frac{l}{l_{\text{max}}}$. [ ]{}$C$ [ ]{} $P_0$[ ]{}$P_0$ [ ]{}$P_1$[ ]{}$P_\beta$[ ]{}$\log \frac{Z_1}{Z_0} \approx {\left\langle \log \frac{P_1^*({\bf v})}{P_0^*({\bf v})} \right\rangle}_{ {\bf v} \sim P_0 }$[ ]{}$P_1$[ ]{}$P_0$[ ]{}$C=20$ [ ]{}$n_\beta = 5\times 10^4$ [ ]{}$P_0$[ ]{} $P_{MSA}$. [ ]{} ${\bf w}^{\beta =0} = 0$. [ ]{}
### Explicit formula for sampling and training RBM
[ ]{}$P({\bf v}|{\bf h})$, [ ]{} $P({\bf h}|{\bf v})$[ ]{} $E_{\text{eff}}({\bf v})$ [ ]{}
1. [ ]{} $i$ [ ]{}$v_i \in [0,20]$[ ]{}$N \times 21$ [ ]{}$M \times N \times 21$ [ ]{}$P({\bf v} | {\bf h}) = \prod_i P({\bf v_i} | {\bf h})$[ ]{} $P({\bf v} | {\bf h})$ [ ]{} $P(v_i |{\bf h})$ [ ]{}
2. [ ]{} $P({\bf h} | {\bf v} )$ [ ]{} $\bf v$[ ]{}$P(h_\mu | {\bf v})$[ ]{} ${\cal U}(h)=\frac12 \gamma h^2 + \theta h$, [ ]{} $\mathcal{U}(h)$ [ ]{} $\Phi(x) = \exp(\frac{x^2}{2}) \left[1- \text{erf}(\frac{x}{\sqrt{2}} ) \right] \sqrt{\frac{\pi}{2}}$
[ ]{}$\Phi$ [ ]{}
- $\Phi(x) \sim_{x \rightarrow -\infty} \exp(\frac{x^2}{2}) \sqrt{2\pi} $
- $\Phi(x) \sim_{x \rightarrow \infty} \frac{1}{x} - \frac{1}{x^3} + \frac{3}{x^5} + \mathcal{O}(\frac{1}{x^7})$
- $\Phi'(x) = x \Phi(x) - 1$
[ ]{}$\Phi$ [ ]{}$x<5$ [ ]{}$\mathcal{TN}(\mu,\sigma^2,a,b)$ [ ]{} $\mu$[ ]{}$\sigma$ [ ]{}$[a,b]$.
[ ]{} $P(h|I)$ [ ]{} $$P(h| I) = p^+ \mathcal{TN} \left(\frac{I - \theta^+}{\gamma_+},\frac{1}{\gamma_+}, 0,+\infty \right) + p^- \mathcal{TN} \left(\mu = \frac{I - \theta^-}{\gamma^-},\sigma^2 = \frac{1}{\gamma^-},- \infty,0 \right)$$ where $Z^\pm = \Phi \left( \frac{\mp (I - \theta^\pm)}{\sqrt{\gamma^\pm}} \right) \frac{1}{\sqrt{\gamma^\pm}}$, and $p^\pm = \frac{Z^\pm}{Z^++Z^-}$.
3. [ ]{} $E_{\text{eff}}$ [ ]{} $\Gamma(I)$. [ ]{} $\Gamma(I)$ [ ]{}$\Gamma(I) = \log (Z^+ + Z^-)$ [ ]{}$Z^\pm$ [ ]{}.
### Computational complexity
The computational complexity is of the order of $M \times N \times B$, with more accurate variants taking more time. The algorithm scales reasonably to large protein sizes, and was tested successfully for $N$ up to $\sim 700$, taking of the order of 1-2 days on an Intel Xeon Phi processor with 2 $\times$ 28 cores.
Sampling procedure {#sectraining}
------------------
Sampling from $P$ in Eqn. (5) is done with Markov Chain Monte Carlo methods, with the standard alternate Gibbs sampler described in the main text and in [@fischer2012introduction]. Conditional sampling, *i.e.* sampling from $P({\bf v} | h_\mu = h_\mu^c)$ is straightforward with RBM: it is achieved by the same Gibbs sampler while keeping $h_\mu$ fixed.
The RBM architecture can be modified to generate sequences with high probabilities, such as in Figs. 5E&F. The trick is to duplicate the hidden units, the weights, and the local potentials acting on the visible units, as shown in Methods, Fig. 1. By doing so, the sequences $\bf v$ are distributed according to $$P_2({\bf v}) \propto \int \prod_{\mu} dh_{\mu1}\,dh_{\mu2} \; P({\bf v} | {\bf h}_1)\, P({\bf v} | {\bf h}_2) = P({\bf v})^2 \ .$$ Hence, with the duplicated RBM, sequences with high probabilities in the original RBM model are given a boost compared to low-probability sequences. Note that more subtle biases can be introduced by duplicating some (but not all) of the hidden units in order to give more importance in the sampling to the associated statistical features.
{width=".6\columnwidth"} \[figMethods2\]
Contact map estimation
----------------------
RBM can be used for contact prediction in a manner similar to pairwise coupling models, after derivation of an effective coupling matrix $J_{ij}^{\text{eff}}(a,b)$. Consider a sequence $\bf v$, and two sites $i,j$. Define the set of mutated sequences ${\bf v}^{a,b}$ with amino acid content: $v^{a,b}_k = v_k$ if $k \neq i,j$, $a$ if $k = i$, $b$ if $k=j$ (Fig. 6A). The differential likelihood ratio $$\label{epistasis}
\Delta \Delta R_{ij}({\bf v}; a,a',b,b') \equiv \log \left[ \frac{P( {\bf v}^{a,b}) \, P({\bf v}^{a',b'})}{P({\bf v}^{a',b}) \, P({\bf v}^{a,b'})} \right] \ ,$$ where $P$ is the marginal distribution in Eqn. (\[marginal\]), measures epistatic contributions to the double mutation $a \rightarrow a'$ and $b \rightarrow b'$ on, respectively, sites $i$ and $j$ in the background defined by sequence $\bf v$, see Fig. 6A. The effective coupling matrix is then defined as $$\label{couplings_from_epistasis}
J_{ij}^{\text{eff}}(a,b) = {\left\langle \frac{1}{q^2}\sum_{a',b'} \Delta \Delta R_{ij}({\bf v}; a,a',b,b')\right\rangle}_{MSA} \ ,$$ where the average is taken over the sequences $\bf v$ in the MSA. For a pairwise model, $\Delta \Delta R_{ij}$ does not depend on the background sequence ${\bf v}$, and Eqn. (\[couplings\_from\_epistasis\]) coincides with the true coupling in the zero-sum gauge. Contact estimators are based on the Frobenius norms of $J^{\text{eff}}$, with the Average Product Correction, see [@cocco2018inverse].
Code availability
-----------------
The Python 2.7 package for training and visualizing RBMs, used to obtained the results reported in this work, is available at <https://github.com/jertubiana/ProteinMotifRBM>. In addition, Jupyter notebooks are provided for reproducing most figures of the article.
Acknowledgments
===============
We thank D. Chatenay for useful comments on the manuscript and L. Posani for his help on lattice proteins. This work was partly funded by the ANR project RBMPro CE30-0021-01.
Supporting files
================
- Supporting file 1 (pdf): Weight logo for all hidden units inferred from the Kunitz domain MSA.
- Supporting file 2: Weight logo for all hidden units inferred from the WW domain MSA.
- Supporting file 3: Weight logo for all hidden units inferred from the LP MSA.
- Supporting file 4: Weight logo of 12 Hopfield-Potts pattern inferred from the Hsp70 protein MSA. Same format as Appendix 1 Figures 14-16.
- Supporting file 5: Weight logo and associated structures of the 10 weights with highest norms, excluding the gap modes for each of the 16 additional domains shown in Fig. 9.
- Supporting file 6: Weight logo and associated structures of the 10 sparse (i.e. within the 30% most sparse weights of the RBM) weights with highest norms, excluding the gap modes for each of the 16 additional domains shown in Fig. 9.
Supporting Methods and Figures
==============================
.5cm
Lattice-protein synthetic sequences
===================================
LP models have been introduced in the $'90$ to investigate the uniqueness of folding shared by the majority of real proteins [@shakhnovich1990enumeration], and have been more recently used to benchmark graphical models inferred from sequence data [@jacquin2016benchmarking]. There are ${\cal N}=103,406$ possible folds, *i.e.* self-avoiding path of the 27 amino-acid-long chains, on a $3 \times 3\times 3$ lattice cube [@shakhnovich1990enumeration]. The probability that the protein sequence ${\bf v}=(v_1,v_2,...,v_{27})$ folds in one of these, say, $S$, is $$P_{nat}({\bf v} ; S)=\frac{\displaystyle e^{-{\cal E}({\bf v} ; S)}}{\displaystyle \sum_{S'=1}^{\cal N} e^{-{\cal E}({\bf v}; S')}} \ ,
\label{pnat}$$ where the energy of sequence ${\bf v}$ in structure $S$ is given by $${\cal E}({\bf v};S) = \sum_{i<j} c_{ij}^{(S)}\; E(v_i,v_j)\ .
\label{energy}$$ In the formula above, $c^{(S)}$ is the contact map: $c^{(S)}_{ij}=1$ if the pair of sites ${ij}$ is in contact, *i.e.* $i$ and $j$ are nearest neighbors on the lattice and zero otherwise. The pairwise energy $E(v_i,v_j)$ represents the amino-acid physico-chemical interactions, given by the the Miyazawa-Jernigan (MJ) knowledge-based potential [@miyazawa1996residue].
A collection of 36,000 sequences that specifically fold on structure $S_A$ (Fig. 7A, Main Text) with high probability $P_{nat}({\bf v};S_A)>0.995$ were generated by Monte Carlo simulations as described in [@jacquin2016benchmarking]. As real MSA, Lattice Protein data feature short- and long-range correlations between amino-acid on different sites, as well as high-order interactions that arise from competition between folds [@jacquin2016benchmarking].
Model selection
===============
We discuss here the choice of parameters (strength of regularization, number of hidden units, shape of hidden-unit potentials, ...) for the RBM used in main text. Our goal is to achieve good generative performances and to learn biologically interpretable representations. We estimate the accuracy of the fit to the data distribution using the average log-likelihood, divided by the number of visible units $\frac{1}{N} \langle \log P({\bf v}) \rangle _{MSA}$. For visible-unit variables with $q=21$ possible values (20 amino acids + gap symbol), this number typically ranges from $-\log 21\simeq -3.04$ (uniform distribution) to 0. Evaluating $P({\bf v})$ (Methods Eqn. 1) requires knowledge of the partition function $Z = \sum_{\bf v} \exp \left( \sum_{i=1}^N g_i(v_i) + \sum_{\mu=1}^M \Gamma_\mu(I_\mu({\bf v})) \right)$, see Section \[evaluate\_Z\].
Number of hidden units
----------------------
The number of hidden units is critical for the generative performance. We trained RBMs on the Lattice Protein data set for various potentials (Bernoulli, quadratic and dReLU), number of hidden units (1-400) and regularizations ($\lambda_1^2 =0$, $\lambda_1^2=0.025$). The likelihood estimation shows that, as expected, the larger $M$, the better the ability to fit the training data (Appendix 1, Fig. 1). Overfitting *i.e.* a decrease in test set performance may occur for large $M$. For the regularized case, the likelihood saturates with about 100 hidden units. Similar results are obtained on WW, see Appendix 1, Fig. 2.
 \[model\_selection\_LP\]
 \[model\_selection\_WW\]
Besides generative performance, the representation also changes as $M$ increases. For very low values of $M$, each hidden unit tries to explain as much covariation as possible and its corresponding weight vector is extended, similarly to PCA. For larger numbers of hidden units, weights tend to become more sparse; they stabilize at some point, after which new hidden units simply duplicate previous ones.
Sparse regularization
---------------------
We first investigate the importance of the sparsifying penalty term. Our study shows that, unlike the case of MNIST digit data [@tubiana2017emergence], sparsity does not arise naturally from training RBM on protein sequences but requires the introduction of a specific sparsifying regularization, see Fig. 8, Main Text. On the one hand, sparse weights, as the ones shown in Fig. 2, 3, 4 and 7 of the Main Text, are easier to interpret, but, on the other hand, regularization generally leads to a decrease in the generative performance. We show below that the choice of regularization strength used in main text is a good compromise between sparsity and generative performance.
 \[sparsity\_performance\]
We train several RBM on the Lattice Proteins MSA, with a fixed number of hidden units ($M=100$), fixed potential, and varying strength of the sparse penalty $\lambda_1^2$ (defined in Methods, Eqn. (8)), and evaluate their likelihoods. We repeat the same procedure using the standard $L_1$ regularization ($\lambda_1 \sum_{i,v,\mu} |w_{i\mu}(v)|$) instead of $L_1^2$. Results are shown in Appendix 1, Fig. 3. In both cases, the likelihood on test set decreases mildly with the regularization strength. However, for $L_1$ regularization, several hidden units become disconnected (*i.e.* $w_{i\mu}(v) = 0$ for all $i,v$) as we increase the penalty strength. The $L_1^2$ penalty achieves sparse weights without disconnecting hidden units when the penalty is too large, hence it is more robust and requires less fine tuning.
Hidden-unit potentials
----------------------
Lastly, we discuss the choice of the hidden-unit potentials. A priori, the major difference between Bernoulli, quadratic and dReLU potentials are that (i) Bernoulli hidden unit take discrete values whereas quadratic and dReLU take continuous ones and (ii) After marginalization, quadratic potentials create pairwise effective interactions whereas Bernoulli and dReLU create non-pairwise ones. It was shown in the context of image processing and text mining that non-pairwise models are more efficient in practice, and theoretical arguments also highlight the importance of high-order interactions [@tubiana2017emergence].
In terms of generative performance, our results on Lattice Proteins and WW domain MSAs show that, for the same number of parameters, dReLU RBM perform better than Gaussian and Bernoulli RBM. Similar results, not shown, were obtained for the Kunitz domain MSA. Although RBM with Bernoulli hidden units are known to be universal approximators as $M \rightarrow \infty$ [@le2008representational], they require more hidden units than the other types; hence more data. This can be intuitively explained by the fact that Bernoulli units cannot naturally express modulation in the degree of presence of a feature. To overcome this issue, one needs more than one hidden unit to encode each feature, as in [@nair2010rectified]. This is consistent with the heavier distribution of hidden units correlations observed in all data sets, see Appendix 1, Fig. 4. For example, for RBM for Bernoulli potentials, 51 out of 100 hidden units encode gap stretches, as opposed to 23 for quadratic and 15 for dReLU potentials; on WW, the numbers are respectively 18, 15 and 9. For both data sets, dReLU encode more efficiently the gap modes.
 \[correlation\_hiddens\]
One of the key aspect that explains the difference of performance between dReLU and Gaussian RBM is the ability of the former to better model ’outlier’ sequences, with rare extended features such as Bikunin-AMBP (Weight 5 in Main Text, Fig. 2) or non-aromatic W28-substitution feature (Weight 3 in Main Text, Fig. 3). Indeed, thanks to the thresholding effect of the average activity, dReLU can account for outliers, without altering the distribution for the bulk of other sequences - unlike quadratic potentials. To illustrate this property, we compare in Appendix 1, Fig. 5 the likelihoods for all sequences of two RBMs trained with quadratic (resp. dReLU) potentials, $M=100$, $\lambda_1^2=0.25$ on the Kunitz domain MSA. The color code measures the degree of anomaly of the sequence, which is obtained as follows:
1. Compute average activity $h_\mu^l$ of dReLU RBM for all data sequences ${\bf v}^l$,
2. Normalize (z-score) each dimension: $\hat{h}_\mu = \frac{h_\mu - {\left\langle h_\mu\right\rangle}_{MSA} }{\sqrt{{\text{Var}[h_\mu]}_{MSA}}}$,
3. Define: $$\label{color_code}
c^l = \arg \max_\mu |\hat{h}_\mu^l|$$
For instance, a sequence ${\bf v}^l$ with $c^l = 10$ has at least one hidden-unit average activity that is 10 standard deviations away from the mean. Clearly, most sequences have very similar likelihood but the outlier sequences are better modeled by dReLU potentials.
 \[dReLU\_vs\_Gaussian\]
The features extracted are fairly robust with respect to the choice of potential when regularization is used. Clearly, the nature of the potentials does not matter for finding contacts features because for any potential, a hidden unit connected to only two sites will create only pairwise effective interaction. For larger collective modes, some difference arise. As discussed above, Bernoulli features are more redundant, and Gaussian RBM tend to miss outlier features.
Summary
-------
To summarize, the systematic study suggests that:
- More general potentials like dReLU perform better than the simpler quadratic and Bernoulli potentials;
- There exist values of sparsity regularization penalties allowing for both good generative performance and interpretability;
- As the number of hidden units increases, more features are captured and generative performance improve. Beyond some point, increasing $M$ simply adds duplicate hidden units and does not enhance performance.
Sequence generation
===================
We use Lattice Proteins to check that our RBM is a good generative model, [*i.e.*]{} is able to generate sequences that have both high fitness and high diversity (far away from one another and from the sequences provided in the training data set), as was done for Boltzmann Machines [@jacquin2016benchmarking]. Various RBM are trained, sequences are generated for each RBM and scored using the ground truth $p_{nat}$, see Appendix 1, Fig. 6. We find that (i) RBMs with low likelihood (Bernoulli and/or small $M$) generate low quality sequences; (ii) Unregularized BMs and RBMs, which tend to overfit, generate sequences with higher fitness but low diversity; (iii) The true fitness function is well predicted by the inferred log probability. Moreover, conditional sampling also generates high-quality sequences, even when conditioning on unseen combination of features.
 \[generative\_LP\]
For RBMs trained on real proteins sequences, no ground-truth fitness is available and sequence quality cannot be assessed numerically. Appendix 1, Fig. 7 shows nonetheless that the generated sequences, including the ones with recombined features that do not appear in nature are consistent with a pairwise model trained on the same data.
 \[generative\_natural\]
[ ]{} $S = - \sum_{ {\bf v} } P({\bf v}) \log P({\bf v})$ [ ]{} $e^{S} \sim 10^{12}$ [ ]{}$10^{18}$ [ ]{}
 \[generative\_regularization\]
Contact predictions
===================
Since RBMs learn a full energy landscape, they can predict epistatic interactions, see Methods, and therefore contacts, as shown in Main Text Fig. 6. The effective couplings derived with RBM are consistent with the ones inferred from a pairwise model, see Appendix 1, Fig. 9. Predictions for distant contacts in the Kunitz domain are shown in Appendix 1, Fig. 10, and are slightly worse than with DCA.
We discuss briefly the best set of parameters for contact prediction. As seen from Appendix 1, Fig. 11, all RBMs can predict more or less accurately contacts maps on Lattice Proteins. As for the likelihood and generative performance, increasing the number of hidden units significantly improves contact prediction. The best hidden unit potentials for predicting contacts are dReLU and quadratic.
We have also studied how constraints on the sparsity of weights, tuned by the regularization penalty $\lambda_1^2$, influenced the performance. Since weights are never exactly zero, proxies are required for an appropriate definition of sparsity. In order to avoid arbitrary thresholds, we use Participation Ratios. The Participation Ratio $(PR_e)$ of a vector ${\bf x}=\{x_i \}$ is $$PR_e(\bf x) = \frac{(\sum_{i} |x_i|^e)^2}{\sum_{i} |x_i|^{2e} }$$ If $\bf x$ has $K$ nonzero and equal (in modulus) components PR is equal to $K$ for any $e$. In practice we use the values $e=2$ and 3: the higher $e$ is, the more small components are discounted against strong components in $\bf x$. Note also that it is invariant under rescaling of ${\bf x}$. We then define the weight sparsity $p_\mu$ of a hidden unit, through $$\label{sparsity}
p_\mu = \frac{1}{N} PR_3( \bf{x_\mu}) \quad \text{with}\quad ({\bf x}_{\mu})_i \equiv \sqrt{\sum_v w_{i\mu}(v)^2} \\$$ and average it over $\mu$ to get a unique estimator of weight sparsity across the RBM. Results are reported in Appendix 1, Fig. 12, and shows that performance strongly worsen when increasing sparsity, both in Lattice Proteins and in real families.
 \[scatter\_JJ\]
 \[Contact\_map\_PF14\]
 \[Contacts\_Predictions\_LP\]
 \[Contacts\_Predictions\_others\]
Feature robustness
==================
[ ]{}
 \[WW\_half\_training1\]
Comparison with the Hopfield-Potts model
========================================
[ ]{}
- [ ]{} [@cocco2013principal], [ ]{}
- [ ]{}
- [ ]{} $p_{\text{nat}}$,[ ]{}-$P({\bf v})$ [ ]{} [@jacquin2016benchmarking].[ ]{}
 \[HP\_Kunitz\]
 \[HP\_WW\]
 \[HP\_LP\]
 \[HP\_LP2\]
 \[HP\_contacts\]
Additional figure: hidden-input distribution for Kunitz domain, separated by phylogenetic identity and genes
============================================================================================================
 \[additional\_figures\_PF14\]
 \[additional\_figures\_PF14\_bis\]
Additional figure: Weight logos, 3D visualizations, input distributions of 10 hidden units for Hsp70
====================================================================================================
Hidden unit numbering: 1 = short vs long loop 2 = function feature on SBD. 3 = LID/SBD interdomain. 4 = NBD/SBD interdomain and non-allosteric specific. 5 = Unstructured tail 6 = short/long vs very short loop 7 = Long loop variant 8 = ER-specific 9 = second non-allosteric specific 10 = Dimer contacts
![Truncated Weight logo of 10 selected HSP70 hidden units (2/2)[]{data-label="additional_figures_HSP70_1"}](appendix1_figure22.png)
![Corresponding structures (3/3). Left: ADP-bound conformation (PDB: 2kho). Right: ATP-bound conformation (PDB: 4jne). For the last hidden unit, we show the structure of the dimer Hsp70/Hsp70 in ATP conformation (PDB: 4JNE), highlighting dimeric contacts.[]{data-label="additional_figures_HSP70_2"}](appendix1_figure25.png)
![Corresponding input distributions. Note that both hidden unit 4 and 9 discriminate the non-allosteric subfamily from the rest; and that hidden unit 8 discriminates Eukaryotic Hsp expressed in the Endoplasmic Reticulum from the rest.[]{data-label="additional_figures_HSP70_3"}](appendix1_figure26.png)
![Some scatter plots of inputs for the 10 hidden units shown.[]{data-label="additional_figures_HSP70_4"}](appendix1_figure27.png)
![[**Statistics of length and amino-acid content of the unstructured tail of Hsp70**]{}. Hidden unit 5 defines a set of sites, mostly located on the unstructured tail of Hsp70; its sequence logo and input distribution suggests that for a given sequence, the tail can be enriched either in tiny (A,G) or hydrophilic amino-acids (E,D,K,R,T,S,N,Q). This is qualitatively confirmed by the non-gaussian statistics of the distributions of fraction of tiny and hydrophilic amino-acids in the tail (blue histograms and top left contour plots). This effect could however be due to the variable length of the loop (bottom histogram). To assess this enrichment, we build a null model where the tail size is random (same statistics as Hsp70), and each amino-acid is drawn randomly, independent from the others, using the same amino-acid frequency as in the tail of Hsp70. The null model statistics (orange histograms and lower left contour plots) are clearly different, validating the collective mode.[]{data-label="additional_figures_HSP70_5"}](appendix1_figure28.png)
[^1]: This can be seen from the gradient of the regularization term, which reads $\lambda_1^2 \left(\sum_{i,v'} | w_{i\mu}(v')|/qN \right) \text{sign}(w_{i\mu}(v) )$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
This report contains revision and extension of some results about RBO from [[@DBLP:journals/corr/abs-1108-5095]]{}. RBO is a simple and efficient broadcast scheduling of $n = 2^k$ uniform frames for battery powered radio receivers. Each frame contains a key from some arbitrary linearly ordered universe. The broadcast cycle – a sequence of frames sorted by the keys and permuted by $k$-bit reversal – is transmitted in a round robin fashion by the broadcaster. At arbitrary time during the transmission, the receiver may start a simple protocol that reports to him all the frames with the keys that are contained in a specified interval of the key values $[\kappa', \kappa'']$. RBO receives at most $2 k + 1$ other frames’ keys before receiving the first key from $[\kappa', \kappa'']$ or noticing that there are no such keys in the broadcast cycle. As a simple corollary, $4 k + 2$ is upper bound the number of keys outside $[\kappa', \kappa'']$ that will ever be received. In unreliable network the expected number of efforts to receive such frames is bounded by $(8 k + 4) / p + 2 (1 - p) / p^2$, where $p$ is probability of successful reception, and the reception rate of the frames with the keys in $[\kappa', \kappa'']$ is $p$ – the highest possible.
The receiver’s protocol state consists of the values $k$, $\kappa'$ and $\kappa''$, one wake-up timer and two other $k$-bit variables. Its only nontrivial computation – the computation of the next wake-up time slot – can be performed in $O (k)$ simple operations, such as arithmetic/bit-wise operations on $k$-bit numbers, using only constant number of $k$-bit variables.
author:
- |
Marcin Kik[^1]\
Faculty of Fundamental Problems of Technology\
Wroc[ł]{}aw University of Technology\
ul. Wybrzeże Wyspiańskiego 27\
PL-50-370 Wroc[ł]{}aw\
Poland
title: 'Notes on Bit-reversal Broadcast Scheduling'
---
=1
Introduction
============
RBO [[@DBLP:journals/corr/abs-1108-5095]]{} is a simple and efficient method of periodic broadcasting of a large sequence of uniform radio messages for radio receivers with a limited source of energy. Examples of such receivers are battery powered sensors or portable devices. In modern devices, the receiver can save the energy by keeping it’s radio device switched off for long periods of time.
The broadcaster transmits in a round robin fashion a large sequence of frames. Such sequence is called a [[[*broadcast cycle*]{}]{}]{}. Each frame is of the same length (we call it a [[[*time slot*]{}]{}]{}) and contains in its header a [[[*key*]{}]{}]{} from an arbitrary linearly ordered universe of key values.
The receiver may decide at arbitrary time (usually somewhere in the middle of the broadcast cycle) to locate and receive all the frames in the stream that contain the keys from some specified range $[\kappa', \kappa'']$. The receiver may [[[*wake-up*]{}]{}]{} (switch on its radio) at arbitrary time slot to receive the transmitted frame. However, the radio consumes energy while it is switched on. We want to minimize the energy dissipated by the receiver, i.e. to minimize the number of the wake-ups. In RBO, the receiver is able to receive all the requested frames transmitted since that moment. Roughly speaking: the receiver listens to some keys of the broadcast cycle and learns the interval of positions in the sorted sequence with the keys in $[\kappa',
\kappa'']$. After that, it only listens in the time slots that contain the keys from these positions.
RBO requires that the length of the broadcast cycle is an integer power of two. This can be achieved by duplicating some of the frames. If $n'$ denotes the number of frames that must be transmitted, then the length of the broadcast cycle is $n = 2^k$, where for integer $k$, $k \geq \lceil \log_2 n' \rceil$.
We assume that the length of each frame is the same, i.e. a single time slot. However, the same key may be repeated many times in the broadcast cycle. Thus, as single long information attributed with some key can be split among many frames with the same key. We can also repeat many times, the frames that that should be delivered more frequently to the receivers. (The frames with the same key are scattered uniformly over the transmission cycle).
The keys may be arbitrary values from arbitrary linearly ordered domain. The receiver does not have any knowledge of the distribution of the keys in the cycle. RBO is energetically efficient for the receiver (Section \[Section-analysis\]), robust to the radio interferences (Section \[Section-unreliable\]), and its implementation is very simple and efficient and requires little memory (Section \[Section-implementation\]), thus it is suitable even for very weak sensor devices (see e.g. [[@TinyOSProgramming]]{}).
This report updates [[@DBLP:journals/corr/abs-1108-5095]]{} as follows:
New, simpler proof of the main theorem (Theorem \[first-hit-Theorem\]) is based on a simpler decomposition of the time-slots sequence.
We focus on the application of the RBO to filtering the frames with the keys from specified interval $[\kappa', \kappa'']$. In Corollary \[Corollary-ee\] we show that the receiver has to listen to no more than $4 k + 2$ frames with keys outside $[\kappa', \kappa'']$, to learn which are the time-slots of the frames with keys in $ [\kappa', \kappa'']$.
The expected energetic costs for the receiver in unreliable network has been estimated in Section \[Section-unreliable\].
A simpler and more efficient algorithm for computing the next wake-up time slot has been proposed in Section \[Section-implementation\].
Example Applications
--------------------
The protocol can be applied to the dissemination of information or to centralized controlling or synchronizing of large populations of energy constrained devices. Some examples are following:
- The keys may be identifiers of records from a huge database transmitted in the stream.
- The keys may be identifiers of the receiver. The broadcaster may send commands or messages to individual receivers.
- The keys may be identifiers of groups of mutually non-interfering sensors. Each frame with such key would contain only the header, while the rest of the time slot can be used for transmission by the sensors from this group.
- The keys may be coordinates of the objects on the plane encoded by Morton z-ordering [[@ZorderMorton]]{}. In such ordering the receiver may limit an approximately square region containing the objects that are interesting to him.
Diverse applications could be mixed within a single stream of frames by assigning to them disjoint intervals of key values. The sorted sequence of keys is permuted by bit-reversal permutation, which scatters the keys from each interval uniformly over the whole stream.
Related Work
-------------
Broadcast scheduling for radio receivers with low [[[*access time*]{}]{}]{} (i.e. the delay to the reception of the required record) and low average [[[*tuning time*]{}]{}]{} (i.e. the energetic cost) was considered by Imielinski, Viswanathan, and Badrinath (see e.g. [[@DBLP:conf/sigmod/ImielinskiVB94]]{}, [[@DBLP:conf/edbt/ImielinskiVB94]]{}, [[@DBLP:journals/tkde/ImielinskiVB97]]{}). In [[@DBLP:conf/edbt/ImielinskiVB94]]{}, [[[*hashing*]{}]{}]{} and [[[*flexible indexing*]{}]{}]{} for finding single records in broadcast cycle have been proposed and compared. In [[@DBLP:journals/tkde/ImielinskiVB97]]{}, a distributed index based on a ordered balanced tree has been proposed. The broadcast sequence consists of two kinds of [[[*buckets*]{}]{}]{}. Groups of [[[*index buckets*]{}]{}]{}, containing parts of the index tree, are interleaved with the groups of [[[*data buckets*]{}]{}]{} containing proper data and a pointer (i.e. time offset) to the next index bucket. Each group of index buckets consists of the copy of upper part of the index tree together with the relevant fragment of the lower part of the tree. This mechanism has found useful application even in more complex scenarios of delivering data to mobile users [[@DBLP:journals/tods/DattaVCK99]]{}.
Khanna and Zhou [[@Khanna2000575]]{} proposed a sophisticated version of the index tree aimed at minimizing [[[*mean*]{}]{}]{} access and tuning time, for given probability of each data record being requested. The broadcast cycle contains multiple copies of data items, so that spacing between copies of each item is related to the optimal spacing, minimizing mean access time derived in [[@DBLP:journals/winet/VaidyaH99]]{}. However the keys are not arbitrary. The key of the item is determined by its probability of being requested. Indexing of broadcast stream for XML documents [[@DBLP:journals/isci/ChungL07]]{} or for full text search [[@DBLP:journals/tkde/ChungYK10]]{} have also been considered.
If the broadcast cycle contains indexing tree structure, then the reception of data in current broadcast cycle depends on the successful reception of the path to this data. Instead of separate index buckets RBO uses short [[[*headers*]{}]{}]{} of the frames. Each such header contains the key assigned to the frame. As a consequence, in unreliable network the receiver has much more chances of efficient navigation towards the desired frames.
In practical applications, due to imperfect synchronization between the broadcaster and the receiver, the header should also contain either the time-slot number or its bit reversal – the index of the frame. To enable changing the contents and the length of the sequence of the transmitted keys by the broadcaster, the header may also include the parameter $k$, such that $2^k$ is the length of the broadcast cycle, and some bits used to notify the receiver that the that the sequence of keys has been changed. For RBO, these issues have been discussed in [[@DBLP:journals/corr/abs-1108-5095]]{}.
Recall that each step of the classic [[[*binary search*]{}]{}]{} algorithm actually clips the interval of the possible locations of the searched key in the sorted sequence of keys. The customary presentation is that the keys of the sequence are organized in a balanced [[[*binary search tree*]{}]{}]{}, and the searched key is compared with a sequence of keys from subsequent levels of this tree. Bit-reversal permutes the sorted sequence of keys so that the broadcast cycle is a sequence of the subsequent levels of a balanced binary search tree for the keys. Moreover, each level is recursively so permuted. We show that it enables efficient search in the periodic transmission of the broadcast cycle even if the search is started at arbitrary time slot. We also exploit this property in the computation of the next time slot that should be listened by the receiver. Bit-reversal permutation has been found useful in many contexts. Some examples of its applications are in FFT algorithm [[@CooleyTukey]]{} [[@CormenLR89]]{}, lock-free extensible hash arrays [[@DBLP:journals/jacm/ShalevS06]]{} distributed arrays in P2P [[@DBLP:conf/infocom/FukuchiSSH09]]{}, address mapping in SDRAM [[@DBLP:conf/scopes/ShaoD05]]{}, scattering of video bursts in transmission scheduling in mobile TV [[@DBLP:journals/ton/HefeedaH10]]{}. In RBO, bit-reversal emerged from updating the recursive definition of the ${\ensuremath{\operatorname{rbo}}}$ permutation used in the underlying ranking procedure in [[@DBLP:conf/adhoc-now/Kik08]]{} in such a way that zero became a fixed point. The simplicity of bit-reversal computation is a great advantage for practical implementations.
Notation and preliminaries
==========================
Let $\mathbbm{Z}$ denote the set of integers. Let $\mathbbm{R}$ denote the set of real numbers. For simplicity and generality, we assume that the keys are from $\mathbbm{R}$. By $[a, b]$ we denote the interval of real numbers $\{x
\in \mathbbm{R}| a \leq x \leq b\}$. If $a > b$ then $[a, b] = \emptyset$. By $[[a, b]]$ we denote we denote $[a, b] \cap \mathbbm{Z}$ (i.e. interval of integers between $a$ and $b$). For a set $S$, we denote the number of its elements by $|S|$.
For $x \in \mathbbm{Z}$, $x \geq 0$, for $i \geq 0$, let ${\ensuremath{\operatorname{bit}}}_i (x)$ be the $i$th least significant bit of the binary representation of $x$, i.e. ${\ensuremath{\operatorname{bit}}}_i (x) = \lfloor (x {\ensuremath{\operatorname{mod}}} 2^{i + 1}) / 2^i \rfloor$. For $l
\geq 0$, a number with binary representation $x_l \ldots x_0$ is denoted by $(x_l, \ldots, x_0)_2$, i.e. $(x_l, \ldots, x_0)_2 = \sum_{i = 0}^l 2^i \cdot
x_i$.
For $x \in [[0, 2^k - 1]]$ let ${\ensuremath{\operatorname{rev}}}_k (x)$ denote the [*bit-reversal*]{} of $x$, i.e: if $x_i = {\ensuremath{\operatorname{bit}}}_i (x)$ then $x = (x_{k - 1}, x_{k - 2}, \ldots,
x_0)_2$ and ${\ensuremath{\operatorname{rev}}}_k (x) = (x_0, x_1, \ldots, x_{k - 1})_2$.
For a set $S \subseteq [[0, 2^k - 1]]$, ${\ensuremath{\operatorname{rev}}}_k S$ denotes the [[[*image*]{}]{}]{} of $S$ under ${\ensuremath{\operatorname{rev}}}_k$, i.e ${\ensuremath{\operatorname{rev}}}_k S
=\{{\ensuremath{\operatorname{rev}}}_k (x) | x \in S\}$.
Let $n$ denote the length of the broadcast cycle, $n = 2^k$, for integer $k
\geq 0$. Let $\kappa_{- 1}, \kappa_0, \ldots, \kappa_{n - 1}, \kappa_n$ be a sequence defined as follows:
$\kappa_{- 1} = - \infty$
$\kappa_n = + \infty$
$\kappa_0, \ldots, \kappa_{n - 1}$ is a sorted sequence of $n$ finite real values of the keys (i.e. $\kappa_i \leq \kappa_{i + 1}$, for $- 1 \leq i \leq n - 1$).
Let ${\ensuremath{\operatorname{KEYS}}} =\{\kappa_0, \ldots, \kappa_{n - 1} \}$ (the set of the values of the keys in the sequence).
Let $\kappa'$ and $\kappa''$ be finite real key values such that $\kappa' \leq
\kappa''$. $[\kappa', \kappa'']$ is the [*interval of the searched keys*]{}.
$E [X]$ denotes expected value of random variable $X$.
The description of the protocol
-------------------------------
The broadcaster at time-slot $t$ broadcasts the frame with the key $\kappa_{{\ensuremath{\operatorname{rev}}}_k (t {\ensuremath{\operatorname{mod}}} n)}$. The receiver searching for the $[\kappa', \kappa'']$ has two variables ${\ensuremath{\operatorname{lb}}}$ and ${\ensuremath{\operatorname{ub}}}$ initialized to $0$ and $n - 1$, respectively. The receiver may start at arbitrary time slot $s$, and executes the following algorithm:
- While ${\ensuremath{\operatorname{lb}}} \leq {\ensuremath{\operatorname{ub}}}$:
- In time-slot $t$ if ${\ensuremath{\operatorname{lb}}} \leq {\ensuremath{\operatorname{rev}}}_k (t {\ensuremath{\operatorname{mod}}} n) \leq
{\ensuremath{\operatorname{ub}}}$, then the receiver receives the message with the key $\kappa =
\kappa_{{\ensuremath{\operatorname{rev}}}_k (t {\ensuremath{\operatorname{mod}}} n)}$ and
- if $\kappa < \kappa'$ then it sets ${\ensuremath{\operatorname{lb}}}$ to ${\ensuremath{\operatorname{rev}}}_k (t
{\ensuremath{\operatorname{mod}}} n) + 1$, else
- if $\kappa'' < \kappa$ then it sets ${\ensuremath{\operatorname{ub}}}$ to ${\ensuremath{\operatorname{rev}}}_k (t
{\ensuremath{\operatorname{mod}}} n) - 1$, else
- if $\kappa' \leq \kappa\leq \kappa''$ then it reports reception of the key $\kappa$ from $[\kappa', \kappa'']$
- if ${\ensuremath{\operatorname{lb}}} > {\ensuremath{\operatorname{ub}}}$ then the receiver reports that $[\kappa',
\kappa''] \cap {\ensuremath{\operatorname{KEYS}}} = \emptyset$
In the above description we used [[[*broadcaster time slot numbers*]{}]{}]{}. By [[[*receiver time*]{}]{}]{} we mean the number of time slots that elapsed since the start of the receiver’s protocol. Thus, just before the time slot $s$ the receiver time is zero, just after time slot $s$ the receiver time is one, an so on. However, the receiver knows the broadcaster time modulo $n$ (this information may be included in the frame header) and uses it it to compute the timer waking-up the radio for next reception of the frame.
Subsets $Y_{k, s, i}$ and $X_{k, s, i}$
---------------------------------------
In the analysis of the receiver’s protocol (Section \[Section-analysis\]), we split the sequence of the time slots following the starting slot $s$ into segments $Y_{k, s, i}$. The set $X_{k, s, i}$ is the set of indexes of the elements transmitted during time slots $Y_{k, s, i}$. We show that the “density” of initially transmitted indexes bounds the length of $[{\ensuremath{\operatorname{lb}}}, {\ensuremath{\operatorname{ub}}}]$ and the “sparsity” of the set of indexes of the next segment bounds the number of needed receptions. Finally we sum up the bounds on receptions in all segments. In Section \[Section-implementation\], we use this decomposition and also the binary search tree on the elements of $X_{k, s, i} $ embedded on the graph of the permutation ${\ensuremath{\operatorname{rev}}}_k$, for efficient computation of the wake-up timer.
For the starting time slot $s \in [[0, 2^k - 1]]$, for $i \geq 0$, let $t_{k,
s, i}$ and $l_{k, s, i}$ be defined as follows:
- $t_{k, s, 0} = t$ and $l_{k, s, 0} = \max \{l \leq k | t_{k, s, 0}
{\ensuremath{\operatorname{mod}}} 2^l = 0\}$.
- For $i > 0,$ $t_{k, s, i} = (t_{k, s, i - 1} + 2^{l_{k, s, i - 1}})
{\ensuremath{\operatorname{mod}}} n$ and $l_{k, s, i} = \max \{l \leq k | t_{k, s, i} {\ensuremath{\operatorname{mod}}}
2^l = 0\}.$
$l_{k, s, i}$ is the maximal length of of the suffix of the zero bits in binary representation of $t_{k, s, i}$. $t_{k, s, i+1}$ is the next time slot after $t_{k, s, i}$ (modulo $n$), that has longer such suffix. Note that $t_{k, s, 0}, t_{k, s, 1}, t_{k, s, 2}, \ldots$ is a (possibly empty) increasing sequence of some integers from $[[1, 2^k - 1]]$ followed by infinite sequence of zeroes.
Let ${\ensuremath{\operatorname{last}}}_{k, s} = \min \{i \geq 0 | t_{k, s, i} = 0\}$. Note that $l_{k, s, 0}, \ldots, l_{k, s, {\ensuremath{\operatorname{last}}}_{k, s}}$ is an increasing sequence of integers from $[[0, k]]$. For $0 \leq i < {\ensuremath{\operatorname{last}}}_{k, s}$, let $Y_{k,
s, i} = [[t_{k, s, i}, t_{k, s, i + 1} - 1]]$ and let $Y_{k, s, {\ensuremath{\operatorname{last}}}} =
[[0, 2^k - 1]]$. For $0 \leq i \leq {\ensuremath{\operatorname{last}}}_{k, s},$ let $X_{k, s, i} =
{\ensuremath{\operatorname{rev}}}_k Y_{k, s, i}$.
\[X\_i-Lemma\] $X_{k, s, i} =\{{\ensuremath{\operatorname{rev}}}_k (t_{k, s, i}) + 2^{k -
l_{k, s, i}} \cdot x' | x' \in [[0, 2^{l_{k, s, i}} - 1]]\}$ and ${\ensuremath{\operatorname{rev}}}_k (t_{k, s, i}) < 2^{k - l_{k, s, i}}$.
Let $y_j = {\ensuremath{\operatorname{bit}}}_j (t_{k, s, i})$, let $l =
l_{k, s, i}$. Then $Y_{k, s, i}$ is the set of all numbers $(y_{k - 1},
\ldots, y_l, y'_{l - 1}, \ldots, y'_0)_2$ such that $y'_j \in \{0, 1\}$. Thus $X_{k, s, i} = {\ensuremath{\operatorname{rev}}}_k (Y_{k, s, i})$ is the set of all numbers $(x'_{l - 1}, \ldots, x'_0, y_l, \ldots, y_{k - 1})_2$ such that $x_j' \in
\{0, 1\}$. Note that ${\ensuremath{\operatorname{rev}}}_k (t_{k, s, i}) = (0, \ldots, 0, y_l,
\ldots, y_{k - 1})_2 = (0, \ldots, 0, x_{k - l - 1}, \ldots, x_0)_2$, where $x_i = y_{k - 1 - i}$. Thus ${\ensuremath{\operatorname{rev}}}_k (t_{k, s, i}) < 2^{k - l}$. This completes the proof of Lemma \[X\_i-Lemma\].
For $0 \leq i \leq {\ensuremath{\operatorname{last}}}_{k, s}$, for $0 \leq l \leq l_{k, s, i}$, let $Y_{k, s, i, l} = [[t_{k, s, i} + \lfloor 2^{l - 1} \rfloor, t_{k, s, i} + 2^l
- 1]]$. Note that $Y_{k, s, i}$ is a disjoint union of the sets $Y_{k, s, i,
l}$, for $0 \leq l \leq l_{k, s, i}$. For $0 \leq i \leq {\ensuremath{\operatorname{last}}}$, for $0
\leq l \leq l_{k, s, i}$, let $X_{k, s, i, l} = {\ensuremath{\operatorname{rev}}}_k (Y_{k, s, i,
l})$.
\[X\_il-Lemma\] For $l \in [[0, l_{k, s, i}]]$, $X_{k, s, i, l}
=\{{\ensuremath{\operatorname{rev}}}_k (t_{k, s, i} + \lfloor 2^{l - 1} \rfloor) + 2^{k - l + 1}
\cdot x' | x' \in [[0, \lceil 2^{l - 1} \rceil - 1]]\}$ and ${\ensuremath{\operatorname{rev}}}_k
(t_{k, s, i} + \lfloor 2^{l - 1} \rfloor) < 2^{k - l + 1}$.
If $l = 0$, then $Y_{k, s, i, l} =\{t_{k, s, i}
\}$ and, $X_{k, s, i, l} =\{{\ensuremath{\operatorname{rev}}}_k (t_{k, s, i})\}$ and ${\ensuremath{\operatorname{rev}}}_k (t_{k, s, i}) < 2^{k + 1}$.
Consider the case: $l > 0$. $Y_{k, s, i, l} = [[(t_{k, s, i} + 2^{l - 1}), (t_{k,
s, i} + 2^{l - 1}) + 2^{l - 1} - 1]]$. Since $t_{k, s, i} {\ensuremath{\operatorname{mod}}}
2^{l_{k, s, i}} = 0$ and $l - 1 < l_{k, s, i}$, we have $(t_{k, s, i} + 2^{l
- 1}) {\ensuremath{\operatorname{mod}}} 2^{l - 1} = 0$ and $Y_{k, s, i, l}$ is the set of all numbers $(y_{k - 1}, \ldots, y_{l - 1}, y'_{l - 2}, \ldots, y'_0)_2$ such that $y_j = {\ensuremath{\operatorname{bit}}}_j (t_{k, s, i} + 2^{l - 1})$ and $y'_j \in \{0, 1\}$. Thus ${\ensuremath{\operatorname{rev}}}_k (t_{k, s, i} + 2^{l - 1}) < 2^{k - l + 1}$ and $X_{k, s,
i, l}$ is the set of all numbers $(x'_{l - 2}, \ldots, x'_0, x_{k - l},
\ldots, x_0)_2$ such that $x_j = y_{k - 1 - j}$ and $x'_j \in \{0, 1\}$.
\[union-X-Lemma\]$\bigcup_{j = 0}^l X_{k, s, i, j} =\{{\ensuremath{\operatorname{rev}}}_k
(t_{k, s, i}) + 2^{k - l} \cdot x' | x' \in [[0, 2^l - 1]]\}$ and ${\ensuremath{\operatorname{rev}}}_k (t_{k, s, i}) < 2^{k - l}$.
$\bigcup_{j = 0}^l X_{k, s, i, j} =
{\ensuremath{\operatorname{rev}}}_k ( \bigcup_{j = 0}^l Y_{k, s, i, j}) = {\ensuremath{\operatorname{rev}}}_k ([[t_{k, s,
i}, t_{k, s, i} + 2^l - 1]])$. Since $t_{k, s, i} {\ensuremath{\operatorname{mod}}} 2^l = 0$, the proof follows as in the previous lemmas.
\[Section-analysis\][[**[The analysis of the receiver’s process for $[\kappa', \kappa'']$]{}**]{}]{}
====================================================================================================
Let $r'$ and $r''$ be defined as follows:
$r' = \min \{r \in [[- 1, n]] | \kappa' \leq \kappa_r \}$, and
$r'' = \max \{r \in [[- 1, n]] | \kappa_r \leq \kappa'' \}$.
For each $r \in [[r', r'']]$, $\kappa_r \in [\kappa', \kappa'']$. If $[\kappa',
\kappa''] \cap {\ensuremath{\operatorname{KEYS}}} = \emptyset$ then, for some $r \in [[- 1, n - 1]]$, $\kappa_r < \kappa'$ and $\kappa'' < \kappa_{r + 1}$, and $r' = r + 1$ and $r'' = r$. If $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} \not= \emptyset$ then, since $- \infty < \kappa' \leq \kappa'' < + \infty$, we have $0 \leq r' \leq
r'' \leq n - 1$.
Let $s$ be the first time slot of the receiver’s protocol. We assume w.l.o.g. that $s \in [[0, n - 1]]$. For $t \geq 0$: Let ${\ensuremath{\operatorname{lb}}}_t$ and ${\ensuremath{\operatorname{ub}}}_t$ be the values of the variables ${\ensuremath{\operatorname{lb}}}$ and ${\ensuremath{\operatorname{ub}}}$, respectively, at receiver time $t$. (Thus ${\ensuremath{\operatorname{lb}}}_0 = 0$ and ${\ensuremath{\operatorname{ub}}}_0 = n - 1$.) Let $x_t = {\ensuremath{\operatorname{rev}}}_k ((s + t) {\ensuremath{\operatorname{mod}}} n)$. Let ${\ensuremath{\operatorname{used}}}_t = 1$ if ${\ensuremath{\operatorname{lb}}}_t \leq x_t \leq {\ensuremath{\operatorname{ub}}}_t$ and ${\ensuremath{\operatorname{used}}}_t = 0$ otherwise. (${\ensuremath{\operatorname{used}}}_t = 1$ if the receiver wakes-up the radio at receiver time $t$.) Let ${\ensuremath{\operatorname{hit}}}_t = 1_{}$ if $r' \leq x_t \leq r''$ and ${\ensuremath{\operatorname{hit}}}_t = 0$ otherwise. (${\ensuremath{\operatorname{hit}}}_t = 1$ if the requested frame is received at receiver time $t$.) The [[[*energy*]{}]{}]{} used in the initial $t$ time slots is ${\ensuremath{\operatorname{en}}} (t) = \sum_{j = 1}^t {\ensuremath{\operatorname{used}}}_j$. The [[[*extra energy*]{}]{}]{} is the energy used for the reception of messages with the keys outside $[\kappa',
\kappa'']$: ${\ensuremath{\operatorname{ee}}} (t) = {\ensuremath{\operatorname{en}}} (t) - \sum_{j = 1}^t {\ensuremath{\operatorname{hit}}}_j$. Let ${\ensuremath{\operatorname{HY}}}_t =\{(s + y) {\ensuremath{\operatorname{mod}}} n | y \in [[0, t - 1]]\}$. ${\ensuremath{\operatorname{HY}}}_t$ is the set of the broadcaster time-slot numbers modulo $n$ of the receiver’s initial $t$ slots. Let ${\ensuremath{\operatorname{HX}}}_t = {\ensuremath{\operatorname{rev}}}_k {\ensuremath{\operatorname{HY}}}_t$. Note that ${\ensuremath{\operatorname{HX}}}_0 = {\ensuremath{\operatorname{HY}}}_0 = \emptyset$. A [[[*history of the lower (respectively, upper) bounds up to time $t$ for $[\kappa', \kappa'']$*]{}]{}]{} is the sequence ${\ensuremath{\operatorname{HL}}} (\kappa', \kappa'', t) = ({\ensuremath{\operatorname{lb}}}_0, \ldots,
{\ensuremath{\operatorname{lb}}}_t)$ (respectively, ${\ensuremath{\operatorname{HU}}} (\kappa', \kappa'', t) = ({\ensuremath{\operatorname{ub}}}_0,
\ldots, {\ensuremath{\operatorname{ub}}}_t)$).
Let $\tau = \min \{t \geq 0 | {\ensuremath{\operatorname{hit}}}_t = 1 \vee {\ensuremath{\operatorname{lb}}}_{t + 1} >
{\ensuremath{\operatorname{ub}}}_{t + 1} \}$. Note that $\tau$ is the time until the first hit or noticing that $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} = \emptyset$. By the [[[*first receiver cycle*]{}]{}]{} we mean the first $n$ slots of the receiver time. For $y \in [[0, n]]$, let ${\ensuremath{\operatorname{tt}}} (y)$ denote the receiver time just before the transmission of the broadcast time slot $y {\ensuremath{\operatorname{mod}}} n$, in the first receiver cycle, i.e. ${\ensuremath{\operatorname{tt}}} (y) = \min \{t \geq 0 | (s + t)
{\ensuremath{\operatorname{mod}}} n = y\}$. Note that, since ${\ensuremath{\operatorname{HX}}}_n = [[0, n - 1]]$, we have $\tau < n$.
\[first-hit-Theorem\]We have $\tau < n$ and ${\ensuremath{\operatorname{en}}} (\tau) \leq 2
\cdot k + 1$.
We prove Lemmas \[clipped-Lemma\], \[first-tree-Lemma\], \[next-tree-Lemma\], \[total-Lemma\], to show the theorem in the case $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} = \emptyset$, and then conclude the general case.
\[clipped-Lemma\] If $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} =
\emptyset$, then, for $t \geq 0$, $[[{\ensuremath{\operatorname{lb}}}_t, {\ensuremath{\operatorname{ub}}}_t]]
\subseteq [[0, n - 1]] \setminus {\ensuremath{\operatorname{HX}}}_t $.
Note that the Lemma follows directly from the algorithm: ${\ensuremath{\operatorname{lb}}}_0 = 0$, ${\ensuremath{\operatorname{ub}}}_0 = n - 1$, and, since $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} = \emptyset$, for each $x \in
{\ensuremath{\operatorname{HX}}}_t$, either $x < {\ensuremath{\operatorname{lb}}}_t$ or ${\ensuremath{\operatorname{ub}}}_t < x$.
Since $k$ and $s$ are fixed, we use the following notation: ${\ensuremath{\operatorname{last}}} =
{\ensuremath{\operatorname{last}}}_{k, s}$, $t_i = t_{k, s, i}$, $k_i = l_{k, s, i}$, $Y_i = Y_{k,
s, i}$, $X_i = X_{k, s, i}$, $Y_{i, j} = Y_{k, s, i, j}$, and $X_{i, j} =
X_{k, s, i, j}$.
\[first-tree-Lemma\] If $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} =
\emptyset$, then $\sum_{(s + t) {\ensuremath{\operatorname{mod}}} n \in Y_0} {\ensuremath{\operatorname{used}}}_t \leq
k_0 + 1$.
Since $Y_{0, 0} =\{t_0 \}$, and only the first time slot congruent modulo $n$ to $t_0$ is used, we have $\sum_{(s +
t) {\ensuremath{\operatorname{mod}}} n \in Y_{0, 0}} {\ensuremath{\operatorname{used}}}_t = 1$.
For $0 < l \leq k_0$, we show that $\sum_{(s + t) {\ensuremath{\operatorname{mod}}} n \in Y_{0,
l}} {\ensuremath{\operatorname{used}}}_t \leq 1$: By Lemma \[clipped-Lemma\], $[[{\ensuremath{\operatorname{lb}}}_{2^{l - 1}}, {\ensuremath{\operatorname{ub}}}_{2^{l - 1}}]] \subseteq [[0, n - 1]]
\setminus {\ensuremath{\operatorname{HX}}}_{2^{l - 1}}$, and ${\ensuremath{\operatorname{HX}}}_{2^{l - 1}} = \bigcup_{j
= 0}^{l - 1} X_{0, j}$, which, by Lemma \[union-X-Lemma\], contains [[[*all*]{}]{}]{} the integers from $[[0, n - 1]]$ congruent modulo $2^{k -
(l - 1)}$ to ${\ensuremath{\operatorname{rev}}} (t_0)$. Hence ${\ensuremath{\operatorname{ub}}}_{2^{l - 1}} + 1 \leq
{\ensuremath{\operatorname{lb}}}_{2^{l - 1}} - 1 + 2^{k - (l - 1)}$. By Lemma \[X\_il-Lemma\], $X_{0, l}$ contains [[[*only*]{}]{}]{} the integers from $[[0, n - 1]]$ congruent modulo $2^{k - (l - 1)}$ to ${\ensuremath{\operatorname{rev}}}_k (t_0 + 2^{l - 1})$. Hence, $|X_{0, l} \cap [[{\ensuremath{\operatorname{lb}}}_{2^{l - 1}}, {\ensuremath{\operatorname{ub}}}_{2^{l - 1}}]] |
\leq 1$, and, since only the first time slot congruent modulo $n$ is used, we have $\sum_{(s + t) {\ensuremath{\operatorname{mod}}} n \in Y_{0, l}} {\ensuremath{\operatorname{used}}}_t \leq 1$. Since $Y_0 = \bigcup_{l = 0}^{k_0} Y_{0, l}$, the Lemma follows.
\[next-tree-Lemma\] If $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} =
\emptyset$, then, for $1 \leq i \leq {\ensuremath{\operatorname{last}}}$, $\sum_{(s + t)
{\ensuremath{\operatorname{mod}}} n \in Y_i} {\ensuremath{\operatorname{used}}}_t \leq k_i - k_{i - 1} + 1$.
Let $t'' = {\ensuremath{\operatorname{tt}}} (t_i)$. By Lemma \[clipped-Lemma\], $[[{\ensuremath{\operatorname{lb}}}_{t''}, {\ensuremath{\operatorname{ub}}}_{t''}]] \subseteq
[[0, n - 1]] \setminus {\ensuremath{\operatorname{HX}}}_{t''}$. We have $X_{i - 1} \subseteq
{\ensuremath{\operatorname{HX}}}_{t''}$, and, by Lemma \[X\_i-Lemma\], $X_{i - 1}$ contains [[[*all*]{}]{}]{} the integers from $[[0, n - 1]]$ congruent modulo $2^{k -
k_{i - 1}}$ to ${\ensuremath{\operatorname{rev}}} (t_{i - 1})$. Thus, ${\ensuremath{\operatorname{ub}}}_{t''} + 1 \leq
{\ensuremath{\operatorname{lb}}}_{t''} - 1 + 2^{k - k_{i - 1}}$. By Lemma \[union-X-Lemma\], $\bigcup_{j = 0}^{k_{i - 1}} X_{i, j}$ contains [[[*only*]{}]{}]{} the integers from $[[0, n - 1]]$ congruent modulo $2^{k - k_{i - 1}}$ to ${\ensuremath{\operatorname{rev}}}_k (t_i)$. Hence, we have $| \bigcup_{j = 0}^{k_{i - 1}} X_{i,
j} \cap [{\ensuremath{\operatorname{lb}}}_{t''}, {\ensuremath{\operatorname{ub}}}_{t''}] | \leq 1$, and $\sum_{(t' + t)
{\ensuremath{\operatorname{mod}}} n \in \bigcup_{0 \leq j \leq k_{i - 1}} Y_{i, j}} {\ensuremath{\operatorname{used}}}_t
\leq 1$.
For $k_{i - 1} < l \leq k_i$, we show that $\sum_{(s + t) {\ensuremath{\operatorname{mod}}} n
\in Y_{i, l}} {\ensuremath{\operatorname{used}}}_t \leq 1$: We have $[[{\ensuremath{\operatorname{lb}}}_{t'' + 2^{l -
1}}, {\ensuremath{\operatorname{ub}}}_{t'' + 2^{l - 1}}]] \subseteq [[0, n - 1]] \setminus
{\ensuremath{\operatorname{HX}}}_{t'' + 2^{l - 1}}$ and ${\ensuremath{\operatorname{HX}}}_{t'' + 2^{l - 1}}$ is a super-set of $\bigcup_{j = 0}^{l - 1} X_{i, j}$, which, by Lemma \[union-X-Lemma\], contains [[[*all*]{}]{}]{} the integers from $[[0, n
- 1]]$ congruent modulo $2^{k - (l - 1)}$ to ${\ensuremath{\operatorname{rev}}} (t_i)$. By Lemma \[X\_il-Lemma\], $X_{i, l}$ contains [[[*only*]{}]{}]{} the integers from $[[0, n - 1]]$ congruent modulo $2^{k - (l - 1)}$ to ${\ensuremath{\operatorname{rev}}}_k
(t_i + \lfloor 2^{l - 1} \rfloor)$. Hence $|X_{i, l} \cap [[{\ensuremath{\operatorname{lb}}}_{t''
+ 2^{l - 1}}, {\ensuremath{\operatorname{ub}}}_{t'' + 2^{l - 1}}]] | \leq 1$.
Thus $\sum_{(s + t)
{\ensuremath{\operatorname{mod}}} n \in \bigcup_{k_{i - 1} < j \leq k_i} Y_{i, j}} {\ensuremath{\operatorname{used}}}_t
\leq k_i - k_{i - 1}$ and the lemma follows.
\[total-Lemma\] If $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} = \emptyset$, then $\sum_{t > 0} {\ensuremath{\operatorname{used}}}_t \leq 2 k + 1$.
$[[0, n - 1]] = \bigcup_{i = 0}^{{\ensuremath{\operatorname{last}}}}
Y_i$, and $\sum_{t > 0} {\ensuremath{\operatorname{used}}}_t = $ $\sum_{(s + t) {\ensuremath{\operatorname{mod}}} n \in
Y_0} {\ensuremath{\operatorname{used}}}_t + \sum_{i = 1}^{{\ensuremath{\operatorname{last}}}} ( \sum_{(s + t) {\ensuremath{\operatorname{mod}}}
n \in Y_i} {\ensuremath{\operatorname{used}}}_t) \text{}$. Thus, by Lemma \[first-tree-Lemma\] and Lemma \[next-tree-Lemma\], $\sum_{t > 0} {\ensuremath{\operatorname{used}}}_t \leq k_0 + 1
+ \sum_{i = 1}^{{\ensuremath{\operatorname{last}}}} (k_i - k_{i - 1} + 1) = k_{{\ensuremath{\operatorname{last}}}} +
{\ensuremath{\operatorname{last}}} + 1$. Since $k_0, \ldots, k_{{\ensuremath{\operatorname{last}}}}$ is increasing sequence of values from $[[0, k]]$, we have $k_{{\ensuremath{\operatorname{last}}}} \leq k$ and ${\ensuremath{\operatorname{last}}} \leq k$.
In Lemma \[total-Lemma\] we assumed that $[\kappa', \kappa''] \cap
{\ensuremath{\operatorname{KEYS}}} = \emptyset$. Note that we have:
- $0 \leq {\ensuremath{\operatorname{lb}}}_{\tau} \leq {\ensuremath{\operatorname{ub}}}_{\tau} \leq n - 1$, and
- $\kappa_{{\ensuremath{\operatorname{lb}}}_{\tau } - 1} < \kappa' \leq \kappa'' <
\kappa_{{\ensuremath{\operatorname{ub}}}_{\tau } + 1}$ (since ${\ensuremath{\operatorname{hit}}}_t = 0$, for $0 < t
\leq \tau - 1$), and
- ${\ensuremath{\operatorname{HX}}}_{\tau} \cap [{\ensuremath{\operatorname{lb}}}_{\tau}, {\ensuremath{\operatorname{ub}}}_{\tau}] = \emptyset$.
Since $[\kappa_{{\ensuremath{\operatorname{lb}}}_{\tau} - 1}, \kappa'] \cap {\ensuremath{\operatorname{KEYS}}}$ is finite, we can choose real number $\gamma$ such that $\kappa_{{\ensuremath{\operatorname{lb}}}_{\tau} - 1} < \gamma < \kappa'$ and $\gamma \not\in
{\ensuremath{\operatorname{KEYS}}}$. Since $\kappa_{{\ensuremath{\operatorname{lb}}}_{\tau} - 1} < \gamma <
\kappa_{{\ensuremath{\operatorname{ub}}}_{\tau} + 1}$ , the respective histories of the bounds up to the time $\tau$ for $[\kappa', \kappa'']$ and $[\gamma, \gamma]$ are identical:
- ${\ensuremath{\operatorname{HL}}} (\kappa', \kappa'', \tau) = {\ensuremath{\operatorname{HL}}} (\gamma,
\gamma, \tau)$, and
- ${\ensuremath{\operatorname{UL}}} (\kappa', \kappa'', \tau) = {\ensuremath{\operatorname{UL}}} (\gamma,
\gamma, \tau)$.
Note that, since ${\ensuremath{\operatorname{lb}}}_{\tau} \leq {\ensuremath{\operatorname{ub}}}_{\tau}$, the energy needed to notice that $[\gamma, \gamma]\cap {\ensuremath{\operatorname{KEYS}}} = \emptyset$ is at least ${\ensuremath{\operatorname{en}}} (\tau)$. Therefore, by Lemma \[total-Lemma\], ${\ensuremath{\operatorname{en}}} (\tau) \leq 2 k + 1$. We conclude that ${\ensuremath{\operatorname{en}}} (\tau) \leq 2 k + 1$ also when $[\kappa', \kappa'']
\cap {\ensuremath{\operatorname{KEYS}}} \not= \emptyset$.
\[Corollary-ee\]For arbitrary $t > 0$, ${\ensuremath{\operatorname{ee}}} (t) \leq 4 k + 2$.
If $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} =
\emptyset$, then ${\ensuremath{\operatorname{ee}}} (t) \leq {\ensuremath{\operatorname{en}}} (\tau)$ and, by Theorem \[first-hit-Theorem\], ${\ensuremath{\operatorname{en}}} (\tau) \leq 2 k + 1$.
Consider the case $[\kappa', \kappa''] \cap {\ensuremath{\operatorname{KEYS}}} \not= \emptyset$. Then $- 1 < r' \leq r'' < n$. Let $\gamma'$ and $\gamma''$ be such that $\kappa_{r' - 1} < \gamma' < \kappa_{r'}$ and $\kappa_{r''} < \gamma'' <
\kappa_{r'' + 1}$. Then, for arbitrary $t \geq 0$, ${\ensuremath{\operatorname{HL}}} (\gamma',
\gamma', t) = {\ensuremath{\operatorname{HL}}} (\kappa', \kappa'', t)$ and ${\ensuremath{\operatorname{HU}}} (\gamma'',
\gamma'', t) = {\ensuremath{\operatorname{HU}}} (\kappa', \kappa'', t)$. Any reception of the key that is outside $[\kappa', \kappa'']$ updates either the lower or the upper bound: For $t > 0$, ${\ensuremath{\operatorname{used}}}_t = 1$ and ${\ensuremath{\operatorname{hit}}}_t = 0$ if and only if either ${\ensuremath{\operatorname{lb}}}_{t - 1} < {\ensuremath{\operatorname{lb}}}_t$ or ${\ensuremath{\operatorname{ub}}}_t < {\ensuremath{\operatorname{ub}}}_{t -
1}$. Thus ${\ensuremath{\operatorname{ee}}} (t)$ is equal to the total number of changes in both ${\ensuremath{\operatorname{HL}}} (\kappa', \kappa'', t)$ and ${\ensuremath{\operatorname{HU}}} (\kappa', \kappa'', t)$. Since $[\gamma', \gamma'] \cap {\ensuremath{\operatorname{KEYS}}} = \emptyset$, the number of changes in ${\ensuremath{\operatorname{HL}}} (\gamma', \gamma', t)$ is not greater than $2 k + 1$. Similarly, the number of changes in ${\ensuremath{\operatorname{HU}}} (\gamma'', \gamma'', t)$ is not greater than $2 k + 1$.
\[Section-unreliable\]Unreliable network
----------------------------------------
Consider a model of the network, where the probability of successful reception is $p$, $0 \leq p \leq 1$. Thus the receiver may wake up to listen in some time slot, and still fail to receive the frame with probability $q = 1 - p$. Thus the unit of energy used for the wake-up is lost. We state that in the case of reception failure, the receiver’s protocol leaves its variables ${\ensuremath{\operatorname{lb}}}$ and ${\ensuremath{\operatorname{ub}}}$ unchanged and waits for the next time slot from ${\ensuremath{\operatorname{rev}}}_k [[{\ensuremath{\operatorname{lb}}}, {\ensuremath{\operatorname{ub}}}]]$.
We split the wake-ups of the receiver into [[[*hits*]{}]{}]{} – the wake-ups in the time slots from ${\ensuremath{\operatorname{rev}}}_k [[r', r'']]$, and [[[*misses*]{}]{}]{} – the remaining wake-ups. The hits are unavoidable: the requested keys are transmitted during the hits. The penalty for unreliability here is that the reception rate drops from $1$ to $p$ – which is the highest possible in this model. Another penalty is the increase in the number of the misses. We show the bound on the number of the misses in unreliable network. Recall that the first wake up of the protocol is in time slot $s$. For $t \geq 0$, let ${\ensuremath{\operatorname{success}}} (t)$ be true if the transmission in the $t$th receiver’s time slot is successful, and false – otherwise.
\[Lemma-expected-trailer\]The expected number of misses after the first receiver cycle (i.e. after the initial $n$ time slots) is not greater than $2 \cdot q / p^2$.
The misses in the cycle following the first cycle are the wake-ups during the time slots in ${\ensuremath{\operatorname{rev}}}_k ([[{\ensuremath{\operatorname{lb}}}_n, r' - 1]] \cup [[r'' + 1,
{\ensuremath{\operatorname{ub}}}_n]])$. The values of ${\ensuremath{\operatorname{lb}}}_n - 1$ and ${\ensuremath{\operatorname{ub}}}_n + 1$ are the following random variables:
- ${\ensuremath{\operatorname{lb}}}_n - 1 = \max \{- 1\} \cup \{i \in [[0, r' - 1]] \;|\;
{\ensuremath{\operatorname{success}}} ({\ensuremath{\operatorname{tt}}} ({\ensuremath{\operatorname{rev}}}_k (i))\}$
- ${\ensuremath{\operatorname{ub}}}_n + 1 = \min \{n\} \cup \{i \in [[r'' + 1, n - 1]] \;|\;
{\ensuremath{\operatorname{success}}} ({\ensuremath{\operatorname{tt}}} ({\ensuremath{\operatorname{rev}}}_k (i))\}$
Each of $r' - ({\ensuremath{\operatorname{lb}}}_n - 1)$ and $({\ensuremath{\operatorname{ub}}}_n + 1) - r''$ can be bound by a random variable with geometric distribution (see e.g. [[@CormenLR89]]{}) and expected value $1 / p$. Hence, $\max \{E [r' -
{\ensuremath{\operatorname{lb}}}_n], E [{\ensuremath{\operatorname{ub}}}_n - r'']\} \leq 1 / p - 1 = 1 / (1 - q) - 1$.
After the $j$th cycle, for $j \geq 1$, each position has been tested $j$ times. Thus $\max \{E [r' - {\ensuremath{\operatorname{lb}}}_{j \cdot n}], E [{\ensuremath{\operatorname{ub}}}_{j \cdot n}
- r'']\} \leq 1 / (1 - q^j) - 1$ and the expected number of misses in the $(j + 1)$st cycle is not greater than $2 (1 / (1 - q^j) - 1)$. Finally, note that $\sum_{j = 1}^{\infty} (1 / (1 - q^j) - 1) = \sum_{j = 1}^{\infty} (q^j
/ (1 - q^j)) \leq \frac{1}{1 - q} \sum_{j = 1}^{\infty} q^j = q / (1 - q)^2
$.
The more complex task is to bound the number of misses during the first cycle.
\[Lemma-expected-init\]If ${\ensuremath{\operatorname{KEYS}}} \cap [\kappa', \kappa''] =
\emptyset$, then the expected number of wake-ups (all of them are misses) during the first cycle is not greater than $(4 k + 2) / p$.
Since ${\ensuremath{\operatorname{KEYS}}} \cap [\kappa', \kappa''] = \emptyset$, we have $r' = r''
+ 1$. Let us use the notation from the proof of Theorem \[first-hit-Theorem\].
First consider the time-slots in $Y_0$. There is one wake-up in $Y_{0, 0}
=\{t_0 \}$. For each $l \geq 0$, $\bigcup_{j = 0}^l X_{0, j} \subseteq
{\ensuremath{\operatorname{HX}}}_{{\ensuremath{\operatorname{tt}}} (\min Y_{0, l + 1})}$. Hence, by Lemma \[union-X-Lemma\],
- ${\ensuremath{\operatorname{lb}}}_{{\ensuremath{\operatorname{tt}}} (\min Y_{0, l + 1})} - 1
\geq
\max \{- 1\} \cup \{i \in [[0, r' - 1]] \;|\; (i - {\ensuremath{\operatorname{rev}}}_k (t_0))
{\ensuremath{\operatorname{mod}}} 2^{k - l} = 0 \wedge {\ensuremath{\operatorname{success}}} ({\ensuremath{\operatorname{tt}}} ({\ensuremath{\operatorname{rev}}}_k
(i)))\}$, and
- ${\ensuremath{\operatorname{ub}}}_{{\ensuremath{\operatorname{tt}}} (\min Y_{0, l + 1})} + 1
\leq
\min \{n\}
\cup \{i \in [[r'' + 1, n - 1]] \;|\; (i - {\ensuremath{\operatorname{rev}}}_k (t_0)) {\ensuremath{\operatorname{mod}}} 2^{k -
l} = 0 \wedge {\ensuremath{\operatorname{success}}} ({\ensuremath{\operatorname{tt}}} ({\ensuremath{\operatorname{rev}}}_k (i)))\}$.
Note that $[[0, r' - 1]] \cup [[r'' + 1, n - 1]] = [[0, n - 1]]$. Thus, $E
[({\ensuremath{\operatorname{ub}}}_{{\ensuremath{\operatorname{tt}}} (\min Y_{0, l + 1})} - {\ensuremath{\operatorname{lb}}}_{{\ensuremath{\operatorname{tt}}} (\min
Y_{0, l + 1})}) / 2^{k - l}] < 2 / p$ – the expected number of integers congruent modulo $2^{k - l}$ to ${\ensuremath{\operatorname{rev}}}_k (t_0)$ in \[\[${\ensuremath{\operatorname{lb}}}_{{\ensuremath{\operatorname{tt}}} (\min Y_{0, l + 1})}, {\ensuremath{\operatorname{ub}}}_{{\ensuremath{\operatorname{tt}}} (\min
Y_{0, l + 1})}]]$. Since, by Lemma \[X\_il-Lemma\], all elements of $X_{0,
l + 1}$ are congruent modulo $2^{k - l}$ to ${\ensuremath{\operatorname{rev}}}_k (t_0 + \lfloor 2^l
\rfloor)$, the expected number of wake-ups during time slots $Y_{0, l + 1}$ is bounded by $2 / p$. Thus the expected number of wake-ups in $Y_0$ is not greater than $2 k_0 / p + 1 \leq 2 (k_0 + 1) / p$.
Now consider $Y_i$, for $i \in [[1, {\ensuremath{\operatorname{last}}}]]$. Since $X_{i - 1}
\subseteq {\ensuremath{\operatorname{HX}}}_{{\ensuremath{\operatorname{tt}}} (\min Y_i)}$ and, by Lemma \[X\_i-Lemma\], $X_{i - 1}$ contains all integers congruent modulo $2^{k - k_{i - 1}}$ to $\min X_{i - 1}$ and, by Lemma \[union-X-Lemma\], $\bigcup_{j = 0}^{k_{i -
1}} X_{i, j}$, contains only integers congruent modulo $2^{k - k_{i - 1}}$ to $\min X_i$, the expected number of wake-ups in $ \bigcup_{j = 0}^{k_{i -
1}} Y_{i, j}$ can be bound, as above, by $2 / p$.
For each $l \in [[k_{i - 1} + 1, k_i]]$, we use $\bigcup_{j = 0}^{l - 1}
X_{i, j} \subseteq {\ensuremath{\operatorname{HX}}}_{{\ensuremath{\operatorname{tt}}} (\min Y_{i, l})}$, to bound the expected number of wake-ups in $Y_{i, l}$ by $2 / p$. Thus the expected number of wake-ups in $Y_i$ is not greater than $2 (k_i - k_{i - 1} + 1) /
p$.
Summing up, as in the proof of Lemma \[total-Lemma\], the expected number of wake-ups during the first cycle is at most $\frac{2}{p} (k_0 + 1 +
\sum_{i = 0}^{{\ensuremath{\operatorname{last}}}} (k_i - k_{i - 1} + 1)) \leq (4 k + 2) / p$.
\[Theorem-expected\]The expected number of misses during the infinite execution of the protocol is not greater than $(8 k + 4) / p + 2 (1 - p) /
p^2$.
If ${\ensuremath{\operatorname{KEYS}}} \cap [\kappa', \kappa''] = \emptyset$, then the theorem follows directly from Lemmas \[Lemma-expected-trailer\] and \[Lemma-expected-init\].
Consider the case ${\ensuremath{\operatorname{KEYS}}} \cap [\kappa', \kappa''] \not= \emptyset$. As in Corollary \[Corollary-ee\], let $\gamma'$ and $\gamma''$ be key values such that $\kappa_{r' - 1} < \gamma' < \kappa_{r'}$ and $\kappa_{r''}
< \gamma'' < \kappa_{r'' + 1}$. Let ${\ensuremath{\operatorname{EX}}}_{\gamma}$ denotes the expected number of misses in the first cycle when the protocol is started for interval $[\gamma, \gamma]$. By Lemma \[Lemma-expected-trailer\], $\max \{E_{\gamma'}, E_{\gamma''} \} \leq (4 k + 2) / p$. The expected number of misses during the first cycle of the protocol for $[\kappa',
\kappa'']$ is the sum of the expected number of misses on both sides of $[r', r'']$ which is not greater than $E_{\gamma'} + E_{\gamma''}$.
\[Section-implementation\]Implementation issues
===============================================
We present an efficient algorithm for computing the time slot of the reception of the next frame required by the protocol. The efficiency of this algorithm is based on the observation that elements of $X_{k, s, i}$ are organized by ${\ensuremath{\operatorname{rev}}}_k$ into subsequent levels of an almost balanced binary search tree.
Binary search tree on $X_{k, s, i}$
-----------------------------------
For $d \geq 0$, for any sequence $c = (c_1, \ldots, c_d) \in \{- 1, 1\}^d$, let a [[[*descendant*]{}]{}]{} of $x$ by path $c$ be defined as ${\ensuremath{\operatorname{dsc}}}_k(x, c) = x + \sum_{i = 1}^d 2^{k - i} \cdot c_i$. Note that ${\ensuremath{\operatorname{dsc}}}_k (x,(c_1, c_2, \ldots, c_d)) = {\ensuremath{\operatorname{dsc}}}_{k - 1} ({\ensuremath{\operatorname{dsc}}}_k (x, (c_1)), (c_2,
\ldots, c_d))$. Note that $({\ensuremath{\operatorname{dsc}}}_k (x, (c_1, \ldots, c_d)) - x)
{\ensuremath{\operatorname{mod}}} 2^{k - d} = 0$. Let a [[[*level at depth $d$ rooted at $x$*]{}]{}]{} be defined as $L_{k, d} (x) =\{{\ensuremath{\operatorname{dsc}}}_x (x, c) | c \in \{- 1, 1\}^d \}$. Let a [[[*sub-tree of depth $d$ rooted at $x$*]{}]{}]{} be defined as ${\ensuremath{\operatorname{ST}}}_{k, d}
(x) =\{{\ensuremath{\operatorname{dsc}}}_x (x, c) | \exists_{d' \in [[0, d]]} c \in \{- 1, 1\}^{d'}
\}$. The following properties are easy to note without the proof:
\[BST-properties\]For $k \geq 0$, for $d \in [[0, k]]$, we have the following properties:
$|L_{k, d} (x) | = 2^d$.
\[BST-L-ST\]$L_{k, 0} (x) = {\ensuremath{\operatorname{ST}}}_{k, 0} (x) =\{x\}$ and, for $d > 1$, $L_{k, d} (x) = {\ensuremath{\operatorname{ST}}}_{k, d} (x) \setminus {\ensuremath{\operatorname{ST}}}_{k,
d - 1} (x)$.
$| {\ensuremath{\operatorname{ST}}}_{k, d} (x) | = 2^{d + 1} - 1$.
${\ensuremath{\operatorname{ST}}}_{k, d} (x) =\{x + i \cdot 2^{k - d} | i \in [[- 2^d + 1,
2^d - 1]]\}$.
\[BST-with-right-ST\]If $d \geq 1$ then $\{x\} \cup {\ensuremath{\operatorname{ST}}}_{k
- 1, d - 1} ({\ensuremath{\operatorname{dsc}}}_k (x, (1))) =\{x + i \cdot 2^{k - d} | i \in [[0,
2^d - 1]]\}$.
\[BST-ST-root-ST\] ${\ensuremath{\operatorname{ST}}}_{k, d} (x) = {\ensuremath{\operatorname{ST}}}_{k - 1, d -
1} ({\ensuremath{\operatorname{dsc}}}_k (x, (- 1))) \cup \{x\} \cup {\ensuremath{\operatorname{ST}}}_{k - 1, d - 1} (x,
{\ensuremath{\operatorname{dsc}}}_k (x, (1)))$.
\[BST-inorder\] $\max {\ensuremath{\operatorname{ST}}}_{k - 1, d} ({\ensuremath{\operatorname{dsc}}}_k (x, (-
1))) + 2^{k - 1 - d} = x = \min {\ensuremath{\operatorname{ST}}}_{k - 1, d} ({\ensuremath{\operatorname{dsc}}}_k (x,
(1))) - 2^{k - 1 - d}$.
Lemma \[X-BST\] shows that each $X_{k, s, i}$ is organized by ${\ensuremath{\operatorname{rev}}}_k$ in a binary search tree with the root at $\min X_{k, s, i} = {\ensuremath{\operatorname{rev}}}_k
(t_{k, s, i})$, without the left sub-tree and with a totally balanced right sub-tree, see Figure \[BST-fig\].
Lemma \[X-BST\]\[parent-above-child\] states that the elements of the levels closer to the root have lower values of their $k$-bit reversals than the elements of the more distant levels.
\[X-BST\] For $k \geq 0$, $t \in [[0, 2^k - 1]]$, $i \in [[0,
{\ensuremath{\operatorname{last}}}_{k, s}]]$, let $r = {\ensuremath{\operatorname{rev}}}_k (t_{k, s, i})$ and $l = k_{k,
s, i}$.
Then we have:
\[upper-levels\] $X_{k, s, i, 0} =\{r\}$ and, for $\text{$d \in [[1, l]]$}, \bigcup_{j = 0}^d X_{k, s, i, j} =\{r\} \cup
{\ensuremath{\operatorname{ST}}}_{k - 1, d - 1} ({\ensuremath{\operatorname{dsc}}}_k (r, (1)))$.
\[X-BST-structure\] If $l > 0$ then $X_{k, s, i} =\{r\} \cup
{\ensuremath{\operatorname{ST}}}_{k - 1, l - 1} ({\ensuremath{\operatorname{dsc}}}_k (r, (1)))$. If $l = 0$ then $X_{k,
s, i} =\{r\}$.
\[X-BST-level\]$X_{k, s, i, 0} =\{r\}$ and, for $d \in [[1, l]]$, $X_{k, s, i, d} = L_{k - 1, d - 1} ({\ensuremath{\operatorname{dsc}}}_k (r,
(1)))$.
\[parent-above-child\]If $c' \in \{- 1, 1\}^{d'}$ and $c'' \in
\{- 1, 1\}^{d''}$, where $0 \leq d' < d'' \leq l$, and $x' = {\ensuremath{\operatorname{dsc}}}_k
(r, c')$ and $x'' = {\ensuremath{\operatorname{dsc}}}_k (r, c'')$ and $x', x'' \in X_{k, s, i}$ and $y' = {\ensuremath{\operatorname{rev}}}_k (x')$ and $y'' = {\ensuremath{\operatorname{rev}}}_k (x'')$ then $y' <
y''$.
Lemma \[X-BST\]\[upper-levels\]: By Lemma \[BST-properties\]\[BST-with-right-ST\] $\{r\} \cup {\ensuremath{\operatorname{ST}}}_{k
- 1, d - 1} ({\ensuremath{\operatorname{dsc}}}_k (r, (1))) =\{r + i \cdot 2^{k - d} | i \in [[0,
2^d]]\}$ which, by Lemma \[union-X-Lemma\], is equal to $\bigcup_{j = 0}^d
X_{k, s, i, j}$.
Lemma \[X-BST\]\[X-BST-structure\] follows from $X_{k, s, i} =
\bigcup_{j = 0}^l X_{k, s, i, j}$ and from Lemma \[X-BST\]\[upper-levels\].
Lemma \[X-BST\]\[X-BST-level\] follows from Lemma \[X-BST\]\[upper-levels\] and from Lemma \[BST-properties\]\[BST-L-ST\].
Lemma \[X-BST\]\[parent-above-child\]: If $d' = 0$ then $x' = r$ and the lemma follows, since $t_{k, s, i} = \min Y_{k, s, i}$. Otherwise, we have $0 < d' < d''$, $x' \not= r$ and $x'' \not= r$ and, by Lemma \[X-BST\]\[X-BST-structure\], $x', x'' \in {\ensuremath{\operatorname{ST}}}_{k - 1, l -
1} ({\ensuremath{\operatorname{dsc}}}_k (r, (1)))$. Thus $x' \in L_{k - 1, d' - 1} ({\ensuremath{\operatorname{dsc}}}_k
(r, (1)))$ and $x'' \in L_{k - 1, d'' - 1} ({\ensuremath{\operatorname{dsc}}}_k (r, (1)))$. By Lemma \[X-BST\]\[X-BST-level\], $x' \in X_{k, s, i, d'}$ and $x'' \in
X_{k, s, i, d''}$. To conclude, note that $\max Y_{k, s, i, d'} < \min Y_{k,
s, i, d''}$.
\[Section-nsi\]Implementation of ${\ensuremath{\operatorname{nsi}}}$
--------------------------------------------------------------------
In realistic implementation, after each reception, the receiver has to compute the next time slot with the index of the transmitted key in the interval $[{\ensuremath{\operatorname{lb}}}, {\ensuremath{\operatorname{ub}}}]$, and switch off the radio for the time remaining to this event.
By ${\ensuremath{\operatorname{nsi}}}_k (t, r_1, r_2)$ we denote the next slot number (modulo $2^k$) after the slot $t$ with its $k$-bit reversal in $[r_1, r_2]$: For $r_1,
r_2 \in [[0, 2^k - 1]]$, $r_1 \leq r_2$, and $t \in [[0, 2^k - 1]]$, ${\ensuremath{\operatorname{nsi}}}_k (t, r_1, r_2) = (t + \tau (t, r_1, r_2)) {\ensuremath{\operatorname{mod}}} 2^k$, where $\tau (t, r_1, r_2) = \min \{d > 0 | {\ensuremath{\operatorname{rev}}}_k ((t + d) {\ensuremath{\operatorname{mod}}}
2^k) \in [[r_1, r_2]]\}$.
One could naively test subsequent values after $t$ or all values in ${\ensuremath{\operatorname{rev}}}_k [[r_1, r_2]]$. However, both these methods are time consuming, when both $2^k / (r_2 - r_1)$ and $r_2 - r_1$ are large.
We present an efficient algorithm for the computation of ${\ensuremath{\operatorname{nsi}}}_k (t,
r_1, r_2)$:
$t'' \leftarrow (t + 1) {\ensuremath{\operatorname{mod}}} 2^k$
$l \leftarrow 0$
\[repeat-line\]repeat
$t' \leftarrow t''$
\[internal-while\]while $l < k \wedge t' {\ensuremath{\operatorname{mod}}} 2^{l + 1} =
0$ do $l \leftarrow l + 1$
$x_1 \leftarrow {\ensuremath{\operatorname{rev}}}_k (t')$
$t'' \leftarrow (t' + 2^l) {\ensuremath{\operatorname{mod}}} 2^k$
$x_2 \leftarrow {\ensuremath{\operatorname{rev}}}_k (t' + 2^l - 1)$
\[until-line\]until $r_1 \leq x_2 \wedge r_2 \geq x_1 \wedge
\lceil (r_1 - x_1) / 2^{k - l} \rceil \leq \lfloor (r_2 - x_1) / 2^{k - l}
\rfloor$
\[bin-search-start\]$c \leftarrow 2^{k - 1}$
\[bin-search-while\]while $x_1 < r_1 \vee x_1 > r_2$ do
if $x_1 < r_1$ then $x_1 \leftarrow x_1 + c$ else $x_1 \leftarrow
x_1 - c$
$c \leftarrow c / 2$
return ${\ensuremath{\operatorname{rev}}}_k (x_1)$
Correctness of the algorithm:
Let $s = (t + 1) {\ensuremath{\operatorname{mod}}} 2^k$. Let the iterations of the “repeat-until” loop be numbered starting from zero. After the $i$th iteration, at line \[until-line\], we have $l = l_{k, s, i}$, $t' = t_{k, s, i}$, $x_1 = \min
X_{k, s, i} = {\ensuremath{\operatorname{rev}}}_k (t_{k, s, i})$, $x_2 = \max X_{k, s, i}$, and $t''
= t_{k, s, i + 1}$. Let $i' = \min \{i \geq 0 | X_{k, s, i} \cap [r_1, r_2]
\not= \emptyset\}$. Since $r_1, r_2 \in [[0, 2^k - 1]]$, $r_1 \leq r_2$, and $X_{k, s, {\ensuremath{\operatorname{last}}}_{k, s}} = [[0, 2^k - 1]]$, we have $0 \leq i' \leq
{\ensuremath{\operatorname{last}}}_{k, s}$. Thus, by Lemma \[X\_i-Lemma\], $i'$ is the number of the first iteration, after which$ r_1 \leq x_2 \wedge r_2 \geq x_1 \wedge \min \{j
| x_1 + 2^{k - l} \cdot j \geq r_1 \} \leq \max \{j | x_1 + 2^{k - l} \cdot j
\leq r_2 \}$, which is equivalent to $ r_1 \leq x_2 \wedge r_2 \geq x_1 \wedge
\lceil (r_1 - x_1) / 2^{k - l} \rceil \leq \lfloor (r_2 - x_1) / 2^{k - l}
\rfloor$.
After the “repeat-until” loop finishes, at line \[bin-search-start\], we have $x_1 = {\ensuremath{\operatorname{rev}}}_k (t_{k, s, i'})$ and, by Lemma \[X-BST\]\[X-BST-structure\], $X_{k, s, i'} =\{x_1 \} \cup S$, where either $S = {\ensuremath{\operatorname{ST}}}_{k - 1, l - 1} ({\ensuremath{\operatorname{dsc}}}_k (x_1, (1)))$, if $l >
0$, or $S = \emptyset$, if $l = 0$. Since $X_{k, s, i'} \cap [r_1, r_2]
\not= \emptyset$, we do a binary search in $X_{k, s, i'}$ until we enter the interval $[r_1, r_2]$ for the first time. By the Lemma \[X-BST\]\[parent-above-child\], the returned value is $\min
\{{\ensuremath{\operatorname{rev}}}_k (x) | x \in X_{k, s, i'} \}$.
Complexity of the algorithm:
The memory complexity: Only the constant number of $k$-bit variables are used.
The time complexity: The number of iterations of the “repeat-until” loop is never greater than $k + 1$. Since the value of $l$ never decreases, the total number of iterations of the internal “while” loop (line \[repeat-line\]\[internal-while\]) in all iterations of the “repeat-until” loop is never grater than $k + 1$. The total number of iterations of the binary search loop (starting at line \[bin-search-while\]) is never greater than $k$. Thus the total complexity is $O (k)$ elementary operations on $k$-bit integers.
Multiplication, division and modulo operations by the powers of two can be replaced by shifting or bit-masking operations. The implementation of this algorithm in programming language, with optimizations of bit-wise operations can be found on [[@RBO-WWW]]{}.
Some technical aspects of the implementation, such as dealing with imperfect synchronization and proposed structure of the frame header has been discussed in technical report [[@DBLP:journals/corr/abs-1108-5095]]{}.
[10]{} Yon Dohn Chung and Ji Yeon Lee. An indexing method for wireless broadcast XML data. [[*[Inf. Sci.]{}*]{}]{}, 177(9):1931–1953, 2007.
Yon Dohn Chung, Sanghyun Yoo, and Myoung-Ho Kim. Energy- and latency-efficient processing of full-text searches on a wireless broadcast stream. [[*[IEEE Trans. Knowl. Data Eng.]{}*]{}]{}, 22(2):207–218, 2010.
James Cooley and John Tukey. An algorithm for the machine calculation of complex Fourier series. [[*[Mathematics of Computation]{}*]{}]{}, 19(90):297–301, 1965.
Thomas H. Cormen, Charles E. Leiserson, and Ronald L. Rivest. [[*[Introduction to Algorithms]{}*]{}]{}. The MIT Press and McGraw-Hill Book Company, 1989.
Anindya Datta, Debra E. VanderMeer, Aslihan Celik, and Vijay Kumar. Broadcast protocols to support efficient retrieval from databases by mobile users. [[*[ACM Trans. Database Syst.]{}*]{}]{}, 24(1):1–79, 1999.
Daisuke Fukuchi, Christian Sommer, Yuichi Sei, and Shinichi Honiden. Distributed arrays: A p2p data structure for efficient logical arrays. In [[*[INFOCOM]{}*]{}]{}, pages 1458–1466. IEEE, 2009.
Mohamed Hefeeda and Cheng-Hsin Hsu. On burst transmission scheduling in mobile TV broadcast networks. [[*[IEEE/ACM Trans. Netw.]{}*]{}]{}, 18(2):610–623, 2010.
Tomasz Imielinski, S. Viswanathan, and B. R. Badrinath. Energy efficient indexing on air. In Richard T. Snodgrass and Marianne Winslett, editors, [[*[SIGMOD Conference]{}*]{}]{}, pages 25–36. ACM Press, 1994.
Tomasz Imielinski, S. Viswanathan, and B. R. Badrinath. Power efficient filtering of data an air. In Matthias Jarke, Janis A. Bubenko Jr., and Keith G. Jeffery, editors, [[*[EDBT]{}*]{}]{}, volume 779 of [[*[Lecture Notes in Computer Science]{}*]{}]{}, pages 245–258. Springer, 1994.
Tomasz Imielinski, S. Viswanathan, and B. R. Badrinath. Data on air: Organization and access. [[*[IEEE Trans. Knowl. Data Eng.]{}*]{}]{}, 9(3):353–372, 1997.
Sanjeev Khanna and Shiyu Zhou. On indexed data broadcast. [[*[Journal of Computer and System Sciences]{}*]{}]{}, 60(3):575 – 591, 2000.
Marcin Kik. .
Marcin Kik. Ranking and sorting in unreliable single hop radio network. In David Coudert, David Simplot-Ryl, and Ivan Stojmenovic, editors, [[*[ADHOC-NOW]{}*]{}]{}, volume 5198 of [[*[Lecture Notes in Computer Science]{}*]{}]{}, pages 333–344. Springer, 2008.
Marcin Kik. RBO protocol: Broadcasting huge databases for tiny receivers. [[*[CoRR]{}*]{}]{}, abs/1108.5095, 2011.
Philip Levis and David Gay. [[*[TinyOS Programming]{}*]{}]{}. Cambridge University Press, New York, NY, USA, 2009.
G.M. Morton. A computer oriented geodetic data base and a new technique in file sequencing. [[*[IBM technical report]{}*]{}]{}. Ottawa, Canada, 1966.
Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free extensible hash tables. [[*[J. ACM]{}*]{}]{}, 53(3):379–405, 2006.
Jun Shao and Brian T. Davis. The bit-reversal SDRAM address mapping. In Krishna M. Kavi and Ron Cytron, editors, [[*[SCOPES]{}*]{}]{}, volume 136 of [[*[ACM International Conference Proceeding Series]{}*]{}]{}, pages 62–71, 2005.
Nitin H. Vaidya and Sohail Hameed. Scheduling data broadcast in asymmetric communication environments. [[*[Wireless Networks]{}*]{}]{}, 5(3):171–182, 1999.
[^1]: [[*Web:* `http://www.im.pwr.wroc.pl/ kik/`]{}]{}
|
{
"pile_set_name": "ArXiv"
}
|
[GBK]{}[song]{}
****
****
Huang Yuke Wen Zhiying
****
The Fibonacci sequence $\mathbb{F}$ is the fixed point beginning with $a$ of morphism $\sigma(a,b)=(ab,a)$. In this paper, we get the explicit expressions of all squares and cubes, then we determine the number of distinct squares and cubes in $\mathbb{F}[1,n]$ for all $n$, where $\mathbb{F}[1,n]$ is the prefix of $\mathbb{F}$ of length $n$. By establishing and discussing the recursive structure of squares and cubes, we give algorithms for counting the number of repeated squares and cubes in $\mathbb{F}[1,n]$ for all $n$, and get explicit expressions for some special $n$ such as $n=f_m$ (the Fibonacci number) etc., which including some known results such as in A.S.Fraenkel and J.Simpson[@FS1999; @FS2014], J.Shallit et al[@DMSS2014].
**Key words:** the Fibonacci sequence; square; cube; algorithm; the sequence of return words.
Introduction
============
Let $\mathcal{A}=\{a,b\}$ be a binary alphabet. The concatenation of factors $\nu$ and $\omega$ denoted by $\nu\omega$. The Fibonacci sequence $\mathbb{F}$ is the fixed point beginning with $a$ of the Fibonacci morphism $\sigma$ defined by $\sigma(a)=ab$ and $\sigma(b)=a$. As a classical example over a binary alphabet, $\mathbb{F}$ having many remarkable properties, we refer to M.Lothaire[@L1983; @L2002], J.M.Allouche and J.Shallit[@AS2003], Berstel[@B1966; @B1980].
Let $\omega$ be a factor of $\mathbb{F}$, denoted by $\omega\prec\mathbb{F}$. Since $\mathbb{F}$ is uniformly recurrent, $\omega$ occurs infinitely many times. Let $\omega_p$ be the $p$-th occurrence of $\omega$. If the factor $\omega$ and integer $p$ such that $\omega_p\omega_{p+1}$ (resp. $\omega_p\omega_{p+1}\omega_{p+2}$) is the factor of $\mathbb{F}$, we call it a square (resp. cube) of $\mathbb{F}$. As we know, $\mathbb{F}$ contains no fourth powers. The properties of squares and cubes are objects of a great interest in many aspects of mathematics and computer science etc.
We denote $F_m=\sigma^m(a)$ for $m\geq0$ and define $F_{-1}=b$, $F_{-2}=\varepsilon$ (empty word). Define $f_m=|F_m|$ the $m$-th Fibonacci number, $f_{-2}=0$, $f_{-1}=1$, $f_{m+1}=f_m+f_{m-1}$ for $m\geq -1$. Let $\mathbb{F}[1,n]$ be the prefix of $\mathbb{F}$ of length $n$. In this paper, we consider the four functions below:
$A(n):=\sharp\{\omega:\omega\omega\prec\mathbb{F}[1,n]\}$, the number of distinct squares in $\mathbb{F}[1,n]$;
$B(n):=\sharp\{(\omega,p):\omega_p\omega_{p+1}\prec\mathbb{F}[1,n]\}$, the number of repeated squares in $\mathbb{F}[1,n]$;
$C(n):=\sharp\{\omega:\omega\omega\omega\prec\mathbb{F}[1,n]\}$, the number of distinct cubes in $\mathbb{F}[1,n]$;
$D(n):=\sharp\{(\omega,p):\omega_p\omega_{p+1}\omega_{p+2}\prec\mathbb{F}[1,n]\}$, the number of repeated cubes in $\mathbb{F}[1,n]$.
The methods for counting the four functions have attracted some many authors, but known results are not rich. A.S.Fraenkel and J.Simpson gave the expression of $A(f_m)$ and $B(f_m)$ in 1999[@FS1999] and 2014[@FS2014]. In 2014, C.F.Du, H.Mousavi, L.Schaeffer and J.Shallit gave the expression of $B(f_m)$ and $D(f_m)$ by mechanical methods, see Theorem 58 and Theorem 59 in [@DMSS2014]. In this paper, we give the explicit expressions of $A(n)$, $B(f_m)$, $C(n)$ and $D(f_m)$. Although we haven’t get the explicit expressions of $B(n)$ and $D(n)$, we give fast algorithms for counting $B(n)$ and $D(n)$ for all $n$.
The main tool of this paper is the “structure properties" of the sequence of return words in the Fibonacci sequence, which introduced and studied in [@HW2015-1], also see Property \[G\]. The definition of return words is from F.Durand[@D1998]. Let $\omega$ be a factor of $\mathbb{F}$. For $p\geq1$, let $\omega_p=x_{i+1}\cdots x_{i+n}$ and $\omega_{p+1}=x_{j+1}\cdots x_{j+n}$. The factor $x_{i+1}\cdots x_{j}$ is called the $p$-th return word of $\omega$ and denoted by $r_{p}(\omega)$. The sequence $\{r_p(\omega)\}_{p\geq1}$ is called the sequences of the return words of factor $\omega$.
By the “structure properties" (Property \[G\]), we can determine the positions of all $\omega_p$. By the definition of square (resp. cube) and return word, we have $$\omega_p\omega_{p+1}\prec\mathbb{F}\Leftrightarrow r_p(\omega)=\omega,~~
\omega_p\omega_{p+1}\omega_{p+2}\prec\mathbb{F}\Leftrightarrow r_p(\omega)=r_{p+1}(\omega)=\omega,$$ where the “=" means “have the same expressions". By these relations, we can determine the positions of all squares and cubes, and then get $A(n)$, $B(n)$, $C(n)$ and $D(n)$. But this method is complicated, another improved and fast method is used in this paper.
This paper is organized as follows. Section 2 present some basic notations and known results. Section 3 prove some basic properties of squares. We determine $A(n)$ (distinct squares) in Section 4. Section 5 is devoted to establish the recursive structure of squares, then we determine $B(n)$ (repeated squares) in Section 6. Similarly, we establish the recursive structure of cubes, then determine $C(n)$ (distinct cubes) and $D(n)$ (repeated cubes) in Section 7 to 10.
Preliminaries
=============
Let $\tau=x_1\cdots x_n$ be a finite word (or $\tau=x_1x_2\cdots$ be a sequence). For any $i\leq j\leq n$, define $\tau[i,j]:=x_ix_{i+1}\cdots x_{j-1}x_j$. By convention, we denote $\tau[i]:=\tau[i,i]=x_i$ and $\tau[i,i-1]=\varepsilon$. Notation $\nu\triangleright\omega$ means word $\nu$ is a suffix of word $\omega$.
For $m\geq-1$, let $\delta_m\in\{a,b\}$ be the last letter of $F_m$, then $\delta_m=a$ iff $m$ is even. The $m$-th singular word is defined as $K_m=\delta_{m+1}F_m\delta_m^{-1}=\delta_{m+1}F_m[1,f_{m}-1]$ for $m\geq-1$. By Property 2(9) in [@WW1994], all singular words are palindromes. Let $Ker(\omega)$ be the maximal singular word occurring in factor $\omega$, called the kernel of $\omega$. Then by Theorem 1.9 in [@HW2015-1], $Ker(\omega)$ occurs in $\omega$ only once. Moreover
\[wp\] $Ker(\omega_p)=Ker(\omega)_p$ for all $\omega\in\mathbb{F}$ and $p\geq1.$
This means, let $Ker(\omega)=K_m$, then the maximal singular word occurring in $\omega_p$ is just $K_{m,p}$. For instance, $Ker(aba)=b$, $(aba)_3=\mathbb{F}[6,8]$, $(b)_3=\mathbb{F}[7]$, so $Ker((aba)_3)=(b)_3$, $(aba)_3=a(b)_3a$.
\[G\] For any factor $\omega$, the sequence of return words $\{r_p(\omega)\}_{p\geq1}$ is the Fibonacci sequence over the alphabet $\{r_1(\omega),r_2(\omega)\}$.
Property \[k1\] and \[k2\] are useful in our proofs. Property \[k1\] can be proved by induction. Since all singular words are palindromes, Property \[k2\] holds by the cylinder structure of palindromes in [@HW2016-1].
\[k1\] For $m\geq-1$, (1) $K_{m+3}=K_{m+1}K_{m}K_{m+1}$.
\(2) $K_{m+2}=K_{m}K_{m+1}\delta_{m}^{-1}\delta_{m+1}=\delta_{m}^{-1}\delta_{m+1}K_{m+1}K_{m}$.
\[k2\] $K_{m}\prec K_{m+3}[2,f_{m+3}-1]$, $K_{m+1}\not\!\prec K_{m+3}[2,f_{m+3}-1]$, $K_{m+2}\not\!\prec K_{m+3}[2,f_{m+3}-1]$.
Basic properties of squares
===========================
By Definition 2.9 and Corollary 2.10 in [@HW2015-1], any factor $\omega$ with kernel $K_m$ can be expressed uniquely as $\omega=K_{m+1}[i,f_{m+1}] K_m K_{m+1}[1,j]=K_{m+3}[i,f_{m+2}+j],$ where $2\leq i\leq f_{m+1}+1$ and $0\leq j\leq f_{m+1}-1$. By Property \[wp\], $\omega_p\omega_{p+1}\prec\mathbb{F}$ means $$\omega_p\omega_{p+1}=K_{m+1}[i,f_{m+1}]
\underbrace{K_{m,p} K_{m+1}[1,j]K_{m+1}[i,f_{m+1}]}_{r_p(K_m)}
K_{m,p+1} K_{m+1}[1,j]\prec\mathbb{F}.$$
By Property \[G\], $K_m$ has only two distinct return words $r_1(K_m)=K_mK_{m+1}$ and $r_2(K_m)=K_mK_{m-1}$, so $\omega_p\omega_{p+1}\prec\mathbb{F}$ has two cases as below, and in each case, $|\omega|=|r_p(K_m)|$.
**Case 1.** $r_p(K_m)=r_1(K_m)=K_mK_{m+1}$. Comparing the two expressions of $r_p(K_m)$, we have $$K_mK_{m+1}[1,j]K_{m+1}[i,f_{m+1}]=K_mK_{m+1}\Rightarrow
j=i-1.$$ Comparing the two ranges of $i$ that $2\leq i\leq f_{m+1}+1$ and $0\leq j=i-1\leq f_{m+1}-1$, we have $2\leq i\leq f_{m+1}$ and $m\geq0$. Moreover $|\omega|=|r_1(K_m)|=f_{m+2}$ and $$\begin{split}
\omega\omega=&
K_{m+1}[i,f_{m+1}] K_{m} K_{m+1} K_{m} K_{m+1}[1,i-1]\\
=&K_{m+2}[i,f_{m+2}]\underline{K_{m+1}} K_{m+2}[1,f_{m}+i-1]
=K_{m+4}[i,2f_{m+2}+i-1].
\end{split}$$ The second and third equalities hold by Property \[k1\].
Since $K_{m+1}\prec\omega\omega\prec K_{m+4}[2,f_{m+4}-1]$, by Property \[k2\], $Ker(\omega\omega)=K_{m+1}$.
**Case 2.** $r_p(K_m)=r_2(K_m)=K_mK_{m-1}$. Comparing the two expressions, we have $$K_mK_{m+1}[1,j]K_{m+1}[i,f_{m+1}]=K_mK_{m-1}
\Rightarrow j=i-f_m-1.$$ So $f_m+1\leq i\leq f_{m+1}+1$ and $m\geq-1$. Moreover $|\omega|=|r_2(K_m)|=f_{m+1}$ and $$\begin{split}
\omega\omega=&
K_{m+1}[i,f_{m+1}] K_{m} K_{m-1} K_{m} K_{m+1}[1,i-f_m-1]\\
=&K_{m+1}[i,f_{m+1}] \underline{K_{m+2}} K_{m+1}[1,i-f_m-1]
=K_{m+5}[f_{m+2}+i,f_{m+3}+f_{m+1}+i-1].
\end{split}$$ Since $K_{m+2}\prec\omega\omega\prec K_{m+5}[2,f_{m+4}-1]$, by Property \[k2\], $Ker(\omega\omega)=K_{m+2}$.
By the discussion above, we have: all squares in $\mathbb{F}$ are of length $2f_m$ for some $m\geq0$; for all $m\geq0$, there exists a square of length $2f_m$ in $\mathbb{F}$. This is a known result of P.S$\acute{e}\acute{e}$bold[@S1985].
\[P\] $P(K_m,p)=pf_{m+1}+(\lfloor\phi p\rfloor+1)f_{m}-1$ for $m\geq-1$, $p\geq1$.
\[P1\] $P(a,p)=p+\lfloor\phi p\rfloor$, $P(b,p)=2p+\lfloor\phi p\rfloor$ for $p\geq1$.
For $m,p\geq1$, we define two sets below $$\begin{cases}
\langle1,K_m,p\rangle:=
\{P(\omega\omega,p):Ker(\omega\omega)=K_m,|\omega|=f_{m+1},\omega\omega\prec\mathbb{F}\}\\
\langle2,K_m,p\rangle:=
\{P(\omega\omega,p):Ker(\omega\omega)=K_m,|\omega|=f_{m-1},\omega\omega\prec\mathbb{F}\}
\end{cases}$$ Obviously they correspond the two cases of squares respectively. By Property \[P\] we have $$\begin{split}
\langle1,K_m,p\rangle=&\{P(\omega,p):
\omega=K_{m+1}[i,f_{m+1}]K_{m}K_{m+1}[1,f_{m-1}+i-1],2\leq i\leq f_{m}\}\\
=&\{P(K_m,p)+f_{m-1}+i-1,2\leq i\leq f_{m}\}\\
=&\{pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1},\cdots,pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+2}-2\},\\
\langle2,K_m,p\rangle=&\{P(\omega,p):\omega=K_{m-1}[i,f_{m-1}] K_{m} K_{m-1}[1,i-f_{m-2}-1],f_{m-2}+1\leq i\leq f_{m-1}+1\}\\
=&\{P(K_m,p)+i-f_{m-2}-1,f_{m-2}+1\leq i\leq f_{m-1}+1\}\\
=&\{pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m}-1,\cdots,
pf_{m+1}+\lfloor\phi p\rfloor f_{m}+2f_{m-1}-1\}.
\end{split}$$
$\sharp\langle 1,K_m,p\rangle=f_m-1$ and $\sharp\langle 2,K_m,p\rangle=f_{m-3}+1$ for $m,p\geq1$.
The number of distinct squares in $\mathbb{F}[1,n]$
===================================================
Denote $a(n):=\sharp\{\omega:\omega\omega\triangleright\mathbb{F}[1,n],
\omega\omega\not\!\prec\mathbb{F}[1,n-1]\}$, obversely, $A(n)=\sum_{i=1}^n a(i)$. In order to count $a(n)$, we only need to consider $\langle i,K_m,1\rangle$ where $i=1,2$.
$\langle1,K_m,1\rangle=\{2f_{m+1},\cdots,f_{m+3}-2\}$, $\langle2,K_m,1\rangle=\{f_{m+2}-1,\cdots,
f_{m+1}+2f_{m-1}-1\}$.
It is easy to see that sets $\langle i,K_m,1\rangle$ are pairwise disjoint, and each set contains some consecutive integers. Therefore we get a chain $$\langle2,K_1,1\rangle,\langle1,K_1,1\rangle,\cdots,
\langle1,K_{m-1},1\rangle,\langle2,K_m,1\rangle,
\langle1,K_m,1\rangle,\langle2,K_{m+1},1\rangle,\cdots$$
By this chain, $a(n)=1$ iff $n\in\cup_{m\geq1}(\langle2,K_m,1\rangle\cup\langle1,K_m,1\rangle)$. The “$\cup$" means pairwise disjoint union in this paper. Moreover, we have $\langle1,K_m,1\rangle\cup\langle2,K_{m+1},1\rangle=\{2f_{m+1},\cdots,f_{m+2}+2f_{m}-1\}$.
$a(1)=a(2)=a(3)=0$, $a(4)=1$ and for $n\geq5$ $$a(n)=1\text{ iff }n\in\cup_{m\geq1}\{2f_{m+1},\cdots,f_{m+2}+2f_{m}-1\}.$$
One method for counting $A(n)$ is by $A(n)=\sum_{i=1}^n a(i)$. By consider $A(f_{m+2}+2f_{m}-1)$ for $m\geq1$, we can give a fast algorithm of $A(n)$ for all $n\geq1$. Since $\sum_{i=-1}^mf_i= f_{m+2}-1$, $$\begin{array}{rl}
&A(f_{m+2}+2f_{m}-1)=a(4)+\sum\limits_{i=1}^{m}\sharp\{2f_{i+1},\cdots,f_{i+2}+2f_{i}-1\}\\
=&1+\sum\limits_{i=1}^{m}(f_{i}+f_{i-2})
=1+\sum\limits_{i=-1}^{m}f_{i}-f_0-f_{-1}+\sum\limits_{i=-1}^{m-2}f_{i}
=f_{m+2}+f_{m}-3.
\end{array}$$
\[T4.3\] For all $n\geq1$, let $m$ satisfies $2f_m\leq n<2f_{m+1}$, $$A(n)=\begin{cases}
n-f_{m-1}-2,&n\leq f_{m+1}+2f_{m-1}-1;\\
f_{m+1}+f_{m-1}-3,&otherwise.
\end{cases}$$
When $f_{m+1}+2f_{m-1}\leq n\leq 2f_{m+1}-1$, $a(n)=0$, $A(n)=A(f_{m+1}+2f_{m-1}-1)=f_{m+1}+f_{m-1}-3$.
When $2f_m\leq n\leq f_{m+1}+2f_{m-1}-1$, $a(n)=1$, $A(n)=A(2f_m-1)+n-2f_m+1$. Since $A(2f_m-1)=A(f_{m}+2f_{m-2}-1)$, we have $A(n)=n-f_{m-1}-2$. Thus the conclusion holds.
Since $2f_{m-2}\leq f_{m}\leq f_{m-1}+2f_{m-3}-1$ for $m\geq2$, as a spacial case of Theorem \[T4.3\], $$A(f_{m})=f_{m}-f_{m-3}-2=2f_{m-2}-2.$$ This is a known result of A.S.Fraenkel and J.Simpson, see Theorem 1 in [@FS1999].
The recursive structure of squares
==================================
In this section, we establish a recursive structure of squares. Using it, we will count the number of repeated squares in $\mathbb{F}[1,n]$ (i.e. $B(n)$) in Section 6. For $m,p\geq1$, consider the vectors $$\begin{split}
\Gamma_{1,m,p}:&=[pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1}-1,\langle 1,K_m,p\rangle]\\
&=[pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1}-1,\cdots,pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+2}-2];\\
\Gamma_{2,m,p}:&=[\langle 2,K_m,p\rangle,pf_{m+1}+\lfloor\phi p\rfloor f_{m}+2f_{m-1},\cdots,pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1}-2]\\
&=[pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m}-1,\cdots,pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1}-2].
\end{split}$$ Here vector $[\langle i,K_m,p\rangle]$ means arrange all elements in set $\langle i,K_m,p\rangle$, $i=1,2$.
Obversely, each $\Gamma_{i,m,p}$ contains consecutive integers. The numbers of components in vectors $\Gamma_{1,m,p}$ and $\Gamma_{2,m,p}$ are $f_{m}$ and $f_{m-1}$ respectively. Moreover $\max\Gamma_{2,m,p}+1=\min\Gamma_{1,m,p}$ for $m,p\geq1$.
\[L\] $\lfloor\phi(p+\lfloor\phi p\rfloor+1)\rfloor=p$, $\lfloor\phi(2p+\lfloor\phi p\rfloor+1)\rfloor=p+\lfloor\phi p\rfloor$.
\[R1\] $\Gamma_{1,m,p}=[\Gamma_{2,m-1,P(a,p)+1},\Gamma_{1,m-1,P(a,p)+1}]$ for $m\geq2$, $p\geq1$.
By Corollary \[P1\], $P(a,p)+1=p+\lfloor\phi p\rfloor+1$. By Lemma \[L\], $\lfloor\phi (p+\lfloor\phi p\rfloor+1)\rfloor=p$. $$\begin{split}
&\min\Gamma_{2,m-1,P(a,p)+1}
=(p+\lfloor\phi p\rfloor+1)f_{m}+\lfloor\phi(p+\lfloor\phi p\rfloor+1)\rfloor f_{m-1}+f_{m-1}-1\\
=&(p+\lfloor\phi p\rfloor+1)f_{m}+p f_{m-1}+f_{m-1}-1=pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1}-1=\min\Gamma_{1,m,p};\\
&\max\Gamma_{1,m-1,P(a,p)+1}
=(p+\lfloor\phi p\rfloor+1)f_{m}+\lfloor\phi(p+\lfloor\phi p\rfloor+1)\rfloor f_{m-1}+f_{m+1}-2\\
=&(p+\lfloor\phi p\rfloor+1)f_{m}+p f_{m-1}+f_{m+1}-2=pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+2}-2=\max\Gamma_{1,m,p}.
\end{split}$$
Since $\max\Gamma_{2,m,p}+1=\min\Gamma_{1,m,p}$ for $m,p\geq1$, $\max\Gamma_{2,m-1,P(a,p)+1}+1=\min\Gamma_{1,m-1,P(a,p)+1}$. Thus the conclusion holds.
By an analogous argument, we have
\[R2\] $\Gamma_{2,m,p}=[\Gamma_{2,m-2,P(b,p)+1},\Gamma_{1,m-2,P(b,p)+1}]$ for $m\geq3$, $p\geq1$.
In Property \[R1\] and \[R2\], we establish the recursive relations for any $\Gamma_{1,m,p}$ ($m\geq2$) and $\Gamma_{2,m,p}$ ($m\geq3$). By the one-to-one correspondence between $\Gamma_{i,m,p}$ and $\langle i,K_m,p\rangle$, we can define the recursive structure over $\{\langle i,K_m,p\rangle|~i=1,2;~m,p\geq1\}$ denoted by $\mathcal{S}$. Each $\langle i,K_m,p\rangle$ is an element in $\mathcal{S}$. The recursive structure $\mathcal{S}$ is a family of finite trees with root $\langle i,K_m,1\rangle$ for all $i=1,2$, $m\geq1$; and with recursive relations: $$\begin{cases}
\tau_1\langle1,K_m,p\rangle=\langle2,K_{m-1},P(a,p)+1\rangle\cup\langle1,K_{m-1},P(a,p)+1\rangle&\text{for }m\geq2;\\
\tau_2\langle2,K_m,p\rangle=\langle2,K_{m-2},P(b,p)+1\rangle\cup\langle1,K_{m-2},P(b,p)+1\rangle&\text{for }m\geq3.
\end{cases}$$
Each $\langle i,K_m,p\rangle$ belongs to the recursive structure $\mathcal{S}$, $i=1,2$, $m,p\geq1$.
Each element $\langle i,K_m,1\rangle$ is root of a finite tree in $\mathcal{S}$. For $m,p\geq1$, $$\begin{cases}
\langle1,K_m,P(a,p)+1\rangle\in\tau_1\langle1,K_{m+1},p\rangle\\
\langle1,K_m,P(b,p)+1\rangle\in\tau_2\langle2,K_{m+2},p\rangle
\end{cases}
\text{and }
\begin{cases}
\langle2,K_m,P(a,p)+1\rangle\in\tau_1\langle1,K_{m+1},p\rangle\\
\langle2,K_m,P(b,p)+1\rangle\in\tau_2\langle2,K_{m+2},p\rangle
\end{cases}$$ Since $\mathbb{N}=\{1\}\cup\{P(a,p)+1\}\cup\{P(b,p)+1\}$, the recursive structure $\mathcal{S}$ contains all $\langle i,K_m,p\rangle$.
On the other hand, by the recursive relations $\tau_1$ and $\tau_2$, each element $\langle i,K_m,p\rangle$ has a unique position in $\mathcal{S}$. By Property \[R1\] and \[R2\], the trees in $\mathcal{S}$ are pairwise disjoint. Fig.1 and Fig.2 show the two finite trees in the recursive structure $\mathcal{S}$ with roots $\langle1,K_5,1\rangle$ and $\langle 2,K_5,1\rangle$ respectively.
(155,56) (126,1)[53]{} (120.5,5)[$\langle1,K_1,12\rangle$]{} (120,0)[(1,0)[16]{}]{} (120,8)[(1,0)[16]{}]{} (120,0)[(0,1)[8]{}]{} (136,0)[(0,1)[8]{}]{} (146,9)[51]{} (140.5,13)[$\langle2,K_1,12\rangle$]{} (140,8)[(1,0)[16]{}]{} (140,16)[(1,0)[16]{}]{} (140,8)[(0,1)[8]{}]{} (156,8)[(0,1)[8]{}]{} (95,1)[53]{} (95,5)[52]{} (90.5,9)[$\langle1,K_2,7\rangle$]{} (90,0)[(1,0)[15]{}]{} (90,12)[(1,0)[15]{}]{} (90,0)[(0,1)[12]{}]{} (105,0)[(0,1)[12]{}]{} (110,13)[50]{} (110,17)[49]{} (105.5,21)[$\langle2,K_2,7\rangle$]{} (105.5,12.5)[(1,0)[14]{}]{} (105.5,24)[(1,0)[14]{}]{} (105.5,12.5)[(0,1)[11.5]{}]{} (119.5,12.5)[(0,1)[11.5]{}]{} (65,1)[53]{} (65,5)[52]{} (65,9)[51]{} (65,13)[50]{} (60.5,17)[$\langle1,K_3,4\rangle$]{} (60,0)[(1,0)[15]{}]{} (60,20)[(1,0)[15]{}]{} (60,0)[(0,1)[20]{}]{} (75,0)[(0,1)[20]{}]{} (35,1)[53]{} (35,5)[52]{} (35,9)[51]{} (35,13)[50]{} (35,17)[49]{} (35,21)[48]{} (35,25)[47]{} (30.5,29)[$\langle1,K_4,2\rangle$]{} (30,0)[(1,0)[15]{}]{} (30,32)[(1,0)[15]{}]{} (30,0)[(0,1)[32]{}]{} (45,0)[(0,1)[32]{}]{} (126,21)[48]{} (120.5,25)[$\langle1,K_1,11\rangle$]{} (120.5,20)[(1,0)[15.5]{}]{} (120.5,28)[(1,0)[15.5]{}]{} (120.5,20)[(0,1)[8]{}]{} (136,20)[(0,1)[8]{}]{} (80,25)[47]{} (80,29)[46]{} (75.5,33)[$\langle2,K_3,4\rangle$]{} (75,24)[(1,0)[14.5]{}]{} (75,36)[(1,0)[14.5]{}]{} (75,24)[(0,1)[12]{}]{} (89.5,24)[(0,1)[12]{}]{} (146,29)[46]{} (140.5,33)[$\langle2,K_1,11\rangle$]{} (140,28)[(1,0)[16]{}]{} (140,36)[(1,0)[16]{}]{} (140,28)[(0,1)[8]{}]{} (156,28)[(0,1)[8]{}]{} (95,33)[45]{} (95,37)[44]{} (90.5,41)[$\langle1,K_2,6\rangle$]{} (90.5,32)[(1,0)[14.5]{}]{} (90.5,44)[(1,0)[14.5]{}]{} (90.5,32)[(0,1)[12]{}]{} (105,32)[(0,1)[12]{}]{} (146,41)[43]{} (140.5,45)[$\langle2,K_1,10\rangle$]{} (140,40)[(1,0)[16]{}]{} (140,48)[(1,0)[16]{}]{} (140,40)[(0,1)[8]{}]{} (156,40)[(0,1)[8]{}]{} (126,33)[45]{} (120.5,37)[$\langle1,K_1,10\rangle$]{} (120,32)[(1,0)[16]{}]{} (120,40)[(1,0)[16]{}]{} (120,32)[(0,1)[8]{}]{} (136,32)[(0,1)[8]{}]{} (50,41)[43]{} (50,45)[42]{} (50,49)[41]{} (45.5,53)[$\langle2,K_4,2\rangle$]{} (45,40)[(1,0)[15]{}]{} (45,56)[(1,0)[15]{}]{} (45,40)[(0,1)[16]{}]{} (60,40)[(0,1)[16]{}]{} (110,45)[42]{} (110,49)[41]{} (105.5,53)[$\langle2,K_2,6\rangle$]{} (105.5,44)[(1,0)[14.5]{}]{} (105.5,56)[(1,0)[14.5]{}]{} (105.5,44)[(0,1)[12]{}]{} (120,44)[(0,1)[12]{}]{} (5,1)[53]{} (5,5)[52]{} (5,9)[51]{} (5,13)[50]{} (5,17)[49]{} (5,21)[48]{} (5,25)[47]{} (5,29)[46]{} (5,33)[45]{} (5,37)[44]{} (5,41)[43]{} (5,45)[42]{} (0.5,49)[$\langle1,K_5,1\rangle$]{} (0,0)[(1,0)[15]{}]{} (0,52)[(1,0)[15]{}]{} (0,0)[(0,1)[52]{}]{} (15,0)[(0,1)[52]{}]{} (63,52)[$\tau_2$]{} (61,50)[(3,-1)[28]{}]{} (61,50)[(1,0)[43]{}]{} (18,38)[$\tau_1$]{} (16,34)[(1,-1)[13]{}]{} (16,34)[(2,1)[28]{}]{} (108,37)[$\tau_1$]{} (106,42)[(2,-1)[13]{}]{} (106,42)[(1,0)[33]{}]{} (93,29.5)[$\tau_2$]{} (91,26)[(1,0)[28]{}]{} (91,26)[(3,1)[15]{}]{} (106,31)[(1,0)[33]{}]{} (48,29)[$\tau_1$]{} (46,27)[(1,-1)[13]{}]{} (46,27)[(1,0)[28]{}]{} (78,19)[$\tau_1$]{} (76,17)[(1,-1)[13]{}]{} (76,17)[(1,0)[28]{}]{} (108,4)[$\tau_1$]{} (106,10)[(2,-1)[13]{}]{} (106,10)[(1,0)[33]{}]{}
Fig.1: The finite tree in the recursive structure $\mathcal{S}$ with root $\langle1,K_5,1\rangle$.
(150,40) (5,17)[36]{} (5,21)[35]{} (5,25)[34]{} (5,29)[33]{} (0.5,33)[$\langle2,K_5,1\rangle$]{} (0,16)[(1,0)[15]{}]{} (0,36)[(1,0)[15]{}]{} (0,16)[(0,1)[20]{}]{} (15,16)[(0,1)[20]{}]{} (19,24)[$\tau_2$]{} (16,30)[(2,-1)[28]{}]{} (16,30)[(1,0)[43]{}]{} (50,1)[40]{} (50,5)[39]{} (50,9)[38]{} (50,13)[37]{} (45.5,17)[$\langle1,K_3,3\rangle$]{} (45,20)[(1,0)[15]{}]{} (45,0)[(1,0)[15]{}]{} (45,0)[(0,1)[20]{}]{} (60,0)[(0,1)[20]{}]{} (64,13)[$\tau_1$]{} (61,10)[(2,-1)[13]{}]{} (61,10)[(3,1)[28]{}]{} (65,25)[34]{} (65,29)[33]{} (60.5,33)[$\langle2,K_3,3\rangle$]{} (60,24)[(1,0)[15]{}]{} (60,36)[(1,0)[15]{}]{} (60,24)[(0,1)[12]{}]{} (75,24)[(0,1)[12]{}]{} (79,37)[$\tau_2$]{} (76,35)[(3,-1)[30]{}]{} (76,35)[(1,0)[48]{}]{} (80,1)[40]{} (80,5)[39]{} (75.5,9)[$\langle1,K_2,5\rangle$]{} (75,0)[(1,0)[15]{}]{} (75,12)[(1,0)[15]{}]{} (75,0)[(0,1)[12]{}]{} (90,0)[(0,1)[12]{}]{} (94,1)[$\tau_1$]{} (91,4)[(1,0)[15]{}]{} (91,4)[(3,1)[33]{}]{} (95,13)[37]{} (95,17)[36]{} (90.5,21)[$\langle2,K_2,5\rangle$]{} (90,12.5)[(1,0)[15]{}]{} (90,24)[(1,0)[15]{}]{} (90,12.5)[(0,1)[11.5]{}]{} (105,12.5)[(0,1)[11.5]{}]{} (112,1)[40]{} (107.5,5)[$\langle1,K_1,9\rangle$]{} (107,0)[(1,0)[15]{}]{} (107,8)[(1,0)[15]{}]{} (107,0)[(0,1)[8]{}]{} (122,0)[(0,1)[8]{}]{} (112,21)[35]{} (107.5,25)[$\langle1,K_1,8\rangle$]{} (107,20)[(1,0)[15]{}]{} (107,28)[(1,0)[15]{}]{} (107,20)[(0,1)[8]{}]{} (122,20)[(0,1)[8]{}]{} (130,9)[38]{} (125.5,13)[$\langle2,K_1,9\rangle$]{} (125,8)[(1,0)[15]{}]{} (125,16)[(1,0)[15]{}]{} (125,8)[(0,1)[8]{}]{} (140,8)[(0,1)[8]{}]{} (130,29)[33]{} (125.5,33)[$\langle2,K_1,8\rangle$]{} (125,28)[(1,0)[15]{}]{} (125,36)[(1,0)[15]{}]{} (125,28)[(0,1)[8]{}]{} (140,28)[(0,1)[8]{}]{}
Fig.2: The finite tree in the recursive structure $\mathcal{S}$ with root $\langle2,K_5,1\rangle$.
By the recursive structure $\mathcal{S}$, we have the relation between the number of squares ending at position $\Gamma_{1,m,p}[i]=pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1}+i-1$ and $\Gamma_{1,m,1}[i]=2f_{m+1}+i-1$, see Property \[P5.5\]. Similarly, we have the relation between the number of squares ending at position $\Gamma_{2,m,p}[i]=pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m}+i-2$ and $\Gamma_{2,m,1}[i]=f_{m+2}+i-2$, see Property \[P5.6\].
\[P5.5\] For $1\leq i\leq f_m-1$, $$\begin{split}
&\{\omega:\omega\omega\triangleright\mathbb{F}[1,2f_{m+1}+i-1]\}\\
=&\{\omega:\omega\omega\triangleright\mathbb{F}[1,pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1}+i-1],Ker(\omega)=K_j,1\leq j\leq m\}.
\end{split}$$
\[P5.6\] For $1\leq i\leq f_{m-3}+1$, $$\begin{split}
&\{\omega:\omega\omega\triangleright\mathbb{F}[1,f_{m+2}+i-2]\}\\
=&\{\omega:\omega\omega\triangleright\mathbb{F}[1,pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m}+i-2],Ker(\omega)=K_j,1\leq j\leq m\}.
\end{split}$$
For instance, taking $m=3$, $p=3$ and $i=2$ in the property above. All squares ending at position 13 are $\{aabaab\}$. All squares ending at position 34 are $\{aabaab,abaabaababaabaab\}$. Since $Ker(aabaab)=aabaa=K_{3}$ and $Ker(abaabaababaabaab)=aabaababaabaa=K_5$, only $\{aabaab\}$ is square with kernel $K_{j}$, $1\leq j\leq 3$. Fig.3 shows the relation:
(150,48) (5,2)[36]{} (5,6)[35]{} (5,10)[34]{} (5,14)[33]{} (0.5,18)[$\langle2,K_5,1\rangle$]{} (0,1)[(1,0)[15]{}]{} (0,21)[(1,0)[15]{}]{} (0,1)[(0,1)[20]{}]{} (15,1)[(0,1)[20]{}]{} (19,9)[$\tau_2$]{} (16,15)[(2,-1)[28]{}]{} (16,15)[(1,0)[43]{}]{} (48,0)[$\cdots\cdots$]{} (65,10)[34]{} (65,14)[33]{} (60.5,18)[$\langle2,K_3,3\rangle$]{} (60,9)[(1,0)[15]{}]{} (60,21)[(1,0)[15]{}]{} (60,9)[(0,1)[12]{}]{} (75,9)[(0,1)[12]{}]{} (79,22)[$\tau_2$]{} (76,20)[(3,-1)[30]{}]{} (76,20)[(1,0)[48]{}]{} (112,6)[35]{} (107.5,10)[$\langle1,K_1,8\rangle$]{} (107,5)[(1,0)[15]{}]{} (107,13)[(1,0)[15]{}]{} (107,5)[(0,1)[8]{}]{} (122,5)[(0,1)[8]{}]{} (130,14)[33]{} (125.5,18)[$\langle2,K_1,8\rangle$]{} (125,13)[(1,0)[15]{}]{} (125,21)[(1,0)[15]{}]{} (125,13)[(0,1)[8]{}]{} (140,13)[(0,1)[8]{}]{} (67,24)[$\vdots$]{} (114,24)[$\vdots$]{} (132,24)[$\vdots$]{} (65,35)[13]{} (65,39)[12]{} (60.5,43)[$\langle2,K_3,1\rangle$]{} (60,34)[(1,0)[15]{}]{} (60,46)[(1,0)[15]{}]{} (60,34)[(0,1)[12]{}]{} (75,34)[(0,1)[12]{}]{} (79,47)[$\tau_2$]{} (76,45)[(3,-1)[30]{}]{} (76,45)[(1,0)[48]{}]{} (112,31)[14]{} (107.5,35)[$\langle1,K_1,3\rangle$]{} (107,30)[(1,0)[15]{}]{} (107,38)[(1,0)[15]{}]{} (107,30)[(0,1)[8]{}]{} (122,30)[(0,1)[8]{}]{} (130,39)[12]{} (125.5,43)[$\langle2,K_1,3\rangle$]{} (125,38)[(1,0)[15]{}]{} (125,46)[(1,0)[15]{}]{} (125,38)[(0,1)[8]{}]{} (140,38)[(0,1)[8]{}]{}
Fig.3: An example of the graph embedding in the recursive structure $\mathcal{S}$.
From Fig.3 we can see that: in the tree with root $\langle2,K_5,1\rangle$, the branch from node $\langle2,K_3,3\rangle$ is the graph embedding of the tree with root $\langle2,K_3,1\rangle$.
The number of repeated squares in $\mathbb{F}[1,n]$
===================================================
Denote $b(n):=\sharp\{(\omega,p):\omega_p\omega_{p+1}\triangleright\mathbb{F}[1,n]\}$ the number of squares ending at position $n$. By the definition of $\langle i,K_m,p\rangle$, $b(n)$ is equal to the number of integer $n$ occurs in the recursive structure $\mathcal{S}$. Thus we can calculate $b(n)$ by the property below.
\[b\] $b([1,2,3])=[0,0,0]$, $b(\Gamma_{2,1,1})=b([4])=[1]$, $b(\Gamma_{1,1,1})=b([5,6])=[0,1]$, $b(\Gamma_{2,2,1})=b([7,8])=[1,1]$, $b(\Gamma_{1,2,1})=b([9,10,11])=[1,1,2]$, for $m\geq3$, $$\begin{cases}
~b(\Gamma_{1,m,1})
=[b(\Gamma_{2,m-1,1}),b(\Gamma_{1,m-1,1})]+[0,\underbrace{1,\cdots,1}_{f_m-1}];\\
~b(\Gamma_{2,m,1})
=[b(\Gamma_{2,m-2,1}),b(\Gamma_{1,m-2,1})]
+[\underbrace{1,\cdots,1}_{f_{m-3}+1},\underbrace{0,\cdots,0}_{f_{m-2}-1}].
\end{cases}$$
The first few values of $b(n)$ are $b([1,2,3])=[0,0,0]$, $b([4])=[1]$, $b([5,6])=[0,1]$,
$b([7,8])=[1,1]$, $b([9,10,11])=[1,1,2]$, $b([12,13,14])=[2,1,1]$, $b([15,\cdots,19])=[1,2,2,2,3]$,
$b([20,\cdots,24])=[2,2,2,1,2]$, $b([25,\cdots,32])=[2,2,2,2,3,3,3,4]$.
For $m\geq3$, the immediately corollaries are $$\begin{cases}
\sum b(\Gamma_{1,m,1})=\sum b(\Gamma_{2,m-1,1})+\sum b(\Gamma_{1,m-1,1})+f_m-1;\\
\sum b(\Gamma_{2,m,1})=\sum b(\Gamma_{2,m-2,1})+\sum b(\Gamma_{1,m-2,1})+f_{m-3}+1.
\end{cases}$$
\[P6.1\] For $m\geq1$, (1) $\sum b(\Gamma_{1,m,1})=\frac{2m+5}{5}f_{m}+\frac{2m-6}{5}f_{m-2}-1$,
\(2) $\sum b(\Gamma_{2,m,1})=\frac{2m-2}{5}f_{m-1}+\frac{2m-3}{5}f_{m-3}+1$.
Since $\Gamma_{1,m,1}=[2f_{m+1}-1,\cdots,f_{m+3}-2]$ and $\Gamma_{2,m,1}=[f_{m+2}-1,\cdots,2f_{m+1}-2]$, we have $B(f_{m+3}-2)
=B(f_{m+2}-2)+\sum b(\Gamma_{2,m,1})+\sum b(\Gamma_{1,m,1})$ and $B(2f_{m+1}-2)=B(f_{m+2}-2)+\sum b(\Gamma_{2,m,1})$. Thus by induction and Property \[P6.1\],
\[B\](1) $B(f_{m+3}-2)=\frac{2m-4}{5}f_{m+3}+\frac{2m}{5}f_{m+1}+4$ for $m\geq-1$.
\(2) $B(2f_{m+1}-2)=\frac{4m-11}{5}f_{m+1}+\frac{4m-3}{5}f_{m-1}+5$ for $m\geq0$.
Property \[C6.4\] can be proved by induction and Property \[b\].
\[C6.4\] $b(f_m-1)=\lfloor\frac{m-1}{2}\rfloor$, $b(f_m)=\lfloor\frac{m}{2}-1\rfloor$, $b(f_m-1)+b(f_m)=m-2$ for $m\geq2$.
Since $B(f_m)=B(f_{m}-2)+b(f_m-1)+b(f_m)$, $m\geq2$. By Property \[B\] and \[C6.4\], $$B(f_m)=\tfrac{4}{5}(m+1)f_{m}-\tfrac{2}{5}(m+7)f_{m-1}-4f_{m-2}+m+2
=\tfrac{4m-16}{5}f_{m}-\tfrac{2m-6}{5}f_{m-1}+m+2.$$ This is a known result of A.S.Fraenkel and J.Simpson[@FS2014].
Obversely we can calculate $B(n)$ by $B(n)=\sum_{i=4}^n b(i)$. But when $n$ is large, this method is complicated. Now we turn to give a fast algorithm. For any $n\geq4$, let $m$ such that $f_m\leq n+1< f_{m+1}$. Since we already determine the expression of $B(f_m-2)$ and $B(2f_{m-1}-2)$ for $m\geq2$, in order to give a fast algorithm of $B(n)$, we only need to calculate $\sum_{i=f_{m}-1}^nb(i)$ or $\sum_{i=2f_{m-1}-1}^nb(i)$. One method is calculating $b(n)$ by Property \[b\], the other method is using the corollaries as below.
\[c1\] For $n\geq4$, let $m$ such that $f_m\leq n+1\leq2f_{m-1}-1$, then $m\geq3$ and $$\sum_{i=f_{m}-1}^nb(i)
=\begin{cases}
\sum\limits_{i=f_{m-2}-1}^{n-f_{m-1}}b(i)+n-f_{m}+2,&n+1\leq f_{m}+f_{m-5}-1;\\
\sum\limits_{i=f_{m-2}+f_{m-5}-1}^{n-f_{m-1}}b(i)
+\frac{2m-5}{5}f_{m-5}+\frac{2m-11}{5}f_{m-7}+2,&otherwise.
\end{cases}$$
By Property \[b\], when $f_m\leq n+1\leq f_{m}+f_{m-5}-1$, $$\begin{array}{rl}
\sum\limits_{i=f_{m}-1}^nb(i)=\sum\limits_{i=f_{m-2}-1}^{n-f_{m-1}}[b(i)+1]
=\sum\limits_{i=f_{m-2}-1}^{n-f_{m-1}}b(i)+n-f_{m}+2.
\end{array}$$
When $f_{m}+f_{m-5}\leq n+1\leq2f_{m-1}-1$, $\sum\limits_{i=f_{m}-1}^nb(i)=
\sum\limits_{i=f_{m}-1}^{f_{m}+f_{m-5}-2}b(i)+\sum\limits_{i=f_{m}+f_{m-5}-1}^{n}b(i)$, where $$\left\{\begin{array}{rl}
\sum\limits_{i=f_{m}-1}^{f_{m}+f_{m-5}-2}b(i)
=&\sum\limits_{i=f_{m-2}-1}^{f_{m-2}+f_{m-5}-2}[b(i)+1]
=\sum b(\Gamma_{2,m-4,1})+f_{m-5}\\
=&\frac{2m-5}{5}f_{m-5}+\frac{2m-11}{5}f_{m-7}+1;\\
\sum\limits_{i=f_{m}+f_{m-5}-1}^{n}b(i)
=&\sum\limits_{i=f_{m-2}+f_{m-5}-1}^{n-f_{m-1}}b(i)+1.
\end{array}\right.$$ Thus $\sum\limits_{i=f_{m}-1}^nb(i)=
\sum\limits_{i=f_{m-2}+f_{m-5}-1}^{n-f_{m-1}}b(i)
+\frac{2m-5}{5}f_{m-5}+\frac{2m-11}{5}f_{m-7}+2$. The conclusion holds.
\[c2\]For $n\geq9$, let $m$ such that $2f_{m-1}\leq n+1\leq f_{m+1}-1$, then $m\geq4$ and $$\sum_{i=2f_{m-1}-1}^nb(i)
=\begin{cases}
\sum\limits_{i=f_{m-1}-1}^{n-f_{m-1}}b(i)+n-2f_{m-1}+1,
~~~~~~~~~~~~~~~~~~n+1\leq f_{m}+f_{m-2}-1;\\
\sum\limits_{i=2f_{m-2}-1}^{n-f_{m-1}}b(i)+n-2f_{m-1}
+\frac{2m-8}{5}f_{m-4}+\frac{2m-9}{5}f_{m-6}+2,~otherwise.
\end{cases}$$
By Property \[b\], when $2f_{m-1}\leq n+1\leq 2f_{m-1}+f_{m-4}-1=f_{m}+f_{m-2}-1$, $$\begin{array}{c}
\sum\limits_{i=2f_{m-1}-1}^nb(i)=\sum\limits_{i=f_{m-1}-1}^{n-f_{m-1}}b(i)+\sum\limits_{i=f_{m-1}}^{n-f_{m-1}}1
=\sum\limits_{i=f_{m-1}-1}^{n-f_{m-1}}b(i)+n-2f_{m-1}+1.
\end{array}$$ When $f_{m}+f_{m-2}\leq n+1\leq f_{m+1}-1$, $\sum\limits_{i=2f_{m-1}-1}^nb(i)
=\sum\limits_{i=2f_{m-1}-1}^{f_{m}+f_{m-2}-2}b(i)
+\sum\limits_{i=f_{m}+f_{m-2}-1}^nb(i)$, where $$\left\{\begin{array}{rl}
\sum\limits_{i=2f_{m-1}-1}^{f_{m}+f_{m-2}-2}b(i)
=&\sum\limits_{i=f_{m-1}-1}^{2f_{m-2}-2}b(i)+\sum\limits_{i=f_{m-1}}^{2f_{m-2}-2}1
=\sum\limits_{i=f_{m-1}-1}^{2f_{m-2}-2}b(i)+f_{m-4}-1\\
=&\sum b(\Gamma_{2,m-3,1})+f_{m-4}-1
=\frac{2m-3}{5}f_{m-4}+\frac{2m-9}{5}f_{m-6};\\
\sum\limits_{i=f_{m}+f_{m-2}-1}^nb(i)
=&\sum\limits_{i=2f_{m-2}-1}^{n-f_{m-1}}[b(i)+1]
=\sum\limits_{i=2f_{m-2}-1}^{n-f_{m-1}}b(i)+n-2f_{m-1}-f_{m-4}+2.
\end{array}\right.$$ Thus $\sum\limits_{i=2f_{m-1}-1}^nb(i)
=\sum\limits_{i=2f_{m-2}-1}^{n-f_{m-1}}b(i)+n-2f_{m-1}
+\frac{2m-8}{5}f_{m-4}+\frac{2m-9}{5}f_{m-6}+2$. The conclusion holds.
**Example.** One method to calculate $\sum_{i=20}^{23}b(i)$ is by Property \[b\]. Since $b(\Gamma_{2,4,1})=b([20,\cdots,24])=[2,2,2,1,2]$, $\sum_{i=20}^{23}b(i)=7$. The other method is using Corollary \[c1\] and \[c2\]: $$\begin{array}{c}
\sum\limits_{i=20}^{23}b(i)
=\sum\limits_{i=f_{4}+f_{1}-1}^{23-f_{5}}b(i)
+\frac{7}{5}f_{1}+\frac{1}{5}f_{-1}+2
=\sum\limits_{i=9}^{10}b(i)+5=7.
\end{array}$$
Step 1. For $n\leq3$, $B(n)=0$; for $n\leq4$, find the $m$ such that $f_m\leq n+1< f_{m+1}$.
Step 2. Compare $n$ with $2f_{m-1}-1$.
\(1) If $n<2f_{m-1}-1$, calculate $B(f_m-2)$ by Property \[B\]; calculate $\sum_{i=f_{m}-1}^{n}b(i)$ by Property \[b\] or by Corollary \[c1\] and \[c2\]. Then $B(n)=B(f_{m}-2)+\sum_{i=f_{m}-1}^nb(i)$.
\(2) If $n\geq2f_{m-1}-1$, calculate $B(2f_{m-1}-2)$ by Property \[B\]; calculate $\sum_{i=2f_{m-1}-1}^nb(i)$ by Property \[b\] or by Corollary \[c1\] and \[c2\]. Then $B(n)=B(2f_{m-1}-2)+\sum_{i=2f_{m-1}-1}^nb(i)$.
When $m$ is large (resp. small), Corollary \[c1\] and \[c2\] (resp. Property \[b\]) is faster.
**Example.** We calculate $B(23)$. Since $f_{6}=21\leq 23+1<f_{7}=34$, $m=6$. Moreover $23<2f_{5}-1$.
By Property \[B\], $B(f_{6}-2)=B(19)=\frac{2}{5}f_{6}+\frac{6}{5}f_{4}+4=22$. By Property \[b\] or by Corollary \[c1\] and \[c2\], $\sum_{i=20}^{23}b(i)=7$. Thus $B(23)=B(19)+\sum_{i=20}^{23}b(i)=22+7=29$.
Basic properties of cubes
=========================
Let $\omega$ be a factor with kernel $K_m$, by an analogous argument as Section 3 and by Proposition 4.8 in [@HW2015-1], $\omega_{p}\omega_{p+1}\omega_{p+2}\prec\mathbb{F}$ has only one case: $r_p(K_m)=r_{p+1}(K_m)=r_1(K_m)=K_mK_{m+1}$. In this case, $|\omega|=f_{m+2}$. Moreover $2\leq i\leq f_{m+1}$ and $m\geq0$, $$\begin{split}
\omega\omega\omega=
&K_{m+1}[i,f_{m+1}] K_{m} K_{m+1}K_{m} K_{m+1} K_{m} K_{m+1}[1,i-1]\\
=&K_{m+2}[i,f_{m+2}] \underline{K_{m+3}} K_{m+2}[1,i+f_{m}-1]
=K_{m+6}[i+f_{m+3},i+f_{m+5}+f_{m}-1].
\end{split}$$ Since $K_{m+3}\prec\omega\omega\omega\prec K_{m+6}[2,f_{m+6}-1]$, by Property \[k2\], $Ker(\omega\omega\omega)=K_{m+3}$.
By the discussion above, we have: all cubes in $\mathbb{F}$ are of length $3f_m$ for some $m\geq2$, and a cube of each such length occurs. This is Theorem 8 in J.Shallit et al[@DMSS2014].
For $m\geq3$ and $p\geq1$, we define a set below: $$\langle K_m,p\rangle:=
\{P(\omega\omega\omega,p):Ker(\omega\omega\omega)=K_m,|\omega|=f_{m-1},
\omega\omega\omega\prec\mathbb{F}\}.$$ Obviously it contains all cubes. By Property \[P\] we have $$\begin{split}
\langle K_m,p\rangle=&\{P(\omega,p):
\omega=K_{m-1}[i,f_{m-1}] K_{m} K_{m-1}[1,i+f_{m-3}-1],2\leq i\leq f_{m-2}\}\\
=&\{P(K_m,p)+f_{m-3}+i-1,2\leq i\leq f_{m-2}\}\\
=&\{pf_{m+1}+\lfloor\phi p\rfloor f_{m}+2f_{m-1},\cdots,pf_{m+1}+\lfloor\phi p\rfloor f_{m}+f_{m+1}-2\}.
\end{split}$$
$\sharp\langle K_m,p\rangle=f_{m-2}-1$ for $m\geq3$, $p\geq1$.
The number of distinct cubes in $\mathbb{F}[1,n]$
=================================================
Denote $c(n):=\sharp\{\omega:\omega\omega\omega\triangleright\mathbb{F}[1,n],
\omega\omega\omega\not\!\prec\mathbb{F}[1,n-1]\}$. Obversely, $C(n)=\sum_{i=1}^n c(i)$.
$\langle K_m,1\rangle=\{f_{m+1}+2f_{m-1},\cdots,2f_{m+1}-2\}$ for $m\geq3$.
Sets $\langle K_m,1\rangle$ are pairwise disjoint, and each set contains some consecutive integers. We get a chain $\langle K_3,1\rangle=\{14\},\langle K_4,1\rangle=\{23,24\},\cdots,
\langle K_m,1\rangle,\cdots$. So $c(n)=1$ iff $n\in\cup_{m\geq3}\langle K_m,1\rangle$. Thus
$c(n)=1$ iff $n\in\cup_{m\geq3}\{f_{m+1}+2f_{m-1},\cdots,2f_{m+1}-2\}$.
By consider $C(2f_{m+1}-2)$ for $m\geq3$, we can give a fast algorithm of $C(n)$ for all $n\geq1$. Since $\sum_{i=-1}^mf_i= f_{m+2}-1$, $C(2f_{m+1}-2)=\sum_{i=3}^{m}\sharp\langle K_i,1\rangle
=\sum_{i=3}^{m}(f_{i-2}-1)=f_{m}-m-1$.
For $n<14$, $C(n)=0$; for $n\geq14$, let $m$ s.t. $f_{m+1}+2f_{m-1}\leq n<f_{m+2}+2f_{m}-1$, then $m\geq3$ and $$C(n)=\begin{cases}
n-f_{m+1}-f_{m-1}-m+1,&n\leq 2f_{m+1}-2;\\
f_{m}-m-1,&otherwise.
\end{cases}$$
When $2f_{m+1}-1\leq n\leq f_{m+2}+2f_{m}-1$, $c(n)=0$, $C(n)=C(2f_{m+1}-2)=f_{m}-m-1$.
When $f_{m+1}+2f_{m-1}\leq n\leq 2f_{m+1}-2$, $c(n)=1$, $$C(n)=C(f_{m+1}+2f_{m-1}-1)+n-f_{m+1}-2f_{m-1}+1.$$ Since $C(f_{m+1}+2f_{m-1}-1)=C(2f_{m}-2)=f_{m-1}-m$, $C(n)=n-f_{m+1}-f_{m-1}-m+1.$ Thus the conclusion holds.
Since $2f_{m-2}-1\leq f_m\leq f_{m-1}+2f_{m-3}-1$ for $m\geq6$, we have
$C(f_m)=0$ for $m\leq5$, $C(f_{m})=f_{m-3}-m+2$ for $m\geq6$.
The recursive structure of cubes
================================
In this section, we establish a recursive structure of cubes. Using it, we will count the number of repeated cubes in $\mathbb{F}[1,n]$ (i.e. $D(n)$) in Section 10.
\[R3\] For $m\geq5$, $\min\langle K_m,p\rangle-2=\max\langle K_{m-2},P(b,p)+1\rangle$.
Since $P(b,p)=2p+\lfloor\phi p\rfloor$, $\lfloor\phi(2p+\lfloor\phi p\rfloor+1)\rfloor=p+\lfloor\phi p\rfloor$, for $m\geq3$, $f_{m-1}+f_{m-4}=2f_{m-2}$, $$\begin{split}
&\max\langle K_{m-2},P(b,p)+1\rangle+2=(2p+\lfloor\phi p\rfloor+1)f_{m-1}+\lfloor\phi (2p+\lfloor\phi p\rfloor+1)\rfloor f_{m-2}+f_{m-1}\\
=&(2p+\lfloor\phi p\rfloor+1)f_{m-1}+(p+\lfloor\phi p\rfloor)f_{m-2}+f_{m-1}
=pf_{m+1}+\lfloor\phi p\rfloor f_{m}+2f_{m-1}=\min\langle K_m,p\rangle.
\end{split}$$ This means $\max\langle K_{m-2},P(b,p)+1\rangle+2=\min\langle K_m,p\rangle$, so the conclusion holds.
By an analogous argument, we have
\[R4\] For $m\geq4$, $\max\langle K_m,p\rangle+f_{m-4}+2=\min\langle K_{m-1},P(a,p)+1\rangle$.
In Property \[R3\] and \[R4\], we establish the recursive relations for any $\langle K_m,p\rangle$, $m\geq3$. Thus we can define the recursive structure over $\{\langle K_m,p\rangle|~m\geq3,p\geq1\}$ denoted by $\mathcal{C}$. Each $\langle K_m,p\rangle$ is an element in $\mathcal{C}$. The recursive structure $\mathcal{C}$ is a family of finite trees with roots $\langle K_m,1\rangle$ for all $m\geq3$; and with recursive relations: $$\begin{cases}
\tau_3\langle K_{m},p\rangle=\langle K_{m-2},P(b,p)+1\rangle\cup\langle K_{m-1},P(a,p)+1\rangle\text{ for }m\geq5;\\
\tau_4\langle K_4,p\rangle=\langle K_{m-1},P(a,p)+1\rangle.
\end{cases}$$ Since $\max\langle K_{m-2},P(b,p)+1\rangle<\min\langle K_{m-1},P(a,p)+1\rangle$, the “$\cup$" is a disjoint union.
Each $\langle K_m,p\rangle$ belongs to the recursive structure $\mathcal{C}$, for $m\geq3$ and $p\geq1$.
Each element $\langle K_m,1\rangle$ is root of a finite tree in $\mathcal{C}$. For $p\geq1$, $$\begin{cases}
\langle K_m,P(a,p)+1\rangle\in\tau_3\langle K_{m+1},p\rangle~(m\geq4)
\text{ and }\langle K_3,P(a,p)+1\rangle\in\tau_4\langle K_{4},p\rangle;\\
\langle K_m,P(b,p)+1\rangle\in\tau_3\langle K_{m+2},p\rangle~(m\geq3).
\end{cases}$$ Since $\mathbb{N}=\{1\}\cup\{P(a,p)+1\}\cup\{P(b,p)+1\}$, the recursive structure $\mathcal{C}$ contains all $\langle K_m,p\rangle$.
On the other hand, by the recursive relations $\tau_3$ and $\tau_4$, each element $\langle K_m,p\rangle$ has a unique position in $\mathcal{C}$. Fig.4 show the finite tree in the recursive structure $\mathcal{C}$ with root $\langle K_6,1\rangle$.
(75,72) (65,1)[82]{} (61.5,5)[$\langle K_3,7\rangle$]{} (61,0)[(1,0)[11]{}]{} (61,0)[(0,1)[9]{}]{} (72,0)[(0,1)[9]{}]{} (61,9)[(1,0)[11]{}]{} (65,29)[69]{} (61.5,33)[$\langle K_3,6\rangle$]{} (61,28)[(1,0)[11]{}]{} (61,28)[(0,1)[9]{}]{} (72,28)[(0,1)[9]{}]{} (61,37)[(1,0)[11]{}]{} (65,53)[61]{} (61.5,57)[$\langle K_3,5\rangle$]{} (61,52)[(1,0)[11]{}]{} (61,52)[(0,1)[9]{}]{} (72,52)[(0,1)[9]{}]{} (61,61)[(1,0)[11]{}]{} (45,5)[79]{} (45,9)[78]{} (41.5,13)[$\langle K_4,4\rangle$]{} (41,4)[(1,0)[11]{}]{} (41,4)[(0,1)[13]{}]{} (52,4)[(0,1)[13]{}]{} (41,17)[(1,0)[11]{}]{} (45,61)[58]{} (45,65)[57]{} (41.5,69)[$\langle K_4,3\rangle$]{} (41,60)[(1,0)[11]{}]{} (41,60)[(0,1)[13]{}]{} (52,60)[(0,1)[13]{}]{} (41,73)[(1,0)[11]{}]{} (25,13)[74]{} (25,17)[73]{} (25,21)[72]{} (25,25)[71]{} (21.5,29)[$\langle K_5,2\rangle$]{} (21,12)[(1,0)[11]{}]{} (21,12)[(0,1)[21]{}]{} (32,12)[(0,1)[21]{}]{} (21,33)[(1,0)[11]{}]{} (5,33)[66]{} (5,37)[65]{} (5,41)[64]{} (5,45)[63]{} (5,49)[62]{} (5,53)[61]{} (5,57)[60]{} (1.5,61)[$\langle K_6,1\rangle$]{} (1,32)[(1,0)[11]{}]{} (1,32)[(0,1)[33]{}]{} (12,32)[(0,1)[33]{}]{} (1,65)[(1,0)[11]{}]{} (14,45)[$\tau_3$]{} (13,40)[(1,1)[27]{}]{} (13,40)[(1,-2)[7]{}]{} (34,24)[$\tau_3$]{} (33,22)[(3,1)[27]{}]{} (33,22)[(1,-1)[7]{}]{} (54,10)[$\tau_4$]{} (53,10)[(1,-1)[7]{}]{} (54,65)[$\tau_4$]{} (53,65)[(1,-1)[7]{}]{}
Fig.4: The finite tree in the recursive structure $\mathcal{C}$ with root $\langle K_6,1\rangle$.
For $m\geq1$, (1) $P(a,f_m-1)=f_{m+1}-2$, $\lfloor\phi (f_m-1)\rfloor=f_{m-1}-1$.
\(2) $\lfloor\phi f_m\rfloor=f_{m-1}$ if $m$ is odd; $\lfloor\phi f_m\rfloor=f_{m-1}-1$ if $m$ is even. (3) $P(b,f_{2m})=f_{2m+2}-1$.
Denote by $|\omega|_a$ (resp. $|\omega|_b$) the number of letter $a$ (resp. $b$) occurring in $\omega$.
\(1) Since $|F_{m+1}|_a=f_{m}$, $aba\triangleright F_{2m}$ and $aab\triangleright F_{2m+1}$, we have $P(a,f_m-1)=f_{m+1}-2$. On the other hand, by Corollary \[P1\], $P(a,f_m-1)=f_m-1+\lfloor\phi (f_m-1)\rfloor$. Comparing the two expressions of $P(a,f_m-1)$, we have $\lfloor\phi (f_m-1)\rfloor=f_{m-1}-1$ for $m\geq1$.
\(2) By Corollary \[P1\], $P(a,f_m)=f_m+\lfloor\phi f_m\rfloor$. By the analogous argument in (1), we have: when $m$ is odd, $P(a,f_m)=f_{m+1}$, then $P(a,f_m)=f_{m+1}=f_m+\lfloor\phi f_m\rfloor\Rightarrow\lfloor\phi f_m\rfloor=f_{m-1}$; when $m$ is even, $P(a,f_m)=f_{m+1}-1$, then $P(a,f_m)=f_{m+1}-1=f_m+\lfloor\phi f_m\rfloor\Rightarrow\lfloor\phi f_m\rfloor=f_{m-1}-1$.
\(3) Since $|F_m|_b=f_{m-2}$, $aba\triangleright F_{2m}$, we have $P(b,f_{2m})=f_{2m+2}-1$ for $m\geq1$.
$f_{m}f_{k}+f_{m-1}f_{k-1}=f_{m+k+1}$ for $m,k\geq-1$.
Since $f_{m}f_{k}+f_{m-1}f_{k-1}=f_{m}(f_{k-1}+f_{k-2})+f_{m-1}f_{k-1}
=f_{m}f_{k-2}+(f_{m}+f_{m-1})f_{k-1}$, using it repeatedly, $f_{m}f_{k}+f_{m-1}f_{k-1}=f_{m}f_{k-2}+f_{m+1}f_{k-1}=\cdots=f_{m+k-1}f_{-1}+f_{m+k}f_{0}=f_{m+k+1}$.
For $m\geq3$, we define the vectors $\Gamma_{m}:=[f_{m+2}-1,\cdots,f_{m+3}-2]$, then
\[P9.6\]The finite tree with root $\langle K_m,1\rangle$ belongs to $\Gamma_{m}$ for $m\geq3$.
\(1) Since $P(a,f_m-1)=f_{m+1}-2$, the maximal of the recursive structure from $\langle K_{m},1\rangle$ is $$\max\{\max\langle K_{m},1\rangle,\max\langle K_{m-1},f_2-1\rangle,\max\langle K_{m-2},f_3-1\rangle,\cdots,\max\langle K_3,f_{m-2}-1\rangle\}$$ By Property \[R4\], $\max\langle K_{m},p\rangle<\min\langle K_{m-1},P(a,p)+1\rangle$, so $\max\langle K_{m-i},f_{i+1}-1\rangle$ is strictly increasing for $0\leq i\leq m-3$. Thus the maximal integer in the tree is $\max\langle K_3,f_{m-2}-1\rangle$. $$\begin{split}
&\max\langle K_3,f_{m-2}-1\rangle=(f_{m-2}-1)f_{4}+\lfloor\phi(f_{m-2}-1)\rfloor f_{3}+f_{4}-2\\
=&(f_{m-2}-1)f_4+(f_{m-3}-1)f_3+6=f_{m-2}f_4+f_{m-3}f_3-7=f_{m+3}-7<\max\Gamma_m.
\end{split}$$
\(2) Similarly, since $P(b,f_{2m})=f_{2m+2}-1$, $\min\langle K_{m-2i},f_{2i}\rangle$ is strictly decreasing for $0\leq i\leq[\frac{m-4}{2}]$. So the minimal integer in the tree is $$\min\{\min\langle K_{m},1\rangle,\min\langle K_{m-2},f_2\rangle,\min\langle K_{m-4},f_4\rangle,\cdots\}=
\begin{cases}
\min\langle K_4,f_{m-4}\rangle& \text{if $m$ is even;}\\
\min\langle K_3,f_{m-3}\rangle& \text{if $m$ is odd.}
\end{cases}$$
When $m$ is even, $\min\langle K_4,f_{m-4}\rangle=f_{m-4}f_{5}+\lfloor\phi f_{m-4}\rfloor f_{4}+2f_{3}$, $\lfloor\phi f_{m-4}\rfloor=f_{m-5}-1$, so $$\min\langle K_4,f_{m-4}\rangle
=f_{m-4}f_{5}+(f_{m-5}-1)f_{4}+2f_{3}=f_{m+2}+2>\min\Gamma_m.$$
When $m$ is odd, $\min\langle K_3,f_{m-3}\rangle=f_{m-3}f_{4}+\lfloor\phi f_{m-3}\rfloor f_{3}+2f_{2}$, $\lfloor\phi f_{m-3}\rfloor=f_{m-4}-1$, so $$\min\langle K_3,f_{m-3}\rangle
=f_{m-3}f_{4}+(f_{m-4}-1)f_{3}+2f_{2}
=f_{m+2}+1>\min\Gamma_m.$$
In each case, the minimal integer in the tree is larger than $\min\Gamma_m$, so the conclusion holds.
By Property \[P9.6\] and the definition of $\Gamma_m$, the finite trees in recursive structure $\mathcal{C}$ with different roots $\langle K_{m},1\rangle$ are disjoint.
The number of repeated cubes in $\mathbb{F}[1,n]$
=================================================
Denote $d(n):=\sharp\{(\omega,p):\omega_p\omega_{p+1}\omega_{p+2}\triangleright\mathbb{F}[1,n]\}$, the number of cubes ending at position $n$. Obversely, $D(n)=\sum_{i=1}^n d(i)$. By the definition of $\langle K_m,p\rangle$, $d(n)$ is equal to the number of integer $n$ occurs in the recursive structure $\mathcal{C}$. Thus we can calculate $d(n)$ by the property below.
\[d\] For $m\geq3$, $$\begin{split}
&d([f_{m+4}-1,\cdots,f_{m+5}-2])\\
=&d([f_{m+2}-1,\cdots,f_{m+3}-2,f_{m+3}-1,\cdots,f_{m+4}-2])
+[\underbrace{0,\cdots,0}_{f_{m-1}+1},\underbrace{1,\cdots,1}_{f_m-1},\underbrace{0,\cdots,0}_{f_{m+2}}].
\end{split}$$
The first few values of $d(n)$ are $d([f_{5}-1,\cdots,f_{6}-2])=[d(12),\cdots,d(19)]=[0,0,1,0,0,0,0,0]$,
$d([f_{6}-1,\cdots,f_{7}-2])=[d(20),\cdots,d(32)]=[0,0,0,1,1,0,0,1,0,0,0,0,0]$,
$d([f_{7}-1,\cdots,f_{8}-2])=[d(33),\cdots,d(53)]=[0,0,1,0,1,1,1,1,0,0,0,1,1,0,0,1,0,0,0,0,0]$.
By Property \[d\], $\sum d(\Gamma_{m+2})=\sum d(\Gamma_{m})+\sum d(\Gamma_{m+1})+f_{m-2}-1$. By induction, we have
$\sum\limits_{f_{m+2}-1}^{f_{m+3}-2} d(n)
=\frac{m-5}{5}f_{m}+\frac{m+2}{5}f_{m-2}+1$ for $m\geq3$.
By the definition of $D(n)$, $D(f_{m+1}-2)=D(f_{m}-2)+\sum_{f_{m}-1}^{f_{m+1}-2} d(n)$. By induction, we have
\[D\] $D(f_m-2)
=\frac{m-11}{5}f_{m-1}+\frac{m+1}{5}f_{m-3}+m+1$ for $m\geq6$.
By Property \[d\], we get $d(f_m-1)=d(f_m)=0$ easily by induction, thus $D(f_m)=D(f_m-2)$.
\[T10.4\] $D(f_m)
=\frac{m-11}{5}f_{m-1}+\frac{m+1}{5}f_{m-3}+m+1$ for $m\geq6$.
Theorem 59 in [@DMSS2014] shows the number of cube occurrences in $F_m$ as $$D(f_m)=[d_1(m+2)+d_2]\alpha^{m+2}+[d_3(m+2)+d_4]\beta^{m+2}+m+1.$$ where $\alpha=\frac{1+\sqrt{5}}{2}$, $\beta=\frac{1-\sqrt{5}}{2}$, $d_1=\frac{3-\sqrt{5}}{10}$, $d_2=\frac{17}{50}\sqrt{5}-\frac{3}{2}$, $d_3=\frac{3+\sqrt{5}}{10}$, $d_4=-\frac{17}{50}\sqrt{5}-\frac{3}{2}$. Since $$\begin{array}{c}
f_m=\frac{\alpha^{m+2}-\beta^{m+2}}{\alpha-\beta},~
\alpha\beta=-1,~\frac{1}{\alpha}=\frac{\sqrt{5}-1}{2},~
\frac{1}{\beta}=\frac{-1-\sqrt{5}}{2},~ (\frac{1}{\alpha})^3=\sqrt{5}-2,~(\frac{1}{\beta})^3=-\sqrt{5}-2,
\end{array}$$ we can prove the two expressions are same. By our expression in Theorem \[T10.4\], $$\begin{array}{rl}
&D(f_m)-m-1=\frac{m-11}{5}\times\frac{\alpha^{m+1}-\beta^{m+1}}{\alpha-\beta}
+\frac{m+1}{5}\times\frac{\alpha^{m-1}-\beta^{m-1}}{\alpha-\beta}\\
=&\frac{m-11}{5\sqrt{5}}\times\left(\frac{\sqrt{5}-1}{2}\alpha^{m+2}
+\frac{1+\sqrt{5}}{2}\beta^{m+2}\right)
+\frac{m+1}{5\sqrt{5}}\times\left((\sqrt{5}-2)\alpha^{m+2}+(\sqrt{5}+2)\beta^{m+2}\right)\\
=&[\frac{m-11}{5\sqrt{5}}\times\frac{\sqrt{5}-1}{2}+\frac{m+1}{5\sqrt{5}}(\sqrt{5}-2)]\alpha^{m+2}
+[\frac{m-11}{5\sqrt{5}}\times\frac{1+\sqrt{5}}{2}+\frac{m+1}{5\sqrt{5}}(\sqrt{5}+2)]\beta^{m+2}\\
=&[\frac{3-\sqrt{5}}{10}m+\frac{7\sqrt{5}-45}{50}]\alpha^{m+2}
+[\frac{3+\sqrt{5}}{10}m+\frac{-7\sqrt{5}-45}{50}]\beta^{m+2}.
\end{array}$$
By J.shallit’s expression in [@DMSS2014], $$\begin{array}{c}
D(f_m)-m-1=[\frac{3-\sqrt{5}}{10}m+\frac{3-\sqrt{5}}{5}+\frac{17}{50}\sqrt{5}-\frac{3}{2}]\alpha^{m+2}
+[\frac{3+\sqrt{5}}{10}m+\frac{3+\sqrt{5}}{5}-\frac{17}{50}\sqrt{5}-\frac{3}{2}]\beta^{m+2}.
\end{array}$$ Comparing the coefficients of $m\alpha^{m+2}$, $\alpha^{m+2}$, $m\beta^{m+2}$ and $\beta^{m+2}$, we have the two expressions are same.
For any $n\geq12$, let $m$ such that $f_m\leq n+1< f_{m+1}$. Since we already determine the expression of $D(f_m-2)$, in order to give a fast algorithm of $D(n)$, we only need to calculate $\sum_{i=f_{m}-1}^nd(i)$. One method is calculating $d(n)$ by Property \[d\], the other method is using the corollaries as below.
\[c3\]For $n\geq12$, let $m$ such that $f_m\leq n+1< f_{m+1}$, then $m\geq5$ and $$\sum_{i=f_{m}-1}^nd(i)
=\begin{cases}
\sum\limits_{i=f_{m-2}-1}^{n-f_{m-1}}d(i),&f_{m}\leq n+1\leq f_{m}+f_{m-5};\\
\sum\limits_{i=f_{m-2}+1}^{n-f_{m-1}}d(i)+n-f_{m}-f_{m-5}+1,&f_{m}+f_{m-5}+1\leq n+1\leq f_{m}+f_{m-3}-1;\\
\sum\limits_{i=f_{m-1}-1}^{n-f_{m-1}}d(i)+\frac{m-4}{5}f_{m-4}+\frac{m-2}{5}f_{m-6},&f_{m}+f_{m-3}\leq n+1< f_{m+1}.
\end{cases}$$
By Property \[d\], when $f_{m}\leq n+1\leq f_{m}+f_{m-5}$, $\sum\limits_{i=f_{m}-1}^nd(i)
=\sum\limits_{i=f_{m-2}-1}^{n-f_{m-1}}d(i)$.
When $f_{m}+f_{m-5}+1\leq n+1\leq f_{m}+f_{m-3}-1$, $$\begin{array}{rl}
&\sum\limits_{i=f_{m}-1}^nd(i)
=\sum\limits_{i=f_{m}-1}^{f_{m}+f_{m-5}-1}d(i)
+\sum\limits_{i=f_{m}+f_{m-5}}^{n}d(i)
=\sum\limits_{i=f_{m-2}+1}^{n-f_{m-1}}d(i)
+\sum\limits_{i=f_{m-2}+f_{m-5}}^{n-f_{m-1}}1\\
=&\sum\limits_{i=f_{m-2}+1}^{n-f_{m-1}}d(i)+n-f_{m}-f_{m-5}+1.
\end{array}$$
When $f_{m}+f_{m-3}\leq n+1< f_{m+1}$, $$\begin{array}{rl}
&\sum\limits_{i=f_{m}-1}^nd(i)
=\sum\limits_{i=f_{m}-1}^{f_{m}+f_{m-5}-1}d(i)
+\sum\limits_{i=f_{m}+f_{m-5}}^{f_{m}+f_{m-3}-2}d(i)
+\sum\limits_{i=f_{m}+f_{m-3}-1}^{n}d(i)\\
=&\sum\limits_{i=f_{m-2}-1}^{f_{m-1}-2}d(i)+f_{m-4}-1
+\sum\limits_{i=f_{m-1}-1}^{n-f_{m-1}}d(i)
=\sum\limits_{i=f_{m-1}-1}^{n-f_{m-1}}d(i)+\frac{m-4}{5}f_{m-4}+\frac{m-2}{5}f_{m-6}.
\end{array}$$
So the conclusion holds.
**Example.** Since $[d(33),\cdots,d(53)]=[0,0,1,0,1,1,1,1,0,0,0,1,1,0,0,1,0,0,0,0,0]$, using Property \[d\], we have $\sum_{i=33}^{48}d(i)=8$. The other method is using Corollary \[c3\]. Since $f_{7}+f_{4}=42\leq n+1=49< f_{8}=55$, $\sum_{i=33}^{48}d(i)
=\sum_{i=f_{6}-1}^{48-f_{6}}d(i)+\frac{3}{5}f_{3}+\frac{5}{5}f_{1}
=\sum_{i=20}^{27}d(i)+5
=\sum_{i=12}^{14}d(i)+7=8.$
Step 1. For $n\leq11$, $D(n)=0$; for $n\leq12$, find the $m$ such that $f_m\leq n+1< f_{m+1}$.
Step 2. Calculate $D(f_m-2)$ by Property \[D\].
Step 3. Calculate $\sum_{i=f_{m}-1}^{n}d(i)$ by Property \[d\] or by Corollary \[c3\].
Step 4. $D(n)=D(f_{m}-2)+\sum_{i=f_{m}-1}^{n}d(i)$.
**Example.** We calculate $D(48)$. Since $f_{7}=34\leq 48+1< f_{8}=55$, $m=7$.
By Property \[D\], $D(32)=D(f_7-2)=\frac{-4}{5}f_{6}+\frac{8}{5}f_{4}+7+1=4$.
By Property \[d\] or by Corollary \[c3\], $\sum_{i=33}^{48}d(i)=8$. Thus $D(48)=D(32)+\sum_{i=33}^{48}d(i)=12$.
****
The research is supported by the Grant NSF No.11431007, No.11271223 and No.11371210.
[AB]{}
J.M.Allouche, J.Shallit. Automatic sequences: Theory, applications, generalizations. Cambridge University Press, Cambridge, 2003.
J.Berstel. Recent results in Sturmian words, in J.Dassow, A.Salomaa (Eds.), Developments in Language Theory, World Scientific, Singapore. (1966) 13-24.
J.Berstel. Mot de Fibonacci, S$\acute{e}$minaire d’informatique th$\acute{e}$rique, L.I.T.P., Paris, 1980/1981, 57-78.
W.-F.Chuan, H.-L.Ho. Locating factors of the infinite Fibonacci word, Theoretical Computer Science. 349 (2005) 429-442.
W.-T.Cao, Z.-Y.Wen. Some properties of the factors of Sturmian sequences, Theoretical Computer Science. 304 (2003) 365-385.
F.Durand. A characterization of substitutive sequences using return words, Discrete Math. 179 (1998) 89-101.
C.-F.Du, H.Mousavi, L.Schaeffer, J.Shallit. Decision Algorithms for Fibonacci-Automatic Words, with Applications to Pattern Avoidance. Eprint Arxiv, 2014.
A.S.Fraenkel, J.Simpson. The exact number of squares in Fibonacci words, Theoretical Computer Science. 218 (1999) 95-106.
A.S.Fraenkel, J.Simpson. Corrigendum to “The exact number of squares in Fibonacci words”, Theoretical Computer Science. 547 (2014) 122. A.Glen. On Sturmian and Episturmian Words, and Related Topics, PhD thesis, The University of Adelaide, Australia. 2006.
Y.-K.Huang, Z.-Y.Wen. The sequence of return words of the Fibonacci sequence, Theoretical Computer Science. 593 (2015) 106-116.
Y.-K.Huang, Z.-Y.Wen. Kernel words and gap sequence of the Tribonacci sequence, Acta Mathematica Scientia (Series B). 36.1 (2016) 173-194.
Y.-K.Huang, Z.-Y.Wen. The structure of palindromes in the Fibonacci sequence. arXiv: 1601.04391.
M.Lothaire. Combinatorics on words, in: Encyclopedia of Mathematics and its applications, Vol.17, Addison-Wesley, Reading, MA, 1983.
M.Lothaire. Algebraic combinatorics on words, Cambridge Univ. Press, Cambridge, 2002.
P.S$\acute{e}\acute{e}$bold. Propri$\acute{e}$t$\acute{e}$s combinatoires des mots infinis engendr$\acute{e}$s par certains morphismes (Th$\grave{e}$se de $3^e$ cycle). PhD thesis, Universit$\acute{e}$ P. et M. Curie, Institut de Programmation, Paris, 1985.
B.Tan, Z.-X.Wen. Invertible substitutions and Sturmian sequences. European Journal of Combinatorics, 24.8 (2003) 983-1002.
Z.-X.Wen, Z.-Y.Wen. Some properties of the singular words of the Fibonacci word, European J. Combin. 15 (1994) 587-598.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '[We use the super version of RS construction method and Gauss decomposition to obtain Drinfel$^{\prime}$d’s currents realization of super Yangian double $DY(gl(1|1))$ with center extension]{}'
---
‘=11 =23.5cm =-20truemm =0.0cm
**Super Yangian Double $DY(gl(1|1))$ With Center Extension**
$^a$Jin-fang Cai, $^a$Guo-xing Ju $^{bc}$S.K. Wang and $^a$Ke Wu
$^a$ Institute of Theoretical Physics, Academia Sinica,
Beijing, 100080, P. R. China
$^b$ CCAST (World Laboratory), P.O. Box 3730, Beijing, 100080, P. R. China
$^c$ Institute of Applied Mathematics, Academia Sinica,
Beijing, 100080, P. R. China
Yangians and quantum affine algebras are quantum algebras which are related respectively with rational and trigonometric solutions of Yang-Baxter equation [@DRI1]. The explicit isomorphism between two realizations of quantum affine algebras and Yangians given by Drinfel$^{\prime}$d [@DRI2] and RS [@RS] was established by Ding and Frenkel [@DF] using Gauss decomposition. In this paper, we use super version of RS construction method to define super Yangian double $DY(gl(1|1))$ [@C1] with center extension, and get its Drinfeld’s currents realization by Guass decomposition.
By $V$ we denote a ${\bf Z}_2$-graded two-dimension vector space (graded auxiliary space), and set the first and second basis of $V$ is even and odd respectively. In graded case, we must use this tensor product form: $(A \otimes B)(C \otimes D) =(-1)^{P(B)P(C)}
AC \otimes BD $, here $P(A)=0,1$ when $A$ is bosonic and Fermionic respectively. Then the graded(super) Yang-Baxter equation(YBE) takes this form[@LS]: $$\begin{aligned}
&&\eta _{12}R_{12}(u) \eta _{13}R_{13}(u+v) \eta _{23}R_{23}(v)
= \eta _{23}R_{23}(v) \eta _{13}R_{13}(u+v) \eta _{12}R_{12}(u),\end{aligned}$$ where $R(u) \in End(V\otimes V) $ and it must obey weight conservation condition: $R_{ij,kl}\neq 0 $ only when $i+j = k+l $. $\eta _{ik,jl}=(-1)^{(i-1)(k-1)} \delta _{ij} \delta _{lk}$ . It’s easily proved that the following $R(u)$ matrix is a rational solution of super YBE:
$$R(u)=\frac{1}{u+\hbar}(uI+{\cal P})=\left(\begin{array}{cccc}
1&0&0&0\\
0&\frac{u}{u+\hbar}&\frac{\hbar}{u+\hbar}&0\\
0&\frac{\hbar}{u+\hbar}&\frac{u}{u+\hbar}&0\\
0&0&0&\frac{u-\hbar}{u+\hbar} \end{array}\right)$$
This sulution of graded YBE satisfied unitarity condition: $R(u)R_{21}(-u)=I$. From this rational solution of graded YBE, we can define the super Yangian double $DY(gl(1|1))$ with a central extension employing super version of RS method [@RS]. Super Yangian Double\
$DY\left( gl(1|1)\right)$ is a Hopf algebra generated by $ \{ l_{ij}^k \vert 1\leq i,j \leq 2, k\in {\bf Z }\} $ which obey the following relations: $$\begin{aligned}
&&R(u-v)L_1^{\pm}(u)\eta L_2^{\pm}(v)\eta=\eta L_2^{\pm}(v)
\eta L_1^{\pm}(u)R(u-v) \label{rll1} \\
&&R(u_--v_+)L_1^+(u)\eta L_2^-(v)\eta=\eta L_2^-(v)
\eta L_1^+(u)R(u_+-v_-) \label{rll2}\end{aligned}$$ here $u_{\pm}=u\pm \frac{\hbar}{4}c$ and we have used standard notation: $L_1^{\pm}(u)=L^{\pm}(u)\otimes {\bf 1},
L_2^{\pm}(u)={\bf 1} \otimes
L^{\pm}(u) $ and $L^{\pm}(u)=\left( l_{ij}^{\pm}(u) \right)
_{1 \leq i,j \leq 2}$, $l_{ij}^{\pm}(u)$ are generate functions of $l_{ij}^k$ : $$l_{ij}^{+}(u)=\delta _{ij}-\hbar\sum_{k\geq 0}l_{ij}^k u^{-k-1},
{\hspace{0.5cm}}
l_{ij}^{-}(u)=\delta _{ij}+\hbar\sum_{k < 0}l_{ij}^k u^{-k-1}.$$ Hopf structure of $DY\left(gl(1\vert 1)\right)$ are defined as follows: $$\begin{aligned}
&&\triangle \left(l_{ij}^{\pm}(u)\right) =\sum_{k=1,2} l_{kj}^{\pm}(u \pm \frac{\hbar}{4}c_2)
\otimes l_{ik}^{\pm}(u\mp \frac{\hbar}{4}c_1)(-1)^{(k+i)(k+j)}, \label{cop} \\
&& \epsilon\left(l_{ij}^{\pm}(u)\right)=\delta _{ij} ,{\hspace{1.cm}}
S\left( ^{st} L^{\pm}(u) \right) =\left[ ^{st} L^{\pm}(u) \right]^{-1}.
\label{es}\end{aligned}$$ here $\left[^{st} L^{\pm}(u) \right]_{ij} =(-1)^{i+j}l_{ji}^{\pm}(u) $ .
From Ding-Frenkel theorem[@DF] on two realizations of Yangians and quantum affine algebras, $L^{\pm}(u)$ have the following decomposition: $$\begin{aligned}
L^{\pm}(u)&=&\left(\begin{array}{cc}1&0\\f^{\pm}(u)&1\end{array}\right)
\left(\begin{array}{cc}k_1^{\pm}(u)&0\\0&k_2^{\pm}(u)\end{array}\right)
\left(\begin{array}{cc}1&e^{\pm}(u)\\0&1\end{array}\right) \nonumber\\
&=&\left(\begin{array}{cc}k_1^{\pm}(u)&k_1^{\pm}(u)e^{\pm}(u)\\f^{\pm}(u)
k_1^{\pm}(u)&k_2^{\pm}(u)+f^{\pm}(u)k_1^{\pm}(u)e^{\pm}(u)\end{array}\right) \end{aligned}$$ and the inversions of $L^{\pm}(u)$ can be writed as: $$\begin{aligned}
L^{\pm}(u)^{-1}&=&\left(\begin{array}{cc}1&-e^{\pm}(u)\\0&1\end{array}\right)
\left(\begin{array}{cc}k_1^{\pm}(u)^{-1}&0\\0&k_2^{\pm}(u)^{-1}\end{array}\right)
\left(\begin{array}{cc}1&0\\-f^{\pm}(u)&1\end{array}\right) \nonumber\\
&=&\left(\begin{array}{cc}k_1^{\pm}(u)^{-1}+
e^{\pm}(u)k_2^{\pm}(u)^{-1}f^{\pm}(u)&-e^{\pm}(u)k_2^{\pm}(u)^{-1}\\
-k_2^{\pm}(u)^{-1}f^{\pm}(u)&k_2^{\pm}(u)^{-1}\end{array}\right) \end{aligned}$$ Let $$X^+(v)=e^+(v_-)-e^-(v_+) {\hspace{1cm}} X^-(v)=f^+(v_+)-f^-(v_-)$$
To calculate the (anti-)commutation relations of $X^{\pm}(u)$ and $k_i^{\pm}(v)$ (i=1,2), we will use the following equivalent forms of eq.(\[rll1\])(\[rll2\]): $$\begin{aligned}
L_1^{\pm}(u)^{-1}\eta L_2^{\pm}(v)^{-1}\eta R(u-v)&=
&R(u-v)\eta L_2^{\pm}(v)^{-1} \eta L_1^{\pm}(u)^{-1} \label{rll3} \\
L_1^+(u)^{-1}\eta L_2^-(v)^{-1}\eta R(u_--v_+)&=&R(u_+-v_-)\eta L_2^-(v)^{-1}
\eta L_1^+(u)^{-1} \label{rll4} \\
L_1^{\pm}(u)R(u-v)\eta L_2^{\pm}(v)^{-1} \eta &=&\eta L_2^{\pm}(v)^{-1}
\eta R(u-v)L_1^{\pm}(u)\label{rll5} \\
L_1^+(u)R(u_+-v_-)\eta L_2^-(v)^{-1} \eta &=&\eta L_2^-(v)^{-1}
\eta R(u_--v_+)L_1^+(u) \label{rll6}\\
L_1^{\pm}(u)^{-1}R_{21}(v-u)\eta L_2^{\pm}(v) \eta&=
&\eta L_2^{\pm}(v) \eta R_{21}(v-u)L_1^{\pm}(u)^{-1} \label{rll7}\\
L_1^+(u)^{-1}R_{21}(v_+-u_-)\eta L_2^-(v) \eta&=
&\eta L_2^-(v) \eta R_{21}(v_--u_+)L_1^+(u)^{-1} \label{rll8}\end{aligned}$$
From (\[rll1\]) (\[rll2\]) and (\[rll3\])—(\[rll8\]), we can get all relations between $k_1^{\pm}(u)$ and $ k_2^{\pm}(v) $ : $$\begin{aligned}
&&[k_1^{\pm}(u) ~,~k_1^{\pm}(v)]=[k_1^+(u) ~,~k_1^-(v)]=0 \label{k1}\\
&&[k_1^{\pm}(u) ~,~k_2^{\pm}(v)]=[k_2^{\pm}(u) ~,~k_2^{\pm}(v)]=0 \label{k2}\\
&&\frac{u_{\pm}-v_{\mp}}{u_{\pm}-v_{\mp}+\hbar}k_1^{\pm}(u)k_2^{\mp}(v)^{-1}
=\frac{u_{\mp}-v_{\pm}}{u_{\mp}-v_{\pm}+\hbar}k_2^{\mp}(v)^{-1}
k_1^{\pm}(u) \label{k3}\\
&&\frac{u_--v_+-\hbar}{u_--v_++\hbar}k_2^+(u)^{-1}k_2^-(v)^{-1}
=\frac{u_+-v_--\hbar}{u_+-v_-+\hbar}k_2^-(v)^{-1}k_2^+(u)^{-1} \label{k4}\end{aligned}$$
From (\[rll1\])(\[rll2\]) and unitarity of $R$-matrix, we have the following relations between $k_1^{\pm}(u)$ and $e^{\pm}(v), f^{\pm}(v)$:
$$\begin{aligned}
&&k_1^{\pm}(u)k_1^{\pm}(v)e^{\pm}(v)-\frac{u-v}{u-v+\hbar}k_1^{\pm}(v)
e^{\pm}(v)k_1^{\pm}(u) \nonumber\\
&&{\hspace{5cm}}-\frac{\hbar}{u-v+\hbar}k_1^{\pm}(v)k_1^{\pm}(u)
e^{\pm}(u)=0 \\
&&k_1^{\pm}(u)k_1^{\mp}(v)e^{\mp}(v)-\frac{u_{\pm}-v_{\mp}}{u_{\pm}-v_{\mp}
+\hbar}k_1^{\mp}(v)e^{\mp}(v)k_1^{\pm}(u) \nonumber\\
&&{\hspace{5cm}}-\frac{\hbar}{u_{\pm}-v_{\mp}+\hbar}
k_1^{\mp}(v)k_1^{\pm}(u)e^{\pm}(u)=0 \\
&&f^{\pm}(v)k_1^{\pm}(v)k_1^{\pm}(u)-\frac{u-v}{u-v+\hbar}k_1(u)f^{\pm}(v)
k_1^{\pm}(v) \nonumber\\ &&{\hspace{5cm}}
-\frac{\hbar}{u-v+\hbar}f^{\pm}(u)k_1^{\pm}(u)k_1^{\pm}(v)=0 \label{kf1} \\
&&f^{\mp}(v)k_1^{\mp}(v)k_1^{\pm}(u)-\frac{u_{\mp}-v_{\pm}}{u_{\mp}-v_{\pm}+
\hbar}k_1^{\pm}(u)f^{\mp}(v)k_1^{\mp}(v)
\nonumber\\ &&{\hspace{5cm}}
-\frac{\hbar}{u_{\mp}-v_{\pm}+\hbar}
f^{\pm}(u)k_1^{\pm}(u)k_1^{\mp}(v)=0 \label{kf2}\end{aligned}$$
thus $$\begin{aligned}
&&(u_{\pm}-v+\hbar)e^{\pm}(v_{\mp})-(u_{\pm}-v)k_1^{\pm}(u)^{-1}e^{\pm}(v_{\mp})
k_1^{\pm}(u)-\hbar e^{\pm}(u)=0 \\
&&(u_{\pm}-v+\hbar)e^{\mp}(v_{\pm})-(u_{\pm}-v)k_1^{\pm}(u)^{-1}e^{\mp}(v_{\pm})
k_1^{\pm}(u)-\hbar e^{\pm}(u)=0 \\
&&(u_{\mp}-v+\hbar)f^{\pm}(v_{\pm})-(u_{\mp}-v)k_1^{\pm}(u)f^{\pm}(v_{\pm})
k_1^{\pm}(u)^{-1}-\hbar f^{\pm}(u)=0 \\
&&(u_{\mp}-v+\hbar)f^{\mp}(v_{\mp})-(u_{\mp}-v)k_1^{\pm}(u)f^{\mp}(v_{\mp})
k_1^{\pm}(u)^{-1}-\hbar f^{\pm}(u)=0 \end{aligned}$$ so $$\begin{aligned}
&&k_1^{\pm}(u)^{-1}X^+(v)k_1{\pm}(u)=\frac{u_{\pm}-v+\hbar}{u_{\pm}-v}X^+(v) \\
&&k_1^{\pm}(u)X^-(v)k_1^{\pm}(u)^{-1}=\frac{u_{\mp}-v+\hbar}{u_{\mp}-v}X^-(v)\end{aligned}$$
Now, we derive relations between $k_2^{\pm}(u)$ and $X^{\pm}(v)$. From(\[rll3\])(\[rll4\]) and unitarity of $R$-matirx, we have $$\begin{aligned}
&&\frac{u-v-\hbar}{u-v+\hbar}e^{\pm}(u)k_2^{\pm}(u)^{-1}k_2^{\pm}(v)^{-1}-
\frac{u-v}{u-v+\hbar}k_2^{\pm}(v)^{-1}e^{\pm}(u)k_2^{\pm}(u)^{-1}
\nonumber\\ &&{\hspace{5cm}}
+\frac{\hbar}{u-v+\hbar}e^{\pm}(v)k_2^{\pm}(v)^{-1}k_2^{\pm}(u)^{-1}=0 \\
&&\frac{u_{\mp}-v_{\pm}-\hbar}{u_{\mp}-v_{\pm}+\hbar}e^{\pm}(u)
k_2^{\pm}(u)^{-1}k_2^{\mp}(v)^{-1}-\frac{u_{\pm}-v_{\mp}}{u_{\pm}-
v_{\mp}+\hbar}k_2^{\mp}(v)^{-1}e^{\pm}(u)k_2^{\pm}(u)^{-1} \nonumber\\ &&{\hspace{5cm}}+\frac{\hbar}{u_{\pm}-v_{\mp}+\hbar}e^{\mp}(v)k_2^{\mp}(v)^{-1}
k_2^{\pm}(u)^{-1}=0 \\
&&\frac{u-v-\hbar}{u-v+\hbar}k_2^{\pm}(v)^{-1}k_2^{\pm}(u)^{-1}f^{\pm}(u)
-\frac{u-v}{u-v+\hbar}k_2^{\pm}(u)^{-1}f^{\pm}(u)k_2^{\pm}(v)^{-1} \nonumber\\ &&{\hspace{5cm}}
+\frac{\hbar}{u-v+\hbar}k_2^{\pm}(u)^{-1}k_2^{\pm}(v)^{-1}f^{\pm}(v)=0 \\
&&\frac{u_{\pm}-v_{\mp}-\hbar}{u_{\pm}-v_{\mp}+\hbar}k_2^{\mp}(v)^{-1}
k_2^{\pm}(u)^{-1}f^{\pm}(u)-\frac{u_{\mp}-v_{\pm}}{u_{\mp}-v_{\pm}+\hbar}
k_2^{\pm}(u)^{-1}f^{\pm}(u)k_2^{\mp}(v)^{-1} \nonumber\\
&&{\hspace{5cm}}+\frac{\hbar}{u_{\mp}-v_{\pm}
+\hbar}k_2^{\pm}(u)^{-1}k_2^{\mp}(v)^{-1}f^{\mp}(v)=0 \end{aligned}$$ Using relations between $k_2^+(u)$ and $k_2^-(v)$ (\[k2\])(\[k4\]), we have $$\begin{aligned}
&&(u_{\mp}-v-\hbar)e^{\pm}(u_{\mp})-(u_{\mp}-v)k_2^{\pm}(v)^{-1}e^{\pm}(u_{\mp})
k_2^{\pm}(v)+\hbar e^{\pm}(v)=0 \\
&&(u-v_{\mp}-\hbar)e^{\pm}(u_{\mp})-(u-v_{\mp})k_2^{\mp}(v)^{-1}e^{\pm}(u_{\mp})
k_2^{\mp}(v)+\hbar e^{\mp}(v)=0 \\
&&(u_{\pm}-v-\hbar)f^{\pm}(u_{\pm})-(u_{\pm}-v)k_2^{\pm}(v)f^{\pm}(u_{\pm})
k_2^{\pm}(v)^{-1}+\hbar f^{\pm}(v)=0 \\
&&(u-v_{\pm}-\hbar)f^{\pm}(u_{\pm})-(u-v_{\pm})k_2^{\mp}(v)f^{\pm}(u_{\pm})
k_2^{\mp}(v)^{-1}+\hbar f^{\mp}(v)=0\end{aligned}$$ then $$\begin{aligned}
&&k_2^{\pm}(u)^{-1}X^+(v)k_2^{\pm}(u)=\frac{u_{\pm}-v+\hbar}{u_{\pm}-v}X^+(v) \\
&&k_2^{\pm}(u)X^-(v)k_2^{\pm}(u)^{-1}=\frac{u_{\mp}-v+\hbar}{u_{\mp}-v}X^-(v)\end{aligned}$$
Now, we calculate the relations between $X^{\pm}(u)$ and $X^{\pm}(v)$ . From (\[rll1\])(\[rll2\]) and unitarity of $R$-matirx , we have $$\begin{aligned}
&&k_1^{\pm}(u)e^{\pm}(u)k_1^{\pm}(v)e^{\pm}(v)+\frac{u-v-\hbar}
{u-v+\hbar}k_1^{\pm}(v)e^{\pm}(v)k_1^{\pm}(u)e^{\pm}(u)=0 \\
&&k_1^{\pm}(u)e^{\pm}(u)k_1^{\mp}(v)e^{\mp}(v)+\frac{u_{\pm}-v_{\mp}-\hbar}
{u_{\pm}-v_{\mp}+\hbar}k_1^{\mp}(v)e^{\mp}(v)k_1^{\pm}(u)e^{\pm}(u)=0 \\
&&\frac{u-v-\hbar}{u-v+\hbar}f^{\pm}(u)k_1^{\pm}(u)f^{\pm}(v)k_1^{\pm}(v)
+f^{\pm}(v)k_1^{\pm}(v)f^{\pm}(u)k_1^{\pm}(u)=0 \\
&&\frac{u_{\mp}-v_{\pm}-\hbar}{u_{\mp}-v_{\pm}+\hbar}f^{\pm}(u)k_1^{\pm}
(u)f^{\mp}(v)k_1^{\mp}(v)+f^{\mp}(v)k_1^{\mp}(v)f^{\pm}(u)k_1^{\pm}(u)=0 \end{aligned}$$ Using the above relations between $X^{\pm}(u)$ (or $ e^{\pm}(u)$ and $f^{\pm}(u)$) and $k_1^{\pm}(v)$, it’s easily to get the following relations; $$\begin{aligned}
&&X^+(u)X^+(v)+X^+(v)X^+(u)=0 \\
&&X^-(u)X^-(v)+X^-(v)X^-(u)=0\end{aligned}$$
From (\[rll1\])(\[rll2\]) and unitarity of $R$-matirx, we have following relations : $$\begin{aligned}
&&\frac{u-v}{u-v+\hbar}k_1^{\pm}(u)e^{\pm}(u)f^{\pm}(v)k_1^{\pm}(v)
+\frac{u-v}{u-v+\hbar}f^{\pm}(v)k_1^{\pm}(v)k_1^{\pm}(u)e^{\pm}(u) \nonumber \\
&&{\hspace{4cm}}-\frac{\hbar}{u-v+\hbar}\left(k_2^{\pm}(u)+f^{\pm}(u)
k_1^{\pm}(u)e^{\pm}(u)
\right)k_1^{\pm}(v) \nonumber \\
&&{\hspace{4cm}}-\frac{\hbar}{u-v+\hbar}\left(k_2^{\pm}(v)+f^{\pm}(v)k_1^{\pm}(v)
e^{\pm}(v)\right)k_1^{\pm}(u)=0 \\
&&\frac{u_{\mp}-v_{\pm}}{u_{\mp}-v_{\pm}+\hbar}k_1^{\pm}(u)e^{\pm}(u)
f^{\mp}(v)k_1^{\mp}(v)
+\frac{u_{\pm}-v_{\mp}}{u_{\pm}-v_{\mp}+\hbar}f^{\mp}(v)k_1^{\mp}(v)
k_1^{\pm}(u)e^{\pm}(u) \nonumber \\
&&{\hspace{4cm}}-\frac{\hbar}{u_{\mp}-v_{\pm}+\hbar}\left(k_2^{\pm}(u)
+f^{\pm}(u)k_1^{\pm}(u)e^{\pm}(u)\right)k_1^{\mp}(v) \nonumber \\
&&{\hspace{4cm}}-\frac{\hbar}{u_{\pm}-v_{\mp}+\hbar}\left(k_2^{\mp}(v)
+f^{\mp}(v)k_1^{\mp}(v)e^{\mp}(v)\right)k_1^{\pm}(u)=0 \end{aligned}$$
then using (\[kf1\])(\[kf2\]) and the following relations which also from (\[rll1\])(\[rll2\]) and unitarity of $R$-matirx: $$\begin{aligned}
&&k_1^{\pm}(u)e^{\pm}(u)k_1^{\pm}(v)-\frac{u-v}{u-v+\hbar}k_1^{\pm}(v)
k_1^{\pm}(u)e^{\pm}(u) \nonumber \\
&&{\hspace{4cm}}-\frac{\hbar}{u-v+\hbar}k_1^{\pm}(v)e^{\pm}(v)
k_1^{\pm}(u)=0 \\
&&k_1^{\pm}(u)e^{\pm}(u)k_1^{\mp}(v)-\frac{u_{\pm}-v_{\mp}}{u_{\pm}-v_{\mp}
+\hbar}k_1^{\mp}(v)k_1^{\pm}(u)e^{\pm}(u) \nonumber \\
&&{\hspace{4cm}}-\frac{\hbar}{u_{\pm}-v_{\mp}
+\hbar}k_1^{\mp}(v)e^{\mp}(v)k_1^{\pm}(u)=0 \end{aligned}$$ we obtain: $$\begin{aligned}
&&e^{\pm}(u)f^{\pm}(v)+f^{\pm}(v)e^{\pm}(u)=\frac{\hbar}{u-v}k_1^{\pm}
(u)^{-1}k_2^{\pm}(u)-\frac{\hbar}{u-v}k_1^{\pm}(v)^{-1}k_2^{\pm}(v) \\
&&e^{\pm}(u)f^{\mp}(v)+f^{\mp}(v)e^{\pm}(u)=\frac{\hbar}{u_{\mp}-v_{\pm}}
k_1^{\pm}(u)^{-1}k_2^{\pm}(u)-\frac{\hbar}{u_{\pm}-v_{\mp}}k_1^{\mp}(v)^{-1}
k_2^{\mp}(v)\end{aligned}$$ then $$\begin{aligned}
&&\{X^+(u) ~,~ X^-(v)\}=\hbar\left[ \delta(u_--v_+)k_1^+(u_-)^{-1}k_2^+(u_-)
-\delta(u_+-v_-)k_1^-(v_-)^{-1}k_2^-(v_-)\right]\end{aligned}$$ where $ \delta(u-v)=\sum _{k \in Z}u^{k} v^{-k-1} $.
In a word, we get all relations between $k_i^{\pm}(u)$ (i=1,2) and $ X^{\pm}(v)$: $$\begin{aligned}
&&[k_1^{\pm}(u) ~,~k_1^{\pm}(v)]=[k_1^+(u) ~,~k_1^-(v)]=0 \\
&&[k_1^{\pm}(u) ~,~k_2^{\pm}(v)]=[k_2^{\pm}(u) ~,~k_2^{\pm}(v)]=0 \\
&&\frac{u_{\pm}-v_{\mp}}{u_{\pm}-v_{\mp}+\hbar}k_1^{\pm}(u)k_2^{\mp}(v)^{-1}
=\frac{u_{\mp}-v_{\pm}}{u_{\mp}-v_{\pm}+\hbar}k_2^{\mp}(v)^{-1}k_1^{\pm}(u) \\
&&\frac{u_--v_+-\hbar}{u_--v_++\hbar}k_2^+(u)^{-1}k_2^-(v)^{-1}
=\frac{u_+-v_--\hbar}{u_+-v_-+\hbar}k_2^-(v)^{-1}k_2^+(u)^{-1} \\
&&k_i^{\pm}(u)^{-1}X^+(v)k_i^{\pm}(u)=\frac{u_{\pm}-v+\hbar}{u_{\pm}-v}X^+(v)
\hspace{1cm} (i=1,2) \\
&&k_i^{\pm}(u)X^-(v)k_i^{\pm}(u)^{-1}=\frac{u_{\mp}-v+\hbar}{u_{\mp}-v}X^-(v)
\hspace{1cm} (i=1,2) \\
&&\{X^+(u) ~,~X^+(v)\}=\{X^-(u) ~,~X^-(v)\}=0 \\
&&\{X^+(u) ~,~ X^-(v)\}=\hbar\left[ \delta(u_--v_+)k_1^+(u_-)^{-1}k_2^+(u_-)
-\delta(u_+-v_-)k_1^-(v_-)^{-1}k_2^-(v_-)\right]\end{aligned}$$
Introducing a transformation for currents: $$\begin{aligned}
&K^{\pm}(u)=k_1^{\pm}(u+\frac{\hbar}{2})^{-1}k_2^{\pm}(u+\frac{\hbar}{2})
&\hspace{1cm} H^{\pm}(u)=k_1^{\pm}(u-\frac{\hbar}{2})k_2^{\pm}
(u+\frac{\hbar}{2}) \\
&E(u)=X^+(u+\frac{\hbar}{2})&\hspace{1cm} F(u)=X^-(u+\frac{\hbar}{2})\end{aligned}$$ then all relations among $K^{\pm}(u)$, $H^{\pm}(u)$, $E(u)$ and $F(u)$ can be calculated as follows:
$$\begin{aligned}
&&[K^{\pm}(u) ~,~ K^{\pm}(v)]=[K^+(u) ~,~K^-(v)]=0 \\
&&[H^{\pm}(u) ~,~ H^{\pm}(v)]=[K^{\pm}(u) ~,~ H^{\pm}(v)]=0 \\
&&\frac{u_{\mp}-v_{\pm}-\hbar}{u_{\mp}-v_{\pm}+\hbar}H^{\pm}(u)K^{\mp}(v)
=K^{\mp}(v)H^{\pm}(u)\frac{u_{\pm}-v_{\mp}-\hbar}{u_{\pm}-v_{\mp}+\hbar} \\
&&\left(\frac{u_{\mp}-v_{\pm}-\hbar}{u_{\mp}-v_{\pm}+\hbar}\right)^2
H^{\pm}(u)H^{\mp}(v)
=H^{\mp}(v)H^{\pm}(u)\left(\frac{u_{\pm}-v_{\mp}-\hbar}{u_{\pm}-v_{\mp}+\hbar}
\right)^2 \\
&&[K^{\pm}(u) ~,~E(v)]=[K^{\pm}(u) ~,~F(v)]=0 \\
&&E(v)H^{\pm}(u)=\frac{u_{\pm}-v+\hbar}{u_{\pm}-v-\hbar}H^{\pm}(u)E(v) \\
&&H^{\pm}(u)F(v)=\frac{u_{\mp}-v+\hbar}{u_{\mp}-v-\hbar}F(v)H^{\pm}(u) \\
&&\{E(u) ~,~ F(v)\}=\hbar\left[ \delta(u_--v_+)K^+(u_-)
-\delta(u_+-v_-)K^-(v_-)\right]\end{aligned}$$
The co-product structure can be derived directly from (\[cop\])(\[es\]) :
$$\begin{aligned}
&&\triangle \left(E^{\pm}(u)\right)= E^{\pm}(u)\otimes 1+
H^{\pm}(u \mp \frac{\hbar}{4}c_1)\otimes E^{\pm}(u\mp \frac{\hbar}{2}c_1) \\
&&\triangle \left(F^{\pm}(u)\right)=1\otimes F^{\pm}(u)+
F^{\pm}(u \pm \frac{\hbar}{2} c_2)
\otimes H^{\pm}(u \pm \frac{\hbar}{4}c_2) \\
&&\triangle \left(H^{\pm}(u)\right)= H^{\pm}(u\pm \frac{\hbar}{4}c_2)
\otimes H^{\pm}(u\mp \frac{\hbar}{4}c_1)-2F^{\pm}(u \pm \frac{\hbar}{4}c_2
\mp \frac{\hbar}{4}c_1-\hbar) H^{\pm}(u \pm \frac{\hbar}{4}c_2)\otimes
\nonumber \\
&&{\hspace{3cm}}H(u\mp \frac{\hbar}{4}c_1)^{\pm}E^{\pm}
(u \pm \frac{\hbar}{4}c_2 \mp \frac{\hbar}{4}c_1-\hbar)) \\
&&\triangle \left(K^{\pm}(u)\right)=K^{\pm}(u \pm \frac{\hbar}{4}c_2)\otimes
K^{\pm}(u\mp \frac{\hbar}{4}c_1) \\
&&\epsilon (K^{\pm}(u)) =\epsilon (H^{\pm}(u)) =1 \\
&&\epsilon (E^{\pm}(u)) =\epsilon (F^{\pm}(u)) =0 \end{aligned}$$
The quantum affine superalgebra $U_q\widehat{(gl(1|1))}$ corresponding to the trigonometric solution of super YBE have also been studied in our paper [@C2] and the general case $U_q\widehat{(gl(m|n))}$ have been independently worked out by Zhang [@YZZ].
[30]{}
V.G. Drinfel$^{\prime}$d, “Quantum Groups”, [*Proc.ICM-86(Berkeley)*]{}. Vol.1. NewYork Academic Press 1986, 789
V.G. Drinfel$^{\prime}$d: [*Soviet Math. Dokl.*]{} [**36**]{} (1988) 212
N.Yu. Reshetikhin and M.A. Semenov-Tian-Shansky, [*Lett. Math. Phys.*]{} [**19**]{} (1990) 133
J. Ding and I.B. Frenkel, [*Commun. Math. Phys.*]{} [**156**]{} (1993) 277
J.F. Cai, G.X. Ju, K. Wu and S.K. Wang, “Super Yangian Double $DY\left( gl(1|1)\right)$ and Its Gauss Decomposition ”, preprint, q-alg/[**9701011**]{}
L. Liao and X.C. Song, [*Mod. Phys. Lett.* ]{} [**A6**]{} (1991) 959
J.F. Cai, S.K. Wang, K. Wu and W.Z. Zhao, “Drinfel$^{\prime}$d Realization of Quantum Affine Superalgebra $U_q\widehat{(gl(1|1))}$”, preprint, q-alg/[**9703022**]{}
Y.Z. Zhang “Comments on Drinfeld Realization of Quantum Affine Superalgebra $U_q[gl(m|n)^{(1)}]$ and Its Hopf Algebra Structure”, preprint, q-alg/[**9703020**]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The precision measurements of the strong coupling constant, $\as$, and its energy-scale dependence carried out at HERA by the H1 and ZEUS Collaborations are reviewed. An average value of\
$\overline\asz=0.1186\pm 0.0011\ ({\rm exp.})\pm 0.0050\ ({\rm th.})$
\
is obtained from these measurements. The combined HERA determinations of the energy-scale dependence of $\as$ clearly show the running of $\as$ from jet data alone and are in agreement with the running of the coupling as predicted by QCD.
author:
- 'Claudia Glasman[^1]'
title: 'Precision Measurements of $\as$ at HERA[^2]'
---
[address=[Universidad Autónoma de Madrid]{} ]{}
Introduction
============
The strong coupling constant, $\as$, is one of the fundamental parameters of QCD. However, its value is not predicted by the theory and must be determined by experiment. Many precise and consistent determinations of $\as$ from diverse phenomena underlie the success of perturbative QCD (pQCD). At HERA, $\as$ has been determined from many observables, which include jet cross sections and structure functions, by the H1 and ZEUS Collaborations. All the available determinations [@dijetzeus; @incdiszeus; @incdish1; @nlofitzeus; @nlofitzeusn; @nlofith1; @subnczeus; @subcczeus; @incgpzeus; @multijetzeus; @shanczeus] are shown in Fig. \[fig1\]a. They are in good agreement with each other and are consistent with the current world average ($\overline\asz^{WA}=0.1182\pm 0.0027$ [@wa]). These determinations, most of which come from observables which involve jet algorithms, lead to determinations of $\as$ some of which are as precise as those from more inclusive measurements. The uncertainty in these determinations is dominated by the theoretical contributions, which amount to $4\%$ for jet cross sections and fits of structure functions and $8\%$ for the internal structure of jets, whereas the experimental uncertainties amount to $\sim 3\%$.
(10.0,8.0) (-2.0,0.0) (5.0,-0.5) (-2.5,6.0)[**(a)**]{} (5.0,6.0)[**(b)**]{}
Averaging the $\as$ determinations from HERA
============================================
To make a proper average of all these diverse measurements, the correlation among the different determinations has to be taken into account. The experimental contribution to the uncertainty due to that of the energy scale of the jets, which is the dominant source in the jet measurements, is correlated among the determinations from each experiment. On the theoretical side, the uncertainty coming from the proton parton distribution functions (PDFs) is certainly correlated whereas that coming from the hadronisation corrections is only partially correlated. The uncertainty coming from the terms beyond NLO is correlated up to a certain, a priori unknown, degree; since these uncertainties are dominant, special care must be taken in the treatment of these uncertainties when making an average of the HERA determinations.
Several methods have been used to obtain an average value of $\as$ from the HERA measurements and its uncertainty. Using a naive method in which all uncertainties are assumed to be uncorrelated, the average value and its uncertainty are:
$\overline\asz=0.1188\pm 0.0020\ ({\rm ZEUS+H1}).$
The second method used is that developed by M. Schmelling [@schmelling] to average correlated data when correlations are present but hard to quantify. In this method, an error-weighted average and an optimised correlation error were obtained from the error covariance matrix by assuming an overall correlation factor between the total errors of all measurements; the overall factor was determined by the condition that the overall $\chi^2/{\rm dof}$ is equal to unity. First, an error-weighted average was done separately for the ZEUS and H1 measurements, and then the two averages were combined:
$\overline\asz=0.1196\pm 0.0060\ ({\rm ZEUS})\hspace{1cm}{\rm and}\hspace{1cm}\overline\asz=0.1166\pm 0.0053\ ({\rm H1}),$
$\overline\asz=0.1188\pm 0.0057\ ({\rm ZEUS+H1}).$
The averages from the ZEUS and H1 determinations are compatible within the uncertainties. The uncertainty of the combined average is $\sim 5\%$. This procedure gives rise to relatively large uncertainties when there are large correlations among some of the measurements, as it is the case here. To overcome this effect, the method has been repeated by restricting to the most accurate measurements [@bethke]. The result of applying the procedure to those measurements with a total error $\Delta\asz<0.006$ [@incdiszeus; @incdish1; @nlofitzeusn; @nlofith1; @incgpzeus] is:
$\overline\asz=0.1192\pm 0.0047\ (\Delta\as^i<0.006)\ ({\rm ZEUS+H1}).$
Finally, a more reliable, but conservative, approach has been used in which the known correlations from the determinations of $\as$ coming from the same experiment were taken into account (“correlation method”). The theoretical uncertainties arising from the terms beyond NLO were assumed to be (conservatively) fully correlated. Error-weighted averages were obtained separately for the ZEUS and H1 measurements:
$\overline\asz=0.1200\pm 0.0023\ ({\rm exp.})_{-0.0049}^{+0.0058}\ ({\rm th.})\ ({\rm ZEUS}),$
$\overline\asz=0.1160\pm 0.0016\ ({\rm exp.})_{-0.0049}^{+0.0048}\ ({\rm th.})\ ({\rm H1}).$
A HERA average was obtained by using the error-weighted average method on the ZEUS and H1 averages, assuming the experimental uncertainties to be uncorrelated and taking the overall theoretical uncertainty as the linear average of its contribution in each experiment. As a result, the average of the HERA measurements and its uncertainty are:
$\overline\asz=0.1186\pm 0.0011\ ({\rm exp.})\pm 0.0050\ ({\rm th.})\ ({\rm ZEUS+H1}).$
This average, together with the individual values considered, is shown in Fig. \[fig1\]b. It is found to be in good agreement with the current world average, which does not include any of these determinations. The results of applying Schmelling’s and the correlation methods are very similar, giving confidence on the average obtained and its estimated uncertainty.
Energy-scale dependence of $\as$
================================
The H1 and ZEUS Collaborations have tested the pQCD prediction for the energy-scale dependence of the strong coupling constant by determining $\as$ from the measured differential jet cross sections at different $\etjet$ [@dijetzeus; @incdiszeus; @incdish1; @incgpzeus]. Figure \[fig2\]a shows the determinations of the energy-scale dependence of $\as$ as a function of $\etjet$ from H1 and ZEUS. The determinations are consistent with the running of $\as$ as predicted by pQCD over a large range in $\etjet$.
The determinations of $\as(\etjet)$ from H1 and ZEUS at similar $\etjet$ have been combined using the correlation method explained above. The combined HERA determinations of the energy-scale dependence of $\as$ are shown in Fig. \[fig2\]b, in which the running of $\as$ from HERA jet data alone is clearly observed.
(10.0,6.0) (-3.0,-1.0) (5.0,-1.0) (3.6,4.6)[**(a)**]{} (11.6,4.6)[**(b)**]{}
Summary
=======
A comprehensive average of $\as$ and its energy-scale dependence from HERA data has been performed taking into account the known correlations in each experiment and assuming conservatively that the theoretical uncertainties arising from the terms beyond NLO are fully correlated. The HERA average is
$\overline\asz=0.1186\pm 0.0011\ ({\rm exp.})\pm 0.0050\ ({\rm th.}).$
The experimental uncertainty of this average is $\sim 0.9\%$ and the theoretical uncertainty amounts to $\sim 4\%$. There is still room for improvement when the next-to-NLO (NNLO) calculations needed for the determination of the PDFs are included and when the NNLO calculations needed for jet-based observables are finished.
This work has been carried out in collaboration with Juan Terrón. I would like to thank my colleagues from H1 and ZEUS for their help in the preparation of this report.
[99]{}
, J. Breitweg , .
, S. Chekanov , .
, C. Adloff , .
, S. Chekanov , .
, S. Chekanov , DESY 05-050 (March 2005), hep-ph/0503274.
, C. Adloff , .
, S. Chekanov , .
, S. Chekanov , .
, S. Chekanov , .
, S. Chekanov , DESY 05-019 (January 2005), hep-ex/0502007.
, S. Chekanov , .
S. Bethke, hep-ex/0407021.
M. Schmelling, .
S. Bethke, .
[^1]: Ramón y Cajal Fellow.
[^2]: Talk given in the XIII International Workshop on Deep Inelastic Scattering, April $27^{\rm th}$ - May $1^{\rm st}$, 2005, Madison, Wisconsin USA.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Manifolds without boundary, and manifolds with boundary, are universally known and loved in Differential Geometry, but [*manifolds with corners*]{} (locally modelled on $[0,\iy)^k\t\R^{n-k}$) have received comparatively little attention. The basic definitions in the subject are not agreed upon, there are several inequivalent definitions in use of manifolds with corners, of boundary, and of smooth map, depending on the applications in mind.
We present a theory of manifolds with corners which includes a new notion of smooth map $f:X\ra Y$. Compared to other definitions, our theory has the advantage of giving a category $\Manc$ of manifolds with corners which is particularly well behaved as a category: it has products and direct products, boundaries $\pd X$ behave in a functorial way, and there are simple conditions for the existence of fibre products $X\t_ZY$ in $\Manc$.
Our theory is tailored to future applications in Symplectic Geometry, and is part of a project to describe the geometric structure on moduli spaces of $J$-holomorphic curves in a new way. But we have written it as a separate paper as we believe it is of independent interest.
author:
- Dominic Joyce
title: On manifolds with corners
---
\#1 \#1 \#1[[(\[\#1\])]{}]{} \[section\] \[thm\][Lemma]{} \[thm\][Proposition]{} \[thm\][Corollary]{} \[thm\][Definition]{} \[thm\][Example]{} \[thm\][Remark]{} \[thm\][Convention]{} \#1[\#1 ]{} \#1[\#1 ]{} \#1[\#1 ]{}
Introduction {#mc1}
============
Most of the literature in Differential Geometry discusses only manifolds without boundary (locally modelled on $\R^n$), and a smaller proportion manifolds with boundary (locally modelled on $[0,\iy)\t\R^{n-1}$). Only a few authors have seriously studied [*manifolds with corners*]{} (locally modelled on $[0,\iy)^k\t\R^{n-k}$). They were first developed by Cerf [@Cerf] and Douady [@Doua] in 1961, who were primarily interested in their Differential Geometry. Jänich [@Jani] used manifolds with corners to classify actions of transformation groups on smooth manifolds. Melrose [@Melr1; @Melr2] and others study analysis of elliptic operators on manifolds with corners. Laures [@Laur] defines a cobordism theory for manifolds with corners, which has been applied in Topological Quantum Field Theory, by Lauda and Pfeiffer [@LaPf] for instance. Margalef-Roig and Outerelo Dominguez [@MaOu] generalize manifolds with corners to infinite-dimensional Banach manifolds.
How one sets up the theory of manifolds with corners is not universally agreed, but depends on the applications one has in mind. As we explain in Remarks \[mc2rem\], \[mc3rem\] and \[mc6rem2\] which relate our work to that of other authors, there are at least four inequivalent definitions of manifolds with corners, two inequivalent definitions of boundary, and (including ours) four inequivalent definitions of smooth map in use in the literature. The purpose of this paper is to carefully lay down the foundations of a theory of manifolds with corners, which includes a new notion of [*smooth map*]{} $f:X\ra Y$ between manifolds with corners.
The main issue here is that (in our theory) an $n$-manifold with corners $X$ has a boundary $\pd X$ which is an $(n\!-\!1)$-manifold with corners, and so by induction the $k$-fold boundary $\pd^kX$ is an $(n\!-\!k)$-manifold with corners. How to define smooth maps $f:X\ra Y$ in the interiors $X^\ci,Y^\ci$ is clear, but one must also decide whether to impose compatibility conditions on $f$ over $\pd^kX$ and $\pd^lY$, and it is not obvious how best to do this. Our definition gives a nicely behaved category $\Manc$ of manifolds with corners, and in particular we can give simple conditions for the existence of [*fibre products*]{} in $\Manc$.
The author’s interest in manifolds with corners has to do with applications in Symplectic Geometry. Moduli spaces $\oM_{g,h}^{l,m}(M,L,J,\be)$ of stable $J$-holomorphic curves in a symplectic manifold $(M,\om)$ with boundary in a Lagrangian $L$ are used in Lagrangian Floer cohomology, open Gromov–Witten invariants, and Symplectic Field Theory. What geometric structure should we put on $\oM_{g,h}^{l,m}(M,L,J,\be)$? Hofer, Wysocki and Zehnder [@HWZ] make $\oM_{g,h}^{l,m}(M,L,J,\be)$ into a [*polyfold*]{} (actually, the zeroes of a Fredholm section of a polyfold bundle over a polyfold). Fukaya, Oh, Ohta and Ono [@FOOO] make $\oM_{g,h}^{l,m}(M,L,J,\be)$ into a [*Kuranishi space*]{}. But the theory of Kuranishi spaces is still relatively unexplored, and in the author’s view even the definition is not satisfactory.
In [@Joyc] the author will develop theories of [*d-manifolds*]{} and [*d-orbifolds*]{}. Here [*d-manifolds*]{} are a simplified version of Spivak’s [*derived manifolds*]{} [@Spiv], a differential-geometric offshoot of Jacob Lurie’s Derived Algebraic Geometry programme, and d-orbifolds are the orbifold version of d-manifolds. We will argue that the ‘correct’ way to define Kuranishi spaces is as [*d-orbifolds with corners*]{}, which will help to make [@FOOO] more rigorous. In future the author hopes also to show that polyfolds can be truncated to d-orbifolds with corners, building a bridge between the theories of Fukaya et al. [@FOOO] and Hofer et al. [@HWZ].
To define d-manifolds and d-orbifolds with corners we first need a theory of manifolds with corners, and we develop it here in a separate paper as we believe it is of independent interest. For [@Joyc] and later applications in Symplectic Geometry, it is important that boundaries and fibre products in $\Manc$ should be well-behaved. The author strongly believes that the theory we set out here, in particular our definition of smooth map, is the ‘right’ definition for these applications. As evidence for this, note that in [@Joyc] we will show that if $\bs X,\bs Y$ are d-manifolds with corners, $Z$ is a manifold with corners, and $f:\bs X\ra Z$, $g:\bs Y\ra Z$ are arbitrary smooth maps, then a fibre product $\bs
X\t_{f,Z,g}\bs Y$ exists in the 2-category $\dManc$ of d-manifolds with corners. The fact that this works is crucially dependent on the details of our definition of smooth map.
We begin in §\[mc2\] with definitions and properties of manifolds with corners $X$, their boundaries $\pd X$, $k$-boundaries $\pd^kX$ and $k$-corners $C_k(X)\cong \pd^kX/S_k$. Section \[mc3\] defines and studies smooth maps $f:X\ra Y$ of manifolds with corners, and §\[mc4\] explains two ways to encode how a smooth $f:X\ra Y$ relates $\pd^kX$ and $\pd^lY$ for $k,l\ge 0$. Sections \[mc5\]–\[mc7\] discuss submersions, transversality and fibre products of manifolds with corners, and orientations and orientation conventions. The proofs of Theorems \[mc6thm1\] and \[mc6thm2\], two of our main results on fibre products, are postponed to §\[mc8\] and §\[mc9\].
The author would like to thank Franki Dillen for pointing out reference [@MaOu].
Manifolds with corners, and boundaries {#mc2}
======================================
We define [*manifolds without boundary*]{}, [*with boundary*]{}, and [*with corners*]{}.
Let $X$ be a paracompact Hausdorff topological space.
- An [*$n$-dimensional chart on $X$ without boundary*]{} is a pair $(U,\phi)$, where $U$ is an open subset in $\R^n$, and $\phi:U\ra X$ is a homeomorphism with a nonempty open set $\phi(U)$ in $X$.
- An [*$n$-dimensional chart on $X$ with boundary*]{} for $n\ge 1$ is a pair $(U,\phi)$, where $U$ is an open subset in $\R^n$ or in $[0,\iy)\t\R^{n-1}$, and $\phi:U\ra X$ is a homeomorphism with a nonempty open set $\phi(U)$.
- An [*$n$-dimensional chart on $X$ with corners*]{} for $n\ge 1$ is a pair $(U,\phi)$, where $U$ is an open subset in $\R^n_k=[0,\iy)^k\t\R^{n-k}$ for some $0\le k\le n$, and $\phi:U\ra
X$ is a homeomorphism with a nonempty open set $\phi(U)$.
These are increasing order of generality, that is, (i) $\Rightarrow$ (ii) $\Rightarrow$ (iii).
For brevity we will use the notation $\R^n_k=[0,\iy)^k\t\R^{n-k}$, following [@Melr2 §1.1].
Let $A\subseteq\R^m$ and $B\subseteq\R^n$ and $\al:A\ra B$ be continuous. We call $\al$ [*smooth*]{} if it extends to a smooth map between open neighbourhoods of $A,B$, that is, if there exists an open subset $A'$ of $\R^m$ with $A\subseteq A'$ and a smooth map $\al':A'\ra\R^n$ with $\al'\vert_A\equiv\al$. If $A$ is open we can take $A'=A$ and $\al'=\al$. When $m=n$ we call $\al:A\ra B$ a [*diffeomorphism*]{} if it is a homeomorphism and $\al:A\ra B$, $\al^{-1}:B\ra A$ are smooth.
Let $(U,\phi),(V,\psi)$ be $n$-dimensional charts on $X$, which may be without boundary, or with boundary, or with corners. We call $(U,\phi)$ and $(V,\psi)$ [*compatible*]{} if $\psi^{-1}\ci\phi:\phi^{-1}\bigl(\phi(U)\cap\psi(V)\bigr)\ra
\psi^{-1}\bigl(\phi(U)\cap\psi(V)\bigr)$ is a diffeomorphism between subsets of $\R^n$, in the sense above.
An $n$-[*dimensional atlas*]{} for $X$ [*without boundary*]{}, [*with boundary*]{}, or [*with corners*]{}, is a system $\{(U^i,\phi^i):i\in I\}$ of pairwise compatible $n$-dimensional charts on $X$ with $X=\bigcup_{i\in I}\phi^i(U^i)$, where the $(U^i,\phi^i)$ are with boundary, or with corners, respectively. We call such an atlas [*maximal*]{} if it is not a proper subset of any other atlas. Any atlas $\{(U^i,\phi^i):i\in I\}$ is contained in a unique maximal atlas of the same type, the set of all charts $(U,\phi)$ of this type on $X$ which are compatible with $(U^i,\phi^i)$ for all $i\in I$.
An $n$-[*dimensional manifold without boundary*]{}, or [*with boundary*]{}, or [*with corners*]{}, is a paracompact Hausdorff topological space $X$ equipped with a maximal $n$-dimensional atlas without boundary, or with boundary, or with corners, respectively. Usually we refer to $X$ as the manifold, leaving the atlas implicit, and by a [*chart $(U,\phi)$ on the manifold*]{} $X$, we mean an element of the maximal atlas. When we just say manifold, we will usually mean a manifold with corners. \[mc2def1\]
Here is some notation on (co)tangent spaces.
Let $X$ be an $n$-manifold with corners. A map $f:X\ra\R$ is called [*smooth*]{} if whenever $(U,\phi)$ is a chart on the manifold $X$ then $f\ci\phi:U\ra\R$ is a smooth map between subsets of $\R^n,\R$, in the sense of Definition \[mc2def1\]. Write $C^\iy(X)$ for the $\R$-algebra of smooth functions $f:X\ra\R$.
Following [@KoNo p. 4], for each $x\in X$ define the [*tangent space*]{} $T_xX$ by $$\begin{aligned}
T_xX=\bigl\{v:\,&\text{$v$ is a linear map $C^\iy(X)\ra\R$
satisfying}\\
&\text{$v(fg)=v(f)g(x)+f(x)v(g)$ for all $f,g\in C^\iy(X)$}\bigr\},\end{aligned}$$ and define the [*cotangent space*]{} $T_x^*X=(T_xX)^*$. Both are vector spaces of dimension $n$. For each $x\in X$, we define the [*inward sector*]{} $IS(T_xX)$ of vectors which ‘point into $X$’, as follows: let $(U,\phi)$ be a chart on $X$ with $U\subseteq\R^n_k$ open and $0\in U$ with $\phi(0)=x$. Then $\d\phi\vert_0:T_0\R^n_k=\R^n\ra T_xX$ is an isomorphism. Set $IS(T_xX)=\d\phi\vert_0(\R^n_k)\subseteq T_xX$. This is independent of the choice of $(U,\phi)$. \[mc2def2\]
We now study the notion of [*boundary*]{} $\pd X$ for $n$-manifolds $X$ with corners.
Let $U\subseteq\R^n_k$ be open. For each $u=(u_1,\ldots,u_n)$ in $U$, define the [*depth*]{} $\depth_Uu$ of $u$ in $U$ to be the number of $u_1,\ldots,u_k$ which are zero. That is, $\depth_Uu$ is the number of boundary faces of $U$ containing $u$.
Let $X$ be an $n$-manifold with corners. For $x\in X$, choose a chart $(U,\phi)$ on the manifold $X$ with $\phi(u)=x$ for $u\in U$, and define the [*depth*]{} $\depth_Xx$ of $x$ in $X$ by $\depth_Xx=\depth_Uu$. This is independent of the choice of $(U,\phi)$. For each $k=0,\ldots,n$, define the [*depth $k$ stratum*]{} of $X$ to be $$S^k(X)=\bigl\{x\in X:\depth_Xx=k\bigr\}.$$ \[mc2def3\]
The proof of the next proposition is elementary.
Let $X$ be an $n$-manifold with corners. Then
- $X=\coprod_{k=0}^nS^k(X),$ and $\overline{S^k(X)}=\bigcup_{l=k}^n S^l(X);$
- Each $S^k(X)$ has the structure of an $(n-k)$-manifold without boundary;
- $X$ is a manifold without boundary if and only if $S^k(X)=\es$ for $k>0;$
- $X$ is a manifold with boundary if and only if $S^k(X)=\es$ for $k>1;$ and
- If $x\in S^k(X)$ then $IS(T_xX)$ in $T_xX$ is isomorphic to $\R^n_k$ in $\R^n$. Also the intersection $IS(T_xX)\cap -IS(T_xX)$ in $T_xX$ is $T_xS^k(X)\cong\R^{n-k}$.
\[mc2prop1\]
Let $X$ be a manifold with corners, and $x\in X$. A [*local boundary component $\be$ of $X$ at*]{} $x$ is a local choice of connected component of $S^1(X)$ near $x$. That is, for each sufficiently small open neighbourhood $V$ of $x$ in $X$, $\be$ gives a choice of connected component $W$ of $V\cap S^1(X)$ with $x\in\overline W$, and any two such choices $V,W$ and $V',W'$ must be compatible in the sense that $x\in\overline{(W\cap W')}$. \[mc2def4\]
The meaning of local boundary components in coordinate charts is easy to explain. Suppose $(U,\phi)$ is a chart on $X$ with $\phi(u)=x$, where $U$ is an open set in $\R^n_k$, and write $u=(u_1,\ldots,u_n)$. Then $$S^1(U)=\ts\coprod_{i=1}^k\bigl\{(x_1,\ldots,x_n)\in U:\text{$x_i=0$,
$x_j\ne 0$ for $j\ne i$, $j=1,\ldots,k$}\bigr\}.$$ If $u_i=0$ for some $i=1,\ldots,k$, then $\bigl\{(x_1,\ldots,x_n)\in
U:x_i=0$, $x_j\ne 0$ for $j\ne i$, $j=1,\ldots,k\bigr\}$ is a subset of $S^1(U)$ whose closure contains $u$, and the intersection of this subset with any sufficiently small open ball about $u$ is connected, so this subset uniquely determines a local boundary component of $U$ at $u$, and hence a local boundary component of $X$ at $x$. Thus, the local boundary components of $X$ at $x$ are in 1-1 correspondence with those $i=1,\ldots,k$ with $u_i=0$. But the number of such $i$ is the depth $\depth_Xx$. Hence [*there are exactly $\depth_Xx$ distinct local boundary components $\be$ of $X$ at $x$ for each*]{} $x\in X$.
Let $X$ be a manifold with corners. As a set, define the [*boundary*]{} X={(x,):}. \[mc2eq1\] Define a map $i_X:\pd X\ra X$ by $i_X:(x,\be)\mapsto x$. Note that $i_X$ [*need not be injective*]{}, as $\bmd{i_X^{-1}(x)}=\depth_Xx$ for all $x\in X$. \[mc2def5\]
If $(U,\phi)$ is a chart on the manifold $X$ with $U\subseteq\R^n_k$ open, then for each $i=1,\ldots,k$ we can define a chart $(U_i,\phi_i)$ on $\pd X$ by
&U\_i={(x\_1,…,x\_[n-1]{})\^[n-1]{}\_[k-1]{}: (x\_1,…,x\_[i-1]{},0,x\_i,…,x\_[n-1]{})U\^n\_k},\
&\_i:(x\_1,…,x\_[n-1]{})((x\_1,…,x\_[i-1]{}, 0,x\_i,…,x\_[n-1]{}),\_\*({x\_i=0})).
\[mc2eq2\] All such charts on $\pd X$ are compatible, and induce a manifold structure on $\pd X$. Thus as in Douady [@Doua §6] we may prove:
Let $X$ be an $n$-manifold with corners. Then $\pd
X$ is naturally an $(n\!-\!1)$-manifold with corners for $n>0,$ and $\pd X=\es$ if $n=0$. \[mc2prop2\]
The map $i_X:\pd X\ra X$ is continuous, and we will see in §\[mc3\] that it is smooth. By considering the local models $\R^n_k$ for $X$, we see:
As a map between topological spaces, $i_X:\pd X\ra X$ in Definition [\[mc2def5\]]{} is continuous, finite (that is, $i_X^{-1}(x)$ is finite for all $x\in X$), and proper (that is, if $S\subseteq X$ is compact then $i_X^{-1}(S)\subseteq\pd X$ is compact). \[mc2lem\]
As $\pd X$ is a manifold with corners we can iterate the boundary construction to obtain $\pd X,\pd^2X,\ldots,\pd^nX$, with $\pd^kX$ an $(n-k)$-manifold with corners.
Let $X$ be an $n$-manifold with corners. Then for $k=0,\ldots,n$ there are natural identifications
\^kX{(x,\_1,…,\_k):&\
&}.
\[mc2eq3\] \[mc2prop3\]
Consider first the case $k=2$. Points of $\pd^2X$ are of the form $\bigl((x,\be_1),\ti\be_2\bigr)$, where $x\in X$, and $\be_1$ is a local boundary component of $X$ at $x$, and $\ti\be_2$ is a local boundary component of $\pd X$ at $(x,\be_1)$. Suppose $(U,\phi)$ is a chart for $X$ with $x=\phi(u)$ for some $u\in U$, where $U$ is open in $\R^n_l$ for $l\ge 2$, and $u=(u_1,\ldots,u_n)$, with $u_{i_1}=0$ and $\phi^{-1}(\be_1)$ the local boundary component $x_{i_1}=0$ in $U$. Then gives a chart $(U_{i_1},\phi_{i_1})$ for $\pd X$ with $\phi_{i_1}\bigl((u_1,\ldots,u_{i_1-1},u_{i_1+1},\ldots,
u_n)\bigr)=(x,\be_1)$. Thus $\phi_{i_1}^{-1}(\ti\be_2)$ is a local boundary component for $U_{i_1}$, and so is of the form $x_j=0$ for $j=1,\ldots,l-1$. Write $i_2=j$ if $j<i_1$ and $i_2=j+1$ if $j\ge
i_1$. Then $u_{i_2}=0$, as $u_{i_2}$ is the $j^{\rm th}$ coordinate of $(u_1,\ldots,u_{i_1-1},u_{i_1+1},\ldots,u_n)$.
Let $\be_2$ be the local boundary component $\phi_*(\{x_{i_2}=0\})$ of $X$ at $x$. Then $\be_2\ne \be_1$ as $i_2\ne i_1$. We have constructed a 1-1 correspondence between local boundary components $\ti\be_2$ of $\pd X$ at $(x,\be_1)$ and local boundary components $\be_2$ of $X$ at $x$ with $\be_2\ne\be_1$. This 1-1 correspondence is independent of the choice of chart $(U,\phi)$. Identifying $\bigl((x,\be_1),\ti\be_2\bigr)$ with $(x,\be_1,\be_2)$ gives for $k=2$.
We prove the general case by induction on $k$. The case $k=0$ is trivial, $k=1$ is , and $k=2$ we have proved above. For the inductive step, having proved for $k\le l<n$, we show that at a point of $\pd^lX$ identified with $(p,\be_1,\ldots,\be_l)$ under , local boundary components of $\pd^lX$ are in 1-1 correspondence with local boundary components $\be_{l+1}$ of $X$ at $p$ distinct from $\be_1,\ldots,\be_l$.
Write $S_k$ for the symmetric group on $k$ elements, the group of bijections $\si:\{1,\ldots,k\}\ra\{1,\ldots,k\}$. From we see that $\pd^kX$ has a natural, free action of $S_k$ by permuting $\be_1,\ldots,\be_k$, given by $$\si:(x,\be_1,\ldots,\be_k)\longmapsto
(x,\be_{\si(1)},\ldots,\be_{\si(k)}).$$ Each $\si\in S_k$ acts on $\pd^kX$ as an isomorphism of $(n-k)$-manifolds with corners (a diffeomorphism). Thus, if $G$ is a subgroup of $S_k$, then the quotient $(\pd^kX)/G$ is also an $(n-k)$-manifold with corners.
In particular, taking $G=S_k$, we define the $k$-[*corners*]{} $C_k(X)$ of $X$ to be
C\_k(X)={(x,&{\_1,…,\_k}):\
&}\^kX/S\_k,
\[mc2eq4\] an $(n-k)$-manifold with corners. Note that $\be_1,\ldots,\be_k$ are unordered in , but ordered in . We have isomorphisms $C_0(X)\cong X$ and $C_1(X)\cong\pd X$. \[mc2def6\]
We review how our definitions so far relate to those in use by other authors. For manifolds without or with boundary, all definitions the author has found (see Kobayashi and Nomizu [@KoNo] or Lang [@Lang], for instance) are equivalent to Definition \[mc2def1\]. However, for manifolds with corners, there are four main [*inequivalent*]{} definitions. Our terminology for (a),(b),(d) follows Jänich [@Jani §1.1], and for (c) follows Monthubert [@Mont §2.2].
- [*Manifolds with corners*]{} are as in Definition \[mc2def1\].
- A manifold with corners $X$ is called a [*manifold with faces*]{} if each $x\in X$ lies in the image of $\depth_Xx=\bmd{i_X^{-1}(x)}$ different connected components of $\pd X$ under $i_X:\pd X\ra X$.
- A manifold with corners $X$ is called a [*manifold with embedded corners*]{} if there exists a decomposition $\pd
X=\pd_1X\amalg \pd_2 X\amalg \cdots\amalg\pd_NX$ for finite $N\ge 0$ with $\pd_iX$ open and closed in $\pd X$, such that $i_X\vert_{\pd_iX}:\pd_iX\ra X$ is injective for $i=1,\ldots,N$. We allow $\pd_iX=\es$.
- For each integer $N\ge 0$, an $\an{N}$-[*manifold*]{} is a manifold with corners $X$ together with a given decomposition $\pd X=\pd_1X\amalg\cdots\amalg\pd_NX$ with $\pd_iX$ open and closed in $\pd X$, such that $i_X\vert_{\pd_iX}:\pd_iX\ra X$ is injective for $i=1,\ldots,N$. We allow $\pd_iX=\es$. Note that $N$ has no relation to $\dim X$. A $\an{0}$-manifold is a manifold without boundary, and a $\an{1}$-manifold is a manifold with boundary.
Note that (c) implies (b) implies (a), and (d) becomes (c) after forgetting the decomposition $\pd X=\pd_1X\amalg\cdots\amalg
\pd_NX$. An example satisfying (a) but not (b)–(d) is the [*teardrop*]{} $T=\bigl\{(x,y)\in\R^2:x\ge 0$, $y^2\le x^2-x^4\bigr\}$, shown in Figure \[fig1\]. For compact $X$ (b) also implies (c), but one can find pathological examples of noncompact $X$ which satisfy (b), but not (c) or (d). Cerf [@Cerf], Douady [@Doua], Margalef-Roig and Outerelo Dominguez [@MaOu], and others define manifolds with corners as in (a). Melrose [@Melr1; @Melr2] and authors who follow him define manifolds with corners as in (c); Melrose [@Melr2 §1.6] calls manifolds with corners in sense (a) [*t-manifolds*]{}. Jänich [@Jani §1.1] defines manifolds with corners in senses (a),(b) and (d), but is primarily interested in (d). Laures [@Laur] also works with $\an{N}$-manifolds, in sense (d).
,(-1.5,0)\* ,<6cm,-1.5cm>;<6.7cm,-1.5cm>: ,(3,.3)\*[x]{} ,(-1.2,2)\*[y]{} ,(-1.5,0)\* ,(-1.5,0); (1.5,0) \*\* ?(.06)=“a” ?(.12)=“b” ?(.2)=“c” ?(.29)=“d” ?(.4)=“e” ?(.5)=“f” ?(.6)=“g” ?(.7)=“h” ?(.83)=“i” ,(-1.5,0); (1.5,0) \*\* ?(.06)=“j” ?(.12)=“k” ?(.2)=“l” ?(.29)=“m” ?(.4)=“n” ?(.5)=“o” ?(.6)=“p” ?(.7)=“q” ?(.83)=“r” ,“a”;“j”\*\*@[.]{} ,“b”;“k”\*\*@[.]{} ,“c”;“l”\*\*@[.]{} ,“d”;“m”\*\*@[.]{} ,“e”;“n”\*\*@[.]{} ,“f”;“o”\*\*@[.]{} ,“g”;“p”\*\*@[.]{} ,“h”;“q”\*\*@[.]{} ,“i”;“r”\*\*@[.]{} (-1.5,0);(3,0) (-1.5,0);(-3,0) (-1.5,0);(-1.5,2) (-1.5,0);(-1.5,-2)
The [*boundary*]{} $\pd X$ of a manifold with corners $X$ is also defined in different ways in the literature. In our picture, $\pd X$ is a manifold with corners, with an immersion $i_X:\pd X\ra X$ which is not necessarily injective, so that $\pd X$ may not be a subset of $X$. This follows Douady [@Doua §6], who defines $\pd^kX$ (in his notation) to be equivalent to our $C_k(X)$ in , so that his $\pd^1X$ agrees with our $\pd X$. All the other authors cited define $\pd X$ to be $i_X(\pd X)$ in our notation, so that $\pd X$ is a subset of $X$, but is not necessarily a manifold with corners. But in (c),(d) above, the $\pd_iX$ are both subsets of $X$ and manifolds with corners. \[mc2rem\]
If $X,Y$ are manifolds with corners of dimensions $m,n$, there is a natural way to make the product $X\t Y$ into a manifold with corners, of dimension $m+n$. The following result on boundary and $k$-corners of $X\t Y$ is easy to prove by considering local models $\R^{m+n}_{a+b}\cong \R^m_a\t\R^n_b$ for $X\t Y$.
Let $X,Y$ be manifolds with corners. Then there are natural isomorphisms of manifolds with corners (XY)&(XY)(XY), \[mc2eq5\]\
C\_k(XY)&\_[i,j0, i+j=k]{}C\_i(X)C\_j(Y). \[mc2eq6\] \[mc2prop4\]
Note that and imply that \^k(XY)&\_[i=1]{}\^k \^iX\^[k-i]{}Y, \[mc2eq7\]\
\_[k=0]{}\^[XY]{}C\_k(XY)&. \[mc2eq8\] We will see in §\[mc4\] that if $X$ is a manifold with corners then we can make $\coprod_{i=0}^{\dim X}C_i(X)$ behave functorially under smooth maps.
The map $X\mapsto C_k(X)$ commutes with boundaries. The proof is again an easy exercise by considering local models $\R^m_a$ for $X$.
Let $X$ be a manifold with corners and $k\ge 0$. Then there are natural identifications, with the first a diffeomorphism:
(C\_k(X))C\_k(X){(x,\_1,{\_2,…,\_[k+1]{}}):&\
&}.
\[mc2eq9\] \[mc2prop5\]
The next definition will be used in defining smooth maps in §\[mc3\].
Let $X$ be a manifold with corners, and $(x,\be)\in\pd
X$. A [*boundary defining function for $X$ at*]{} $(x,\be)$ is a pair $(V,b)$, where $V$ is an open neighbourhood of $x$ in $X$ and $b:V\ra[0,\iy)$ is a map, such that $b:V\ra\R$ is smooth in the sense of Definition \[mc2def2\], and $\d b\vert_x:T_xV\ra
T_0[0,\iy)$ is nonzero, and there exists an open neighbourhood $\ti
V$ of $(x,\be)$ in $i_X^{-1}(V)\subseteq\pd X$, with $b\ci
i_X\vert_{\ti V}\equiv 0$, and $i_X\vert_{\ti V}:\ti
V\longra\bigl\{x'\in V:b(x')=0\bigr\}$ is a homeomorphism between $\ti V$ and an open subset of $\bigl\{x'\in V:b(x')=0\bigr\}$.
Thus the boundary $\pd X$ is defined near $(x,\be)$ by the equation $b=0$ in $X$ near $x$. Using the ideas on fibre products of manifolds with corners in §\[mc6\], one can say more: $\pd X$ near $(x,\be)$ is naturally isomorphic to the fibre product of manifolds with corners $V\t_{b,[0,\iy),i}\{0\}$ near $(x,0)$, where $i:\{0\}\ra[0,\iy)$ is the inclusion, and the fibre product exists near $(x,0)$. \[mc2def7\]
Here are some properties of such $(V,b)$. The proofs are elementary.
Let $X$ be an $n$-manifold with corners, and $(x,\be)\in\pd X$.
- There exists a boundary defining function $(V,b)$ for $X$ at $(x,\be)$.
- Let $(V,b)$ and $(V',b')$ be boundary defining functions for $X$ at $(x,\be)$. Then there exists an open neighbourhood $V''$ of $x$ in $V\cap V'$ and a smooth function $g:V''\ra(0,\iy)\subset\R$ such that $b'\vert_{V''}\equiv
b\vert_{V''}\cdot g$.
- Let $(V,b)$ be a boundary defining function for $X$ at $(x,\be)$. Then there exists a chart $(U,\phi)$ on the manifold $X,$ such that $U$ is open in $\R^n_k$ for $0<k\le n$ and $0\in U$ with $\phi(0)=x,$ and $\be$ is the image of the local boundary component $x_1=0$ of $U$ at $0,$ and $\phi(U)\subseteq V,$ and $b\ci\phi\equiv
x_1:U\ra[0,\iy)$.
- Let $(U,\phi)$ be a chart on the manifold $X,$ such that $U$ is open in $\R^n_k$ and $u\in U$ with $\phi(u)=x,$ and $\be$ is the image of the local boundary component $x_i=0$ of $U$ at $u$ for $i\le k$. Then $\bigl(\phi(U),x_i\ci\phi^{-1}\bigr)$ is a boundary defining function for $X$ at $(x,\be)$.
\[mc2prop6\]
Smooth maps of manifolds with corners {#mc3}
=====================================
Here is our definition of [*smooth maps*]{} $f:X\ra Y$ of manifolds with corners $X,Y$.
Let $X,Y$ be manifolds with corners of dimensions $m,n$. A continuous map $f:X\ra Y$ is called [*weakly smooth*]{} if whenever $(U,\phi),(V,\psi)$ are charts on the manifolds $X,Y$ then $$\psi^{-1}\ci f\ci\phi:(f\ci\phi)^{-1}(\psi(V))\longra V$$ is a smooth map from $(f\ci\phi)^{-1}(\psi(V))\subset\R^m$ to $V\subset\R^n$, where smooth maps between subsets of $\R^m,\R^n$ are defined in Definition \[mc2def1\].
A weakly smooth map $f:X\ra Y$ is called [*smooth*]{} if it satisfies the following additional condition over $\pd X,\pd Y$. Suppose $x\in X$ with $f(x)=y\in Y$, and $\be$ is a local boundary component of $Y$ at $y$. Let $(V,b)$ be a boundary defining function for $Y$ at $(y,\be)$. Then $f^{-1}(V)$ is an open neighbourhood of $x$ in $X$, and $b\ci f:f^{-1}(V)\ra[0,\iy)$ is a weakly smooth map. We require that either $b\ci f\equiv 0$ on an open neighbourhood of $x$ in $f^{-1}(V)$, or $(f^{-1}(V),b\ci f)$ is a boundary defining function for $X$ at $(x,\ti\be)$, for some unique local boundary component $\ti\be$ of $X$ at $x$. \[mc3def1\]
We also define five special classes of smooth maps:
Let $X,Y$ be manifolds with corners of dimensions $m,n$, and $f:X\ra Y$ a weakly smooth map. If $x\in X$ with $f(x)=y$ then in the usual way there is an induced linear map on tangent spaces $\d f\vert_x:T_xX\ra T_yY$. In the notation of Definition \[mc2def2\], $\d f\vert_x:T_xX\ra T_yY$ maps $IS(T_xX)\ra
IS(T_yY)$, that is, $\d f\vert_x$ maps inward-pointing vectors to inward-pointing vectors.
Let $x\in S^k(X)$ and $y\in S^l(Y)$. Then the inclusion $T_x(S^k(X))\subseteq IS(T_xX)\subseteq T_xX$ is modelled on $\{0\}\t\R^{n-k}\subseteq [0,\iy)^k\t\R^{n-k}\subseteq \R^n$. Hence $T_x(S^k(X))=IS(T_xX)\cap -IS(T_xX)$, and similarly $T_y(S^l(Y))=IS(T_yY)\cap -IS(T_yY)$. Since $\d f\vert_x$ maps $IS(T_xX)\ra IS(T_yY)$ it maps $IS(T_xX)\cap -IS(T_xX)\ra IS(T_yY)
\cap -IS(T_yY)$, that is, $\d f\vert_x$ maps $T_x(S^k(X))\ra
T_y(S^l(Y))$. Hence there is an induced linear map (f\_x)\_\*:T\_xX/T\_x(S\^k(X))T\_yY/T\_y(S\^l(Y)). \[mc3eq1\]
Now let $f:X\ra Y$ be a smooth map.
- We call $f$ a [*diffeomorphism*]{} if $f$ has a smooth inverse $f^{-1}:Y\ra X$.
- We call $f$ an [*immersion*]{} if $\d f\vert_x:T_xX\ra
T_{f(x)}Y$ is injective for all $x\in X$.
- We call $f$ an [*embedding*]{} if it is an injective immersion.
- We call $f$ a [*submersion*]{} if $\d f\vert_x:T_xX\ra
T_{f(x)}Y$ and $\d f\vert_x:T_x(S^k(X))\ra T_{f(x)}(S^l(Y))$ are surjective for all $x\in X$, where $x\in S^k(X)$, $f(x)\in S^l(Y)$.
- We call $f$ [*boundary-submersive*]{}, or [*b-submersive*]{}, if $(\d f\vert_x)_*$ in is surjective for all $x\in X$. Note that $\d f\vert_x$ surjective implies $(\d f\vert_x)_*$ surjective, so submersions are automatically b-submersive.
\[mc3def2\]
Here is how Definition \[mc3def1\] relates to other definitions in the literature:
Weakly smooth maps $f:X\ra Y$ are just the obvious generalization of the usual definition [@KoNo §I.1] of smooth maps for manifolds without boundary. If $\pd Y=\es$ the additional condition in Definition \[mc3def1\] is vacuous, and weakly smooth maps are smooth. Note that the definition of smooth maps $f:X\ra\R$ in Definition \[mc2def2\] is equivalent to Definition \[mc3def1\] when $Y=\R$.
Our definition of smooth maps between manifolds with corners is not equivalent to any other definition that the author has found in the literature, though it is related. Most authors, such as Cerf [@Cerf §I.1.2], define smooth maps of manifolds with corners to be weakly smooth maps, in our notation. But there are also two more complex definitions. Firstly, Monthubert [@Mont Def. 2.8] defines [*morphisms of manifolds with corners*]{} $f:X\ra Y$. One can show that these are equivalent to [*b-submersive smooth maps*]{}, in our notation. We prefer our definition, as b-submersive smooth maps do not have all the properties we want. In particular, Theorem \[mc3thm\](iv),(vi) below fail for b-submersive smooth maps.
Secondly, Melrose [@Melr2 §1.12] defines [*b-maps*]{} between manifolds with corners. Let $f:X\ra Y$ be a weakly smooth map. We call $f$ a [*b-map*]{} if the following holds. Let $x\in
X$ with $f(x)=y$, and let the local boundary components of $X$ at $x$ be $\ti\be_1,\ldots,\ti\be_k$, and of $Y$ at $y$ be $\be_1,\ldots,\be_l$. Suppose $(\ti V_i,\ti b_i)$ is a boundary defining function for $X$ at $(x,\ti\be_i)$, $i=1,\ldots,k$, and $(V_j,b_j)$ a boundary defining function for $Y$ at $(y,\be_j)$, $j=1,\ldots,l$. Then for all $j=1,\ldots,l$ either $b_j\ci f$ should be zero near $x$ in $X$, or there should exist $e_{1j},\ldots,e_{kj}\in\N$ such that $b_j\ci f\equiv
G_j\cdot\prod_{i=1}^k\ti b{}_i^{e_{ij}}$ near $x$ in $X$ for smooth $G_j>0$. Thus, a smooth map in the sense of Definition \[mc3def1\] is exactly a b-map $f:X\ra Y$ such that for all such $x,y$ and $j=1,\ldots,l$, one of $e_{1j},\ldots,e_{kj}$ is 1 and the rest are zero. So our smooth maps are a special class of Melrose’s b-maps. \[mc3rem\]
Here are some properties of smooth maps. The proofs are elementary.
Let $W,X,Y,Z$ be manifolds with corners.
- If $f:X\ra Y$ and $g:Y\ra Z$ are smooth then $g\ci f:X\ra Z$ is smooth.
- The identity map $\id_X:X\ra X$ is smooth.
- Diffeomorphisms $f:X\ra Y$ are equivalent to isomorphisms of smooth manifolds, that is, to homeomorphisms of topological spaces $f:X\ra Y$ which identify the maximal atlases on $X$ and $Y$.
- The map $i_X:\pd X\ra X$ in Definition [\[mc2def5\]]{} is a smooth immersion.
- If $f:W\ra Y$ and $g:X\ra Z$ are smooth, the
product
$f\t g:W\t X\ra Y\t Z$ given by $(f\t g)(w,x)=\bigl(f(w),g(x)\bigr)$ is smooth.
- If $f:X\ra Y$ and $g:X\ra Z$ are smooth, the
direct product
$(f,g):X\ra Y\t Z$ given by $(f,g)(x)=\bigl(f(x),g(x)\bigr)$ is smooth.
- Regarding the empty set $\es$ as a manifold and the point $\{0\}$ as a $0$-manifold, the unique maps $\es:\es\ra X$ and $\pi:X\ra\{0\}$ are smooth.
\[mc3thm\]
Theorem \[mc3thm\](i),(ii) show that manifolds with corners form a [*category*]{}, which we write $\Manc$, with objects manifolds with corners $X$ and morphisms smooth maps $f:X\ra Y$. We write $\Manb$ for the full subcategory of $\Manc$ whose objects are manifolds with boundary, and $\Man$ for the full subcategory of $\Manc$ whose objects are manifolds without boundary, so that $\Man
\subset\Manb\subset\Manc$. Theorem \[mc3thm\](v),(vi) have a category-theoretic interpretation in terms of products in $\Manc$, and (vii) says that $\es$ is an [*initial object*]{} in $\Manc$, and $\{0\}$ is a [*terminal object*]{} in $\Manc$. Here are some examples.
[**(a)**]{} If $X$ is a manifold with corners, the diagonal map $\De_X:X\ra X\t X$, $\De_X:x\mapsto(x,x)$, is a smooth embedding. This follows from Theorem \[mc3thm\](ii),(vi), as $\De_X=(\id_X,\id_X)$. If $\pd X\ne\es$ then $\De_X$ is not b-submersive, so it is not a morphism of manifolds in the sense of Monthubert [@Mont Def. 2.8].
[**(b)**]{} If $X,Y$ are manifolds with corners then the projection $\pi_X:X\t Y\ra X$ is a smooth submersion. This follows from Theorem \[mc3thm\](ii),(v),(vii), by identifying $\pi_X:X\t
Y\ra X$ with $\id_X\t\pi:X\t Y\ra X\t\{0\}$.
[**(c)**]{} The inclusion $i:[0,\iy)\ra\R$ is smooth, but it is not a submersion, since at $0\in[0,\iy)$ the map $\d
i\vert_0:T_0S^0\bigl([0,\iy)\bigr)\ra T_0S^1(\R)$ is not surjective.
[**(d)**]{} The map $f:\R\ra[0,\iy)$, $f(x)=x^2$ is weakly smooth but [*not*]{} smooth, as the additional condition in Definition \[mc3def1\] fails at $x=0$.
[**(e)**]{} The map $f:[0,\iy)^2\ra[0,\iy)$, $f(x,y)=x+y$ is weakly smooth but [*not*]{} smooth, as Definition \[mc3def1\] fails at $(x,y)=(0,0)$.
[**(f)**]{} The map $f:[0,\iy)^2\ra[0,\iy)$, $f(x,y)=xy$ is weakly smooth but [*not*]{} smooth, as Definition \[mc3def1\] fails at $(x,y)=(0,0)$. However, $f$ is a b-map in the sense of Melrose [@Melr2 §1.12], with $e_{11}=e_{21}=1$. \[mc3ex\]
Describing how smooth maps act on corners {#mc4}
=========================================
If $f:X\ra Y$ is a smooth map of manifolds with corners, then $f$ may relate $\pd^kX$ to $\pd^lY$ for $k,l\ge 0$ in complicated ways. We now explain two different ways to describe these relations. The first involves a decomposition $X\t_Y\pd Y=\Xi^f_+\amalg\Xi^f_-$ and maps $\xi^f_+:\Xi^f_+\ra X$ and $\xi^f_-:\Xi^f_-\ra\pd X$. This will be important in [@Joyc] when we define [*d-manifolds with corners*]{} and [*d-orbifolds with corners*]{}, which are ‘derived’ versions of manifolds and orbifolds with corners. To make this generalization we find it helpful to replace a manifold with corners $X$ by the triple $(X,\pd X,i_X)$, so we need to characterize smooth maps $f:X\ra Y$ in terms of the triples $(X,\pd X,i_X),(Y,\pd
Y,i_Y)$.
Let $X,Y$ be manifolds with corners, and $f:X\ra Y$ a smooth map. Consider the smooth maps $f:X\ra Y$ and $i_Y:\pd Y\ra Y$ as continuous maps of topological spaces. Then we may form the fibre product of topological spaces $X\t_Y\pd Y=X\t_{f,Y,i_Y}\pd Y$, given explicitly by $$X\t_Y\pd Y=\bigl\{\bigl(x,(y,\be)\bigr)\in X\t\pd
Y:f(x)=y=i_Y(y,\be)\bigr\}.$$ This is a closed subspace of the topological space $X\t\pd Y$, since $X,\pd Y$ are Hausdorff, and so it is a topological space with the subspace topology.
By Definition \[mc3def1\], for each $\bigl(x,(y,\be)\bigr)\in
X\t_Y\pd Y$, if $(V,b)$ is a boundary defining function for $Y$ at $(y,\be)$, then either $b\ci f\equiv 0$ near $x$, or $(f^{-1}(V),b\ci f)$ is a boundary defining function for $X$ at some $(x,\ti\be)$. Define subsets $\Xi^f_+,\Xi^f_-$ of $X\t_Y\pd Y$ by $\bigl(x,(y,\be)\bigr)\in\Xi^f_+$ if $b\ci f\equiv 0$ near $x$, and $\bigl(x,(y,\be)\bigr)\in\Xi^f_-$ otherwise. Define maps $\xi^f_+:\Xi^f_+\ra X$ by $\xi^f_+\bigl(x,(y,\be)\bigr)=x$ and $\xi^f_-:\Xi^f_-\ra\pd X$ by $\xi^f_-\bigl(x,(y,\be)\bigr)
=(x,\ti\be)$, for $(x,\ti\be)$ as above. It is easy to show that $\Xi^f_\pm,\xi^f_\pm$ can also be defined solely in terms of $\bigl(x,(y,\be)\bigr)$ and $\d f\vert_x$, and so they are independent of the choice of $(V,b)$, and are well-defined. \[mc4def1\]
Here are some properties of these $\Xi^f_\pm,\xi^f_\pm$. A continuous map $g:X\ra Y$ is a [*finite covering map*]{} if every $y\in Y$ has an open neighbourhood $U$ such that $g^{-1}(U)$ is homeomorphic to $U\t T$ for some finite set $T$ with the discrete topology, and $g:\xi_f^{-1}(U)\ra U$ corresponds to the projection $U\t S\ra U$.
Let $f:X\ra Y$ be a smooth map of manifolds with corners, and $\Xi^f_\pm,\xi^f_\pm$ be as in Definition [\[mc4def1\],]{} and set $n=\dim X$. Then
- $\Xi^f_\pm$ are open and closed subsets of $X\t_Y\pd Y,$ with $X\t_Y\pd Y=\Xi^f_+\amalg\Xi^f_-$.
- $\xi^f_+:\Xi^f_+\ra X$ and $\xi^f_-:\Xi^f_-\ra\pd
X$ are proper, finite covering maps of topological spaces, with $\smash{\xi^f_+\equiv\pi_X\vert_{\Xi^f_+}}$ and $\smash{i_X\ci
\xi^f_-\equiv\pi_X\vert_{\Xi^f_-}}$.
- Part [(b)]{} implies there is a unique way to make $\smash{\Xi^f_+}$ into an $n$-manifold with corners and $\Xi^f_-$ into an $(n\!-\!1)$-manifold with corners so that $\xi^f_\pm$ are local diffeomorphisms, and so covering maps of manifolds. Then the projections $\pi_X:\Xi^f_\pm\ra X$ and $\pi_{\pd Y}:\Xi^f_\pm\ra\pd Y$ are smooth maps.
\[mc4prop1\]
For (a), clearly $X\t_Y\pd Y=\Xi^f_+\amalg\Xi^f_-$. Let $\bigl(x,(y,\be)\bigr)\in X\t_Y\pd Y$, and $(V,b)$ be a boundary defining function for $Y$ at $(y,\be)$. Then $(V,b)$ is also a boundary defining function for $Y$ at any $(y',\be')$ sufficiently close to $(y,\be)$ in $\pd Y$. Hence if $\bigl(x',(y',\be')\bigr)$ is sufficiently close to $\bigl(x,(y,\be)\bigr)\in X\t_Y\pd Y$ then $(V,b)$ is also a boundary defining function for $Y$ at $(y',\be')$. We have $\bigl(x,(y,\be)\bigr)\in\Xi^f_+$ if $b\ci f\equiv 0$ near $x$. Fixing $(V,b)$ this is an open condition in $x$, so $\Xi^f_+$ is open in $X\t_Y\pd Y$, and thus $\Xi^f_-=(X\t_Y\pd Y)\sm\Xi^f_+$ is closed in $X\t_Y\pd Y$. Similarly, $\bigl(x,(y,\be)\bigr)
\in\Xi^f_-$ if $(f^{-1}(V),b\ci f)$ is a boundary defining function for $X$ at some $(x,\ti\be)$. Fixing $(V,b)$ this is an open condition in $(x,\ti\be)$, so $\Xi^f_-$ is open, and $\Xi^f_+=(X\t_Y\pd Y)\sm\Xi^f_-$ is closed, proving (a).
For (b), the identities $\smash{\xi^f_+\equiv\pi_X\vert_{\Xi^f_+}}$ and $\smash{i_X\ci \xi^f_-\equiv\pi_X\vert_{\Xi^f_-}}$ are immediate. First consider $\xi^f_+:\Xi^f_+\ra X$. Since $i_Y:\pd
Y\ra Y$ is proper and finite by Lemma \[mc2lem\], $\pi_X:X\t_Y\pd
Y\ra X$ is proper and finite by properties of topological fibre products, and so $\xi^f_+=\pi_X\vert_{\smash{\Xi^f_+}}:\Xi^f_+\ra X$ is proper and finite as $\Xi^f_+$ is closed in $X\t_Y\pd Y$. To see that $\xi^f_+$ is a covering map, note that it is a local homeomorphism, since as above, given $\bigl(x,(y,\be)\bigr)\in\Xi^f_+$, if $x'$ is close to $x$ in $X$ then setting $y'=f(x')$, $(V,b)$ is a boundary defining function for $Y$ at $(y',\be')$ for some unique local boundary component $\be'$ of $Y$ at $\be'$, and then $\bigl(x',(y',\be')\bigr)\in\Xi^f_+$ with $\xi^f_+\bigl(x',(y',\be')\bigr)=x'$. We have constructed a local inverse $x'\mapsto\bigl(x',(y',\be')\bigr)$ for $\xi^f_+$ which is clearly continuous, so $\xi^f_+$ is a local homeomorphism, and thus a finite covering map, as it is finite.
Next consider $\xi^f_-:\Xi^f_-\ra\pd X$. As above, given $\bigl(x,(y,\be)\bigr)\in\Xi^f_-$ we may fix a boundary defining function $(V,b)$ for $Y$ at $(y,\be)$, and then for $\bigl(x',(y',\be')\bigr)$ near $\bigl(x,(y,\be)\bigr)$ in $\Xi^f_-$ we have $\xi^f_-\bigl(x',(y',\be')\bigr)=(x',\ti\be')$, where $\ti\be'$ is the unique local boundary component of $X$ at $x'$ such that $(f^{-1}(V),b\ci f)$ is a boundary defining function for $X$ at $(x',\ti\be')$. Therefore $\xi^f_-$ is continuous, as $\ti\be'$ depends continuously on $x'$. As above $\pi_X:X\t_Y\pd Y\ra X$ is proper and finite, so $i_X\ci\xi^f_-=\pi_X\vert_{\smash{\Xi^f_-}}:\Xi^f_- \ra X$ is proper and finite as $\Xi^f_-$ is closed, and hence $\xi^f_-$ is proper and finite. We show $\xi^f_-$ is a finite covering map by constructing a local inverse $(x',\ti\be')\mapsto\bigl(x',(y',\be')\bigr)$ as for $\xi^f_+$. This proves (b).
For (c), $\pi_X:\Xi^f_+\ra X$ is $\xi^f_+:\Xi^f_+\ra X$ and $\pi_X:\Xi^f_-\ra X$ is $i_X\ci\xi^f_-:\Xi^f_-\ra X$, so $\pi_X:\Xi^f_\pm\ra X$ are smooth as $\xi^f_\pm$ are covering maps of manifolds, and so smooth. To see $\pi_{\pd Y}:\Xi^f_\pm\ra\pd Y$ are smooth, note that $i_Y\ci\pi_{\pd Y}\equiv f\ci\pi_X$ as maps $\Xi^f_\pm\ra Y$, so $i_Y\ci\pi_{\pd Y}:\Xi^f_\pm\ra Y$ is smooth, and it follows that $\pi_{\pd Y}:\Xi^f_\pm\ra\pd Y$ is smooth as $\pi_{\pd Y}$ is continuous and $i_Y:\pd Y\ra Y$ is an immersion.
Using $\Xi^f_\pm,\xi^f_\pm$ we can define a decomposition $\pd
X=\pd_+^fX\amalg\pd_-^fX$.
Let $f:X\ra Y$ be a smooth map of manifolds with corners. Define $\pd_-^fX=\xi_-^f(\Xi_-^f)$ and $\pd_+^fX=\pd
X\sm\xi_-^f(\Xi_-^f),$ so that $\pd X=\pd_+^fX\amalg\pd_-^fX$. Then $\pd_\pm^fX$ are open and closed in $\pd X,$ so they are manifolds with corners. \[mc4prop2\]
As $\xi_-^f:\Xi_-^f\ra\pd X$ is a covering map by Proposition \[mc4prop1\](b), $\xi_-^f(\Xi_-^f)$ is open in $\pd
X$. Since $\xi_-^f$ is proper and $\Xi_-^f,\pd X$ are Hausdorff, $\xi_-^f(\Xi_-^f)$ is closed in $\pd X$. So $\pd_-^fX$, and hence $\pd_+^fX=\pd X\sm\pd_-^fX$, are open and closed in $\pd X$.
We can characterize b-submersive morphisms $f:X\ra Y$ in Definition \[mc3def2\](v) in terms of $\Xi^f_-,\xi^f_-$. The proof is an easy exercise.
Let $f:X\ra Y$ be a smooth map of manifolds with corners. Then $f$ is b-submersive if and only if $\Xi^f_-=X\t_Y\pd
Y,$ so that $\Xi^f_+=\es,$ and $\xi^f_-:\Xi^f_-\ra\pd X$ is injective. \[mc4lem\]
We now move on to our second way of describing how $f$ relates $\pd^kX$ and $\pd^lY$. Equation showed that if $X$ is a manifold with corners then $X\mapsto\coprod_{i\ge 0}C_i(X)$ commutes with products of manifolds. We will explain how to lift a smooth map $f:X\ra Y$ up to a map $C(f):\coprod_{i\ge 0}C_i(X)\ra\coprod_{j\ge
0}C_j(Y)$ which is (in a generalized sense) smooth, and which is functorial in a very strong sense.
Let $X,Y$ be smooth manifolds with corners and $f:X\ra
Y$ a smooth map. Define $C(f):\coprod_{i=0}^{\dim X} C_i(X)\ra
\coprod_{j=0}^{\dim Y}C_j(Y)$ by
&C(f):(x,{\_1,…,\_i})(y,{\_1,…,\_j}),\
&{\_1,…,\_j}={:(x,(y,)) \^f\_-, \^f\_-(x,(y,))=(x,\_l), l=1,…,i}.
\[mc3eq2\] \[mc4def2\]
Let $\{X_i:i\in I\}$ and $\{Y_j:j\in J\}$ be families of manifolds, where $I,J$ are indexing sets. We do not assume that all $X_i$ have the same dimension, or that all $Y_j$ have the same dimension, so $\coprod_{i\in I}X_i$ and $\coprod_{j\in J}Y_j$ need not be manifolds. We call a map $f:\coprod_{i\in I}X_i\ra
\coprod_{j\in J}Y_j$ [*smooth*]{} if $f$ is continuous, and for all $i\in I$ and $j\in J$ the map $$f\vert_{X_i\cap f^{-1}(Y_j)}:X_i\cap f^{-1}(Y_j)\ra Y_j$$ is a smooth map of manifolds. Here $Y_j$ is an open and closed subset of the topological space $\coprod_{j\in J}Y_j$, so $X_i\cap
f^{-1}(Y_j)$ is an open and closed subset of $X_i$ as $f$ is continuous, and thus $X_i\cap f^{-1}(Y_j)$ is a manifold. \[mc4def3\]
The next theorem, in part parallel to Theorem \[mc3thm\], gives properties of these maps $C(f)$. The proofs are elementary. The theorem basically says that mapping $X\mapsto \coprod_{i\ge
0}C_i(X)$ and $f\mapsto C(f)$ yields a [*functor*]{} which preserves smoothness, composition, identities, boundaries $\pd X$, immersions $i_X:\pd X\ra X$, and products and direct products of smooth maps. Theorem \[mc6thm2\] will also show that the functor preserves strongly transverse fibre products.
Let $W,X,Y,Z$ be manifolds with corners.
- If $f:X\ra Y$ is smooth then $C(f):\coprod_{i\ge 0}C_i(X)\ra\coprod_{j\ge 0}C_j(Y)$ is smooth in the sense of Definition [\[mc4def3\]]{}.
- If $f:X\ra Y$ and $g:Y\ra Z$ are smooth then $C(g\ci f)=C(g)\ci C(f):\coprod_{i\ge 0}C_i(X)\ra\coprod_{k\ge
0}C_k(Z)$.
- $C(\id_X)=\id_{\coprod_{k\ge 0}C_k(X)}:
\coprod_{k\ge 0}C_k(X)\ra\coprod_{k\ge 0}C_k(X)$.
- The diffeomorphisms $C_k(\pd X)\cong\pd
C_k(X)$ in identify $$\begin{aligned}
&\ts C(i_X):\coprod_{k\ge 0}C_k(\pd X)\longra \coprod_{k\ge 0}C_k(X)
\qquad\text{with}\\
&i_{\coprod_{k\ge 0}C_k(X)}:=\ts \coprod_{k\ge 0}i_{C_k(X)}=\ts
\coprod_{k\ge 0}\pd C_k(X)\longra \coprod_{k\ge 0}C_k(X).\end{aligned}$$
- Let $f:W\ra Y$ and $g:X\ra Z$ be smooth maps. Then gives \_[m0]{}C\_m(WX)&,\
\_[n0]{}C\_n(YZ)&. \[mc3eq3\] These identify $C(f\t g):\coprod_{m\ge 0}C_m(W\t
X)\ra\coprod_{n\ge 0}C_n(Y\t Z)$ with $C(f)\t
C(g)\!:\!\bigl[\coprod_{i\ge 0}\!C_i(W)
\bigr]\!\t\!\bigl[\coprod_{j\ge 0}\!C_j(X) \bigr]\!\ra\!
\bigl[\coprod_{k\ge 0}\!C_k(Y)\bigr]\!\t\!\bigl[\coprod_{l\ge
0}\!C_l(Z)\bigr]$.
- Let $f:X\ra Y$ and $g:X\ra Z$ be smooth maps. Then identifies $C\bigl((f,g)\bigr):\coprod_{j\ge
0}C_j(X)\ra\coprod_{n\ge 0}C_n(Y\t Z)$ with $\bigl(C(f),C(g)\bigr):\coprod_{j\ge
0}C_j(X)\ra\bigl[\coprod_{k\ge
0}C_k(Y)\bigr]\t\bigl[\coprod_{l\ge 0}C_l(Z)\bigr]$.
- Let $f:X\ra Y$ be a b-submersive smooth map. Then $C(f)$ maps $C_i(X)\ra \coprod_{j\ge 0}^iC_j(Y)$ for all $i\ge 0$.
\[mc4thm\]
Curiously, there is a second way to define a map $\coprod_{i\ge
0}C_i(X)\ra\coprod_{j\ge 0}C_j(Y)$ with the same properties. Define $\hat C(f):\coprod_{i\ge 0}C_i(X)\ra\coprod_{j\ge 0}C_j(Y)$ by $$\begin{aligned}
&\hat C(f):\bigl(x,\{\ti\be_1,\ldots,\ti\be_i\}\bigr)\longmapsto
\bigl(y,\{\be_1,\ldots,\be_j\}\bigr),\quad\text{where}\quad
\{\be_1,\ldots,\be_j\}=\\
&\bigl\{\be:\bigl(x,(y,\be)\bigr)\!\in\!
\Xi^f_-,\; \xi^f_-\bigl(x,(y,\be)\bigr)\!=\!(x,\ti\be_l),\;
1\!\le\! l\!\le\! i\bigr\}\!\cup\!
\{\bigl\{\be:\bigl(x,(y,\be)\bigr)\!\in\!
\Xi^f_+\bigr\}.\end{aligned}$$ Then the analogues of Theorems \[mc4thm\] and \[mc6thm2\] also hold for $\hat C(f),\hat C(g),\ldots$.
Submersions {#mc5}
===========
Definition \[mc3def2\](iv) defined [*submersions*]{} $f:X\ra Y$ between manifolds with corners $X,Y$. We show that submersions are locally isomorphic to projections.
Let $X,Y$ be manifolds with corners, $f:X\ra Y$ a submersion, and $x\in X$ with $f(x)=y$. Then there exist open neighbourhoods $X',Y'$ of $x,y$ in $X,Y$ with $f(X')=Y',$ a manifold $Z'$ with corners with $\dim X=\dim Y+\dim Z',$ and a diffeomorphism $X'\cong Y'\t Z',$ such that $f\vert_{X'}:X'\ra Y'$ is identified with $\pi_{Y'}:Y'\t Z'\ra Y'$. \[mc5prop1\]
Let $x\in X$ and $y=f(x)\in Y$, with $\dim X=m$, $\dim Y=n$ and $x\in S^k(X)$, $y\in S^l(X)$. Choose charts $(U,\phi)$, $(V,\psi)$ on $X,Y$ with $U,V$ open in $\R^m_k,\R^n_l$ and $0\in U$, $0\in V$ with $\phi(0)=x$, $\psi(0)=y$ and $f\ci\phi(U)\subseteq\psi(V)$. Write $(x_1,\ldots,x_m)$, $(y_1,\ldots,y_n)$ for the coordinates on $U,V$ respectively. Write $\ti\be_i$ for the local boundary component $\phi_*(\{x_i=0\})$ for $i=1,\ldots,k$, and $\be_j$ for the local boundary component $\psi_*(\{y_j=0\})$ for $j=1,\ldots,l$.
Lemma \[mc4lem\] implies that $\bigl(x,(y,\be_j)\bigr)\in\Xi_-^f$ with $\xi_-^f\bigl(x,(y,\be_j)\bigr)=(x,\ti\be_{i_j})$ for each $j=1,\ldots,l$ and some $i_j=1,\ldots,k$, and $i_1,\ldots,i_l$ are distinct as $\xi_-^f$ is injective. Thus $l\le k$, and reordering $x_1,\ldots,x_k$ if necessary we suppose that $i_j=j$. By Proposition \[mc2prop6\](d), $\bigl(\psi(V),y_i\ci\psi^{-1}
\bigr)$ is a boundary defining function for $Y$ at $(y,\be_i)$ for $i=1,\ldots,l$, so by Definition \[mc3def1\] $\bigl(f^{-1}(\psi(V)),y_i\ci\psi^{-1}\ci f\bigr)$ is a boundary defining function for $X$ at $(x,\ti\be_i)$ for $i=1,\ldots,l$. But $\bigl(\phi(U),x_i\ci\phi^{-1}\bigr)$ is also a boundary defining function for $X$ at $(x,\ti\be_i)$, so by Proposition \[mc2prop6\](b), making $U$ smaller if necessary we can suppose that $$y_i\ci\psi^{-1}\ci f\ci\phi\equiv x_i\cdot g_i$$ on $U$, for some smooth $g_i:U\ra(0,\iy)$ and all $i=1,\ldots,l$.
Combining this with the surjectivity conditions in Definition \[mc3def2\](iv), we see that we may choose alternative coordinates $(\ti x_1,\ldots,\ti x_n)$ on an open neighbourhood $\ti U$ of 0 in $U$ taking values in $\R^m_k$ and zero at 0, such that $$\ti x_i\equiv \begin{cases} y_i\ci\psi^{-1}\ci f\ci\phi,&
i=1,\ldots,l, \\ x_i, & i=l+1,\ldots,k, \\
y_{i-k+l}\ci\psi^{-1}\ci f\ci\phi,& i=k+1,\ldots,n-k+l, \\
\text{some function of $x_{k+1},\ldots,x_n$,} & i=n-k+l+1,\ldots,m.
\end{cases}$$ Choose small $\ep>0$ so that $[0,\ep)^k\t(-\ep,\ep)^{m-k} \subseteq
\ti U$ in coordinates $(\ti x_1,\ldots,\ti x_n)$. Then defining $X'=\phi\bigl(\{(\ti x_1,\ldots,\ti x_m)\in
[0,\ep)^k\t(-\ep,\ep)^{m-k}\}\bigr)$, $Y'=\psi\bigl([0,\ep)^l\t
(-\ep,\ep)^{n-l}\bigr)$ and $Z'=[0,\ep)^{k-l}\t
(-\ep,\ep)^{m-k-n+l}$, the proposition follows.
Submersions $f:X\ra Y$ are nicely compatible with the boundaries $\pd X,\pd Y$.
Let $f:X\ra Y$ be a submersion, and $\pd
X=\pd_+^fX\amalg\pd_-^fX$ be as in Proposition [\[mc4prop2\]]{}. Then $f_+=f\ci i_X\vert_{\pd_+^fX}:\pd_+^fX\ra Y$ is a submersion. There is a natural submersion $\smash{f_-:\pd_-^fX\ra\pd Y}$ with $\smash{f\ci i_X\vert_{\pd_-^fX}\equiv i_Y\ci f_-}$. \[mc5prop2\]
Lemma \[mc4lem\] shows that $\xi^f_-:\Xi^f_-=
X\t_Y\pd Y\ra\pd X$ is injective. Since $\pd_-^fX=\xi^f_-(\Xi^f_-)$, it follows that $\xi^f_-:X\t_Y\pd Y\ra\pd_-^fX$ is invertible. Define $f_-:\pd_-^fX\ra\pd Y$ by $f_-=\pi_{\pd Y}\ci(\xi^f_-)^{-1}$. Then $i_Y\ci f_-=i_Y\ci\pi_{\pd Y}\ci(\xi^f_-)^{-1}
=f\ci\pi_X\ci(\xi^f_-)^{-1}=f\ci i_X\vert_{\smash{\pd_-^fX}}$, since $i_Y\ci\pi_Y=f\ci\pi_X:X\t_Y\pd Y\ra Y$ and $\pi_X\ci(\xi^f_-)^{-1}=i_X\vert_{\smash{\pd_-^fX}}: \pd_-^fX\ra X$ as $i_X\ci\xi^f_-\equiv\pi_X\vert_{\smash{\Xi^f_-}}$ by Proposition \[mc4prop1\](b).
It remains to check that the maps $f_\pm$ are submersions. Since being a submersion is a local property, by Proposition \[mc5prop1\] it is enough to show $f_\pm$ are submersions when $f:X\ra Y$ is a projection $\pi_{Y'}:Y'\t Z'\ra Y'$. We have a natural isomorphism $\pd(Y'\t Z')\cong(\pd Y'\t Z')\amalg(Y'\t\pd
Z')$. It is easy to see that $\pd_+^f(Y'\t Z')\cong Y'\t\pd Z'$, so that $f_+$ becomes the projection $Y'\t\pd Z'\ra Y'$ which is a submersion, and that $\pd_-^f(Y'\t Z')\cong\pd Y'\t Z'$, so that $f_-$ becomes the projection $\pd Y'\t Z'\ra\pd Y'$ which is a submersion.
Note that we can iterate this construction to decompose $\pd^kX$, so that $$\pd^2X=\pd_+^{f_+}(\pd_+^fX)\amalg \pd_-^{f_+}(\pd_+^fX)\amalg
\pd_+^{f_-}(\pd_-^fX)\amalg \pd_-^{f_-}(\pd_-^fX),$$ for instance, and $f$ lifts to a submersion on every piece.
Transversality and fibre products of manifolds {#mc6}
==============================================
Let $X,Y,Z$ be manifolds with corners and $f:X\ra Z$, $g:Y\ra Z$ be smooth maps. From category theory, a [*fibre product*]{} $X\t_{f,Z,g}Y$ in the category $\Manc$ consists of a manifold with corners $W$ and smooth maps $\pi_X:W\ra X$, $\pi_Y:W\ra Y$ such that $f\ci\pi_X=g\ci\pi_Y:W\ra Z$, satisfying the universal property that if $W'$ is a manifold with corners and $\pi_X':W'\ra X$, $\pi_Y':W'\ra Y$ are smooth maps with $f\ci\pi_X'=g\ci\pi_Y'$, then there exists a unique smooth map $h:W'\ra W$ with $\pi_X'=\pi_X\ci
h$ and $\pi_Y'=\pi_Y\ci h$. We now give sufficient conditions for fibre products of manifolds with corners to exist.
Let $X,Y,Z$ be manifolds with corners and $f:X\ra
Z$, $g:Y\ra Z$ be smooth maps. We call $f,g$ [*transverse*]{} if the following holds. Suppose $x\in X$, $y\in Y$ and $z\in Z$ with $f(x)=z=g(y)$, so that there are induced linear maps of tangent spaces $\d f\vert_x:T_xX\ra T_zZ$ and $\d g\vert_y:T_yY\ra T_zZ$. Let $x\in S^j(X)$, $y\in S^k(Y)$ and $z\in S^l(Z)$, so that as in Definition \[mc3def2\] $\d f\vert_x$ maps $T_x(S^j(X))\ra
T_z(S^l(Z))$ and $\d g\vert_y$ maps $T_y(S^k(Y))\ra T_z(S^l(Z))$. Then we require that $T_zZ=\d f\vert_x(T_xX)+\d g\vert_y(T_yY)$ and $T_z(S^l(Z))=\d f\vert_x(T_x(S^j(X)))+\d g\vert_y(T_y(S^k(Y)))$ for all such $x,y,z$. From Definition \[mc3def2\], if one of $f,g$ is a submersion then $f,g$ are automatically transverse. \[mc6def1\]
If $X,Y,Z$ are manifolds without boundary then $j=k=l=0$ in Definition \[mc6def1\], and both conditions reduce to the usual definition $T_zZ=\d f\vert_x(T_xX)+\d g\vert_y(T_yY)$ of transverse smooth maps. When $X,Y,Z$ are manifolds with corners we believe this definition of transversality is new, since it depends heavily on our definition of smooth maps which is also new.
Definition \[mc6def1\] imposes two transversality conditions on $f,g$ at $x,y,z$, the first on the corners $C_0(X)\cong
X,C_0(Y)\cong Y,C_0(Z)\cong Z$ of $X,Y,Z$ of largest dimension at $x,y,z$, and the second on the corners $C_j(X)\cong S^j(X)$, $C_k(Y)\cong S^k(X)$, $C_l(Z)\cong S^l(Z)$ (locally) of $X,Y,Z$ of smallest dimension at $x,y,z$.
One might think that to prove Theorem \[mc6thm1\] one would need to impose transversality conditions on corners $C_a(X),C_b(Y),C_c(Y)$ of intermediate dimensions $0\le a\le j$, $0\le b\le k$, $0\le c\le l$ as well. In fact these intermediate conditions are implied by our definition of smooth maps, since the requirement for $f,g$ to pull boundary defining functions back to boundary defining functions is a kind of transversality condition at the boundaries. One of the motivations for our definition of smooth maps of manifolds with corners was to have a simple, not too restrictive condition for the existence of fibre products. \[mc6rem1\]
Margalef-Roig and Outerelo Dominguez [@MaOu §7.2] also define transversality of smooth maps between manifolds with corners, and prove their own version of Theorem \[mc6thm1\] below. They work with Banach manifolds and $C^p$ maps for $p=0,1,\ldots,\iy$. For finite-dimensional manifolds, their notion of smooth map (‘map of class $\iy$’) corresponds to our weakly smooth maps. However, their notion of [*transversality*]{} [@MaOu Def. 7.2.1] is very restrictive.
In our notation, if $f:X\ra Z$ and $g:Y\ra Z$ are weakly smooth maps, then $f,g$ are transverse in the sense of [@MaOu Def. 7.2.1] if and only if whenever $x\in X$ and $y\in
Y$ with $f(x)=g(y)=z\in Z$ then $z\in Z^\ci$, and $x\in S^j(X)$, $y\in S^k(Y)$ with $T_zZ=\d f\vert_x(T_x(S^j(X)))+\d
g\vert_y(T_y(S^k(Y)))$. In particular, $f(X)$ and $g(Y)$ cannot intersect in the boundary strata $S^l(Z)$ for $l>0$ but only in the interior $Z^\ci$, so in effect Margalef-Roig and Outerelo Dominguez reduce to the case in which $\pd Z=\es$, and then their [@MaOu Prop. 7.2.7] is a special case of Theorem \[mc6thm1\]. So, for example, $f,g$ are generally not transverse in the sense of [@MaOu Def. 7.2.1] if one of $f,g$ is a submersion, or even if $f=\id_X:X\ra X=Z$. \[mc6rem2\]
For manifolds without boundary the following theorem is well-known, as in Lang [@Lang Prop. II.4]. For manifolds with corners Margalef-Roig and Outerelo Dominguez [@MaOu Prop. 7.2.7] prove it with a stricter notion of transversality, as above. We believe this version is new. The proof is given in §\[mc8\].
Suppose $X,Y,Z$ are manifolds with corners and $f:X\ra Z,$ $g:Y\ra Z$ are transverse smooth maps. Then there exists a fibre product $W=X\t_{f,Y,g}Z$ in the category $\Manc$ of manifolds with corners, which is given by an explicit construction, as follows.
As a topological space $W=\{(x,y)\in X\t Y:f(x)=g(y)\},$ with the topology induced by the inclusion $W\subseteq X\t Y,$ and the projections $\pi_X:W\ra X$ and $\pi_Y:W\ra Y$ map $\pi_X:(x,y)\mapsto x,$ $\pi_Y:(x,y)\mapsto y$. Let $n=\dim X+\dim
Y-\dim Z$, so that $n\ge 0$ if $W\ne\es$. The maximal atlas on $W$ is the set of all charts $(U,\phi),$ where $U\subseteq\R^n_k$ is open and $\phi:U\ra W$ is a homeomorphism with a nonempty open set $\phi(U)$ in $W,$ such that $\pi_X\ci\phi:U\ra X$ and $\pi_Y\ci\phi:U\ra Y$ are smooth maps, and for all $u\in U$ with $\phi(u)=(x,y),$ the following induced linear map of real vector spaces is injective: (\_X)\_u(\_Y)\_u:T\_uU=\^nT\_xXT\_yY. \[mc6eq1\] \[mc6thm1\]
We note one important special case of Theorem \[mc6thm1\], the intersection of submanifolds. Suppose $X,Y$ are embedded submanifolds of $Z$, with inclusions $i:X\hookra Z$ and $j:Y\hookra
Z$. Then we say that $X,Y$ [*intersect transversely*]{} if the smooth embeddings $i,j$ are transverse. Then the fibre product $W=X\t_ZY$ is just the intersection $X\cap Y$ in $Z$, and Theorem \[mc6thm1\] shows that it is also an embedded submanifold of $Z$. If $f,g$ are not transverse, then a fibre product $X\t_{f,Y,g}Z$ may or may not exist in the category $\Manc$. Even if one exists, from the point of view of derived differential geometry [@Spiv], it is in some sense the ‘wrong answer’. Here are some examples.
[**(a)**]{} The inclusion $i:\{0\}\ra\R$ is not transverse to itself. A fibre product $\{0\}\t_{i,\R,i}\{0\}$ does exist in $\Manc$ in this case, the point $\{0\}$. Note however that it does not have the expected dimension: $\{0\}\t_{\R}\{0\}$ has dimension 0, but Theorem \[mc6thm1\] predicts the dimension $\dim\{0\}+\dim\{0\}-\dim\R=-1$.
[**(b)**]{} Consider the smooth maps $f:\R\ra\R^2$ and $g:\R\ra\R^2$ given by $$f(x)=(x,0)\quad\text{and}\quad g(x,y)=\begin{cases} (x,e^{-x^2}\sin
(\pi/x)), & x\ne 0, \\
(0,0), & x=0. \end{cases}$$ These are not transverse at $f(0)=g(0)=(0,0)$. The fibre product does not exist in $\Manc$. To see this, note that the topological fibre product $\R\t_{f,\R^2,g}\R$ is $\{(1/n,0):0\ne
n\in\Z\}\cup\{(0,0)\}$, which has no manifold structure. \[mc6ex1\]
In the general case of Theorem \[mc6thm1\], the description of $\pd W$ in terms of $\pd X,\pd Y,\pd Z$ is rather complicated, as can be seen from the proof in §\[mc8\]. Here are three cases in which the expression simplifies. The proofs follow from the proof of Theorem \[mc6thm1\] in §\[mc8\], or alternatively from equation below with $i=1$, since $\pd W\cong C_1(W)$ and $f,g$ are strongly transverse in each case.
Let $X,Y$ be manifolds with corners, and $f:X\ra
Y$ a submersion. Then there is a canonical diffeomorphism \_-\^fXX\_[f,Y,i\_Y]{}Y, \[mc6eq2\] which identifies the submersions $\smash{f_-:\pd_-^fX\ra\pd Y}$ and $\pi_{\pd Y}:X\t_{Y}\pd Y\ra\pd Y$. \[mc6prop1\]
Let $X,Y$ be manifolds with corners, $Z$ a manifold without boundary, and $f:X\ra Z,$ $g:Y\ra Z$ be transverse smooth maps. Then $f\ci i_X:\pd X\ra Z,$ $g:Y\ra Z$ are transverse, and $f:X\ra Z,$ $g\ci i_Y:\pd Y\ra Z$ are transverse, and there is a canonical diffeomorphism (X\_[f,Z,g]{}Y)(X\_[fi\_X,Z,g]{}Y) (X\_[f,Z,gi\_Y]{}Y). \[mc6eq3\] \[mc6prop2\]
Let $X,Y,Z$ be manifolds with corners, $f:X\ra
Z$ a submersion and $g:Y\ra Z$ smooth. Then there is a canonical diffeomorphism (X\_[f,Z,g]{}Y)(\_+\^fX \_[f\_+,Z,g]{}Y) (X\_[f,Z,gi\_Y]{}Y). \[mc6eq4\] If both $f,g$ are submersions there is also a canonical diffeomorphism
(&X\_[f,Z,g]{}Y)\
&(\_+\^fX \_[f\_+,Z,g]{}Y)(X \_[f,Z,g\_+]{} \_+\^gY)(\_-\^fX\_[f\_-,Z,g\_-]{}\_-\^gY).
\[mc6eq5\] Equation also holds if $f,g$ are transverse and $f$ is b-submersive, and also holds if $f,g$ are transverse and both are b-submersive. \[mc6prop3\]
We will also discuss a stronger notion of transversality. To introduce it we prove the following lemma.
Let $X,Y,Z$ be manifolds with corners, $f:X\ra Z$ and $g:Y\ra Z$ be transverse smooth maps, and $C(f),C(g)$ be as in . Suppose $(x,\{\be_1,\ldots,\be_j\})\in C_j(X)$ and $(y,\{\ti\be_1,\ldots,\ti\be_k\})\in C_k(Y)$ with $C(f)
(x,\{\be_1,\ab\ldots,\ab\be_j\})\ab=C(g)(y,\{\ti\be_1,\ldots,
\ti\be_k\})= (z,\{\dot\be_1,\ldots,\dot\be_l\})$ in $C_l(Z)$. Then $j+k\ge l$. \[mc6lem\]
Since $C(f)(x,\{\be_1,\ldots,\be_j\})=
(z,\{\dot\be_1,\ldots,\dot\be_l\})$ it follows that $\d f\vert_x$ maps the vector subspace $T_x\be_1\cap\cdots\cap T_x\be_j$ in $T_xX$ to the vector subspace $T_z\dot\be_1\cap\cdots\cap T_z\dot\be_l$ in $T_zZ$, as the restriction of $\d f\vert_x$ to $T_x\be_1\cap
\cdots\cap T_x\be_j$ is naturally identified with $\d
C(f)\vert_{(x,\{\be_1,\ldots,\be_j\})}$. Similarly, $\d g\vert_y$ maps $T_y\ti\be_1\cap\cdots\cap T_y\ti\be_k$ in $T_yY$ to $T_z\dot\be_1\cap\cdots\cap T_z\dot\be_l$ in $T_zZ$. Since $f,g$ are transverse, we have $T_zZ=\d f\vert_x(T_xX)+\d g\vert_y(T_yY)$. Passing to the quotients $T_xX/(T_x\be_1\cap\cdots\cap
T_x\be_j),\ab\ldots,T_zZ/(T_z\dot\be_1\cap\cdots\cap T_z\dot\be_l)$ and using the facts above shows that
(f\_x)\_\*(T\_xX/(T\_x\_1T\_x\_j))+ (g\_y)\_\*(T\_yY/(T\_y\_1T\_y\_k))&\
= T\_zZ/(T\_z\_1T\_z\_l)&.
\[mc6eq6\] As the vector spaces in have dimensions $j,k,l$, it follows that $j+k\ge l$.
Let $X,Y,Z$ be manifolds with corners and $f:X\ra
Z$, $g:Y\ra Z$ be smooth maps. We call $f,g$ [*strongly transverse*]{} if they are transverse, and whenever there are points in $C_j(X),C_k(Y),C_l(Z)$ with C(f)(x,{\_1,…,\_j})=C(g)(y,{\_1,…, \_k})=(z,{\_1,…,\_l}) \[mc6eq7\] we have either $j+k>l$ or $j=k=l=0$. That is, in Lemma \[mc6lem\], equality in $j+k\ge l$ is allowed only if $j=k=l=0$.
Suppose $f,g$ are smooth with $f$ a submersion. Then $f,g$ are automatically transverse, as in Definition \[mc6def1\], and in , Theorem \[mc4thm\](v) implies that $j\ge l$. Hence if $k>0$ then $j+k>l$. If $k=0$ then $l=0$ as $C(g)$ maps $C_0(Y)\ra
C_0(Z)$, so either $j+k>l$ or $j=k=l=0$. So $f,g$ are strongly transverse. Also $f,g$ are strongly transverse if $f,g$ are smooth with $g$ a submersion, or if $f,g$ are transverse and $\pd Z=\es$. \[mc6def2\]
In the situation of Theorem \[mc6thm1\] we have a Cartesian square
@R=10pt[ W \_[\_Y]{} \^[\_X]{} & Y \_g\
X \^f & Z,]{}
&\
&\
&
@R=10pt[ \_[i0]{}C\_i(W) \_[C(\_Y)]{} \^[C(\_X)]{} & \_[k0]{}C\_k(Y) \_[C(g)]{}\
\_[j0]{}C\_j(X) \^[C(f)]{} & \_[l0]{}C\_l(Z).]{}
\[mc6eq8\] Since as in Theorem \[mc4thm\] the transformation $X\mapsto
\coprod_{i\ge 0}C_i(X)$, $f\mapsto C(f)$ has very good functorial properties, it is natural to wonder whether the right hand square in is also Cartesian. The answer is yes if and only if $f,g$ are strongly transverse. The following theorem will be proved in §\[mc9\].
Let $X,Y,Z$ be manifolds with corners, and $f:X\ra
Z,$ $g:Y\ra Z$ be strongly transverse smooth maps, and write $W$ for the fibre product $X\t_{f,Z,g}Y$ in Theorem [\[mc6thm1\]]{}. Then there is a canonical diffeomorphism
C\_i(W)\_
[l]{}j,k,l0:\
i=j+k-l
(C\_j(X)C(f)\^[-1]{}(C\_l(Z))) \_[C(f),C\_l(Z),C(g)]{}&\
(C\_k(Y)C(g)\^[-1]{}(C\_l(Z)))&
\[mc6eq9\] for all $i\ge 0,$ where the fibre products are all transverse and so exist. Hence \_[i0]{}C\_i(W)\_[j0]{}C\_j(X)\_[C(f),\_[l0]{}C\_l(Z),C(g)]{} \_[k0]{}C\_k(Y). \[mc6eq10\] Here the right hand commutative square in induces a map from the left hand side of to the right hand side, which gives the identification . \[mc6thm2\]
Suppose $f:X\ra Z$ and $g:Y\ra Z$ are transverse, but not strongly transverse. Then by Definition \[mc6def2\] there exist points in $C_j(X),C_k(Y),C_l(Z)$ satisfying with $j+k=l$ but $j,k,l$ not all zero. These give a point in the right hand side of with $i=0$ which does not lie in the image of $C_0(W)$ under the natural map, since $C_0(W)$ maps to $C_0(X),C_0(Y)$ and so cannot map to $C_j(X),C_k(Y)$ as $j,k$ are not both zero. Thus and are false if $f,g$ are transverse but not strongly transverse.
Here is an example of $f,g$ which are transverse but not strongly transverse.
Define smooth maps $f:[0,\iy)\ra[0,\iy)^2$ by $f(x)=(x,2x)$ and $g:[0,\iy)\ra[0,\iy)^2$ by $g(y)=(2y,y)$. Then $f(0)=g(0)=(0,0)$. We have $$\begin{gathered}
\d f\vert_0\bigl(T_0[0,\iy)\bigr)+\d g\vert_0\bigl(T_0[0,\iy)\bigr)=
\an{(1,2)}_\R+\an{(2,1)}_\R=\R^2=T_{(0,0)}[0,\iy)^2,\\
\d f\vert_0\bigl(T_0\bigl(S^0([0,\iy))\bigr)\bigr) +\d
g\vert_0\bigl(T_0\bigl(S^0([0,\iy))\bigr)\bigr)=\{0\}=
T_{(0,0)}\bigl(S^0([0,\iy)^2)\bigr),\end{gathered}$$ so $f,g$ are transverse. However we have $$C(f)\bigl(0,\bigl\{\{x=0\}\bigr\}\bigr)=C(g)\bigl(0,\bigl\{\{y=0\}
\bigr\})=\bigl((0,0),\bigl\{\{x=0\},\{y=0\}\bigr\}\bigr),$$ with $j=k=1$ and $l=2$, so $f,g$ are not strongly transverse. The fibre product $W=[0,\iy)_{f,[0,\iy)^2,g}[0,\iy)$ is a single point $\{0\}$. In when $i=0$ the l.h.s. is one point, and the r.h.s. is two points, one from $j\!=\!k\!=\!l\!=\!0$ and one from $j\!=\!k\!=\!1$, $l\!=\!2$, so does not hold. For $i\!\ne\! 0$, both sides of are empty. \[mc6ex2\]
The distinction between transversality and strong transversality will be important in [@Joyc]. There we will define a 2-category $\dManc$ of [*d-manifolds with corners*]{}, a ‘derived’ generalization of manifolds with corners, which contains the 1-category $\Manc$ of manifolds with corners as a full discrete 2-subcategory. If $X,Y,Z$ are manifolds with corners and $f:X\ra Z$, $g:Y\ra Z$ are smooth then a 2-category fibre product $(X\t_ZY)_\dManc$ exists in $\dManc$. If $f,g$ are transverse then a 1-category fibre product $(X\t_ZY)_\Manc$ also exists in $\Manc\subset\dManc$ by Theorem \[mc6thm1\]. However, $(X\t_ZY)_\dManc$ and $(X\t_ZY)_\Manc$ coincide if and only if $f,g$ are strongly transverse.
Orientations and orientation conventions {#mc7}
========================================
Orientations are discussed in [@KoNo §I.1] and [@Lang §VIII.3].
Let $X$ be an $n$-manifold and $E\ra X$ a vector bundle of rank $k$. The [*frame bundle*]{} $F(E)$ is $$F(E)=\bigl\{(x,e_1,\ldots,e_k):\text{$x\in X$, $(e_1,\ldots,e_k)$ is a
basis for $E\vert_x\cong\R^k$}\bigr\}.$$ It is a manifold of dimension $n+k^2$. Define an action of $\GL(k,\R)$ on $F(E)$ by $(A_{ij})_{i,j=1}^k:(x,e_1,\ldots,e_k)
\mapsto\bigl(x,\sum_{j=1}^kA_{1j}e_j,\ldots,\sum_jA_{kj}e_j\bigr)$. This action is smooth and free, and makes $F(E)$ into a principal $\GL(k,\R)$-bundle over $X$, with projection $\pi:F(E)\ra X$ given by $\pi:(x,e_1,\ldots,e_k)\mapsto x$.
Write $\GL_+(k,\R)$ for the subgroup of $A\in\GL(k,\R)$ with $\det
A>0$. It is a normal subgroup of $\GL(k,\R)$ of index 2, and we identify the quotient subgroup $\GL(k,\R)/\GL_+(k,\R)$ with $\{\pm
1\}$ by $A\GL_+(k,\R)\mapsto\det A/\md{\det A}$. The [*orientation bundle*]{} $\Or(E)$ of $E$ is $F(E)/\GL_+(k,\R)$. It is a principal $\GL(k,\R)/\GL_+(k,\R)=\{\pm 1\}$-bundle over $X$. Points of the fibre of $\Or(E)$ over $x\in X$ are equivalence classes of bases $(e_1,\ldots,e_k)$ for $E\vert_x$, where two bases are equivalent if they are related by a $k\t k$ matrix with positive determinant.
An [*orientation $o_E$ for the fibres of*]{} $E$ is a continuous section $o_E:X\ra\Or(E)$ of $\Or(E)$. The pair $(E,o_E)$ is called an [*oriented vector bundle*]{} on $X$. If $E\ra X$, $F\ra X$ are vector bundles on $X$ of ranks $k,l$ and $o_E,o_F$ are orientations on the fibres of $E,F$, we define the [*direct sum orientation*]{} $o_{E\op F}=o_E\op o_F$ on the fibres of $E\op F$ by saying that if $x\in X$, $(e_1,\ldots,e_k)$ is an oriented basis for $E\vert_x$ and $(f_1,\ldots,f_l)$ is an oriented basis for $F\vert_x$, then $(e_1,\ldots,e_k,f_1,\ldots,f_l)$ is an oriented basis for $(E\op
F)\vert_x$.
An [*orientation $o_X$ for*]{} $X$ is an orientation for the fibres of the tangent bundle $TX\ra X$. An [*oriented manifold*]{} $(X,o_X)$ is a manifold $X$ with an orientation $o_X$. Usually we leave $o_X$ implicit, and call $X$ an oriented manifold. If $o_X$ is an orientation on $X$ then the [*opposite orientation*]{} on $X$ is $-o_X$, where $o_X:X\ra\Or(TX)$ is a section, $-1:\Or(TX)\ra\Or(TX)$ comes from the principal $\{\pm 1\}$-action on $\Or(TX)$, and $-o_X=-1\ci o_X$ is the composition. When $X$ is an oriented manifold, we write $-X$ for $X$ with the opposite orientation. \[mc7def1\]
We shall consider issues to do with orientations on manifolds with corners, and orientations on fibre products of manifolds. To do this, we need [*orientation conventions*]{} to say how to orient boundaries $\pd X$ and fibre products $X\t_ZY$ of oriented manifolds $X,Y,Z$. Our conventions generalize those of Fukaya, Oh, Ohta and Ono [@FOOO Conv. 45.1], who restrict to $f,g$ submersions.
[**(a)**]{} Let $(X,o_X)$ be an oriented manifold with corners. Define $o_{\pd X}$ to be the unique orientation on $\pd X$ such that i\_X\^\*(TX)\_[out]{}T(X) \[mc7eq1\] is an isomorphism of oriented vector bundles over $\pd X$, where $i_X^*(TX),T(\pd X)$ are oriented by $o_X,o_{\pd X}$, and $\R_{\rm
out}$ is oriented by an outward-pointing normal vector to $\pd X$ in $X$, and the r.h.s. of has the direct sum orientation.
[**(b)**]{} Let $(X,o_X),(Y,o_Y),(Z,o_Z)$ be oriented manifolds with corners, and $f:X\ra Z$, $g:Y\ra Z$ be transverse smooth maps, so that a fibre product $W=X\t_ZY$ exists in $\Manc$ by Theorem \[mc6thm1\]. Then we have an exact sequence of vector bundles over $W$ @C=8.5pt[0 & TW \^(0.35)[\_X\_Y]{} &&& \_X\^\*(TX)\_Y\^\*(TY) \^(0.55)[\_X\^\*(f)-\_Y\^\*(g)]{} &&&& (f\_X)\^\*(TZ) & 0.]{} \[mc7eq2\] Choosing a splitting of induces an isomorphism of vector bundles TW (f\_X)\^\*(TZ)\_X\^\*(TX)\_Y\^\*(TY). \[mc7eq3\] Define $o_W$ to be the unique orientation on $W$ such that the direct sum orientations in induced by $o_W,o_Z,o_X,o_Y$ differ by a factor $(-1)^{\dim Y\dim Z}$.
Here are two was to rewrite this convention in special cases. Firstly, suppose $f$ is a submersion. Then $\d f:TX\ra f^*(TZ)$ is surjective, so by splitting the exact sequence $0\ra\Ker\d f\ra
TX\,{\buildrel\d f\over\longra}\,f^*(TZ)\ra 0$ we obtain an isomorphism TXff\^\*(TZ). \[mc7eq4\] Give the vector bundle $\Ker\d f\ra X$ the unique orientation such that is an isomorphism of oriented vector bundles, where $TX,f^*(TZ)$ are oriented using $o_X,o_Z$. As $f:X\ra Z$ is a submersion so is $\pi_Y:W\ra Y$, and $\d\pi_X$ induces an isomorphism $\Ker(\d\pi_Y)\ra\pi_X^*(\Ker\d f)$. Thus we have an exact sequence $$\xymatrix@C=15pt{0 \ar[r] & \pi_X^*(\Ker\d f)
\ar[rr]^(0.6){(\d\pi_X)^{-1}} && TW \ar[rr]^(0.45){\d\pi_Y} && \pi_Y^*(TY)
\ar[r] & 0.}$$ Splitting this gives an isomorphism TW\_X\^\*(f) \_Y\^\*(TY). \[mc7eq5\] The orientation on $W$ makes into an isomorphism of oriented vector bundles, using $o_Y$ and the orientation on $\Ker\d
f$ to orient the right hand side.
Secondly, let $g$ be a submersion. Then as for – we have isomorphisms TYg\^\*(TZ)gTW\_X\^\*(TX)\_Y\^\*(g). \[mc7eq6\] We use the first equation of to define an orientation on the fibres of $\Ker\d g$, and the second to define an orientation on $W$. \[mc7conv\]
If $X$ is an oriented manifold with corners then by induction Convention \[mc7conv\](a) gives orientations on $\pd^kX$ for all $k=0,1,2,\ldots$. Now Definition \[mc2def6\] defined a smooth, free action of $S_k$ on $\pd^kX$ for each $k$. By considering local models $\R^n_l$ it is easy to see that the action of each $\si\in
S_k$ multiplies the orientation on $\pd^kX$ by $\sign(\si)=\pm 1$. Since $C_k(X)\cong\pd^kX/S_k$ by and $S_k$ does not preserve orientations for $k\ge 2$, we see that $C_k(X)$ [*does not have a natural orientation for*]{} $k\ge 2$. We show by example that $C_k(X)$ need not even be orientable.
Let $X$ be the 4-manifold with corners $\bigl({\cal
S}^2\t[0,\iy)^2\bigr)/\Z_2$, where $${\cal S}^2\t[0,\iy)^2=\bigl\{(x_2,x_2,x_3,y_1,y_2):\text{$x_j,y_j\in\R$,
$x_1^2+x_2^2+x_3^2=1$, $y_1,y_2\ge 0$}\bigr\},$$ and $\Z_2=\an{\si}$ acts freely on $X$ by $$\si:(x_2,x_2,x_3,y_1,y_2)\mapsto(-x_1,-x_2,-x_3,y_2,y_1).$$ There is an orientation on ${\cal S}^2\t[0,\iy)^2$ which is invariant under $\Z_2$, and so descends to $X$. We have diffeomorphisms $$\pd X\cong C_1(X)\cong {\cal S}^2\t[0,\iy),\quad \pd^2X\cong {\cal S}^2,
\quad C_2(X)\cong \RP^2,$$ and $\pd^kX\!=\!C_k(X)\!=\!\es$ for $k\!>\!2$. Thus $X$ is oriented, but $C_2(X)$ is not orientable. \[mc7ex\]
Given any canonical diffeomorphism between expressions involving boundaries and fibre products of oriented manifolds with corners, we can use Convention \[mc7conv\] to define orientations on each side. These will be related by some sign $\pm 1$, which we can try to compute. Here is how to add signs to –.
In Propositions [\[mc6prop1\], \[mc6prop2\]]{} and [\[mc6prop3\],]{} suppose $X,Y,Z$ are oriented. Then in oriented manifolds, equations – respectively become \_-\^fX&(-1)\^[X+Y]{}X\_[f,Y,i\_Y]{}Y, \[mc7eq7\]\
(X\_[f,Z,g]{}Y)&(X\_[fi\_X,Z,g]{}Y) (-1)\^[X+Z]{}(X\_[f,Z,gi\_Y]{}Y), \[mc7eq8\]\
(X\_[f,Z,g]{}Y)&(\_+\^fX \_[f\_+,Z,g]{}Y) (-1)\^[X+Z]{}(X\_[f,Z,gi\_Y]{}Y), \[mc7eq9\]\
(X\_[f,Z,g]{}Y)&(\_+\^fX \_[f\_+,Z,g]{}Y) (-1)\^[X+Z]{}(X\_[f,Z,g\_+]{}\_+\^gY)\
&(\_-\^fX\_[f\_-,Z,g\_-]{}\_-\^gY).
\[mc7eq10\] \[mc7prop1\]
Here are some more identities involving only fibre products:
[**(a)**]{} If $f:X\ra Z,$ $g:Y\ra Z$ are transverse smooth maps of oriented manifolds with corners then in oriented manifolds we have X\_[f,Z,g]{}Y(-1)\^[(X-Z)(Y-Z)]{}Y\_[g,Z,f]{}X. \[mc7eq11\]
[**(b)**]{} If $d:V\ra Y,$ $e:W\ra Y,$ $f:W\ra Z,$ $g:X\ra
Z$ are smooth maps of oriented manifolds with corners then in oriented manifolds we have V\_[d,Y,e\_W]{}(W\_[f,Z,g]{}X)(V\_[d,Y,e]{}W)\_[f\_W,Z,g]{}X, \[mc7eq12\] provided all four fibre products are transverse.
[**(c)**]{} If $d:V\ra Y,$ $e:V\ra Z,$ $f:W\ra Y,$ $g:X\ra
Z$ are smooth maps of oriented manifolds with corners then in oriented manifolds we have
&V\_[(d,e),YZ,fg]{}(WX)\
&(-1)\^[Z(Y+W)]{} (V\_[d,Y,f]{}W)\_[e\_V,Z,g]{}X,
\[mc7eq13\] provided all three fibre products are transverse. \[mc7prop2\]
[**(i)**]{} Equations , and can be found in Fukaya et al. [@FOOO Lem. 45.3] for the case of Kuranishi spaces.
[**(ii)**]{} The proofs of Propositions \[mc7prop1\] and \[mc7prop2\] are elementary calculations starting from Convention \[mc7conv\]. Here is a way to make these calculations easier. For simplicity, assume all the smooth maps involved are submersions. By Proposition \[mc5prop1\], submersions are locally projections. Since identities like – are local, it is enough to prove the identities for projections.
Let $M,N,Z$ be oriented manifolds with corners, of dimensions $m,n,z$. Set $X=M\t Z$ and $Y=Z\t N$, with the product orientations, and define $f:X\ra Z$, $g:Y\ra Z$ by $f=\pi_Z=g$. Convention \[mc7conv\](b) is arranged so that $W\cong M\t Z\t N$ holds in oriented manifolds. Exchanging the order in a product of oriented manifolds yields $X\t Y\cong(-1)^{\dim X\dim Y}Y\t X$. Thus to compute the sign in , for instance, note that $$\begin{aligned}
&X\t_{f,Z,g}Y=(M\t Z)\t_Z(Z\t N)\cong M\t Z\t N,\\
&Y\t_{g,Z,f}X=(Z\t N)\t_Z(M\t Z)\cong(-1)^{(m+n)z}
(N\t Z)\t_Z(Z\t M)\\
&\quad\cong(-1)^{(m+n)z}N\t Z\t M\cong(-1)^{(m+n)z}(-1)^{mn+mz+nz}M\t Z\t N,\end{aligned}$$ and then substitute in $m=\dim X-\dim Z$, $n=\dim Y-\dim Z$.
[**(iii)**]{} Ramshaw and Basch [@RaBa] prove that there is a [*unique*]{} orientation convention for transverse fibre products of manifolds without boundary satisfying the three conditions: (A) if $X,Y$ are oriented then $X\t_{\{0\}}Y\cong X\t Y$ in oriented manifolds, where $X\t Y$ has the product orientation from $T(X\t Y)\cong \pi_X^*(TX)\op \pi_Y^*(TY)$; (B) if $f:X\ra Y$ is a smooth map of oriented manifolds then $X\cong Y\t_{\id_Y,Y,f}X$ in oriented manifolds; and (C) equation holds. Convention \[mc7conv\](b) satisfies (A)–(C), and so agrees with that of [@RaBa]. \[mc7rem\]
Proof of Theorem \[mc6thm1\] {#mc8}
============================
Theorem \[mc6thm1\] follows from the next two propositions. In the proof we assume Theorem \[mc6thm1\] for manifolds without boundary, since this is well known, as in Lang [@Lang Prop. II.4] for instance. The difference between [*transverse*]{} and [*strongly transverse*]{} $f,g$ in §\[mc6\] appears in part (C) in the proof below: transverse $f,g$ are strongly transverse if and only if there are no $\sim$-equivalence classes $E$ of type (b).
Let $X,Y,Z$ be manifolds with corners, and $f:X\ra Z,$ $g:Y\ra Z$ be transverse smooth maps. Then the construction of Theorem [\[mc6thm1\]]{} defines a manifold with corners $W,$ with $\dim W=\dim X+\dim Y-\dim Z$ if $W\ne\es,$ and the maps $\pi_X:W\ra X,$ $\pi_Y:W\ra Y$ are smooth. \[mc8prop1\]
If $W=\es$ the proposition is trivial, so suppose $W\ne\es$. Then $n=\dim X+\dim Y-\dim Z\ge 0$, since $f,g$ are transverse. Let $(x,y)\in W$, so that $x\in X$ and $y\in Y$ with $f(x)=g(y)=z$ in $Z$. We will first construct a chart $(U,\phi)$ on $W$ satisfying the conditions of Theorem \[mc6thm1\], with $U$ open in $\R^n_d$ and $0\in U$ with $\phi(0)=(x,y)$.
Choose charts $(R,\th),(S,\psi),(T,\xi)$ on $X,Y,Z$ respectively with $0\in R,S,T$ and $\th(0)=x$, $\psi(0)=y$, $\xi(0)=z$, where $R,S,T$ are open in $\R^k_a,\R^l_b,\R^m_c$ with $k=\dim X$, $l=\dim
Y$, $m=\dim Z$. Making $R,S$ smaller if necessary suppose $f\ci\th(R),g\ci\psi(S)\subseteq\xi(T)$. Then $\ti f=\xi^{-1}\ci
f\ci\th:R\ra T$ and $\ti g=\xi^{-1}\ci g\ci\psi:S\ra T$ are smooth maps between subsets of $\R^k,\R^l,\R^m$ in the sense of Definition \[mc2def1\]. So by definition we can choose open subsets $\hat
R\subseteq\R^k$, $\hat S\subseteq\R^l$, $\hat T\subseteq\R^m$ with $R=\hat R\cap\R^k_a$, $S=\hat S\cap\R^l_b$, $T=\hat T\cap \R^m_c$ and smooth maps $\hat f:\hat R\ra\hat T$, $\hat g:\hat S\ra\hat T$ with $\hat f\vert_R=\ti f$, $\hat g\vert_S=\ti g$.
Now $f,g$ are transverse, so $\ti f,\ti g$ are transverse on $R,S$, and as this is an open condition, by making $\hat R,\hat S$ smaller if necessary we can make $\hat f,\hat g$ transverse. Since $\hat
f:\hat R\ra\hat T$, $\hat g:\hat S\ra\hat T$ are transverse smooth maps of manifolds without boundary, by [@Lang Prop. II.4] the fibre product $\hat V=\hat R\t_{\hat f,\hat T,\hat g}\hat S$ exists as an $n$-manifold without boundary, and it is also easy to show that charts on $\hat V$ are characterized by the injectivity of . Define $V=\bigl\{(r,s)\in\hat V:r\in R$, $s\in
S\bigr\}$. We will show that near $(0,0)\in V$, the embedding of $V$ in $\hat V$ is modelled on the inclusion of $\R^n_d$ in $\R^n$, so that $V$ is a manifold with corners.
The local boundary components of $R\subseteq\R^k_a$ at 0 are $\{r_i=0\}$ for $i=1,\ldots,a$, where $(r_1,\ldots,r_k)$ are the coordinates on $R$ and $\hat R$. Write $\be_i=\th_*(\{r_i=0\})$ for the corresponding local boundary component of $X$ at $x$. Then $\be_1,\ldots,\be_a$ are the local boundary components of $X$ at $x$. Similarly, write $(s_1,\ldots,s_l)$ for coordinates on $S,\hat
S$ and $\ti\be_1,\ldots,\ti\be_b$ for the local boundary components of $Y$ at $y$, where $\ti\be_i=\psi_*(\{s_i=0\})$, and $(t_1,\ldots,t_m)$ for coordinates on $T,\hat T$ and $\dot\be_1,\ldots,\dot\be_c$ for the local boundary components of $Z$ at $z$, where $\dot\be_i=\xi_*(\{t_i=0\})$.
Define subsets $P^f,P^g\subseteq\{1,\ldots,c\}$ by $P^f=\bigl\{i:(x,(z,\dot\be_i))\in\Xi_-^f\bigr\}$ and $P^g=\bigl\{i:(y,(z,\dot\be_i))\in\Xi_-^g\bigr\}$. Define maps $\Pi^f:P^f\ra\{1,\ldots,a\}$ and $\Pi^g:P^g\ra\{1,\ldots,b\}$ by $\Pi^f(i)=j$ if $\xi_-^f\bigl(x,(z,\dot\be_i)\bigr)=(x,\be_j)$ and $\Pi^g(i)=j$ if $\xi_-^g\bigl(y,(z,\dot\be_i)\bigr)=(y,\ti\be_j)$. We can express the maps $C(f),C(g)$ of over $x,y,z$ as follows: if $A\subseteq\{1,\ldots,a\}$ and $B\subseteq\{1,\ldots,b\}$ then
C(f):(x,{\_i:iA})&(z,{\_j:jP\^f, \^f(j)A}),\
C(g):(y,{\_i:iB})&(z,{\_j:jP\^g, \^g(j)B}).
\[mc8eq1\]
Lemma \[mc6lem\] on $C(f),C(g)$ over $x,y,z$, which uses $f,g$ transverse, then turns out to be equivalent to the following conditions on $P^f,P^g,\Pi^f,\Pi^g$:
- $\Pi^f(P^f\cap P^g)\cap
\Pi^f(P^f\sm P^g)=\es$ and $\Pi^g(P^f\cap P^g)\cap \Pi^g(P^g\sm
P^f)=\es$.
- $\Pi^f\vert_{P^f\sm P^g}\!:\!P^f\!\sm\! P^g\!\ra\!\{1,
\ldots,a\}$, $\Pi^g\vert_{P^g\sm P^f}\!:\!P^g\!\sm\!P^f\!\ra\!
\{1,\ldots,b\}$ are injective.
- Let $\approx$ be the equivalence relation on $P^f\cap P^g$ generated by $i\approx j$ if $\Pi^f(i)=\Pi^f(j)$ or $\Pi^f(i)=\Pi^f(j)$. Then for each $\approx$-equivalence class $E\subseteq P^f\cap P^g$ we have either
- $\md{\Pi^f(E)}+\md{\Pi^g(E)}=\md{E}+1$, or
- $\md{\Pi^f(E)}+\md{\Pi^g(E)}=\md{E}$.
Here it is automatic that $\md{\Pi^f(E)}+\md{\Pi^g(E)}\le\md{E}+1$, and Lemma \[mc6lem\] implies that $\md{\Pi^f(E)}+\md{\Pi^g(E)} \ge\md{E}$. The number of equivalence classes of type (a) is $\md{\Pi^f(P^f\cap P^g)}+
\md{\Pi^g(P^f\cap P^g)}-\md{P^f\cap P^g}$.
Also, if $i\in\{1,\ldots,c\}\sm (P^f\cup P^g)$ then $\bigl(x,(z,\dot\be_i)\bigr)\in\Xi^f_+$ and $\bigl(y,(z,\dot\be_i)\bigr)\in\Xi^g_+$. These imply that $\d
f\vert_x(T_xX),\d g\vert_y(T_yY)\subseteq T_z\dot\be_i$, so that $\d
f\vert_x(T_xX)+\d g\vert_y(T_yY)\subseteq T_z\dot\be_i\subsetneq
T_zZ$, contradicting $f,g$ transverse. This proves:
- $P^f\cup P^g=\{1,\ldots,c\}$.
Now $V$ is cut out in $\hat V$ by $r_i\ge 0$, $i\le a$, and $s_i\ge
0$, $i\le b$, that is
V={((r\_1,…,r\_k),(s\_1,…,s\_l))V:&\
&}.
\[mc8eq2\] We claim that making $R,S,\hat R,\hat S$ smaller if necessary, the following hold:
- If $i\in P^f\sm P^g$ then the inequality $r_{\Pi^f(i)}\ge
0$ does not change $V$, and can be omitted in .
- If $i\in P^g\sm P^f$ then the inequality $s_{\Pi^g(i)}\ge 0$ does not change $V$, and can be omitted in .
- If $i,j\in P^f\cap P^g$ with $i\approx j$, then the four inequalities $r_{\Pi^f(i)}\ge 0$, $r_{\Pi^f(j)}\ge 0$, $s_{\Pi^g(i)}\ge 0$, $s_{\Pi^g(j)}\ge 0$ have the same effect in $\hat V$. Thus for each $\approx$-equivalence class $E$ in $P^f\cap P^g$, it is sufficient to impose only one of the $2\md{E}$ inequalities $r_{\Pi^f(i)}\ge 0$, $s_{\Pi^g(i)}\ge 0$ in for $i\in E$ to define $V$.
- If $E$ is an $\approx$-equivalence class of type (b) in (C) above, then $r_{\smash{\Pi^f(i)}}\equiv s_{\Pi^g(i)}\equiv
0$ in $\hat V$ for all $i\in E$. Thus we can omit the inequalities $r_{\Pi^f(i)}\ge 0$, $s_{\Pi^g(i)}\ge 0$ in for $i\in E$ to define $V$.
Here we mean that all of the inequalities $r_i\ge 0$, $s_i\ge 0$ which (i)–(iv) allow us to omit may all be omitted simultaneously without changing $V$.
To prove (i), note that as $\bigl(\xi(T),t_i\ci\xi^{-1}\bigr)$ is a boundary defining function for $Z$ at $(z,\dot\be_i)$, and $\bigl(\th(R),r_{\Pi^f(i)}\ci\th^{-1}\bigr)$ is a boundary defining function for $X$ at $(x,\be_{\Pi^f(i)})$, and $\xi_-^f\bigl(x,(z,\dot\be_i)\bigr)=(x,\be_{\Pi^f(i)})$, Definition \[mc3def1\] and Proposition \[mc2prop6\](b) imply that $t_i\ci\xi^{-1}\ci f\equiv (r_{\Pi^f(i)}\ci\th^{-1})\cdot G$ on $\th(T)$ near $x$ for some smooth $G:\th(T)\ra(0,\iy)$. Hence $t_i\ci\hat f\equiv r_{\Pi^f(i)}\cdot\hat G$ on $\hat R$ near 0 for some smooth $\hat G:\hat R\ra(0,\iy)$ defined near 0. Therefore making $R,\hat R$ smaller if necessary, $r_{\Pi^f(i)}\ge 0$ is equivalent to $t_i\ci\hat f\ci\pi_{\hat R}\ge 0$ on $\hat V$. But $t_i\ci\hat f\ci\pi_{\hat R}\!=\!t_i\ci\hat g\ci\pi_{\hat S}$, and $t_i\!\ge\! 0$ on $T$, so $t_i\ci\hat g\!\ge\! 0$ on $S$ as $\hat g$ maps $S\!\ra\! T$.
Hence the inequality $r_{\smash{\Pi^f(i)}}\ge 0$ is unnecessary in provided we restrict to $S$ in $\hat S$, that is, provided we impose all the conditions $s_i\ge 0$. In fact we need more than this: we must also be able to omit conditions $s_{\Pi^g(j)}\ge 0$ when this is allowed by (ii)–(iv). This is possible because the $s_{\Pi^g(j)}\ge 0$ omitted in (ii)–(iv) correspond to different conditions $t_j\ge 0$ in $\hat T$ than the condition $t_i\ge 0$ we are considering, since (i) deals with $i\in
P^f\sm P^g$, (ii) with $j\in P^g\sm P^f$, and (iii)–(iv) with $j\in
P^f\cap P^g$, which are disjoint sets. This proves (i), and also that we can omit $r_{\Pi^f(i)}\ge 0$ in (i) independently of other omissions in (i)–(iv). The proof for (ii) is the same.
For (iii), if $i\in P^f\cap P^g$ then by Definition \[mc3def1\] and Proposition \[mc2prop6\](b) as above we see that making $R,\hat R,S,\hat S$ smaller if necessary, $r_{\smash{\Pi^f(i)}}\ge
0$ is equivalent to $t_i\ci\hat f\ci\pi_{\hat R}\ge 0$ on $\hat V$, which is also equivalent to $s_{\smash{\Pi^g(i)}}\ge 0$ on $\hat V$. Suppose $i,j\in P^f\cap P^g$ with $\Pi^f(i)=\Pi^f(j)$. Then the conditions $r_{\smash{\Pi^f(i)}}\ge 0$, $r_{\smash{\Pi^f(j)}}\ge 0$ are the same, and are equivalent to $s_{\smash{\Pi^g(i)}}\ge 0$ and $s_{\smash{\Pi^g(j)}}\ge 0$. Similarly, if $\Pi^g(i)=\Pi^g(j)$ then the conditions $r_{\smash{\Pi^f(i)}}\ge 0$, $r_{\smash{\Pi^f(j)}}\ge
0$, $s_{\smash{\Pi^g(i)}}\ge 0$ and $s_{\smash{\Pi^g(j)}}\ge 0$ are all equivalent. Since these two cases generate $\approx$, part (iii) follows.
For (iv), let $E$ be an $\approx$-equivalence class of type (b). Define submanifolds $\hat R_E,\hat S_E,\hat T_E$ in $\hat R,\hat
S,\hat T$ by $$\begin{aligned}
\hat R_E&=\bigl\{(r_1,\ldots,r_k)\in\hat R:r_{\smash{\Pi^f(i)}}=0,\;
i\in E\bigr\}, \\
\hat S_E&=\bigl\{(s_1,\ldots,s_l)\in\hat S:s_{\smash{\Pi^g(i)}}=0,\;
i\in E\bigr\}, \\
\hat T_E&=\bigl\{(t_1,\ldots,t_m)\in\hat T:t_i=0,\; i\in E\bigr\}.\end{aligned}$$ Making $R,S,\hat R,\hat S$ smaller if necessary, $\hat f$ maps $\hat
R_E\ra \hat T_E$ and $\hat g$ maps $\hat S_E\ra\hat T_E$. As $\hat
f,\hat g$ are transverse, an argument similar to Lemma \[mc6lem\] shows that $\hat f:\hat R_E\ra\hat T_E$ and $\hat g:\hat S_E\ra\hat
T_E$ are transverse, so the fibre product $\hat R_E\t_{\hat T_E}\hat
S_E$ exists as a manifold, and is a submanifold of $\hat V=\hat
R\t_{\hat T}\hat S$. Now $$\dim \hat R_E\!\t_{\hat T_E}\hat
S_E\!=\!(k\!-\!\md{\Pi^f(E)})\!+\!(l\!-\!\md{\Pi^g(E)})\!-\!
(m\!-\!\md{E})=k\!+\!l\!-\!m\!=\!\dim\hat V,$$ since $E$ is of type (b). Thus $\hat R_E\t_{\hat T_E}\hat S_E$ is open in $\hat V$, as they are of the same dimension, and contains $(0,0)$. So making $R,S,\hat R,\hat S$ smaller if necessary, we have $\hat R_E\t_{\hat T_E}\hat S_E=\hat V$, proving (iv). There are no further issues about simultaneous omissions in (i)–(iv).
Choose a subset $Q\subseteq P^f\cap P^g$ such that $Q$ contains exactly one element of each $\approx$-equivalence class of type (a) in $P^f\cap P^g$, and no elements of $\approx$-equivalence classes of type (b) . Then and (i)–(iv) above imply that
V={((r\_1,&…,r\_k),(s\_1,…,s\_l))V: r\_i0, i{1,…,a}\^f(P\^f),\
&s\_i0, i{1,…,b}\^g(P\^g), r\_[\^f(i)]{}0, iQ}.
\[mc8eq3\] For the first condition $r_i\ge 0$, $i\in\{1,\ldots,a\}
\sm\Pi^f(P^f)$ in , there are $$a-\bmd{\Pi^f(P^f)}=a-\bmd{\Pi^f(P^f\cap P^g)}-\bmd{\Pi^f(P^f\sm
P^g)}=a-\bmd{\Pi^f(P^f\cap P^g)}-\bmd{P^f\sm P^g}$$ inequalities, using (A) above in the first step and (B) in the second. Similarly, for the second condition $s_i\ge 0$, $i\in\{1,\ldots,b\} \sm\Pi^g(P^g)$ there are $b-\bmd{\Pi^g(P^f\cap
P^g)}-\bmd{P^g\sm P^f}$ inequalities. For the third there are $\md{Q}=\md{\Pi^f(P^f\cap P^g)}+\md{\Pi^g(P^f\cap P^g)}-\md{P^f\cap
P^g}$ inequalities by (C) above. Hence in total there are
(&a--)+ (b--)\
&+(+-)\
&=a+b---=a+b-=a+b-c
\[mc8eq4\] inequalities $r_i\ge 0$, $s_i\ge 0$ appearing in , using (D) at the last step.
Define a vector subspace $L$ of $T_{(0,0)}\hat V$ by
L&={((r\_1,…,r\_k),(s\_1,…,s\_l))T\_[(0,0)]{}V:}.
\[mc8eq5\] That is, we replace each inequality $r_i\ge 0$, $s_i\ge 0$ in by $r_i=0$, $s_i=0$. By the proof of the equivalence of and using (i)–(iv), we see that
L&={((r\_1,…,r\_k),(s\_1,…,s\_l))T\_[(0,0)]{}V: r\_[\^f(i)]{}=0 , iQ,\
&r\_i=0, i{1,…,a}\^f(P\^f), s\_i=0, i{1,…,b}\^g(P\^g)},
\[mc8eq6\] replacing inequalities $r_i\ge 0$, $s_i\ge 0$ in by $r_i=0$, $s_i=0$.
Now $T_z(S^c(Z))=\d f\vert_x(T_x(S^a(X)))+\d g\vert_y(T_y(S^b(Y)))$ by Definition \[mc6def1\]. As there is a natural isomorphism $$L\cong \bigl\{u\op v\in T_x(S^a(X))\op T_y(S^b(Y)): \d f\vert_x(u)=
\d g\vert_y(v)\bigr\},$$ we see that
L&=S\^a(x)+S\^b(Y)-S\^c(Z)\
&=(p-a)+(q-b)-(r-c)=V-(a+b-c).
\[mc8eq7\] Since by there are $a+b-c$ equalities $r_i=0$, $s_i=0$ in , equation implies that the conditions $r_i=0$, $s_i=0$ in are transverse, so the inequalities $r_i\ge 0$, $s_i\ge 0$ in are transverse at $(0,0)$ in $\hat V$. That is, the corresponding 1-forms $\d r_i\vert_{(0,0)}$, $\d s_i\vert_{(0,0)}$ are linearly independent at $T^*_{(0,0)}\hat
V$. This is an open condition in $\hat V$. Since $\hat V$ is a manifold without boundary of dimension $n$, it follows that $V$ is near $(0,0)$ a manifold with corners of dimension $n$, locally modelled on $\R^n_{a+b-c}$. Making $R,S$ smaller if necessary, $V$ becomes an $n$-manifold with corners.
Let $(U,\phi')$ be a chart on $V$, with $U$ open in $\R^n_{a+b-c}$ and $0\in U$ with $\phi'(0)=(0,0)$. Define $\phi:U\ra W=X\t_ZY$ by $\phi=(\th\t\psi)\ci\phi'$. Then $\phi$ is a homeomorphism with an open set in $W$, since $\phi':U\ra V$ is a homeomorphism with an open set in $V$ and $\th\t\psi:V\ra W$ is a homeomorphism with an open set in $W$. Also $\phi(0)=(x,y)$ as $\phi'(0)=(0,0)$ and $\th(0)=x$, $\psi(0)=y$. Thus $(U,\phi)$ is a chart on the topological space $W$ whose image contains $(x,y)$.
Now $(U,\phi')$ extends to a chart $(\hat U,\hat\phi')$ on $\hat V$. But $\hat V$ comes from Theorem \[mc6thm1\] for manifolds without boundary, and so $$\d(\pi_{\hat S}\ci\hat\phi')\vert_{\hat u}\op\d(\pi_{\hat T}\ci\hat
\phi')\vert_{\hat u}:T_{\hat u}\hat U\longra T_{\hat s}\hat S\op
T_{\hat t}\hat T$$ is injective for all $\hat u\in\hat U$ with $\hat\phi'(\hat u)=(\hat
s,\hat t)$. Restricting to $U$ shows that (\_S’)\_u(\_T’)\_u:T\_uUT\_sS T\_tT \[mc8eq8\] is injective for all $u\in U$ with $\phi'(u)=(s,t)$. But if $u\in U$ with $\phi'(u)=(s,t)$ and $\th(s)=x'$, $\psi(t)=y'$ then $\d(\pi_X\ci\phi)\vert_u=\d\th\vert_s\ci\d(\pi_S\ci\phi')\vert_u$ and $\d(\pi_Y\ci\phi)\vert_u=\d\psi\vert_t\ci\d(\pi_T\ci\phi')
\vert_u$, where $\d\th\vert_s:T_sS\ra T_{x'}X$ and $\d\psi\vert_t:T_tT\ra T_{y'}Y$ are isomorphisms as $(S,\th),(T,\psi)$ are charts on $X,Y$. So composing with $\d\th\vert_s\t \d\psi\vert_t$ shows that $$\d(\pi_X\ci\phi)\vert_u\op\d(\pi_Y\ci\phi)\vert_u:T_uU\longra
T_{x'}X\op T_{y'}Y$$ is injective, so $(U,\phi)$ satisfies the conditions of Theorem \[mc6thm1\].
We have now shown that $W$ can be covered by charts $(U,\phi)$ satisfying the conditions of Theorem \[mc6thm1\]. For such $(U,\phi)$, observe that actually maps
(\_X)\_[u]{}&(\_Y)\_[u]{}: T\_[u]{}U\^n\
&{(,)T\_xXT\_yY:f\_x()=g\_y() }.
\[mc8eq9\] Now the r.h.s. of has dimension $n$ by transversality of $f,g$, and is injective, so it is an isomorphism. Let $(U,\phi)$ and $(V,\psi)$ be two such charts, and $u\in U$, $v\in V$ with $\phi(u)=\psi(v)=(x,y)$. Since and its analogue for $(V,\psi)$ are isomorphisms, we see that $\psi^{-1} \ci\phi$ is differentiable at $u$ and its derivative is an isomorphism, the composition of with the inverse of its analogue for $\psi$. Using the same argument for all $u\in \phi^{-1}(\psi(V))$, we find that $\psi^{-1}\ci\phi:\phi^{-1}(\psi(V))\ra
\psi^{-1}(\phi(U))$ is a diffeomorphism, and so $(U,\phi),(V,\psi)$ are automatically compatible.
Therefore the collection of all charts on $W$ satisfying the conditions of Theorem \[mc6thm1\] is an atlas. But any chart compatible with all charts satisfying Theorem \[mc6thm1\] also satisfies the conditions of Theorem \[mc6thm1\], so this atlas is maximal. Also the topological space $W=X\t_ZY$ is paracompact and Hausdorff, since $X,Y,Z$ are paracompact and Hausdorff as they are manifolds. Hence the construction of Theorem \[mc6thm1\] does make $W$ into an $n$-manifold with corners.
It remains to show that $\pi_X:W\ra X,$ $\pi_Y:W\ra Y$ are smooth. They are clearly continuous, since $W$ was defined as the topological fibre product. Locally $\pi_X,\pi_Y$ are identified with $\pi_S:V\ra S$ and $\pi_T:V\ra T$ above, which are restrictions to $V$ of $\pi_{\hat S}:\hat V\ra\hat S$ and $\pi_{\hat T}:\hat
V\ra\hat T$. But $\hat V$ is a fibre product of manifolds without boundary, so $\pi_{\hat S},\pi_{\hat T}$ are smooth, which implies that $\pi_S,\pi_T$ are weakly smooth. To prove $\pi_S,\pi_T$ are smooth we note that $(S,s_i)$ for $i=1,\ldots,a$ are boundary defining functions on $S$ at $(s,\{s_i=0\})$, and $(T,t_j)$ for $j=1,\ldots,b$ are boundary defining functions on $T$ at $(t,\{t_j=0\})$, and we show using the discussion of (i)–(iv) that the pullbacks to $V$ satisfy the conditions of Definition \[mc3def1\]. As smoothness is a local condition, $\pi_X,\pi_Y$ are smooth.
In the situation of Proposition [\[mc8prop1\],]{} $W,\pi_X,\pi_Y$ are a fibre product $X\t_{f,Z,g}Y$ in $\Manc$. \[mc8prop2\]
By definition $f\ci\pi_X=g\ci\pi_Y$. Suppose $W'$ is a manifold with corners and $\pi_X':W'\ra X$, $\pi_Y':Y'\ra Y$ are smooth maps with $f\ci\pi_X'=g\ci\pi_Y'$. We must show there exists a unique smooth map $h:W'\ra W$ with $\pi_X'=\pi_X\ci h$ and $\pi_Y'=\pi_Y\ci h$. Since $W$ is a fibre product at the level of topological spaces, there is a unique continuous map $h:W'\ra W$ given by $h(w')=\bigl(\pi_X'(w'),\pi_Y'(w')\bigr)$ with $\pi_X'=\pi_X\ci h$ and $\pi_Y'=\pi_Y\ci h$. We must show $h$ is smooth.
Let $w'\in W'$, with $\pi'_X(w')=x\in X$ and $\pi'_Y(w')=y\in Y$, so that $f(x)=g(y)=z\in Z$. Let $(R,\th),(S,\psi),\ldots$ be as in the proof of Proposition \[mc8prop1\]. Then $\hat V=\hat R\t_{\hat
T}\hat S$ is a fibre product of manifolds without boundary, and $V=R\t_TS\subseteq\hat V$ is a manifold with corners, and $\th\t\psi:V\ra W$ is a diffeomorphism with an open subset of $W$. Let $U'$ be an open neighbourhood of $w'\in W'$ such that $\pi_X'(U')\subseteq\th(S)$ and $\pi_Y'(U')\subseteq\psi(T)$. Consider the map $\ti h=(\th\t\psi)^{-1}\ci h\vert_{U'}:U'\ra
V\subseteq\hat V$. We will show $\ti h$ is smooth. This implies $h\vert_{U'}=(\th\t\psi)\ci\ti h$ is smooth, so $h$ is smooth as this is a local condition.
As $\pi_{\hat S}\ci\ti h=\th^{-1}\ci\pi'_X$ and $\pi_{\hat T}\ci\ti
h=\psi^{-1}\ci\pi'_Y$ are smooth, and $\hat V=\hat R\t_{\hat T}\hat
S$ is a fibre product of manifolds, we see that $\ti h:U'\ra \hat V$ is smooth, and therefore $\ti h:U'\ra V$ is weakly smooth. To show $\ti h:U'\ra V$ is smooth, we must verify the additional condition in Definition \[mc3def1\]. It is enough to do this at $w'\in W$ and $(0,0)\in V$. The proof of Proposition \[mc8prop1\] shows that $V$ is given by , and the inequalities $r_i\ge 0$, $s_i\ge 0$ in are transverse. Therefore, if $\be$ is a local boundary component of $V$ at $(0,0)$, then either (a) $(V,r_i)$ is a boundary defining function for $V$ at $((0,0),\be)$ for some $r_i\ge 0$ appearing in , or (b) $(V,s_i)$ is a boundary defining function for $V$ at $((0,0),\be)$ for some $s_i\ge
0$ in .
In case (a), as $\bigl(\th(U), r_i\ci\th^{-1}\bigr)$ is a local boundary defining function for $X$ at $(x,\th_*(\{r_i=0\})$, and $\pi_X':W'\ra X$ is smooth, by Definition \[mc3def1\] either $r_i\ci\th^{-1}\ci\pi_X'\equiv 0$ near $w'$ in $W$ or $\bigl((\pi_X')^{-1}(\th(U)), r_i\ci\th^{-1}\ci\pi_X'\bigr)$ is a boundary defining function for $W'$ at some $(w',\ti\be)$. Since $r_i\ci\ti h=r_i\ci\th^{-1}\ci\pi'_X\vert_{U'}$ and $U'$ is an open neighbourhood of $w'$ in $W$ it follows that either $r_i\ci\ti
h\equiv 0$ near $w'$, or $(V,r_i\ci\ti h)$ is a boundary defining function for $U'$ at some $(w',\ti\be)$. This proves the additional condition in Definition \[mc3def1\] in case (a). The proof for (b) is the same, using $\pi_Y':W'\ra Y$ smooth. Thus $\ti h$, and hence $h$, is smooth.
Proof of Theorem \[mc6thm2\] {#mc9}
============================
We first construct bijections –. Let $X,Y,Z$ be manifolds with corners, and $f:X\ra Z,$ $g:Y\ra Z$ be strongly transverse smooth maps, and write $W$ for the fibre product $X\t_{f,Z,g}Y$, which we proved exists as a manifold with corners in §\[mc8\]. Let $(x,y)\in W$, so that $x\in X$ and $y\in Y$ with $f(x)=g(y)=z\in Z$. Use all the notation of Proposition \[mc8prop1\], so that the local boundary components of $X$ at $x$ are $\be_1,\ldots,\be_a$, of $Y$ at $y$ are $\ti\be_1,\ldots,
\ti\be_b$, and of $Z$ at $z$ are $\dot\be_1,\ldots,\dot\be_c$. Then over $x,y,z$, the maps $C(f),C(g)$ are given explicitly by in terms of $P^f,P^g\subseteq\{1,\ldots,c\}$ and maps $\Pi^f:P^f\ra\{1,\ldots,a\}$ and $\Pi^g:P^g\ra\{1,\ldots,b\}$.
Properties of these $P^f,P^g,\Pi^f,\Pi^g$ are given in (A)–(D) of the proof of Proposition \[mc8prop1\]. In addition, as $f,g$ are strongly transverse, there are no $\approx$-equivalence classes $E$ of type (b) in part (C), since if $E$ is such a class then gives C(f)(x,{\_i:i \^f(E)})= C(g)(y,{\_i:i \^g(E)})= (z,{\_j:j E}), \[mc9eq1\] and $j=\md{\Pi^f(E)}$, $k=\md{\Pi^g(E)}$, $l=\md{E}$ satisfy $j,k,l>0$ and $j+k=l$, contradicting Definition \[mc6def2\]. Using and these properties of $P^f,P^g,\Pi^f,\Pi^g$ we can describe the points of the r.h.s. of over $x,y,z$ explicitly when $i=1$. We divide such points into three types:
- $\bigl((x,\{\be_i\}),(y,\es)\bigr)$ for $i\in\{1,\ldots,a\}\sm\Pi^f(P^f)$ lies in the term on the r.h.s. of with $i\!=\!j\!=\!1$ and $k\!=\!l\!=\!0$, as $C(f)(x,\{\be_i\})=C(g)(y,\es)=(z,\es)$.
- $\bigl((x,\es),(y,\{\ti\be_i\})\bigr)$ for $i\in\{1,\ldots,b\}\sm\Pi^g(P^g)$ lies in the term on the r.h.s. of with $i\!=\!k\!=\!1$ and $j\!=\!l\!=\!0$, as $C(f)(x,\es)=C(g)(y,\{\ti\be_i\})=(z,\es)$.
- $\bigl((x,\{\be_i:i\in \Pi^f(E)\}),(y,\{\ti\be_i:i\in
\Pi^g(E)\})\bigr)$ for $E$ a $\approx$-equivalence class of type (a) in part (C) lies in the term on the r.h.s. of with $i=1$, $j=\md{\Pi^f(E)}$, $k=\md{\Pi^g(E)}$, $l=\md{E}$, since then holds, and $i=1=j+k-l$.
Now $W$ near $(x,y)$ is diffeomorphic to $V$ near $(0,0)$, where $V$ is given in , and the inequalities $r_i\ge 0$, $s_i\ge 0$ in are transverse. Thus, the local boundary components of $W$ at $(x,y)$ correspond to the inequalities $r_i\ge 0$, $s_i\ge
0$ appearing in . These in turn correspond to points in the r.h.s. of with $i=1$ as follows:
- The local boundary component $r_i=0$ for $i\in\{1,\ldots,a\}\sm\Pi^f(P^f)$ of $V$ at $(0,0)$ corresponds to the point $\bigl((x,\{\be_i\}),(y,\es)\bigr)$ of type (i).
- The local boundary component $s_i=0$ for $i\in\{1,\ldots,b\}\sm\Pi^g(P^g)$ of $V$ at $(0,0)$ corresponds to the point $\bigl((x,\es),(y,\{\ti\be_i\})\bigr)$ of type (ii).
- The local boundary component $r_{\smash{\Pi^f(i)}}=0$ for $i\in
Q$ of $V$ at $(0,0)$ corresponds to the point $\bigl((x,\{\be_i:i\in
\Pi^f(E)\}),(y,\{\ti\be_i:i\in \Pi^g(E)\})\bigr)$ of type (iii), where $E$ is the unique $\approx$-equivalence class containing $i$.
(Note that $Q$ contains one element of each $\approx$-equivalence class.)
The natural map $\coprod_{i\ge 0}C(W)\ra \coprod_{j\ge
0}C(X)\t_{\coprod_{l\ge 0}C(Z)}\coprod_{k\ge 0}C(Y)$ referred to in the last part of Theorem \[mc6thm2\] agrees with this correspondence.
This proves that is a bijection for $i=1$. For the general case, suppose $(w,\{\hat\be_1,\ldots,\hat\be_i\})\in
C_i(W)$. Let $(w,\{\hat\be_a\})$ correspond to $\bigl((x,J_a),(y,K_a)\bigr)$ as above for $a=1,\ldots,i$. Then $C(f)(x,J_a)=C(g)(y,K_a)=(z,L_a)$ for some $L_a$. It is easy to show that as $\hat\be_1,\ldots,\hat\be_i$ are distinct, the subsets $J_1,\ldots,J_i$ are disjoint, and $K_1,\ldots,K_i$ are disjoint, and $L_1,\ldots,L_i$ are disjoint. Also $C(
f)(x,J_1\amalg\cdots\amalg J_i)=C(g)(y,K_1\amalg\cdots\amalg
K_i)=(z,L_1\amalg\cdots\amalg L_i)$. Hence $$\begin{aligned}
&\bigl((x,J_1\amalg\cdots\amalg J_i),(y,K_1\amalg\cdots\amalg
K_i)\bigr)\in \\
&\bigl(C_j(X)\cap C(f)^{-1}(C_l(Z))\bigr)\t_{C(f),C_l(Z),C(g)}\!
\bigl(C_k(Y)\cap C(g)^{-1}(C_l(Z))\bigr),\end{aligned}$$ where $j=\md{J_1}+\cdots+\md{J_i}$, $k=\md{K_1}+\cdots+\md{K_i}$ and $l=\md{L_1}+\cdots+\md{L_i}$. As $1=\md{J_a}+\md{K_a}-\md{L_a}$ for $a=1,\ldots,i$, we have $i=j+k-l$. So mapping $(w,\{\hat\be_1,
\ldots,\hat\be_i\})$ to $\bigl((x,J_1\amalg\cdots\amalg J_i),(y,K_1
\amalg\cdots\amalg K_i)\bigr)$ takes the l.h.s. of to the r.h.s. of . Generalizing the argument in the $i=1$ case proves that this map is a bijection, so is a bijection, and thus is a bijection.
Now let $(w,\{\hat\be_{a_1},\ldots,\hat\be_{a_i}\})\in C_i(W)$ with $$\begin{aligned}
C(\pi_X):(w,\{\hat\be_{a_1},\ldots,\hat\be_{a_i}\})&\longmapsto
(x,\{\be_{b_1},\ab\ldots,\ab\be_{b_j}\})\in C_j(X),\\
C(\pi_Y):(w,\{\hat\be_{a_1},\ldots,\hat\be_{a_i}\})&\longmapsto
(y,\{\ti\be_{c_1},\ldots,\ti\be_{c_k}\})\in C_k(Y),\\
C(f):(x,\{\be_{b_1},\ldots,\be_{b_j}\})&\longmapsto
(z,\{\dot\be_{d_1},\ldots,\dot\be_{d_l}\})\in C_l(Z),
\quad\text{and}\\
C(g):\bigl(y,\{\ti\be_{c_1},\ldots,\ti\be_{c_k}\}\bigr)&\longmapsto
(z,\{\dot\be_{d_1},\ldots,\dot\be_{d_l}\})\in C_l(Z).\end{aligned}$$ Define an immersion $\io^i_W:C_i(W)\ra W$ by $\io^i_W:(w,\{\hat\be_{a_1},\ldots,\hat\be_{a_i}\})\mapsto w$, and similarly for $X,Y,Z$. Then $i_Z^l\ci C(f)\equiv f\ci i_X^j$ near $(x,\{\be_{b_1},\ldots,\be_{b_j}\})$ in $C_j(X)$, and so on, so we have a commutative diagram
@R=5pt@C=17pt[ T\_[(w,{\_[a\_1]{},…,\_[a\_i]{}})]{}C\_i(W) \_[C(\_Y)]{} \_[C(\_X)]{} \_(0.65)[\_W\^i]{} && T\_[(y,{\_[c\_1]{},…,\_[c\_k]{}})]{}C\_k(Y) \^(0.3)[C(g)]{} \^(0.6)[\_Y\^k]{}\
& T\_wW \^(0.2)[\_Y]{} \^(0.3)[\_X]{} && T\_yY \_(0.45)[g]{}\
T\_[(x,{\_[b\_1]{},…,\_[b\_j]{}})]{}C\_j(X) \^(0.4)[C(f)]{} \_(0.6)[\_X\^j]{} && T\_[(z,{\_[d\_1]{},…, \_[d\_l]{}})]{}C\_l(Z) \^(0.65)[\_Z\^l]{}\
& T\_xX \^[f]{} && T\_zZ.]{}
\[mc9eq2\]
Since $W=X\t_ZY$ in $\Manc$, in we have an isomorphism \_X\_Y: T\_wW ((f-g):T\_xXT\_yYT\_zZ). \[mc9eq3\] As $\io_W^i,\io_X^j,\io_Y^k,\io_Z^l$ are immersions, the diagonal maps $\d\io_W^i,\d\io_X^j,\d\io_Y^k,\d\io_Z^l$ in are injective. Thus we can identify $T_{(w,\{\hat\be_{a_1},\ldots,
\hat\be_{a_i}\})}C_i(W)$ with its image in $T_wW$ under $\d\io_W^i$, and similarly for $X,Y,Z$. The proof of Proposition \[mc8prop1\] implies that for each local boundary component $\hat\be_a$ of $W$ at $w$, the tangent space $T_w\hat\be_a\subset T_wW$ is the pullback under $\d f$ or $\d g$ of $T_x\be_b$ or $T_y\ti\be_c$ for appropriate local boundary components $\be_b$ of $X$ at $x$ or $\ti\be_c$ of $Y$ at $y$. So using that is an isomorphism and $\d\io_W^i$ is injective, we see that
&C(\_X)C(\_Y):T\_[(w,{\_[a\_1]{},…,\_[a\_i]{}})]{} C\_i(W)\
&(C(f)-C(g):T\_[(x,{\_[b\_1]{},…,\_[b\_j]{} })]{}C\_j(X)T\_[(y,{\_[c\_1]{},…,\_[c\_k]{}})]{}C\_k(Y)\
&T\_[(z,{\_[d\_1]{},…, \_[d\_l]{}})]{}C\_l(Z))
\[mc9eq4\] is an isomorphism. That is, is injective as it is a restriction of which is injective, and it is surjective as the equations defining $T_{(w,\{\hat\be_{a_1},\ldots,\hat\be_{a_i}\})} C_i(W)$ in $T_wW$ are pullbacks of equations defining $T_{(x,\{\be_{b_1},\ldots,\be_{b_j}\})} C_j(X)$ in $T_xX$ or defining $T_{(y,\{\ti\be_{c_1},\ldots,\ti\be_{c_k}\})}C_k(Y)$ in $T_yY$.
As $W=X\t_ZY$ we have $\dim W=\dim X+\dim Y-\dim Z$, and $i=j+k-l$ from above. So $\dim C_i(W)=\dim C_j(X)+\dim C_k(Y)-\dim C_l(Z)$ as $\dim C_i(W)=\dim W-i$,…. Together with an isomorphism this implies
&T\_[(z,{\_[d\_1]{},…, \_[d\_l]{}})]{}C\_l(Z)=\
&C(f)(T\_[(x,{\_[b\_1]{},…,\_[b\_j]{} })]{}C\_j(X) )+C(g)(T\_[(y,{\_[c\_1]{},…,\_[c\_k]{}})]{}C\_k(Y) ).
\[mc9eq5\]
Let $x\in S^p(X)$, $y\in S^q(Y)$ and $z\in S^r(Z)$. Then as $f$ is transverse we have T\_z(S\^r(Z))=f\_x(T\_x(S\^p(X)))+g\_y(T\_y(S\^q(Y))) \[mc9eq6\] Clearly $(x,\{\be_{b_1},\ldots,\be_{b_j}\})\in S^{p-j}(C_j(X))$ and $\d\io^j_X\bigl(T_{(x,\{\be_{b_1},\ldots,\be_{b_j}\})}
S^{p-j}C_j(X)\bigr)\ab =T_xS^p(X)$. So pulling back using $\io^j_X,\io^k_Y,\io^l_Z$ yields
T\_[(z,{\_[d\_1]{},…,\_[d\_l]{}})]{}&(S\^[r-l]{}(C\_l(Z)))=\
&C(f)\_[(x,{\_[b\_1]{},…,\_[b\_j]{}})]{} (T\_[(x,{\_[b\_1]{},…,\_[b\_j]{}})]{} (S\^[p-j]{}(C\_j(X))))+\
&C(g)\_[(y,{\_[c\_1]{},…,\_[c\_k]{}})]{} (T\_[(y,{\_[c\_1]{},…,\_[c\_k]{}})]{}(S\^[q-k]{}(C\_k(Y)))).
\[mc9eq7\] Equations and imply that the fibre product in is transverse, as we have to prove. So the fibre products in exist in $\Manc$ by Theorem \[mc6thm1\], and the natural map from the left hand side of to the right hand side is smooth. We have already shown it is a bijection, and an isomorphism implies that this natural map induces isomorphisms on tangent spaces. Therefore is a diffeomorphism. This completes the proof of Theorem \[mc6thm2\].
[99]{}
J. Cerf, [*Topologie de certains espaces de plongements*]{}, Bull. Soc. Math. France 89 (1961), 227–380.
A. Douady, [*Variétés à bord anguleux et voisinages tubulaires*]{}, Séminaire Henri Cartan 14 (1961-2), exp. 1, 1–11.
K. Fukaya, Y.-G. Oh, H. Ohta and K. Ono, [*Lagrangian intersection Floer theory – anomaly and obstruction*]{}, preprint, final version, 2008. 1385 pages.
H. Hofer, K. Wysocki and E. Zehnder, [*A general Fredholm theory I: A splicing-based differential geometry*]{}, J. Eur. Math. Soc. 9 (2007), 841–876. math.FA/0612604.
K. Jänich, [*On the classification of $O(n)$-manifolds*]{}, Math. Ann. 176 (1968), 53–76.
D. Joyce, [*D-manifolds and d-orbifolds: a theory of derived differential geometry*]{}, in preparation, 2010.
S. Kobayashi and K. Nomizu, [*Foundations of Differential Geometry, volume I*]{}, John Wiley & sons, New York, 1963.
S. Lang, [*Differentiable manifolds*]{}, Addison-Wesley, Reading, MA, 1972.
Lauda, A.D. and Pfeiffer, H., [*Open-closed strings: two-dimensional extended TQFTs and Frobenius algebras*]{}, Topology Appl. 155 (2008), 623–666. math.AT/0510664.
G. Laures, [*On cobordism of manifolds with corners*]{}, Trans. A.M.S. 352 (2000), 5667–5688.
J. Margalef-Roig and E. Outerelo Dominguez, [*Differential Topology*]{}, North-Holland Math. Studies 173, North-Holland, Amsterdam, 1992.
R.B. Melrose, [*The Atiyah–Patodi–Singer Index Theorem*]{}, A.K. Peters, Wellesley, MA, 1993.
R.B. Melrose, [*Differential Analysis on Manifolds with Corners*]{}, unfinished book available at [http://math.mit.edu/$\sim$rbm]{}, 1996.
B. Monthubert, [*Groupoids and pseudodifferential calculus on manifolds with corners*]{}, J. Funct. Anal. 199 (2003), 243–286.
L. Ramshaw and J. Basch, [*Orienting transverse fibre products*]{}, Hewlett Packard Technical Report HPL-2009-144, 2009.
D.I. Spivak, [*Derived smooth manifolds*]{}, arXiv:0810.5174, 2008. To appear in Duke Mathematical Journal.
[The Mathematical Institute, 24-29 St. Giles, Oxford, OX1 3LB, U.K.]{}
[E-mail: [email protected]]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the problem of “algebraic reconstruction” of linear combinations of shifts of several signals $f_1,\ldots,f_k$ from the Fourier samples. For each $r=1,\ldots,k$ we choose sampling set $S_r$ to be a subset of the common set of zeroes of the Fourier transforms ${\cal F}(f_\l), \ \l \ne r$, on which ${\cal F}(f_r)\ne 0$. We show that in this way the reconstruction system is reduced to $k$ separate systems, each including only one of the signals $f_r$. Each of the resulting systems is of a “generalized Prony” form. We discuss the problem of unique solvability of such systems, and provide some examples.'
author:
- '[^1]'
title: Decoupling of Fourier Reconstruction System for Shifts of Several Signals
---
Introduction
============
In this paper we consider reconstruction of signals of the following a priori known form:
\[equation\_decoupling\_model\] F(x)=\_[j=1]{}\^k \_[q=1]{}\^[q\_j]{} a\_[jq]{} f\_j(x-x\_[jq]{}), with $a_{jq}\in \mathbb{R}, \ x_{jq}=(x^1_{jq},\ldots,x^n_{jq}) \in {\mathbb R}^n.$ We assume that the signals $f_1,\dots, f_k: \mathbb{R}^n\to\mathbb{R}$ are known (in particular, their Fourier transforms ${\cal F}(f_j)$ are known), while $a_{jq}, \ x_{jq}$ are the unknown signal parameters, which we want to find from Fourier samples of $F$. We explicitly assume here that $k\geq 2$. So the usual methods which allow one to solve this problem “in closed form” in the case of shifts of a single function (see [@Vet; @Bat.Sar.Yom; @Sar]) are not directly applicable. Still, we shall show that in many cases an explicit reconstruction from a relatively small collection of Fourier samples of $F$ is possible. Practical importance of signals as above is well recognized in the literature: for some discussions and similar settings see, e.g. [@Vet; @gedalyahu2011multichannel; @peter2011nonlinear].
We follow a general line of the “Algebraic Sampling” approach (see [@Vet; @sig_ack; @Bat.Yom1] and references therein), i.e. we reconstruct the values of the unknown parameters, solving a system of non-linear equations, imposed by the measurements (system (\[equation\_uncoupled\_system\]) below). The equations in this system appear as we equate the “symbolic” expressions of the Fourier samples, obtained from (\[equation\_decoupling\_model\]), to their actual measured values.
Our specific strategy is as follows: we choose a sampling set $S_r \subset {\mathbb R}^n, \ r=1,\ldots,k,$ in a special way, in order to “decouple” (\[equation\_uncoupled\_system\]), and to reduce it to $k$ separate systems, each including only one of the signals $f_r$. To achieve this goal we take $S_r$ to be a subset of the common set of zeroes of the Fourier transforms ${\cal F}(f_\l), \ \l\ne r$.
The decoupled systems turn out to be of a “generalized Prony” type: \[equation\_decoupled1\] \_[j=1]{}\^[N]{} a\_j y\_j\^[s\_ł]{} = m\_ł, ł=1,2,…, s\_łS\^n. The standard Prony system, where the sample set $S$ is the set of integer points in a cube of a prescribed size, allows for a solution “in closed form” (see, for example, [@Bat.Sar.Yom; @rao1992mbp; @Sar; @stoica2005spectral] and references therein). We are not aware of any method for an explicit solution of generalized Prony systems. However, “generic” solution methods can be applied. Their robustness can be estimated via Turán-Nazarov inequality for exponential polynomials and its discrete version ([@Fri.Yom; @Naz]). Some initial results in this direction have been presented in [@Sar; @Bat.Sar.Yom]. Below we further extend these results, restricting ourselves to the uniqueness problem only.
Reconstruction System and its Decoupling {#subsection_decoupling_system}
========================================
For $F$ of the form (\[equation\_decoupling\_model\]) and for any $s=(s^1,\ldots,s^n)\in {\mathbb R}^n$ we have for the sample of the Fourier transform ${\cal F}(F)$ at $s$
$$\begin{aligned}
&{\cal F}(F)(s)&= \int_{{\mathbb R}^n} e^{-2\pi isx} F(x)dx\\
&&=\sum_{j=1}^k \sum_{q=1}^{q_j} a_{jq} e^{-2\pi isx_{jq}}{\cal F}(f_j)(s).\end{aligned}$$
So taking samples at the points $s_\l=(s^1_\l,\ldots,s^n_\l)$ of the sample set $S=\{s_1,\dots,s_m\}$, and denoting the vector $e^{-2\pi ix_{jq}}=(e^{-2\pi ix^1_{jq}},\dots,e^{-2\pi ix^n_{jq}})$ by $y_{jq}=(y^1_{jq},\dots, y^n_{jq})$ we get our reconstruction system in the form
\[equation\_uncoupled\_system\] \_[j=1]{}\^k \_[q=1]{}\^[q\_j]{} a\_[jq]{} [F]{}(f\_j)(s\_ł)y\_[jq]{}\^[s\_ł]{} = [F]{}(F)(s\_ł), ł=1,…,m, in the standard multi-index notations. In system (\[equation\_uncoupled\_system\]) the right hand sides ${\cal F}(F)(s_\l)$ are the known measurements, while the Fourier samples ${\cal F}(f_j)(s_\l)$ are known by our assumptions. The unknowns in (\[equation\_uncoupled\_system\]) are the amplitudes $a_{jq}$ and the shifts $x_{jq}$, encoded in the vectors $y_{jq}$.
In the case $k=1$ we could divide the equations in (\[equation\_uncoupled\_system\]) by ${\cal F}(f_1)(s_\l)$ and obtain directly a Prony-like system. However, for $k\geq 2$ this transformation usually is not applicable. Instead we “decouple" system (\[equation\_uncoupled\_system\]) with respect to the signals $f_1,\ldots,f_k$ using the freedom in the choice of the sample set $S$. Let $$Z_\l=\bigl\{x\in {\mathbb R}^n, \ {\cal F}(f_\l)(x)=0\bigr\}$$ denote the set of zeroes of the Fourier transform ${\cal F}(f_\l)$. For each $r=1,\dots,k$ we take the sampling set $S_r$ to be a subset of the set $$W_r=(\bigcap_{\l\ne r} Z_\l)\setminus Z_r$$ of common zeroes of the Fourier transforms ${\cal F}(f_\l), \ \l\ne r$, but not of ${\cal F}(f_r)$. For such $S_r$ all the equations in (\[equation\_uncoupled\_system\]) vanish, besides those with $j=r$. Hence we obtain:
\[proposition\_coupling\] Let for each $r=1,\dots,k$ the sampling set $S_r$ satisfy $$S_r=\{s_{r1},\dots,s_{rm_r}\}\subset W_r.$$ Then for each $r$ the corresponding system (\[equation\_uncoupled\_system\]) on the sample set $S_r$ takes the form
\[equation\_decoupled\] \_[q=1]{}\^[q\_r]{} a\_[rq]{} y\_[rq]{}\^[s\_[rł]{}]{} = c\_[rł]{}(F), ł=1,…,m\_r, where $c_{r\l}(F)= {{{\cal F}(F)(s_{r\l})}/{{\cal F}(f_r)(s_{r\l})}}$. $\square$
So (\[equation\_uncoupled\_system\]) is decoupled into $k$ generalized Prony systems (\[equation\_decoupled\]), each relating to the shifts of the only signal $f_r$. The problem is that some (or all) of the sets $W_r$ may be too small, and the resulting systems (\[equation\_decoupled\]) will not allow us to reconstruct the unknowns $a_{rq}$ and $y_{rq}$. Another problem is instability of zero finding, which may lead to only approximate zeroes of Fourier transforms. We have at present only initial results outlying applicability of the Fourier decoupling method ([@Sar]). In a “good” case where the zero sets $Z_\l$ of the Fourier transforms ${\cal F}(f_\l), \ \l=1,\ldots,k,$ are nonempty $n-1$-dimensional hypersurfaces meeting one another transversally, still for $k>n+1$ the intersection of $Z_\l, \ \l\ne r,$ is empty. So the resulting systems (\[equation\_decoupled\]) contain no equations. Hence we can apply the above decoupling only for $k \leq n+1$.
Some specific examples, as well as investigation of the conditions on $f_1,\ldots,f_k$ which provide solvability of systems (\[equation\_decoupled\]) were presented in [@Sar]. In one-dimensional case $(n=1, k=2)$ these conditions can be given explicitly. In this case $W_1=W_1(f_1,f_2)$ consists of zeroes of ${\cal F}(f_2)$ which are not zeroes of ${\cal F}(f_1)$, and $W_2=W_2(f_1,f_2)$ consists of zeroes of ${\cal F}(f_1)$ which are not zeroes of ${\cal F}(f_2)$. The following result has been proved (for real Prony systems) in [@Sar]. Here we extend it to the case of system (\[equation\_decoupled\]) which has purely imaginary exponents. The constant $2N$ below is sharp, in contrast with the constant $C(n,d)$ in (multidimensional) Theorem \[span\] below.
Let in (\[equation\_decoupling\_model\]) $n=1,k=2,$ and let $q_1=q_2=N$. Assume that for the signals $f_1,f_2$ in (\[equation\_decoupling\_model\]) each of the sets $W_1$ and $W_2$ contains at least $2N$ elements. Let $D_j, \ j=1,2,$ be the length of the shortest interval $\Delta_j$ such that $S_j=\Delta_j\cap W_j$ contains exactly $2N$ elements, and let $\rho_j={1\over {D_j}}$.
\[solv.cond\] For shifts $x_{jq}$ in the interval $[0,\rho_j), \ j=1,2,$ systems (\[equation\_decoupled\]) with the sampling sets $S_1,S_2$ are uniquely solvable. Let us fix $j=1$. The proof for $j=2$ is the same. Substituting $y_{1q}=e^{-2\pi i x_{1q}}$ associates to a solution $(a_{1q},y_{1q}), \ q=1,\ldots,N,$ of (\[equation\_decoupled\]) an exponential polynomial $H(s)=\sum_{q=1}^N a_{1q}e^{-2\pi i x_{1q}s}$ with purely imaginary exponents. If (\[equation\_decoupled\]) has two different solutions, the corresponding exponential polynomials $H_1(s)$ and $H_2(s)$ are equal for each $s \in S_1.$ Hence $S_1$ is a set of zeroes of $H_2(s)-H_1(s)$, which is an exponential polynomial of the order at most $2N$. On the other hand, by Langer’s lemma (Lemma 1.3 in [@Naz]) such polynomial can have in each interval of length $D$ at most $2N-1+{{\rho D}\over {2\pi}}$ zeroes, where $\rho$ is the maximum of the absolute values of the exponents. In our case $D=D_1$ and $\rho < 2\pi\rho_1={{2\pi}\over {D_j}}$. Hence ${{\rho D}\over {2\pi}}$ is strictly less than $1$, and so the number of zeroes of $H_2-H_1$ is at most $2N-1$, in contradiction with the assumptions. $\square$
Examples
========
Some examples of Fourier decoupling have been presented in [@Sar]. In these examples the sets $W_r$ are “large enough” to reduce the problem (with the number of allowed shifts fixed but arbitrarily large) to a set of decoupled standard Prony systems.
In dimension one we can take, for example, $f_1$ to be the characteristic function of the interval $[-1,1],$ while $f_2(x)=\delta(x-1)+\delta(x+1).$ So we consider signals of the form
\[equation\_decoupling\_model1\] F(x)= \_[q=1]{}\^[N]{} \[a\_[1q]{} f\_1(x-x\_[1q]{})+a\_[2q]{} f\_2(x-x\_[2q]{})\]. Easy computations show that
$${\cal F}(f_1)(s)=\sqrt{\frac{2}{\pi}}\frac{\sin s}{s}$$ and $${\cal F}(f_2)(s)=\sqrt\frac{2}{\pi}\cos s.$$ So the zeros of the Fourier transform of $f_1$ are the points $\pi n,\ n\in {\mathbb Z}\setminus \{0\}$ and those of $f_2$ are the points $({1\over 2}+n)\pi,\ n\in {\mathbb Z}$. These sets do not intersect, so $W_1=\{\pi n\}$, and $W_2=\{({1\over 2}+n)\pi\}$. Since $W_1$ and $W_2$ are just shifted integers ${\mathbb Z}$, the generalized Prony systems in (\[equation\_decoupled\]) are actually the standard ones. For $f_2$ the system (\[equation\_decoupled\]) takes the form $$\frac{{\cal F}(F)(\pi n)}{\sqrt\frac{2}{\pi}(-1)^n}=\sum_{q=1}^Na_{2q}({y_{2q}})^{ \pi n}, \ n\in {\mathbb Z}.$$ If we denote $M_n=\frac{{\cal F}(F)(\pi n)}{\sqrt\frac{2}{\pi}(-1)^n}$ , $A_q=a_{2q}(y_{2q})^\pi$ and $\eta_q=(y_{2q})^\pi$ we get the usual Prony system $$M_n=\sum_{q=0}^NA_q\eta_q^n\ , \ n\in {\mathbb Z}.$$ For $f_1$ we get $$\frac{{\cal F}(F)(({1\over 2} +n)\pi)}{\sqrt\frac{2}{\pi}\frac{(-1)^{n+1}}
{({1\over 2} +n)\pi}}=\sum_{q=1}^Na_{1q}({y_{1q}})^{({1\over 2} +n)\pi} \ , \ n\in {\mathbb Z}.$$ In this case we denote $\mu_n=\frac{{\cal F}(F)(({1\over 2} +n)\pi)}{\sqrt\frac{2}{\pi}
\frac{(-1)^{n+1}}{({1\over 2} +n)\pi}},\ \alpha_q=a_{1q}(y_{1q})^{\frac\pi2}$ and $\xi_q=(y_{1q})^\pi$ and we get again the usual Prony system $$\mu_n=\sum_{q=1}^N\alpha_q\xi_q^n,\ n\in {\mathbb Z}.$$ Solving these two systems by any standard method will give us the translations and amplitudes of the functions $f_1, f_2$. Notice that a possible non-uniqueness of the solutions is imposed here by the substitutions $\eta_q=(y_{2q})^\pi$ and $\xi_q=(y_{1q})^\pi$.
In dimension two we can take, in particular, $f_1,f_2,f_3$ to be the characteristic functions of the three squares: $Q_1=[-3,3]^2, \ Q_2=[-5,5]^2,$ and $Q_3$ which is the rotation of the square $[-\sqrt 2,\sqrt 2]^2$ by ${\pi \over 4}$. So we put
\_j(x) ={
[lr]{}1&xQ\_j\
0&xQ\_j
. and consider signals of the form
\[equation\_decoupling\_model2\] F(x)=\_[j=1]{}\^3 \_[q=1]{}\^[q\_j]{} a\_[jq]{} \_j(x-x\_[jq]{}), a\_[jq]{}, x\_[jq]{} \^3. The following result is proved in [@Sar]:
The zero sets $Z_1,Z_2$ and $Z_3$ of the Fourier transforms of the three functions $\chi_1,\chi_2$ and $\chi_3$ intersect each other in such a way that the decoupling procedure based on the sets $W_1=(Z_2\cap Z_3)\setminus Z_1, W_2=(Z_3\cap Z_1)\setminus Z_2$ and $W_3=(Z_1\cap Z_2)\setminus Z_3$ provides three standard Prony systems for the shifts of each of the functions.
[**Sketch of the proof:**]{} Simple calculation gives
c [F]{}(\_1)(ø,)=4\
[F]{}(\_2)(ø,)=4\
[F]{}(\_3)(ø,)=8.
So $Z_1$ is the union of horizontal or vertical lines crossing the Fourier plane’s axes at $(0,\frac{n\pi}{3})$ or $(\frac{n\pi}{3},0)$ respectively, for all non zero integer $n$. Similarly for $Z_2$, with the only difference that the lines cross the axes at $(0,\frac{n\pi}5)$ or $(\frac{n\pi}5,0)$.\
$Z_3$ is the union of lines with slopes $1$ or $-1$ crossing the $\o$ axis at $2\pi n$ for some non zero integer $n$. Hence for any two integers $n$ and $m$ we have $(\frac{1+5n}5,\frac{1+5n}5)\in S_1, (\frac{1+3m}3,\frac{1+3m}3)\in S_2$ and since $\frac{1+3m}3\pm\frac{1+5n}5$ is not an integer, $(\frac{1+3m}3,\frac{1+5n}5)\in S_3$. These three points form a triangle which repeats itself as a periodic pattern. Appropriate transformations now bring the decoupled systems (\[equation\_decoupled\]) to the form of the standard two-dimensional Prony system. See [@Sar; @Bat.Sar.Yom] for a new approach to solving such systems and for the results of numerical simulations. $\square$
Uniqueness of Reconstruction {#Non-uniform}
============================
Application of Proposition \[proposition\_coupling\] prescribes the choice of sample points from the common zeroes of the Fourier transforms ${\cal F}(f_j)$. So the geometry of the sample sets $S_r$ may be complicated, and the known results on unique solvability of the standard Prony system ([@Bat.Sar.Yom; @Bat.Yom; @rao1992mbp; @stoica2005spectral]) are not directly applicable. Non-Uniform Sampling in Prony-type systems is also essential in other problems of algebraic signal reconstruction. In particular, recently it appeared as a key point in a proof of the Eckhoff conjecture, related to the accuracy of reconstruction of piecewise-smooth functions from their Fourier samples ([@Bat]).
There are results on a behavior of exponential polynomials on arbitrary sets, which can provide important information on unique solvability and robustness of the generalized Prony system. In particular, this concerns the Turan-Nazarov inequality ([@Naz]), and its extension to discrete sets obtained in [@Fri.Yom]. In this last paper for each set $S$ a quantity $\o_D(S)$ has been introduced, measuring, essentially, the robustness of solvability of a generalized Prony system with the sample points $s_\l\in S$. Here $D$ comprises the “discrete” parameters of the Prony system to be solved. $\o_D(S)$ can be explicitly estimated in terms of the metric entropy of $S$ (see below), and we expect that in many important cases the quantity $\o_D(W_r)$ for the zeroes sets $W_r$ of the Fourier transforms ${\cal F}(f_j)$ can be effectively bounded from below. Some initial results and discussions in this direction, mainly in dimension one, are presented in [@Sar; @Bat.Yom1]. In the present paper we do not consider robustness of the Prony system, but provide a new multi-dimensional result on the uniqueness of solutions, in the lines of [@Sar; @Fri.Yom] and Theorem \[solv.cond\] above.
Let us recall that for $Z$ a bounded subset of ${\mathbb R}^n,$ and for $\e>0$ the covering number $M(\e,Z)$ is the minimal number of $\e$-balls in ${\mathbb R}^n,$ covering $Z$. The $\e$-entropy $H(\e,Z)$ is the binary logarithm of $M(\e,Z)$.
Let $H(s)=\sum_{j=1}^d a_je^{\lambda_j\cdot s},$ with $a_j\in {\mathbb R}, \ \lambda_j=(\lambda_{j1},\ldots,\lambda_{jn})\in {\mathbb R}^n,$ be a real exponential polynomial in $s\in {\mathbb R}^n.$ Denote $Z(H)$ the set of zeroes of $H$ in ${\mathbb R}^n$, and let $Q^n_R$ be the cube in ${\mathbb R}^n$ with the edge $R$. The following result is a special case of Lemma 3.3 proved in [@Fri.Yom]:
\[entropy\] For each $R>0,$ and $\e$ with $R>\e>0$ we have $M(\e,Z(H)\cap Q^n_R)\leq C(d,n)({R\over \e})^{n-1}$. $\square$
The explicit expression for $C(d,n)$ is given in [@Fri.Yom], via Khovanski’s bound ([@Kho]) for “fewnomial” systems. Consider now a generalized Prony system (\[equation\_decoupled1\]) with a finite set $S$ of samples allowed:
\[Pron\] \_[j=1]{}\^[N]{} a\_j y\_j\^[s\_ł]{} = m\_ł, s\_łS={s\_1,…,s\_m }\^n. We shall consider only real solutions of (\[Pron\]) with $y_j$ having all its coordinates positive.
\[span\] Let $S=\{s_1,\ldots,s_m \}\subset Q_R^n$ be given, such that for a certain $\e>0$ we have $M(\e,S)> C(2N,n)({R\over \e})^{n-1}.$ Then system (\[Pron\]) has at most one solution. Associate to a solution $(a_j,y_j), \ j=1,\ldots,N,$ of (\[Pron\]) an exponential polynomial $H(s)=\sum_{j=1}^N a_je^{\lambda_j\cdot s},$ where $y_j=e^{\lambda_j}, \ \lambda_j\in {\mathbb R}^n.$ If (\[Pron\]) has two different solutions, the corresponding exponential polynomials $H_1(s)$ and $H_2(s)$ are equal for each $s=s_\l\in S.$ Hence $S$ is a set of zeroes of $H_2(s)-H_1(s)$, which is an exponential polynomial of order at most $2N$. By Proposition \[entropy\] we have $M(\e,S) \leq C(2N,n)({R\over \e})^{n-1}$ for each $\e>0$, in contradiction with the assumptions of the theorem. $\square$
Informally, Theorem \[span\] claims that finite sets $S$ which cover (in a “resolution $\e$”, for some $\e>0$), a significant part of the cube $Q_R^n$, are uniqueness sets of the Prony system. The condition of Theorem \[span\] on the sampling set $S$ is quite robust with respect to the geometry of $S$, so we can explicitly verify it in many cases. In particular, for non-regular lattices we get the following result:
For fixed positive $\alpha < {1\over 2}$ and $h>0,$ a set $Z' \subset {\mathbb R}^n$ is called an $(\alpha,h)$-net if it possesses the following property: there exists a regular grid $Z$ with the step $h$ in ${\mathbb R}^n$ such that for each $z'\in Z'$ there is $z\in Z$ with $||z'-z||\leq \alpha h,$ and for each $z\in Z$ there is $z'\in Z'$ with $||z'-z||\leq \alpha h.$ \[Nonreg\] Let $Z' \subset {\mathbb R}^n$ be an $(\alpha,h)$-net. Then for $R> C(2N)h(1-2\alpha)^{1-n}$ the set $S=Z\cap Q_R^n$ is a uniqueness set of the Prony system . By definition, for each $z\in Z$ we can find $z'\in Z'$ inside the $\alpha h$-ball around $z$. Clearly, any two such points are $h'=(1-2\alpha)h$-separated. So for each $\e < h'$ we have $M(\e,S)\geq |Z\cap Q_R^n|=({R\over h})^n.$ We conclude that the inequality $({R\over h})^n > C(2N)({R\over h'})^{n-1},$ or $R > C(2N)h(1-2\alpha)^{1-n}$ implies the condition of Theorem \[span\]. $\square$
The condition of Theorem \[span\] can be verified in many other situations, under natural assumptions on the sample set $S$. In particular, using integral-geometric methods developed in [@Com.Yom], it can be checked for the zero sets of Fourier transforms of various types of signals. We plan to present these results separately.
[**Remark**]{} The restriction to only positive solutions of Prony system is very essential for the result of Theorem \[span\]. Indeed, consider the Prony system
\[Skolem\] a\_1x\_1\^k+a\_2x\_2\^k=m\_k, k=0,1,…. If we put $a_1=1, \ x_1=1, \ a_2=-1, \ x_2=-1$, then $m_k=1^k-(-1)^k=0$ for each even $k$. So the regular grid of even integers is not a uniqueness set for system (\[Skolem\]). This fact is closely related to the classical Skolem-Mahler-Lech Theorem (see [@Lec; @Mey.vdP; @Tao] and references therein) which says that the integer zeros of an exponential polynomial are the union of complete arithmetic progressions and a finite number of exceptional zeros. So such sets may be non-uniqueness sample sets for complex Prony systems.
The proof of the Skolem-Mahler-Lech Theorem is relied on non-effective arithmetic considerations. Recently the problem of obtaining effective such theorem was discussed in [@Tao]. This problem may turn to be important for understanding of complex solutions of Prony systems. One can wonder whether the methods of Khovanskii ([@Kho]) and Nazarov ([@Naz]), as well as their combination in [@Fri.Yom], can be applied here.
[10]{}
D. Batenkov. .
D. Batenkov, N. Sarig, and Y. Yomdin. . , 19(1-2):9–26, 2012.
D. Batenkov and Y. Yomdin. . , 2011.
D. Batenkov and Y. Yomdin. . , 73(1):134–154, 2013.
G. Comte and Y. Yomdin. , 601–630.
P.L. Dragotti, M. Vetterli and T. Blu. , IEEE Transactions on Signal Processing, Vol. 55, Nr. 5, Part 1, pp. 1741-1757, 2007.
O. Friedland and Y. Yomdin. . , 2011.
K. Gedalyahu, R. Tur, and Y.C. Eldar. Multichannel sampling of pulse streams at the rate of innovation. , 59(4):1491–1504, 2011.
A. G. Khovanskii. . Translated from the Russian by Smilka Zdravkovska. , 88. American Mathematical Society, Providence, RI, 1991. viii+139 pp.
C. Lech. . .
G. Myerson and A. J. van der Poorten, . 102, 698-705, 1995.
F.L. Nazarov. . , 5(4):663–718, 1994.
T. Peter, D. Potts, and M. Tasche. Nonlinear approximation by sums of exponentials and translates. , 33(4):1920, 2011.
B.D. Rao and K.S. Arun. . , 80(2):283–309, 1992.
N. Sarig and Y. Yomdin. . , 29(4):97–114, 2008.
N. Sarig. . PhD thesis, [Weizmann Institute of Science]{}, 2010.
P. Stoica and R.L. Moses. . Pearson/Prentice Hall, 2005.
T. Tao. .
[^1]: This research was supported by the Adams Fellowship Program of the Israeli Academy of Sciences and Humanities, ISF Grant No. 639/09, and by the Minerva foundation. We would like to thank the referees for useful corrections and remarks.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we provide some simple characterizations for the spherical harmonics coefficients of an isotropic random field on the sphere. The main result is a characterization of isotropic gaussian fields through independence of the coefficients of their development in spherical harmonics.'
author:
- |
Paolo Baldi\
Dipartimento di Matematica\
Università di Roma - Tor Vergata\
Via della Ricerca Scientifica\
00133 Roma, Italy.\
[email protected]
- |
Domenico Marinucci\
Dipartimento di Matematica\
Università di Roma - Tor Vergata\
Via della Ricerca Scientifica\
00133 Roma, Italy.\
[email protected]
bibliography:
- 'bibbase.bib'
title: SOME CHARACTERIZATIONS OF THE SPHERICAL HARMONICS COEFFICIENTS FOR ISOTROPIC RANDOM FIELDS
---
[*Key words and phrases*]{} Spherical Random Fields, Spherical Harmonics, Characterization of Gaussian Random Fields.
[*AMS 2000 subject classification:*]{} Primary 60B15; secondary 62M15,62M40.
Introduction
============
In recent years the study of real random fields on the sphere has received much attention, as this topic is a necessary tool in the statistical study of the CMB (Cosmic Microwave Background) radiation. The existing physical literature is huge, with particular emphasis on testing for Gaussianity and isotropy (e.g., for a small sample, [@DCB:2003], [@cruz:05], [@Pa:2004]) as both these issues have deep implications for cosmological physics. See for reviews on this subject [@MR2106384] and [@MA:2005], where many more references can be found. We also mention [@JSDAF:05], [@AGMNW:04] and [@MR2065205] for a mathematical treatment of the subject.
A natural tool for this kind of enquiry is the development of the random field in a series of spherical harmonics. This raises some simple, but not so obvious questions. For an isotropic random field $T$, given its development in spherical harmonics (see §2), $$\label{develop}
T(x)=\sum_{\ell=0}^\infty\sum_{m=-\ell}^\ell a_{\ell m} Y_{\ell m}(x)$$ what properties must be expected to be satisfied by the coefficient $a_{\ell m}$? Or, from a different point of view, which conditions must be verified by the coefficients in order that the development above defines an isotropic random field?
These questions are of interest both from a statistical point of view (for instance in order to devise a test suitable to detect non-Gaussianity or anisotropy) or from a probabilistic point of view (how to sample an isotropic random field).
In this paper we provide some results in this perspective. More precisely, first we prove that, for an isotropic random field, the coefficients are necessarily uncorrelated. This fact is well known, but we give a proof that holds without the assumption that the field is mean square continuous.
Then it is proved that each of the complex r.v.’s $a_{\ell m}$ has a distribution whose phase is uniform on $[-\pi,\pi]$. This implies in particular that, for any isotropic field, the ratio $\func{Re} a_{\ell m}/\func{Im} a_{\ell m}$ is necessarily distributed accordingly to a Cauchy distribution.
Finally, having remarked that if the field $T$ is gaussian then the $a_{\ell m}$’s, $\ell=0,1,\dots$, $m=0,\dots,\ell$ are independent, we prove that also the converse is true. Thus the only isotropic fields such that the $a_{\ell m}$’s, $\ell=0,1,\dots$, $m=0,\dots,\ell$ are independent are those that are gaussian. This, which is the main result of this paper, is not a consequence of the central limit theorem, but results from some classical characterization of gaussian random variables, through independence of some linear statistics.
This result gives a rigorous proof of claims that can be occasionally found in the cosmological literature (see [@CoMa:2001] e.g.)
Isotropic random fields
=======================
In this section we recall some well known facts about isotropic random fields $T$ defined on the unit sphere $S^{2}=\{ x\in
{\mathbb R}^{3}:| x | =1\}$. We assume that these fields are isotropic in the strong sense, that is, their probability law is invariant with respect to the action of $
SO(3)$. The isotropy assumption can then be stated as follows: for all $g\in SO(3)$ and $x_{1},\dots ,x_{p}\in S^{2}$, the two vectors $$(T(gx_{1}),\dots ,T(gx_{p}))\quad \mbox{and}\quad(T(x_{1}),\dots ,T(x_{p}))\text{ , }$$have the same distribution. Throughout this paper by “isotropic” we mean isotropic in this sense.
We shall use the spherical coordinates on $S^2$ i.e. $x =(\vartheta ,\varphi )$, where $0\leq \vartheta \leq \pi $ , $0\leq
\varphi <2\pi$. Also we shall use for $g\in SO(3)$ the parameterization through the Euler angles $0\leq \alpha ,\gamma <2\pi $ and $0\leq \beta \leq \pi $. Assuming the right-hand side exists, for each $\ell=1,2,\dots $we can define the random vector $$a_{\ell.}=\int_{S^{2}}T(x )Y_{\ell.}(x )\,dx \label{basdef}$$where $dx =\sin \vartheta\, d\varphi\, d\vartheta $ denotes the Lebesgue measure on $S^{2}$ and $Y_{\ell.}$ the vector of spherical harmonics defined by $$\begin{aligned}
Y_{\ell.}(\theta ,\varphi )& =\left( Y_{\ell\ell}(\theta ,\varphi ),\dots ,Y_{\ell,-\ell}(\theta ,\varphi
)\right) ^{\prime }\text{ ,} \notag \\
Y_{\ell m}(\theta ,\varphi )& =\sqrt{\frac{2l+1}{4\pi }\frac{(\ell-m)!}{(\ell+m)!}}P_{\ell m}(\cos \theta )e^{im\varphi}\text{ , for }m\ge 0\text{ ,} \notag \\
Y_{\ell m}(\theta ,\varphi )& =(-1)^{m}\overline{Y_{\ell,-m}(\theta ,\varphi )}\text{
, for }m<0\text{ }; \label{spharm}\end{aligned}$$here $P_{\ell m}(\cos \theta )$ denotes the associated Legendre functions i.e. $$\begin{aligned}
P_{\ell m}(x)& =(-1)^{m}(1-x^{2})^{m/2}\frac{d^{m}}{dx^{m}}P_{\ell}(x)\text{ , }P_{\ell}(x)=\frac{1}{2^{\ell}l!}\frac{d^{\ell}}{dx^{\ell}}(x^{2}-1)^{\ell}, \\
m& =0,1,2,\dots ,\ell\text{ , }\ell=1,2,3,\dots .\text{ .}\end{aligned}$$A detailed discussion of the properties of the spherical harmonics can be found in Varshalovich, Moskalev and Khersonskii [@MR1022665], chapter 5. Our purpose in this paper is to provide some characterizations of the probability law of the vector $a_{\ell.}$ under the isotropy assumption. It is a standard result that the functions $Y_{\ell m}$ $m=-\ell,\dots,\ell$ form a basis for the vector space of functions on $S^2$ which are restrictions of homogeneous harmonic polynomials of degree $\ell$. This vector space being invariant by the action of $SO(3)$, for any $\ell$ and $g\in SO(3)$ there exist a $(2\ell+1)\times (2\ell+1)$ matrix $D^{\ell}(g)$ such that$$Y_{\ell.}(g x)=D^{\ell}(g)Y_{\ell}(x )\text{ ,} \label{basrel}$$This provides a representation of $SO(3)$ on ${\mathbb C}^{2\ell+1}$ which moreover is irreducible (see Vilenkin and Klymik [@MR1220225], §9.2.6 e.g.) The matrices $D^{\ell}(g)$ are the so-called Wigner’s D-matrices whose entries are, in terms of the Euler angles $$D_{mm^{\prime }}^{\ell}(\alpha ,\beta ,\gamma ) =e^{-im\alpha }\,d_{mm^{\prime
}}^{\ell}(\beta )\,e^{-im^{\prime }\gamma },$$ where $$\begin{aligned}
d_{mm^{\prime }}^{\ell}(\beta ) &=&(-1)^{\ell-m^{\prime} }\left[ (\ell+m)!\,(\ell-m)!\,(\ell+m^{\prime })\,!(\ell-m^{\prime })!\right] ^{1/2} \\
&\times& \sum_{k=0}^{\max (\ell-m,\ell-m^{\prime })}\frac{(\cos \frac{\beta }{2})^{2k-m-m^\prime }(\sin \frac{\beta }{2})^{2\ell-2k-m-m^\prime }}{k!\,(\ell-m-k)!\,(\ell-m^{\prime }-k)!\,(m+m^{\prime }+k)!}\text{ .}\end{aligned}$$Note that $d_{mm^{\prime }}^{\ell}(0 )=\delta
_{m}^{m^\prime }$, where $\delta _{m}^{m\prime }$ denotes the Kronecker delta function. It is immediate that, the Lebesgue measure of $S^2$ being invariant by the action of $SO(3)$, $$\label{invar1}
\begin{array}c
\displaystyle
a_{\ell.} =\int_{S^{2}}T(x )Y_{\ell.}(x )\,dx \overset{d}{=}\int_{S^{2}}T(gx )Y_{\ell.}(x )\,dx\cr
\displaystyle=\int_{S^{2}}T(x)D^{\ell}(g^{-1})Y_{\ell.}(x )\,dx
=D^{\ell}(g^{-1})a_{\ell.}\text{ ,}
\end{array}$$ In particular, in coordinates, $$a_{\ell m}\overset{d}{=}\sum_{m^{\prime }=-\ell}^{\ell}a_{\ell m^{\prime }}D_{m^{\prime
}m}^{\ell}(g)\text{ .} \label{invrot2}$$It is well known that for an isotropic field which is continuous in mean square the development (\[develop\]) holds, the convergence being in $L^2$ (see [@MR1687092] e.g.).
It is also useful to point out that, because of (\[spharm\]), the coefficients satisfy the following identity $$a_{\ell m}=(-1)^m\overline{a_{\ell, -m}}.$$ In particular $a_{\ell 0}$ is real.
General properties of the spherical harmonics coefficients
==========================================================
It is well-known that, for mean square continuous and isotropic random fields, the spherical harmonics coefficients are orthogonal, i.e. $$E[a_{\ell_{i}m_{i}}\overline{a_{\ell_{j}m_{j}}}]=\delta _{\ell_{i}}^{\ell_{j}}\delta
_{m_{j}}^{m_{i}}C_{\ell_{i}}\text{ , } \label{orthocon}$$ the sequence $\{ C_{\ell}\} _{\ell=1,2,\dots }$ denoting the angular power spectrum of these fields. For reasons of completeness we give now a proof of this fact. Actually our statement is slightly stronger. As usual we denote by $A^*$ the complex conjugate of the matrix $A$.
\[prop1\] *Assume $T$ isotropic. Then*
a\) for all $\ell$ such that $E[| a_{\ell.}| ^{2}]<\infty $, $$Ea_{\ell.}a_{\ell.}^{\ast }=C_{\ell}I_{2\ell+1}\text{ ,}$$where $I_{2\ell+1}$ denotes the $(2\ell+1)\times (2\ell+1)$ identity matrix
b\) for all $\ell_{1},\ell_{2}$ such that $E[| a_{\ell_{1}.}|
^{2}]<\infty $, $E[| a_{\ell_{1}.}|
^{2}]<\infty $ $$Ea_{\ell_{1}.}a_{\ell_{2}.}^{\ast }=0$$(in the sense of the $(2\ell_{1}+1)\times (2\ell_{2}+1)$ zero matrix).
[*Proof* ]{} a) Let us denote by $\Gamma_\ell$ the covariance matrix of the random vector $a_{\ell.}$. Since the vectors $a_{\ell.}$ and $D^{\ell}(g)a_{\ell.}$ have the same distribution, they have the same covariance matrix. This gives $$\Gamma_\ell=D^{\ell}(g)\Gamma_\ell D^{\ell}(g)^*=D^{\ell}(g)\Gamma_\ell D^{\ell}(g)^{-1}$$ Since $D^{\ell}$ is an irreducible representation of $SO(3)$, by Schur lemma $\Gamma_\ell$ is of the form $C_{\ell}I_{2\ell+1}$.
b\) The representations $D^{\ell_{1}}$ and $D^{\ell_{2}}$ are not equivalent for $\ell_{1}\not=\ell_{2}$, having different dimensions. Therefore again by Schur lemma, the identity $$Ea_{\ell_{1}.}a_{\ell_{2}.}^{\ast }=D^{\ell_{1}}(g)Ea_{\ell_{1}.}a_{\ell_{2}.}^{\ast }
D^{\ell_{1}}(g)^{-1}$$ can hold only if the right hand side is the zero matrix.
We stress that (for strongly isotropic fields) Proposition 1 is strictly stronger than the standard result on mean square random fields. Indeed, it is immediate to show that $ET^{2}<\infty $ implies $\sum_{l=1}^{\infty }E | a_{l.}| ^{2}<\infty ,$ on the other hand, it is not difficult to find examples where mean square continuity fails but the assumptions of Proposition 1 are fulfilled. Consider for instance the field:$$T(x
)=\sum_{m=-\ell_{1}}^{\ell_{1}}a_{\ell_{1}m_{1}}Y_{\ell_{1}m_{1}}(x
)+\sum_{m=-\ell_{2}}^{\ell_{2}}a_{\ell_{2}m_{2}}Y_{\ell_{2}m_{2}}(x
)+\sum_{m=-\ell_{3}}^{\ell_{3}}b_{\ell_{3}m_{3}}Y_{\ell_{3}m_{3}}(x )\label{ex1}$$ where $b_{\ell_{3}m_{3}}=\eta a_{\ell_{3}m_{3}}$; we assume that the $a_{\ell_{i}m_{i}}$’s $(i=1,2,3)$ satisfy (\[orthocon\]) whereas $\eta $ is a random variable with infinite variance (for instance a Cauchy). It is not difficult to see that the field $T$ is properly defined and strictly isotropic; although (\[ex1\]) is clearly an artificial model, some closely related field may be of interest for practical applications: for instance in CMB data analysis it is often the case that the observed field is a superposition of signal plus foreground contamination, and the latter may be characterized by heavy tails at the highest multipoles (point sources). In such cases, it is of an obvious statistical interest to know that the standard properties of the spherical harmonics coefficients still hold at least for the multipoles where foreground contamination is absent. It is immediate to see that $ET^{2}=\infty ,$ whence the field cannot be mean-square continuous; however (\[basdef\]) is still properly defined for $l=l_{1},l_{2}$ (simply exchange the integral with the finite sum), and therefore Proposition 1 holds for these two vectors of spherical harmonics coefficients.
Our next proposition provides three further characterizations for the spherical harmonics coefficients of isotropic fields.
\[prop2\] *Let $T$ be an isotropic random field (not necessarily mean square continuous); then for all $a_{l.}$ such that $E | a_{\ell.} | ^{2}<\infty $,*
a\) for all $m=1,\dots ,\ell,$ $$\func{Re} a_{\ell m}\overset{d}{=}\func{Im}a_{\ell m}\quad\mbox{ and }\quad \frac{\func{Re}a_{\ell m}}{\func{Im}a_{\ell m}}\sim Cauchy\text{ }$$
b\) for all $m=1,2,\dots ,\ell,$ $\func{Re}a_{\ell m}$ and $\func{Im}a_{\ell m}$ are uncorrelated, with variance $E(\func{Re}a_{\ell m})^{2}=E(\func{Im}a_{\ell m})^{2}=C_{\ell}/2.$
c\) The marginal distribution of $\func{Re}a_{\ell m}$ , $\func{Im}a_{\ell m}$ is always symmetric, that is, $$\func{Re}a_{\ell m}\overset{d}{=}-\func{Re}a_{\ell m}\text{ , }\func{Im}a_{\ell m}\overset{d}{=}-\func{Im}a_{\ell m}\text{ .}$$
[*Proof* ]{} a) For $\beta =\gamma =0,$ (\[invrot2\]) becomes$$a_{\ell m}\overset{d}{=}e^{-im\alpha }a_{\ell m}$$for all $m=-\ell,\dots ,\ell$, $0\leq \alpha <2\pi$. This entails $$\left(
\begin{array}{c}
\func{Re}a_{\ell m} \\
\func{Im}a_{\ell m}\end{array}\right) \overset{d}{=}\left(
\begin{array}{cc}
\cos \varphi & \sin \varphi \\
-\sin \varphi & \cos \varphi
\end{array}\right) \left(
\begin{array}{c}
\func{Re}a_{\ell m} \\
\func{Im}a_{\ell m}\end{array}\right)
\label{rot1}$$for all $m=-\ell,\dots ,\ell$, $0\leq \varphi <2\pi$. Thus the vector $^t(\func{Re}a_{\ell m},\func{Im}a_{\ell m})$ has a distribution that is invariant by rotations, that is, in polar coordinates, they can be written in of the form $$\label{rho1}
R \cos(\Theta)$$ where $R$ is a random variable with values in $\mathbb{R}^+$, whereas $\Theta$ is uniform in $[-\pi,\pi]$. This entails immediately that $\arctan( \func{Re}a_{\ell m}/\func{Im}a_{\ell m}) \sim U(-\frac{\pi }{2},\frac{\pi }{2});$ the result follows immediately.
b\) This property is well-known if $T$ is mean square continuous. From (\[rho1\]), $$E[\func{Re}a_{\ell m}\cdot \func{Im}a_{\ell m}]=\int_0^{+\infty}r^2d\mu_R(r)\int_{-\pi}^{\pi }\cos
\vartheta \sin \vartheta d\vartheta =0\text{ .}$$ c) It suffices to take $\varphi _{0}=\pi $ in (\[rot1\]).
It is interesting to note how Proposition \[prop1\] implies that no information can be derived on the statistical distribution of an isotropic random field by the marginal distribution function of the ratios $( \func{Re}a_{\ell m}/\func{Im}a_{\ell m})$. On the other hand, it may be possible to use these ratios to implement statistical tests of the assumption of isotropy, an issue which has gained a remarkable empirical relevance after the first release of the WMAP data in February 2003
It is clear that if the field $T$ is gaussian, then the r.v.’s $(a_{\ell m})_{\ell,m}$ is a gaussian family. We prove now an independence result for this family of r.v.’s. Thanks to Proposition \[prop1\], these r.v.’s are uncorrelated, but one must be careful, since, in the case of complex r.v.’s, absence of correlation and joint gaussian distribution does not imply independence.
\[prop-if\] *For an isotropic gaussian random field the r.v.’s $a_{\ell m}$, $\ell=0,1,\dots$, $m=0,\dots,\ell$ are independent.*
[*Proof*]{} Let be $(\ell,m )\not=(\ell',m')$, $m>0$, $m'> 0$. Then $a_{\ell m}$ is uncorrelated with both $a_{\ell' m'}$ and $a_{\ell', -m'}=\overline{a_{\ell' m'}}$. Thus $$E[a_{\ell m}\overline{ a_{\ell' m'}}]=0,\qquad E[a_{\ell m}a_{\ell' m'}]=E[a_{\ell m}
\overline{a_{\ell', -m'}}]=0$$ and the statement follows from Lemma \[lemma\]. If one at least among $m$ and $m'$ is equal to $0$, then the r.v. $a_{\ell m}$ (or $a_{\ell' m'}$ is real and independence follows from absence of correlation as for the real case.
\[lemma\]*Let $Z_1$, $Z_2$ be complex r.v.’s, centered and jointly gaussian. Then they are independent if and only if $$\label{complex1}
E[Z_1\overline Z_2]=0,\qquad E[Z_1 Z_2]=0$$*
[*Proof*]{} In one direction the statement is obvious. Let us assume that (\[complex1\]) are satisfied. Then, if we set $Z_k=X_k+iY_k$, $k=1,2$, then $$\displaylines{
E[X_1X_2+Y_1Y_2]+i E[-X_1Y_2+Y_1X_2]=0\cr
E[X_1X_2-Y_1Y_2]+i E[X_1Y_2+Y_1X_2]=0\cr
}$$ From these one obtains $E[X_1X_2]=0$, $E[Y_1Y_2]=0$, $E[X_1Y_2]=0$ and $E[Y_1X_2]=0$. This means that each of the r.v.’s $X_1,Y_1,X_2,Y_2$ is uncorrelated with the other ones, so that, being jointly gaussian, they are independent.
Which is less obvious, is that the converse also holds. The following is the main result of this paper.
\[onlyif\] *For an isotropic random field, let $\ell$ be such that $E | a_{\ell .}|
^{2}<\infty$. Then the coefficients $(a_{\ell 0},a_{\ell 1}\dots , a_{\ell \ell })$ are independent if and only if they are gaussian.*
[*Proof* ]{} We just need to prove the “only if” part. Fix $m_1\ge 0$, $m_2\ge 0$, so that the two complex r.v.’s $a_{\ell m_{1}}$ and $a_{\ell m_{2}}$ are independent. Note that we are not assuming the independence of $\func{Re}a_{\ell m_{1}}$ and $\func{Im}a_{\ell m_{1}}$ or of $\func{Re}a_{\ell m_{2}}$ and $\func{Im}a_{\ell m_{2}}$. Thanks to (\[invar1\]), the two vectors $a_{\ell .}$ and $D^\ell(g)a_{\ell .}$ have the same distribution. Thus the two r.v.’s $$L_1=\sum_{m'=-\ell}^\ell D^\ell_{m',m_1}(g)a_{\ell m'}\quad\mbox{and} \quad
L_2=\sum_{m'=-\ell}^\ell D^\ell_{m',m_2}(g)a_{\ell m'}$$ having the same joint distribution as $a_{\ell m_1}$ and $a_{\ell m_2}$, are independent. Fix $g$ so that the angles $\alpha $ such that $m\alpha
\neq k\pi $ for all integers $m,k$ and $\beta $ such that $d_{mm_{1}}^{\ell }(\beta )$ and $d_{mm_{2}}^{\ell }(\beta )$ are different from zero for all $m=-\ell ,\dots ,\ell $ (note that such a $\beta $ certainly exists, because the functions $d_{m_{2}m}^{\ell }(\beta )$, $m,m'=-\ell,\dots,\ell$ are analytic and can vanish only at a finite number of values $\beta\in[0,\pi]$). For such a choice of $g$, thanks to (\[invar1\]), $$\left(
\begin{array}{c}
\func{Re}L_1 \\
\func{Im}L_1\end{array}\right) {=}\sum_{m=-\ell }^{\ell}d_{mm_{1}}^{\ell }(\beta )\left(
\begin{array}{cc}
\cos m_{1}\alpha & \sin m_{1}\alpha \\
-\sin m_{1}\alpha & \cos m_{1}\alpha
\end{array}\right) \left(
\begin{array}{c}
\func{Re}a_{\ell m} \\
\func{Im}a_{\ell m}\end{array}\right)$$and$$\left(
\begin{array}{c}
\func{Re}L_2 \\
\func{Im}L_1\end{array}\right) {=}\sum_{m=-\ell }^{\ell }d_{mm_{2}}^{\ell }(\beta )\left(
\begin{array}{cc}
\cos m_{2}\alpha & \sin m_{2}\alpha \\
-\sin m_{2}\alpha & \cos m_{2}\alpha
\end{array}\right) \left(
\begin{array}{c}
\func{Re}a_{\ell m} \\
\func{Im}a_{\ell m}\end{array}\right)$$where the $2\times 2$ matrices on the right hand-sides are always full rank. By the Skitovich-Darmois theorem below (see Kagan, Rao and Linnik [@MR0346969] e.g.), it follows that each of the vectors $(\func{Re}a_{\ell m},\func{Im}a_{\ell m})$ is bivariate Gaussian; as $(\func{Re}a_{\ell m},\func{Im}a_{\ell m})$ are uncorrelated and have the same variance by Proposition \[prop1\], then $a_{\ell m}=\func{Re}a_{\ell m}+i\func{Im}a_{\ell m}$ is complex Gaussian.
(Skitovich-Darmois) *Let $X_{1},\dots ,X_{r}$ be mutually independent random vectors in $R^{n}.$ If the linear statistics $$L_{1}=\sum_{j=1}^{r}A_{j}X_{j},\qquad L_{2}=\sum_{j=1}^{r}B_{j}X_{j}\text{
,}$$are independent, for some real nonsingular $n\times n$ matrices $A_{j},B_{j},
$ $j=1,\dots ,r,$ then each of the vectors $X_{1},\dots ,X_{r}$ is normally distributed.*
In particular Theorem \[onlyif\] implies that if an isotropic random field is mean square continuous and the coefficients $a_{\ell m}, \ell=0,1,\dots, m=0,\dots \ell$ are independent, then it is gaussian.
Proposition \[onlyif\] shows that it is not possible to generate isotropic random fields by sampling non-Gaussian, independent complex-valued random variables $a_{\ell m},$ $m=-\ell,\dots,\ell$. This fact shows that, apart from the gaussian case, it is not easy to sample a random field by simulating the values of the random coefficients $a_{\ell m}$. In particular sampling independent values of the $a_{\ell m}$’s with distributions other than the gaussian gives not rise to an isotropic random field.
We wish also to point out that it is indeed possible to construct a non-gaussian random field by choosing the random coefficients $a_{\ell m}$, $\ell=0,1,\dots$ $m=0,\dots,\ell$ independent and with an arbitrary distribution. If they satisfy the conditions of Proposition \[prop1\] and the series (\[develop\]) converges, they certainly define a random field on $S^2$. But Theorem \[onlyif\] states that such a field cannot be isotropic. In particular Theorem \[onlyif\] does not follow by any means from the central limit theorem.
[**Acknowledgment.**]{} The authors wish to thank Professor R. Varadarajan for a very illuminating discussion and useful remarks.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Microservice Architecture (MSA) is a novel service-based architectural style for distributed software systems. Compared to Service-oriented Architecture (SOA), MSA puts a stronger focus on self-containment of services. Each microservice is responsible for realizing exactly one business or technological capability that is distinct from other services’ capabilities. Additionally, on the implementation and operation level, microservices are self-contained in that they are developed, tested, deployed and operated independently from each other.
Next to these characteristics that distinguish MSA from SOA, both architectural styles rely on services as building blocks of distributed software architecture and hence face similar challenges regarding, e.g., service identification, composition and provisioning. However, in contrast to MSA, SOA may rely on an extensive body of knowledge to tackle these challenges. Thus, due to both architectural styles being service-based, the question arises to what degree MSA might draw on existing findings of SOA research and practice.
In this paper we address this question in the field of Model-driven Development (MDD) for design and operation of service-based architectures. Therefore, we present an analysis of existing MDD approaches to SOA, which comprises the identification and semantic clustering of modeling concepts for SOA design and operation. For each concept cluster, the analysis assesses its applicability to MDD of MSA (MSA-MDD) and assigns it to a specific modeling viewpoint. The goal of the presented analysis is to provide a conceptual foundation for an MSA-MDD metamodel.
author:
-
-
bibliography:
- 'IEEEabrv.bib'
- 'literature.bib'
title: 'Analysis of Service-oriented Modeling Approaches for Viewpoint-specific Model-driven Development of Microservice Architecture'
---
Services Architectures, Services Engineering, Modeling of computer architecture
Introduction
============
Microservice Architecture (MSA) [@Newman2015] is an architectural style for service-based software systems. It emphasizes high cohesion, loose coupling and self-containment of services. A *microservice* denotes a component with well-defined interfaces that (i) provides a distinct, cohesive business or technological functionality to consumers; (ii) is autonomously developed and operated by one responsible team; (iii) comprises all technical artifacts for execution and deployment [@Nadareishvili2016; @Francesco2017]. Expected benefits of MSA adoption comprise (i) better flexibility of software system adaptation, as services are independently deployable and replacement is simplified; (ii) increased service quality and safety, because of isolated testability, higher resilience and runtime scalability; (iii) increased development team productivity as the architecture’s structure can be aligned to the team structure and each service being maintained by exactly one team [@Newman2015; @Nadareishvili2016].
Compared to Service-oriented Architecture (SOA) [@Erl2005], which is applied and studied for more than a decade, MSA is relatively young. It started to gain broad attention from practitioners and academia in 2014 and 2015, respectively [@Pahl2016]. However, with both architectural styles relying on services as building blocks for distributed software systems, they face similar challenges, e.g., in identifying, tailoring, composing and providing services [@Papazoglou2007]. Thus, it seems sensible to investigate findings from SOA research and practice with regard to their applicability to MSA. In this paper we focus on assessing the applicability of existing Model-driven Development (MDD) [@RodriguesDaSilva2015] approaches to SOA (SOA-MDD) in the context of MSA (MSA-MDD). MDD is a research area whose adoption to SOA has been object to extensive research [@Ameller2015].
Specifically, the paper aims at providing the conceptual foundation of a metamodel to subsequently implement an MSA-MDD modeling language. Such a language could facilitate design and operation of MSA-based software systems, e.g., by code generation [@Rademacher2017]. It therefore presents an analysis of existing approaches to SOA modeling that yields a threefold contribution. First, our analysis surveys existing approaches to SOA modeling to identify and characterize modeling concepts related to service design and operation. Second, the analysis clusters semantically equivalent concepts of different modeling approaches and identifies clusters applicable to MSA-MDD. Third, it assigns applicable clusters to three modeling viewpoints [@OMG2014] for MSA-MDD, i.e., Data, Service and Operation.
The paper is organized as follows. Section \[sec:soa-msa-mdd\] presents differences between SOA and MSA relevant to service-based MDD. Section \[sec:analysis\] elucidates the analysis. Subsection \[sub:analyzed-approaches\] first introduces the considered SOA modeling approaches. Next, Subsections \[sub:analysis-protocol\], \[sub:identified-concepts\] and \[sub:viewpoint-clustering\] present the analysis method and results, including SOA modeling concepts applicable to MSA-MDD. Section \[sec:discussion\] discusses these results. Section \[sec:related-work\] presents related work and Section \[sec:conclusion\] concludes the paper.
Distinguishing Characteristics of Service-oriented and Microservice Architecture Relevant to Model-driven Development {#sec:soa-msa-mdd}
=====================================================================================================================
MDD is a software engineering approach that considers models as means for abstracting the software to be built [@Combemale2017], as well as first-class citizens in the engineering process [@RodriguesDaSilva2015]. In particular, the development of complex software systems benefits from employing MDD [@France2007; @Whittle2014]. This is due to MDD providing *modeling concepts* on the problem-level that abstract from implementation details. Sets of coherent concepts define *metamodels*, i.e., abstract syntaxes for *modeling languages*, which enable the expression of all models conforming to the underlying metamodel [@RodriguesDaSilva2015]. Another cornerstone of MDD is *model transformation* [@Sendall2003], e.g., the automatic generation of implementation artifacts from models.
The application of MDD to SOA engineering is perceived to exhibit significant potential, because of the naturally high implementation complexity of distributed, service-based software systems [@Ameller2015; @France2007]. Hence, lots of effort has been spent on SOA-MDD research [@Ameller2015] and practice-oriented standardization [@Kreger2009]. SOA modeling puts a strong focus on generating code from models in the Development activity of software engineering [@Ameller2015], which is commonly expected to increase developer productivity [@Hutchinson2011]. SOA-MDD metamodels and languages comprise *modeling concepts* for services, interfaces, messages and ports, and are rather created from scratch than reused across the proposed MDD approaches [@Ameller2015].
While means for MSA-MDD might draw on the outlined knowledge about SOA-MDD, differences between SOA and MSA limit the applicability of existing SOA-MDD approaches to MSA [@Rademacher2017]. Thus, to substantiate the analysis of such approaches with the aim to identify SOA modeling concepts applicable to MDD of MSA design and operation (cf. Section \[sec:analysis\]), Table \[tab:dcs-soa-msa\] lists relevant *distinguishing characteristics* (DCs) of SOA and MSA investigated in a previous work [@Rademacher2017].
[>p[0.3cm]{}|>p[1.8cm]{}|p[15cm]{}]{} **\# & **Distinguishing Characteristic & **Peculiarity\
C1 & Service Granularity & MSA: Alignment of service functionality to a distinct business or technological capability. For business-related services, *bounded contexts* [@Evans2004] for clustering and isolation of related domain concepts may be applied. SOA: No explicit guidance.\
C2 & Interface Abstraction & MSA: Microservices and consumers typically have to use the same message formats/structures. SOA: Interaction of services with consumers using different message formats/structures is enabled by transformation capabilities of an Enterprise Service Bus (ESB).\
C3 & Protocols & MSA: Promotes to apply at most two different communication protocols, one for one-to-one and one for one-to-many service interactions. SOA: ESBs may implement protocol transformations, enabling hypothetic support for an arbitrary amount of protocols.\
C4 & Inter-service Interaction & MSA: For architecture-internal service interaction MSA prefers choreography over orchestration. SOA: SOA may equally apply both interaction patterns.\
C5 & Extra-service Interaction & MSA: For architecture-external interactions *API gateways* [@Balalaie2015], i.e, rather simple façades for abstracting services’ endpoints and granularity, are employed. In particular, they do not implement sophisticated message or protocol transformation means like ESBs.\
C6 & Application Scope & MSA: Mostly applied to (i) realize workflow-based applications with clear process flows; (ii) decompose monoliths with decreased scalability; (iii) realize web applications without generic message formats/structures or protocols. SOA: Typically applied in enterprise-wide or cross-enterprise systems with heterogeneous message formats/structures, protocols or middleware technologies.\
C7 & Practice Orientation & MSA: Higher perceived orientation towards practitioners due to (i) a reduced *service taxonomy* [@Richards2015]; (ii) less complex implementation technologies, e.g., API gateways instead of ESBs; (iii) interoperable frameworks for implementation and provisioning of business-related and various infrastructural MSA components; (iv) focus on communication means perceived as being “lightweight”, e.g., REST instead of SOAP [@Pautasso2008]; (v) teams’ freedom of choice regarding service technologies, i.e., *technology heterogeneity* [@Newman2015].\
C8 & Processes & MSA: Alignment of teams to features facilitates the application of *container-based DevOps* [@Kang2016] within development processes.\
******
Analysis of Service-oriented Modeling Approaches {#sec:analysis}
================================================
This section presents the protocol and results of our analysis of existing service-oriented modeling approaches. Therefore, Subsection \[sub:analyzed-approaches\] introduces the modeling approaches that were subject to our analysis. Following, Subsection \[sub:analysis-protocol\] presents the protocol we applied to perform our analysis, as well as general results of initial protocol steps. Subsection \[sub:identified-concepts\] presents the modeling concepts identified as being applicable to MSA-MDD and their semantic clustering. Subsection \[sub:viewpoint-clustering\] assigns applicable concepts to modeling viewpoints relevant to MSA-MDD, i.e., Data, Service and Operation.
Analyzed Service-oriented Modeling Approaches {#sub:analyzed-approaches}
---------------------------------------------
We considered ten existing SOA modeling approaches for our analysis. As our primary goal was to establish a conceptual foundation for subsequent deduction of an MSA-MDD metamodel that (i) targets two phases of MSA engineering, i.e., design and operation, and (ii) is preferably comprehensive, we selected modeling approaches that (i) aim at enabling holistic SOA modeling, i.e, define modeling concepts for multiple SOA engineering phases, and (ii) exhibit different degrees of abstraction and formality, i.e., comprise conceptual as well as practical elements. Specifically, our analysis comprises modeling approaches of the following types.
#### Reference Models {#reference-models .unnumbered}
A *reference model* is a conceptual framework that identifies relevant concepts of a given problem domain [@Kreger2009]. It further specifies relationships between concepts and is technology-independent.
#### Reference Architectures {#reference-architectures .unnumbered}
*Reference architectures* describe concepts of a problem domain with a focus on software architecture implementation [@Brown2012] and thus, compared to reference models, exhibit a lower degree of abstraction [@Kreger2009].
#### Modeling Languages and Profiles {#modeling-languages-and-profiles .unnumbered}
Modeling languages are based on metamodels that formally describe the structures of valid models expressed with the language (cf. Section \[sec:soa-msa-mdd\]). A metamodel may integrate a mechanism for tailoring or extending deduced modeling languages in *modeling profiles* [@OMG2014]. Because modeling languages and profiles rely on a metamodel, they exhibit a high degree of formality, i.e., modeling concepts have well-defined structures and relationships, which enable applications to automatically process models [@Combemale2017].
#### Architecture Description Languages {#architecture-description-languages .unnumbered}
An *architecture description language* (ADL) is a supportive modeling means for architecture-based software development [@Medvidovic2000]. In contrast to modeling languages and profiles, ADLs must comprise concepts for expressing architectural *components*, *connectors* and their *configurations*. ADL-based configuration modeling also covers *component composition* [@Medvidovic2000], which is a crucial characteristic of service-based architectures [@Papazoglou2007].
Table \[tab:selected-approaches\] lists the service-oriented modeling approaches, which we considered in our analysis. It also states, per approach, publication year, type with respect to abstraction and formality, foundational approaches if any and description.
[c|p[4cm]{}|>p[0.5cm]{}|>p[1.2cm]{}|>p[1.5cm]{}|p[8cm]{}]{} **\# & **Modeling Approach & **Year & **Type & **Foundational Approach & **Description\
A1 & Modeling and Design of Service-Oriented Architecture [@Stojanovic2004] & 2004 & Modeling Language & UML & The approach employs *service components* as SOA building blocks. Interface-based design and UML are used to express capabilities as Business Service Components (BSCs). BSCs are composed from other BSCs and Application Service Components (ASCs) that implement fine-grained operations. Component interaction is contract-based.\
A2 & A Modeling Framework for Service-Oriented Architecture [@Zhang2006] & 2006 & Modeling Language & 3C-modeling [@Szyperski2002] & Proposition of a metamodel with service components as first-class citizens of SOA modeling. The metamodel also comprises modeling concepts for expressing (i) required and provided ports to specify service interfaces; (ii) contracts; (iii) service choreographies.\
A3 & Reference Model for Service Oriented Architecture [@MacKenzie2006] & 2006 & Reference Model & n/a & The reference model identifies essential SOA concepts and their relationships. Next to concepts for service, contract and interaction modeling, it considers execution contexts and visibility of services.\
A4 & A platform independent model for service oriented architectures [@Benguria2007] & 2007 & Modeling Language & MDA [@OMG2014] & The paper introduces the PIM4SOA approach. The underlying metamodel is structured on the basis of the four aspects (i) Information, i.e., modeling of information elements the other aspects rely on; (ii) Service, i.e., technology-independent description of business capabilities; (iii) Processes for modeling message-based service interactions; (iv) Quality of Service (QoS) addressing non-functional aspects.\
A5 & A New Architecture Description Language for Service-Oriented Architecture [@Jia2007] & 2007 & ADL & XML & The presented SOADL language comprises concepts for modeling service interfaces, behavior, semantics and QoS-related aspects. Service composition is addressed by “port bindings”.\
A6 & Service-oriented Modeling Framework (SOMF) [@SOMF2011] & 2011 & Modeling Language & n/a & SOMF comprises a graphical modeling language for SOA-MDD. In addition to modeling concepts for services, interactions and compositions, it also considers modeling of cloud-based deployments as well as service, organizational and deployment boundaries.\
A7 & SOA Reference Architecture [@OpenGroup2011] & 2011 & Reference Architecture & [@OpenGroupOntology2009],approach A3 & The approach clusters SOA modeling concepts in Architectural Building Blocks (ABBs). ABBs are assigned to five functional and four non-functional, cross-cutting layers. Relationships between ABBs and hence modeling concepts are expressed as layered “interactions”.\
A8 & Reference Architecture Foundation for Service Oriented Architecture [@Brown2012] & 2012& Reference Architecture & approach A3 & Ascertainment of the reference model in approach A3 with concept structures and additional concepts. The approach clusters concepts in Service Ecosystem, Realizing SOAs and Owning SOAs viewpoints.\
A9 & Service oriented architecture Modeling Language (SoaML) Specification [@OMG2012] & 2012 & Modeling Profile & UML,approach A3 & The SoaML profile extends UML with concepts for SOA-MDD. It specifically provides means for sophisticated modeling of interfaces and interactions in the context of service-based software systems.\
A10 & Topology and Orchestration Specification for Cloud Applications (TOSCA) [@Palma2013] & 2013 & Modeling Language & XML & TOSCA specifies a metamodel for expressing service deployment and operation. Despite not exclusively targeting SOA, we analyzed TOSCA to strengthen the consideration of operation-related modeling concepts.\
************
Analysis Protocol and General Results {#sub:analysis-protocol}
-------------------------------------
In the following, we outline the protocol for our analysis of the approaches in Table \[tab:selected-approaches\] to identify service-oriented modeling concepts applicable to MSA-MDD and assign them to related modeling viewpoints. Furthermore, we present general results yielded by initial protocol steps.
The analysis protocol comprised the following steps:
1. \[step:1concept-extraction\] Identification and extraction of modeling concepts and concept-specific information from approach publications (cf. Table \[tab:selected-approaches\]).
2. \[step:2concept-characterization\] Further characterization of modeling concepts per approach by surveying extracted concept-specific information. This step comprised three sub-steps.
1. Identification of *relationships* to other modeling concepts of the respective approach.
2. Identification of *structures*, i.e, concepts’ properties not represented as relationships.
3. Identification of formal *constraints* that exceed multiplicity specifications for relationships and structures.
4. Survey of textual concept descriptions that, next to formally expressed relationships, structures or constraints, express further semantics or characteristics of concepts.
3. \[step:3concept-reduction\] Removal of concepts from the extracted set to which at least one of the following exclusion criteria applies:
- Concept is used in an SOA engineering phase other than design or operation.
- Concept enables advanced modeling of (i) architecture structure above service level; (ii) applicable concepts with lower abstraction level; (iii) governance; (iv) policies; (v) QoS; (vi) service-internal behavior.
- Concept lacks relationships, structure, constraints and its textual description is too generic or imprecise.
4. \[step:4semantical-clustering\] Identification and bundling of semantically equivalent remained concepts in *concept clusters* across approaches.
5. \[step:5identification-applicable-concepts\] Assessment of concept clusters’ applicability to MSA-MDD on the basis of SOA and MSA DCs (cf. Table \[tab:dcs-soa-msa\]).
6. \[step:6viewpoint-clustering\] Identification of modeling viewpoints for MSA-MDD and assignment of applicable concept clusters to them.
The following paragraphs describe the executions of the initial steps \[step:1concept-extraction\] to \[step:3concept-reduction\] and their general results. The executions and results of the main protocol steps \[step:4semantical-clustering\] to \[step:6viewpoint-clustering\] are covered in more detail in Subsections \[sub:identified-concepts\] and \[sub:viewpoint-clustering\].
In step \[step:1concept-extraction\] we performed a full reading of the approach publications listed in Table \[tab:selected-approaches\] whereby we identified and extracted 434 modeling concepts. Next, step \[step:2concept-characterization\] yielded that of these concepts (i) 268 have relationships to others; (ii) 93 exhibit formal structure specifications; (iii) 15 comprise formal constraints not expressed in relationship or structure specifications. Table \[tab:numbers-per-approach\] breaks down these numbers per approach.
For reasons of space we do not present the raw results of steps \[step:1concept-extraction\] to \[step:3concept-reduction\], but provide them as supplemental material[^1].
[c|>p[1cm]{}|>p[1.5cm]{}|>p[1.2cm]{}|>p[1.3cm]{}]{} **\# & **Concept Count & **... with Relationships & **... with Structures & **... with Constraints\
A1 & 10 & 7 & 2 & 0\
A2 & 15 & 10 & 3 & 0\
A3 & 25 & 21 & 2 & 0\
A4 & 35 & 35 & 4 & 0\
A5 & 24 & 0 & 11 & 0\
A6 & 45 & 0 & 0 & 0\
A7 & 12 & 10 & 0 & 0\
A8 & 157 & 92 & 4 & 0\
A9 & 34 & 27 & 18 & 15\
A10 & 77 & 48 & 49 & 0\
**$\Sigma$ & **434 & **250 & **93 & **15\
********************
In step \[step:3concept-reduction\] we applied the mentioned exclusion criteria to each of the extracted and characterized modeling concepts. Hence, we filtered out concepts that do not support design and operation, but other phases of the SOA engineering process. For example, these phases comprised requirements’ elicitation with concepts like `Non-Function Requirement` of approach A7 (cf. Table \[tab:selected-approaches\]) and business modeling with concepts like `Business-Goal Use Case` from approach A1 or `MotivationElement` from approach A9.
We further removed concepts from the extracted set for modeling of (i) architecture structure above service level, e.g., `SubArchitecture` from approach A5 or `InterCloud` from approach A6; (ii) concepts at a lower abstraction level without being directly employable for SOA design or operation themselves, e.g., `RelationshipType` from approach A10; (iii) governance, e.g., `SOA Governance` from approach A8; (iv) policies, e.g., `Service Policy` from approach A3 or `Policy` from approach A10; (v) QoS, e.g., `QoSCharacteristic` form approach A4; (vi) internal service behavior, e.g., `Behavior` from approach A5.
The extracted concept set was additionally reduced by removing modeling concepts that did not exhibit relationships to other concepts of their defining approach, structures or constraints in combination with a too generic or imprecise textual description. The approaches A3 and A8 comprise the majority of such concepts, e.g., `Real World Effect` and `Risk`. However, even publications of modeling languages, whose formality is expected to be high (cf. Subsection \[sub:analyzed-approaches\]), partially contain concept descriptions, which were insufficient for our analysis, e.g., `Context` from approach A2.
After finishing the execution of concept reduction step \[step:3concept-reduction\], 100 concepts remained as inputs for protocol steps \[step:4semantical-clustering\] and \[step:5identification-applicable-concepts\]. Figure \[fig:selected-concepts-share\] relates the count of those concepts to the overall concept count per approach. The first percentage under each approach is the share of remained in the overall concept count.
table\[x=approach,y=conceptCount\]; table\[x=approach,y=selectedCount\]; table\[x=approach,y=applicableCount\];
Semantic Clustering of Concepts and Assessment of Clusters’ Applicability to Modeling for Microservice Architecture {#sub:identified-concepts}
-------------------------------------------------------------------------------------------------------------------
The execution of protocol step \[step:4semantical-clustering\] (cf. Subsection \[sub:analysis-protocol\]) included the encapsulation of modeling concepts, which remained after finishing protocol step \[step:3concept-reduction\], in concept clusters across approaches. The clustering was based on the semantics of these modeling concepts, as defined by their characteristics identified in step \[step:2concept-characterization\]. Next, each concept’s applicability to MSA-MDD was assessed in step \[step:5identification-applicable-concepts\].
The results of steps \[step:4semantical-clustering\] and \[step:5identification-applicable-concepts\] are shown in Tables \[tab:fully-applicable-concepts\] and \[tab:partially-applicable-concepts\]. Table \[tab:fully-applicable-concepts\] lists concept clusters in which all concepts were assessed as being fully applicable to MSA-MDD. On the other hand, Table \[tab:partially-applicable-concepts\] includes all clusters that, next to fully applicable concepts, comprise at least one concept assessed as being partially applicable to MSA-MDD. Both tables exhibit an “Approaches” column, in which they state the modeling approaches whose concepts are part of the respective cluster. The column further states concepts’ names that differ in their defining approaches from the respective cluster’s name. Additionally, Table \[tab:partially-applicable-concepts\] differentiates partially (+) from fully (++) applicable modeling concepts, if both peculiarities of applicability occur within a concept cluster. Otherwise all concepts of a cluster in Table \[tab:partially-applicable-concepts\] were assessed as being partially applicable to MSA-MDD. For each concept assessed as being partially applicable, the “Comments” column of Table \[tab:partially-applicable-concepts\] contains a justification of the assessment that, if necessary, is based on the DCs of SOA and MSA (cf. Table \[tab:dcs-soa-msa\]).
------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- --
**\# & **Cluster & **Approaches (cf. Table \[tab:selected-approaches\]) & **\# & **Cluster & **Approaches\
1 & Application Service & A5 & 17 & Message Exchange Pattern & A8\
2 & Architectural Service & A5 & 18 & Message Structure & A8: Structure\
3 & Artifact & A10: ArtifactTemplate, ArtifactType & 19 & MessageType & A9\
4 & Atomic Service & A5: Atomic Service A6: Analysis Atomic Service & 20 & Network Management & A8\
5 & Authentication & A8 & 21 & OutMessage & A5\
6 & Awareness & A3 & 22 & Plan & A10\
7 & Business Service & A5 & 23 & Port & A2, A9: Port A5: Ports\
8 & Categorization & A6: Service Typing Tag A9: CategoryValues & 24 & Protocols & A8\
9 & Consumer & A6, A9: Consumer A8: Service Consumer & 25 & Provider & A4: ServiceProvider A8: Service Provider A9: Provider\
10 & EndPoint & A4 & 26 & Reachability & A3: Reachability A8: Service Reachability\
11 & Execution Context & A3, A8: Execution Context A9: ServicesArchitecture & 27 & RequirePort & A2: RequirePort A9: Request\
12 & FaultMessage & A5 & 28 & Security Management & A8\
13 & Infrastructure & A8 & 29 & Service Description & A3, A8\
14 & InMessage & A5 & 30 & Specification & A1\
15 & Manageable Resource & A8 & 31 & Structure & A3, A8\
16 & Message Exchange & A8 & 32 & Visibility & A3\
************
------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- --
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- --
**\# & **Cluster & **Approaches (cf. Table \[tab:selected-approaches\]) & **Comments on partial Applicability\
33 & Collaboration & A4, A9: Collaboration A4, A9: CollaborationUse & Modeling of complex service collaborations with roles. MSA has a simpler view on service interaction (cf. DCs C4 and C5 in Table \[tab:dcs-soa-msa\]).\
34 & Composite Service & A5: Composite Service A6: Analysis Composite Service A8: Service & Sophisticated means for composing sets of fine-grained services with heterogeneous granularities to coarse-grained services. MSA facilitates composition as it proposes to align services to a self-contained, distinct functionality (cf. DC C1).\
35 & Enabling Technology & A7 & MSA relates technologies to services rather than architectural layers (cf. DC C7).\
36 & Information Model & A3 (++), A7 (++), A8 (+) & A8: In MSA-MDD, service models may not need to comprise abstracted message semantics for transforming exchange formats or structures (cf. DC C2).\
37 & Message & A4 (+), A5 (+), A8 (++) & A4: MSA exhibits a reduced application scope and service taxonomy not requiring role-based interaction modeling (cf. DCs C6, C7). A5: Microservices and consumers typically agree on the employed message formats/structures (cf. DCs C2). Modeling of semantics may hence not be necessary.\
38 & NodeTemplate & A10 & The `minInstances` and `maxInstances` properties denote concept parts applicable to modeling of container-based microservice deployments (cf. DC C8).\
39 & NodeTypeImplementation & A10 & When associated with the concept, the semantics of the `DeploymentArtifact` and `ImplementationArtifact` concepts are applicable to modeling of container-based microservice deployments (cf. DC C8).\
40 & Participant & A9 & Exhibits a high generality. Concept parts related to services are predominantly applicable to designing, e.g., microservice interfaces and contracts.\
41 & ProvidePort & A2: ProvidePort (++) A9: Service (+) & A9: Abstract, technology-independent semantics and constraint models of service provisioning are not mandatory for MSA design and operation (cf. DCs C2, C3).\
42 & Service Component & A1: Service Component A2: ServiceComponent & Both concepts provide, among others, means for sophisticated modeling of role-based service interaction, typically not applied by MSA (cf. DCs C4, C5).\
43 & ServiceContract & A1: Contract A2, A9: ServiceContract A3: Contract & Policy & Concepts comprise conceptual or concrete parts for modeling complex service collaborations with respect to roles. MSA usually does not realize such complex collaborations (cf. DCs C4, C5).\
44 & Service Interface & A2: InterfaceDeclaration (+) A9: Simple Interface (++) & A2: Service interfaces may not need abstract property or constraint descriptions for transforming message formats/structures or protocols (cf. DCs C2, C3, C6).\
45 & Service Operation & A1 (+) A3: Service (++) A5, A8: Operation (++) & A1: Relationship to `Business-Goal Use Case` concept is not a mandatory prerequisite for MSA design and operation.\
46 & Solution Building Block & A7 & Typically, MSA relates technology to services, not architectural layers (cf. DC C7).\
47 & Technical Assumptions & A8 & MSA-MDD may not need to consider modeling of physical limitations, e.g., flow speeds or disk access speeds (cf. DCs C6, C7)\
48 & Usage Management & A8 & Financial resource modeling on the service level may not be needed for MSA-MDD.\
********
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- --
In total, we assessed 80 SOA modeling concepts as fully or partially applicable to MSA-MDD with 54 modeling concepts been being assessed fully applicable (68%). Figure \[fig:selected-concepts-share\] also depicts the counts of applicable concepts per approach, as well as their shares in the overall concept count of the respective approach (second percentage under each approach).
Identification of Viewpoints for Microservice Architecture Modeling and Assignment of Applicable Concept Clusters {#sub:viewpoint-clustering}
-----------------------------------------------------------------------------------------------------------------
The last protocol step \[step:6viewpoint-clustering\] of the analysis of the service-oriented modeling approaches comprised the identification of modeling viewpoints for MDD of MSA design and operation. A *modeling viewpoint* provides a certain type of stakeholder with appropriate criteria to construct, select or present information about a system [@OMG2014]. It hence reduces the complexity practitioners of MDD have to deal with. As microservice teams usually apply DevOps [@Kang2016], typical MSA stakeholder types are *service developer* and *operator*. Hence, we defined a viewpoint for each of these types. The *Service viewpoint* for service developers comprises concept clusters from Tables \[tab:fully-applicable-concepts\] and \[tab:partially-applicable-concepts\], which focus modeling of microservices. The *Operation viewpoint* encapsulates clusters for specifying aspects of service deployment and operation. It thus addresses service operators.
Next to these viewpoints for typical MSA stakeholder types, we also included a *Data viewpoint*. Its addition was perceived to be sensible while executing protocol step \[step:3concept-reduction\], because the analyzed approaches partially comprise explicit concepts for modeling information used by services. For example, `Structure` from approach A3 (cluster 31 in Table \[tab:fully-applicable-concepts\]) enables expression of structural data types. While the number of concepts in the Data viewpoint is comparatively small, their semantics are extensive, e.g., as for `Information Model` from approaches A3, A7 and A8 (cluster 36 in Table \[tab:partially-applicable-concepts\]).
Based on the described semantics and scopes of the viewpoints, we assigned the concept clusters from Tables \[tab:fully-applicable-concepts\] and \[tab:partially-applicable-concepts\] to them as shown in Table \[tab:viewpoints\].
--------------------------------------------------------------------------------------------------------- -- --
**Viewpoint & **Concept Clusters & **Approaches\
Data & 31, 36 & A3, A7, A8\
Service & 1, 2, 4, 7–9, 12, 14, 16–19, 21, 23, 27, 29, 33, 34, 37, 40–45 & A1–A6, A8, A9\
Operation & 3, 5, 6, 10, 11, 13, 15, 20, 22, 24–26, 28, 30, 32, 35, 38, 39, 46–48 & A1, A3, A4, A7–A10\
******
--------------------------------------------------------------------------------------------------------- -- --
: MSA-MDD Viewpoints With Assigned Concept Clusters And Their Defining Approaches.[]{data-label="tab:viewpoints"}
Figure \[fig:applicable-concepts-share\] shows for each viewpoint the count and share of assigned concept clusters.
=\[font=\]
The Data viewpoint comprises concept clusters for modeling data structures (cluster 31) and their integration in services’ information models (36).
The Service viewpoint’s concept clusters enable modeling of (i) services and basic service roles (1, 2, 4, 7–9, 23, 27, 29, 40, 41); (ii) message types and structures (12, 14, 16–19, 21, 37, 42); (iii) interfaces and contracts (43–45); (iv) collaborations and compositions (33, 34).
The Operation viewpoint encapsulates concept clusters for modeling (i) service artifacts (3, 22); (ii) technical infrastructure (5, 11, 13, 15, 20, 28, 35, 38, 39, 46–48); (iii) service provisioning (6, 10, 24–26, 30, 32).
Discussion {#sec:discussion}
==========
Subsections \[sub:metamodel-deduction\] and \[sub:threats\] discuss the impacts of the presented analysis results on subsequent metamodel deduction and elaborate on threats to the analysis’s validity.
Deduction of a Viewpoint-specific Metamodel for Microservice Architecture Modeling {#sub:metamodel-deduction}
----------------------------------------------------------------------------------
The results of our analysis provide the basis for deducing a viewpoint-specific metamodel for MSA-MDD with a focus on microservice design and operation. The concept clusters shown in Tables \[tab:fully-applicable-concepts\] and \[tab:partially-applicable-concepts\] may become initial concepts of the metamodel. Furthermore, it would be necessary to identify those characteristics of the applicable concepts (cf. Subsection \[sub:analysis-protocol\]) that may be adopted to the metamodel.
Another aspect of the metamodel’s deduction is the linkage of the identified viewpoints. For example, it is conceivable that the Service viewpoint needs to refer to the Data viewpoint so that message models may use data structures as types, or that Service and Operation viewpoint have to be associated to enable modeling of microservice deployment. Viewpoint linkage may be achieved by defining relationships between metamodel concepts of different viewpoints, either based on adopted concept characteristics or the introduction of intermediate concepts.
Additionally, we expect the occurrence of contradictory overlaps in concept semantics when deducing the metamodel. In this case it has to be decided for each concept identified as being applicable to MSA-MDD in our analysis (cf. Section \[sec:analysis\]) how such inconsistencies may be dissolved on the metamodel level. It is further likely that the deduction process will reveal clusters that can be merged with others to comprehensively cover an aspect relevant to MSA design or operation in the metamodel. For example, the clusters `InMessage` and `OutMessage` (cf. Table \[tab:fully-applicable-concepts\]) may be represented by a property `direction` of a superior `Message` metamodel concept.
To cope with the mentioned challenges, we plan to deduce the metamodel in an iterative process, which continuously yields a more consistent version of the metamodel.
Threats to Validity {#sub:threats}
-------------------
Our analysis is affected by the following threats to validity.
#### T1 Incomplete Selection of Modeling Approaches {#t1-incomplete-selection-of-modeling-approaches .unnumbered}
The analysis focused on surveying modeling approaches that take a holistic view on SOA-MDD (cf. Subsection \[sub:analyzed-approaches\]). However, it may be possible that we accidentally did not consider additional approaches with possible relevance to MSA design and operation. From our perspective, the impact of this threat is mitigated by the fact that Tables \[tab:fully-applicable-concepts\] and \[tab:partially-applicable-concepts\] comprise clusters that directly correspond to the most recurrent metamodel concepts for modeling SOA design, i.e., `Service`, `Operation`, `Message` and `Port` [@Ameller2015]. In addition, all of these concepts occur in more than one analyzed approach, which makes us confident that they are central to service-oriented modeling in general. Considering the modeling of MSA operation, the threat’s impact is likely to be lowered by the fact that the cluster count of the Operation viewpoint exhibits a scale similar to that of the Service viewpoint (cf. Figure \[fig:applicable-concepts-share\]).
#### T2 Missing Applicable Concepts {#t2-missing-applicable-concepts .unnumbered}
Despite the fact that we are confident to have captured essential modeling concepts applicable to MSA design and operation, we may have missed other relevant concepts of the analyzed approaches. As we executed reduction step \[step:3concept-reduction\] of our protocol (cf. Subsection \[sub:analysis-protocol\]) manually, it is possible that we accidentally filtered out modeling concepts with potential applicability to MSA-MDD. We tried to mitigate this threat’s impact by double checking the results of the reduction step, as well as discussing and jointly deciding on edge cases. Furthermore, when deducing the metamodel, we plan to review further approaches, which specifically focus on a certain aspect of service-oriented modeling, to detail concept characteristics or identify additional concepts. This will apparently become necessary for the Data viewpoint, as it comprises comparatively few concept clusters (cf. Figure \[fig:applicable-concepts-share\]).
#### T3 Lack of Practical Applicability {#t3-lack-of-practical-applicability .unnumbered}
MSA exhibits a high practice orientation (cf. DC C7 in Table \[tab:dcs-soa-msa\]). However our analysis also considered modeling approaches with a rather theoretical focus, i.e., A3, A7 and A8. Hence, concept clusters may comprise elements with a comparatively low or even non-existent practical relevance. For now, we accept that this threat’s impact may result in superfluous metamodel concepts. However, we plan to identify and remove such concepts irrelevant to practical MSA design and operation from the metamodel by surveying (i) related experience reports; (ii) related solution proposals; (iii) MSA developers and operators.
Related Work {#sec:related-work}
============
In the following, we present work related to analyzing SOA modeling approaches and metamodeling for MSA-MDD.
In 2013, Mohammadi et al. published a review of SOA modeling approaches for Enterprise Information Systems [@Mohammadi2013]. The review comprised seven modeling approaches and included the approaches A3, A4, A6, A8 and A9 of our analysis (cf. Table \[tab:selected-approaches\]). The two remaining approaches are the SOA ontology [@OpenGroupOntology2009], on which approach A7 is based, and SOMA [@Arsanjani2008], which leverages SoaML [@OMG2012] for service modeling, i.e., approach A9 of our analysis. Furthermore, the review is mainly focused on identifying and summarizing the major features of the considered modeling approaches. It does not analyze specific modeling concepts. Additionally, MSA in general and aspects of modeling SOA operation are not explicitly covered.
Ameller et al. published results of a comprehensive mapping study on SOA-MDD based on 129 papers in 2015 [@Ameller2015]. Among others, one of their research goals was to investigate characteristics of SOA-MDD approaches (cf. Subsection \[sub:analysis-protocol\]) and the study reported on the seven most recurrent metamodel concepts in SOA-MDD. While considering a wide range of scientific publications for this purpose, the study did not involve an assessment of the concepts’ applicability based on their semantics. Furthermore, MSA and modeling of SOA operation were out of the study’s scope. However, papers from the study addressing modeling of SOA design may be surveyed for populating an initial MSA-MDD metamodel with further concepts or refining existing concepts. This would mitigate the impact of threat T2 (cf. Subsection \[sub:threats\]).
AjiL [@Sorgalla2017] is a graphical language for MSA-MDD. It comprises basic concepts for modeling (i) data structures; (ii) microservices and their interfaces; (iii) technical infrastructure. However, the metamodel lacks several essential concepts of SOA modeling [@Ameller2015] shown in Tables \[tab:fully-applicable-concepts\] and \[tab:partially-applicable-concepts\]. For example, it does not comprise concepts for specifying (i) service contracts and messages; (ii) protocols and endpoints; (iii) artifact and security technologies. Furthermore, AjiL does not define modeling viewpoints. Despite these shortcomings, AjiL exhibits a high practice orientation, as it is based on publicly available MSA implementations. Hence, we will review AjiL when deducing a metamodel on the basis of our analysis to raise its practical applicability and mitigate the impact of threat T3.
Conclusion and Future Work {#sec:conclusion}
==========================
This paper presented an analysis of existing SOA modeling approaches with the aim to identify modeling concepts applicable to viewpoint-specific MDD of MSA design and operation. Therefore, we first elucidated differences between SOA and MSA relevant to MDD (cf. Section \[sec:soa-msa-mdd\]). Next, we introduced the ten SOA modeling approaches (cf. Subsection \[sub:analyzed-approaches\]), which were analyzed following a rigorous protocol (cf. Subsection \[sub:analysis-protocol\]). Its first three steps yielded 100 modeling concepts possibly applicable to MDD of MSA design and operation. The applicability was then assessed considering SOA and MSA differences (cf. Subsection \[sub:identified-concepts\]). This resulted in 48 applicable concept clusters, being assigned to three MSA-MDD viewpoints (cf. Subsection \[sub:viewpoint-clustering\]). The analysis’s discussion covered, among others, challenges of deducing a metamodel from the results (cf. Section \[sec:discussion\]).
In future works we plan to define a metamodel on the basis of the presented results. Our goal is to implement a modeling language for the metamodel to provide practitioners with generative means for microservice design and operation.
[^1]: Link to raw results of protocol steps \[step:1concept-extraction\] and \[step:2concept-characterization\]: <https://fh.do/seaa2018-sm>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Can we predict the future popularity of a song, movie or tweet? Recent work suggests that although it may be hard to predict an item’s popularity when it is first introduced, peeking into its early adopters and properties of their social network makes the problem easier. We test the robustness of such claims by using data from social networks spanning music, books, photos, and URLs. We find a stronger result: not only do predictive models with peeking achieve high accuracy on all datasets, they also generalize well, so much so that models trained on any one dataset perform with comparable accuracy on items from other datasets.
Though practically useful, our models (and those in other work) are intellectually unsatisfying because common formulations of the problem, which involve peeking at the first small-$k$ adopters and predicting whether items end up in the top half of popular items, are both too sensitive to the speed of early adoption and too easy. Most of the predictive power comes from looking at how quickly items reach their first few adopters, while for other features of early adopters and their networks, even the direction of correlation with popularity is not consistent across domains. Problem formulations that examine items that reach $k$ adopters in about the same amount of time reduce the importance of temporal features, but also overall accuracy, highlighting that we understand little about why items become popular while providing a context in which we might build that understanding.
author:
- |
Benjamin Shulman\
Dept. of Computer Science\
Cornell University\
[email protected] Amit Sharma\
Microsoft Research\
New York, NY\
[email protected] Dan Cosley\
Information Science\
Cornell University\
[email protected]
bibliography:
- 'main.bib'
title: 'Predictability of Popularity: Gaps between Prediction and Understanding'
---
How does a book, song, or a movie become popular? The question of how cultural artifacts spread through social networks has captured the imagination of scholars for decades. Many factors are cited as important for an item to spread *virally* through social networks and become popular: its intrinsic quality [@gladwell2006-formula; @simonoff2000], the characteristics of its initial adopters [@gladwell2006], the emotional response it elicits [@berger2012], and so on. Often, explanations are used to justify the popularity of different items after the fact [@berger2013book], making it hard to apply these explanations to new events [@watts2011].
Online social networks allow us to observe individual-level traces of how items are transferred between people, allowing more precise modeling of the phenomenon. Predicting the future popularity of an item based on attributes of the item and the person who introduced it has emerged as a useful problem, both to understand processes of information diffusion and to inform content creation and feed design on social media platforms. For example, Twitter’s managers may want to highlight new tweets that are more likely to become popular, while its users may want to learn from characteristics of popular tweets to imporve their own.
In general, even with detailed information about an item’s content or the person sharing it, it is hard to predict which items will become more popular than others [@bakshy2011; @martin2016]. The problem becomes more tractable when we are allowed to *peek* into the initial spread of an item. The intuition is that early activity data about the speed of adoption, characteristics of people who adopt it and the connections between them might predict the item’s fate. This intuition shows encouraging results for both predicting the final popularity of an item [@szabo2010; @pinto2013; @zhao2015] and whether an item will end up in the top 50% of popular items [@cheng2014; @romero2013; @weng2013].
Buoyed by these successes, one might conclude that the availability of rich features about the item and social network of early adopters has helped us understand why items become popular. However, past work studies individual datasets and varying versions of the prediction problem, making it hard to compare results. For instance, studies disagree on the direction of the effect of network structural features on item popularity [@lerman2010; @romero2013].
In this paper, we try to unify these observations on popularity prediction through studying different problem formulations and kinds of features over a wide range of online social networks. Using an existing formulation that predicts whether the final popularity of items is above the median based on features of the first five adopters [@cheng2014], we confirm past work [@szabo2010] showing that features about those adopters and their social network are at best weak predictors of popularity compared to temporal features. For instance, a single temporal heuristic—the average rate of early adoption—is a better predictor than all non-temporal features combined across all four websites. Further, models trained on one dataset and tested on others using temporal features generalize fairly well, while those that use network structural features generalize badly.
In one reading, this is a useful contribution: peeking-based popularity models that include temporal information achieve up to 83% accuracy on Twitter and generalize well across datasets. From a practical standpoint, we encourage content distributors to use temporal features for predicting the future success of items.
Intellectually, however, our finding is not very satisfying. Rather than identifying features that shed light on why items become popular, we mostly see that items that become popular fast are more likely to achieve higher popularity in the end. Rapid adoption may be a signal of quality, interestingness, and eventual popularity—but doesn’t tell us why. The effect might also be driven by cumulative advantage [@frank2010winner; @watts2011]: items that receive attention early have more chances to spread through via interfaces that highlight popular or trending items.
An alternative formulation of the problem that reduces the effect of temporal features lets us see just what early adopter and network features tell us. This formulation, called Temporal Matching, compares items that achieve similar levels of popularity in the same amount of time, rather than the more common formulation of looking at the first $k$ adopters regardless of the time it takes to reach $k$. Controlling for the average rate of an item’s adoption turns popularity prediction into a hard problem. Using the same features as before, prediction accuracy across all datasets drops below 65%. Such a decrease in accuracy underscores the importance of choosing problem formulations that highlighting relevant phenomena in popularity evolution. Current models may fare well on certain formulations, but there is still much to learn about how items become popular.
---------------------- ------------------------- ------------- ---------------- -------------------- --------------
**Study** **Problem Formulation** **Content** **Structural** **Early Adopters** **Temporal**
Bakshy et al. (2011) Regression (no peeking) n – **Y** –
Martin et al. (2016) Regression (no peeking) n – **Y** –
Szabo et al. (2010) Regression – n – **Y**
Tsur et al. (2012) Regression **Y** **Y** – **Y**
Pinto et al. (2013) Regression – – – **Y**
Yu et al. (2015) Regression – n – **Y**
Romero et al. (2013) Classification – **Y** – –
Cheng et al. (2014) Classification n **Y** **Y** **Y**
Lerman et al. (2008) Classification – **Y** **Y** –
Weng et al. (2013) Classification – **Y** n –
---------------------- ------------------------- ------------- ---------------- -------------------- --------------
Formulations of the prediction problem {#sec:problem-formulation}
======================================
We start by identifying two key dimensions to consider when defining the popularity prediction task: how much peeking into early activity on an item is allowed, and whether the task is a regression or classification. For ease of exposition, we use *item* to denote entities that are consumed in online social networks. *Adoption* refers to an explicit action or endorsement of an item, such as loving a song, favoriting a photo, rating a book highly or retweeting a URL. Finally, we define *popularity* of an item as the number of people who have adopted it.
Predicting apriori versus peeking into early activity
-----------------------------------------------------
Predicting popularity *a priori* for items such as movies [@simonoff2000] or songs [@pachet2012] has long been considered a hard problem. One of the most successful approaches has been to gauge audiences’ interest in an item before it is officially released, such as by measuring the volume of tweets [@asur2010] or search queries [@goel2010]. Such methods can work well for mainstream, popular items for which there might be measurable prior buzz, but are unlikely to be useful for genuinely new items such as tweets or photos uploaded by users.
For such items, popularity prediction is tricky, even when precise data about the content of each tweet and the seed user’s social network is known. On Twitter, models with extensive content features such as the type of content, its source and topic, crowdsourced scores of interestingness, and features about the seed user such as indegree and past popularity of tweets are only able to explain less than half of the variance in popularity [@martin2016]. Further, the content features are usually less important than features of the seed user [@bakshy2011; @martin2016; @jenders2013]. In response, scholars have suggested modified versions of the problem where one peeks into early adoption activity for an item. In studies on networks including Facebook [@cheng2014], Twitter [@lerman2010; @zhao2015; @tsur2012; @kupavskii2013], Weibo [@yu2015], Digg [@lerman2010; @szabo2010] and Youtube [@pinto2013], early activity data consistently predicts future popularity with reasonable accuracy. In light of these results, we focus on the peeking variant of the problem in this paper.
Classification versus regression
--------------------------------
In addition to how much data we look at, we must also specify what to predict. A number of studies have used regression formulations, predicting an item’s exact final popularity: the number of retweets for a URL [@bakshy2011], votes on a Digg post [@lerman2010] or page views of a Youtube video [@szabo2010]. However, we may often be more interested in popularity relative to other items rather than an exact estimate. For example, both marketers and platform owners may want to select ‘up and coming’ items to feature in the interface versus others[^1].
These motivations lead nicely to a classification problem where the goal is to predict whether an item will be more popular then a certain percentage of other items. For instance, Romero et al. predict whether the number of adopters of a hashtag on Twitter will double, given a set of hashtags with the same number of initial adopters [@romero2013]. Cheng et al. generalize this formulation to show that predicting whether an item will double its popularity is equivalent to classifying whether an item becomes more popular than the median and study this question in the case of Facebook photos that received at least five adopters [@cheng2014]. Besides the practical appeal of classifying popular items, classification is also a simpler task than predicting the actual number of adoptions [@bandari2012], thus providing a favorable scenario for evaluating the limits of predictability of popularity. Therefore, we focus on the classification problem in this paper.
Our problem: Peeking-based classification
-----------------------------------------
Based on the above discussion, the general peeking-based classification problem can be stated as:
> **P1:** *Given a set of items and data about their early adoptions, which among them are more likely to become popular?*
This question has a broad range of formulations based on how we define the early activity period, how much activity we are allowed to poke at, and how we define *popular*. The early activity period may be defined in terms of time elapsed $t$ since an item’s introduction [@szabo2010], or in terms of a fixed number $k$ of early adoptions [@romero2013]. Fixing the early activity period in terms of number of adoptions has the useful side-effect of filtering out items with less than $k$ adoptions overall, both making the problem harder and eliminating unpopular (thus often uninteresting) items. For this reason, most past work on peeking-based classification defines early activity in terms of the number of adoptions $k$.
The popularity threshold for what is “popular” may also be set at different percentiles ($n\%$). Table \[tab:past-features\] summarizes past work based on their choices of problem formulation and choice of $(k, n)$. One common approach is to collect all items that have $k$ or more adoptions, then peek into the first $k$ adoptions and predict whether eventual popularity of items lies above or below the median [@cheng2014]. We call this Balanced Classification since there are guaranteed to be an equal number of high and low popularity items. Another variation is to only consider the top-*n* percentile of items as high popularity [@lerman2008], a formulation that is arguably better-aligned with most use cases around content promotion than Balanced Classification. However, it is also harder than Balanced Classification; for this reason, and to continue to align with prior work, we focus on Balanced Classification.
While restricting to items with $k$ adoptions helps to level the playing field because it provides a set of comparably popular items to study, it ignores the *time taken* to reach $k$ adoptions. Based on prior work, our suspicion is that in this formulation temporal features dominate the others. To control for this temporal signal, we later introduce a problem formulation where both $k$ and $t$ are fixed. That is, we collect all items that received exactly $k$ adoptions in a given time period $t$, and then predict which of them would be in the top half of popular items. We call this the Temporally Matched Balanced Classification problem, and as we will see, changing the definition has a profound impact on the quality of the models.
Choosing features {#sec:features}
=================
We now turn to the selection of features for prediction. Part of the allure of modeling is that the features that prove important might give information about *why* some items become popular in ways that could be both practically and scientifically interesting. Features used in prior work can be broadly grouped into four main categories: content, structural, early adopters and temporal [@cheng2014]. Table \[tab:past-features\] shows which feature categories were used in prior studies, with cells in bold representing features that were reported to be useful for prediction. While all feature categories have been reported to be important contributors to prediction accuracy in at least some studies, temporal and structural features are frequently reported as important.
Temporal patterns of early adoption—how quickly the early adopters act—are a major predictor of popularity. Szabo and Huberman show that temporal features alone can predict future popularity reliably [@szabo2010]. When information about the social network or its users is hard to obtain, utilizing temporal features can be fruitful, achieving error rates as low as 15% in a regression formulation [@pinto2013; @zhao2015]. A natural next question is to ask how much these errors can be decreased by adding other features when we do have such information.
Features about the seed user and early resharers—collectively called early adopters—also matter. On Twitter, for example, the number of followers of the seed user and the fraction of her past tweets that received retweets increase the accuracy of predictions [@tsur2012]. Information about other early adopters is also useful for predicting photo cascades in Facebook [@cheng2014].
The structure of the underlying social network also has predictive power [@lerman2008; @romero2013; @cheng2014]. However, these studies do not agree on the direction of effect of these features. For instance, on Digg, low network density is connected with high popularity [@lerman2008], but on Twitter, both very low and very high densities are positively correlated with popularity [@romero2013]. Their intuition is that a lower network density indicates that the item is capable of appealing to a general audience, while a higher network density indicates a tight-knit community supporting the item, both of which can be powerful drivers for an item’s popularity.
Finally, while Tsur et al. report content features to be useful [@tsur2012], most studies find content features to have little predictive power (Table \[tab:past-features\]). Even for domains such as songs or movies where item information is readily available, content features are not significantly associated with item popularity [@pachet2012]. Further, content features do not generalize well; it is hard to compute generalizable content features across different item domains. For these reasons, we do not consider content features in this work.
Features {#sec:our-features}
--------
Based on the above discussion, we use the following categories of features, with the aim of reproducing and extending the features used in past work [@cheng2014]: temporal, structural, and early adopters. To these we add a set of novel features based on preference similarity between early adopters.
### Temporal. {#sec:temporal-features}
These features have to do with the speed of adoptions during the early adoption period between the first and $k$th adoption. This leads to a set of features that focus on the rate of adoption:
- $time_i$: time between the initial adoption and the $i^{th}$ adoption ($2\leq i\leq k$). [@zhao2015; @maity2015; @weng2013]
- $time_{1...k/2}$: Mean time between adoptions for the first half (rounded down) of the adoptions.
- $time_{k/2...k}$: Mean time between adoptions for the last half (rounded up) of the adoptions.
### Structural.
These features have to do with the structure of the network around early adopters and can be broken down into two sub-categories: ego network features that relate the early adopters to their local networks, and subgraph features that consider only connections between the early adopters.
*Early adopters’ ego network features*
- $in_{i}$: Indegree of the $i^{th}$ early adopter ($2 \leq i \leq k$). This is a proxy for the number of people who may be exposed to an early adopter’s activity. For undirected networks, this will simply be the degree, or the number of friends of an early adopter. [@bakshy2011; @zhao2015]
- $reach$: Number of nodes reachable in one step from the early adopters.
- $connections$: Number of edges from early adopters to the entire graph. [@romero2013]
*Early adopters’ subgraph features*
- $indegree_{sub}$: Mean indegree (friends or followers) for each node in the subgraph of early adopters. [@lerman2008]
- $density_{sub}$: Number of edges in the subgraph of early adopters. [@romero2013]
- $cc_{sub}$: Number of connected components in the subgraph of early adopters. [@romero2013]
- $dist_{sub}$: Mean distance between connected nodes in the subgraph of early adopters. This is meant to measure how far the item has spread in the initial early adopters, similar to the cascade depth feature by Cheng et al.
- $sub\_in_{i}$: Indegree of the $i^{th}$ adopter on the subgraph ($1 \leq i \leq k$). [@lerman2008]
### Features of early adopters.
These features capture information about early adopters, such as their popularity, seniority, or activity level, which might be proxies for their influence. They can be divided into two sub-categories: features of the first user to adopt an item (root), and features averaged over other early adopters (resharers).
*Root features*
- $activity_{root}$: Number of adoptions in the four weeks before the end of the early adoption period. This is similar to a measure used by Cheng et al. which measured the number of days a user was active. [@cheng2014; @petrovic2011; @yang2010]
- $age_{root}$: Length of time the user has been registered on the social network.
- $popularity_{root}$: Number of friends or followers on the social network. [@lerman2008; @tsur2012]
*Resharer features*
- $activity_{resharer}$: Mean number of adoptions in the four weeks before the end of the early adoption period.
- $age_{resharer}$: Mean length of time the users have been registered on the social network.
- $popularity_{resharer}$: Mean number of friends or followers on the social network. [@tsur2012]
**Dataset** **Last.fm** **Twitter** **Flickr** **Goodreads**
--------------------- ------------- ------------- ------------ ---------------
Number of users 437k 737k 183k 252k
Number of items 5.8M 64k 10.9M 1.3M
Number of adoptions 44M 2.7M 33M 28M
Mean adoptions 7.6 41.8 3.0 21.4
Median adoptions 1 1 1 1
Maximum adoptions 11062 82507 2762 88027
### Similarity
To these previously tested features, we add features related to preference similarity between the early adopters. As with network density, our intuition is that similarity between early adopters may matter in two ways: high similarity may signify a niche item, or one that people with similar interests are likely to adopt, while low similarity might indicate an item that could appeal to a wide variety of people.
Similarity was computed using the Jaccard index of two users’ adoptions that occurred before the end of the early adoption period of the item in question. We computed the median, mean and maximum of similarity between adopters because these give us an idea of the distribution of the affinity of the early adopters; we do not include users who had less than five adoptions before the item in question because they are likely to have little overlap. The features we extracted are:
- $sim_{count}$: Number of similarities that could be computed between early adopters.
- $sim_{mean}$: Mean similarity between early adopters.
- $sim_{med}$: Median similarity between early adopters.
- $sim_{max}$: Maximum similarity between early adopters.
Data and Method
===============
Datasets from four online social networks
-----------------------------------------
We build models using data from four different online social platforms: Last.fm, Flickr, Goodreads and Twitter. These platforms span a broad range of online activity, including songs, photos, books and URLs; they also have a variety of user interfaces, use cases, and user populations. These variations reduce the risk of overfitting to properties of a particular social network.
- **Last.fm:** A music-focused social network where users can friend one another and love songs. We consider a dataset of 437k users and the songs they loved from their start date until February 2014 [@sharma2016].
- **Flickr:** A photo sharing website where users can friend one another and favorite photos. We use data collected over 104 days in 2006 and 2007 [@cha2009].
- **Goodreads:** A book rating website where users can friend one another and rate books. The dataset consists of 252k users and their ratings before August 2010. Unlike the other sites, Goodreads users rate books; we consider any rating at or above 4 (out of 5) as an endorsement (adoption) of the book [@huang2012].
- **Twitter:** A social networking site where users can form directed edges with one another and broadcast *tweets*, messages no longer than 140 characters (as of 2010). The Twitter dataset consists of URLs tweeted by 737k users for three weeks of 2010 [@hodas2014].
All of these websites have an active social network, providing an activity feed that allows users to explore, like, and reshare the items that their friends shared. The Last.fm feed shows songs that friends have to listened to or *loved*, Flickr shows photos that friends have *favorited*, Goodreads shows books that friends have *rated*, and Twitter shows tweets with URLs that followees have *favorited* or *retweeted*. Thus, like past studies on online social networks such as Facebook, Twitter and Digg, we expect active peer influence processes that should make structural and early adopter features relevant.
Table \[tab:descriptive\] shows descriptive statistics about the datasets, all of which have more than 150k users and millions of items (with the exception of Twitter with 64k URLs). Twitter has the highest mean adoptions per item ($41$), followed by Goodreads ($21$). The maximum number of adoptions for an item also varies, from more than 80k in Twitter and Goodreads to 2.7k in Flickr. The median number of adoptions is consistent, however: at least half of the items have only 1 adoption. The skew in popularity distribution is better shown in Figure \[fig:percent-cumulative\]. The 20% of the most popular items account for over 60% of adoptions in Flickr and over 90% of the adoptions in the other three websites. On Twitter, the skew is extreme: over 81% adoptions are on 4% of items.
![Cumulative percentage of adoptions by items for each dataset. Items on the x-axis are sorted by their popularity; the lines show a step pattern because multiple items may have the same number of adoptions. We observe a substantial skew in popularity. For example, the most popular 20% of items account for 60% of adoptions in Flickr and more than 90% of adoptions in other datasets.[]{data-label="fig:percent-cumulative"}](figures/cumsum_percentile.pdf){width="0.99\linewidth"}
![Boxplot showing the number of adoptions after 28 days (10 for Twitter) for items which have at least 5 adoptions. The bold partial line is the mean number of adoptions. Across datasets, most items receive less than 20 adoptions.[]{data-label="fig:boxplot2"}](figures/boxplots2.pdf){width="0.99\linewidth"}
Classification methodology
--------------------------
We first operationalize the Balanced Classification formulation on these datasets. As a reminder, $k$ is the number of early adoptions that we peek at for each item, and we predict which of these items will end up more popular than the median item.
We measure the final popularity at a time $T$ days after the first adoption of the item. To be consistent with prior work, we follow Cheng et al. and set $k=5$ and $T=28$ days for Last.fm, Flickr and Goodreads. Because the Twitter dataset is only three weeks long, we use a smaller $T=10$ days. To avoid right-censoring, we include only items that had their first adoption at least $T$ days before the last recorded timestamp in each dataset. The parameter $k$ also acts as a filter, allowing only items with at least $k$ adoptions. Figure \[fig:boxplot2\] shows properties of the data thus constructed.
We classify items based on their popularity after $T$ days, labeling those above the median 1 and others as 0. For each item, we extract features from the early adoption period, the time between the first and $k$th adoption. We use 5-fold cross validation to select the items that we train on, then use the trained model to predict final popularity of items in the test set. Since we use median popularity as the classification threshold, the test data has a roughly equal number of items in each class, allowing us to use accuracy as a reasonable evaluation metric. We tried several classification models using Weka [@hall2009weka], including logistic regression, random forests and support vector machines. Logistic regression models generally performed best, so we report results for those models unless otherwise specified.
Balanced classification
=======================
We start by comparing the predictive power of models using different sets of features across the four datasets on the Balanced Classification problem.
![Accuracy for prediction models incorporating different categories of features. The y-axis starts at 50%, the baseline for a random classifier on the balanced formulation. On all datasets, temporal features are the most predictive, almost as accurate as using all available features.[]{data-label="fig:temporal-vs-all"}](figures/temporal_vs_all.pdf){width="0.99\linewidth"}
Temporal features dominate
--------------------------
Figure \[fig:temporal-vs-all\] shows the prediction accuracy of the models. Similar to prior work on Facebook that used peeking [@cheng2014], when using all features we are able to predict whether an item will be above the median popularity around three-fourths of the time: 73% for Goodreads, 75% for Flickr, 81% for Last.fm and 83% for Twitter.
Training models with individual feature categories shows that temporal features are by far most important. Across all four datasets, a model using only temporal features performs almost as well as the full model. The next best feature category, resharer features, is able to predict 71% on Twitter and less than 60% on the other three datasets. Even a model that uses *all* non-temporal features, denoted by the “all-temporal" line in Figure \[fig:temporal-vs-all\], is not very good. For Goodreads and Flickr, this model is not much better than a random classifier. For Last.fm and Twitter, accuracy for non-temporal features improves somewhat, but is still at least 10% worse than when including temporal features.
Even a single temporal feature can be more predictive than models constructed from all non-temporal features. Consider the feature $time_x$, which is the number of days for an item to receive $x$ number of adoptions. At $x=5=k$, the feature $time_5$—time taken for an item to receive 5 adoptions—is the most predictive temporal feature for all datasets. A model based on this single feature achieves more than 70% accuracy on all datasets and accounts for nearly 97% of the accuracy of the full model for each dataset. While past work has highlighted the importance of temporal features as a whole [@szabo2010; @cheng2014], it is interesting to find that we may not even need multiple temporal features: a single measure is able to predict final popularity class label for items in all datasets.
[lcccc]{} Test \\ Train & Last.fm & Flickr & Goodreads & Twitter\
\
Last.fm & **80.6** & 80.7 & 78.0 & 80.0\
Flickr & 73.9 & **74.7** & 70.0 & 73.9\
Goodreads & 70.3 & 69.7 & **71.9** & 70.3\
Twitter & 82.7 & 82.3 & 79.7 & **82.2**\
\
Last.fm & **62.1** & 56.3 & 60.2 & 52.3\
Flickr & 53.0 & **55.1** & 51.8 & 48.2\
Goodreads & 56.0 & 52.1 & **57.1** & 50.6\
Twitter & 45.8 & 44.1 & 56.4 & **73.4**\
Cross-domain prediction
=======================
The analysis in the previous section confirms past findings about the importance of temporal features across a range of websites. We now extend these results to show that temporal features are not only powerful, they are also general: models learnt on one item domain using temporal features are readily transferable to others. In contrast, non-temporal features do not generalize well: even the direction of their effect is not consistent across domains. To show this, we train prediction models separately for each dataset, as before, then apply each model to every dataset.
Temporal features generalize
----------------------------
Table \[tab:train-test-temporal\] shows the accuracy of models trained only on temporal features from one dataset and tested on all four. Reading across the rows shows that regardless of which social network a model was trained on, its accuracy on test data from another network remains within 5% of the accuracy on test data from the same network.
Such consistent prediction accuracy is impressive, especially because the median time to reach 5 adoptions varies, ranging from 1 day in Flickr to 15 days for Goodreads. This suggests that there are general temporal patterns that are associated with future popularity, at least across these particular networks.
Other features have inconsistent effects
----------------------------------------
The story is less rosy for non-temporal features. Table \[tab:train-test-temporal\] shows the cross-domain prediction accuracy for models trained on all non-temporal features (in light of their low accuracy when taken individually, we combine all non-temporal features). Accuracies on the same dataset correspond to the “all-temporal” line in Figure \[fig:temporal-vs-all\]; they are generally low and drop further when tested on a different dataset. In particular, models trained on other websites do poorly when tested on Twitter, with the Last.fm and Flickr models performing worse than a random guesser on Twitter data. Meanwhile, a model trained on Twitter is almost 10 percentage points worse than the Last.fm-trained model for predicting popularity on Last.fm.
Not only does prediction accuracy drop across websites, but fitting single-feature logistic regression models for each feature shows that for 12 of the 25 features, the coefficient term flips between being positive and negative across models fit on different datasets. Similar to the contrasting results found in prior work [@lerman2010; @romero2013], we find that all measures of subgraph structural features of the early adopters, namely $indegree_{sub}$, $density_{sub}$, $cc_{sub}$, $dist_{sub}$ and $sub\_in_i$ (except for $sub\_in_1$ and $sub\_in_4$), can predict either higher or lower popularity depending on the dataset. For example, a higher $density_{sub}$—number of edges in the subgraph of early adopters—is associated with higher popularity on Flickr ($\beta$ coefficient=0.04), whereas on Last.fm, a higher density is associated with lower popularity ($\beta$ coefficient=-0.09). Features from the root, resharer and similarity categories show a similar dichotomous association with final item popularity.
Gaps between prediction and understanding
=========================================
These results show that not only are non-temporal features weak predictors, the direction of their effect on popularity is inconsistent across different domains. Combining this with our observation that a single temporal heuristic is almost as good a predictor as the full model raises questions about what it is that popularity prediction models are predicting and how they contribute to our understanding of popularity.
Temporal features drive predictability
--------------------------------------
While our work may seem contrary to recent work that claims that early adopters and properties of their social network matter for prediction, many of their findings are consistent with our own. Most prior work that uses peeking finds that temporal features are a key predictor [@tsur2012; @szabo2010; @pinto2013; @yu2015]. Further, even though Cheng et al. conclude temporal and structural features are major predictors of cascade size, they report for predicting photos’ popularity on Facebook, accuracy for temporal features alone (78%)is nearly as good as the full model (79.5%) [@cheng2014].
By holding modeling, feature selection and problem formulation consistent, we contribute to this literature by demonstrating the magnitude and generality of the predictive power of temporal features across a range of social networks. Having multiple networks also lets us show that, unlike temporal features, using non-temporal features does not generalize well to new contexts. These features might be useful for understanding the particulars of a given website, but it seems likely that they are capturing idiosyncrasies of that site rather than telling us something general about how items become popular in social networks.
Is cumulative advantage the whole story?
----------------------------------------
If non-temporal features are weakly predictive and not generalizable, and all that matters is the rate of initial adoption, then how do predictive exercises with peeking advance scientific understanding of what drives popularity? In other words, what does it mean when one claims that popularity is predictable once we know about initial adopters?
One answer is that early, rapid adoption is a signal of intrinsic features of an item that help to determine its popularity. Items with better content command a higher initial popularity, and thus the predictive power of early temporal features is simply a reflection of content quality or interestingness to the social network in question. Given increasing evidence from multiple domains that content features are at best weakly connected to an item’s popularity [@salganik2006; @pachet2012; @martin2016], this seems unlikely to be the whole explanation.
Another explanation is that items that get attention early are more likely to be featured in the interface, via feeds, recommendations or ads; they might also be spread through external channels could drive up the rate of early adoption. Those would be interesting questions to explore. Still, whatever be the driving reasons, these models are telling us that once items achieve initial popularity, they are much more likely to become more popular in the future. This is simply a restatement of cumulative advantage, or the rich-get-richer phenomenon [@borghol2012].
Overall, though, we find that neither our results nor other work say much about why or how items become popular, except that items that share temporal patterns of popular items early on tend to be the ones that are more popular in the future, and that making popularity salient and ordering items by popularity can increase this effect [@salganik2006]. While such predictions are practically useful for promoting content, they are not so useful for informing creation of new content or assessing its value, nor for understanding the mechanisms by which items become popular.
Temporally matched balanced classification {#sec:new-problems}
==========================================
In this section, we give a problem formulation that lessens the importance of temporal features by conditioning on the average rate of adoption. That is, instead of considering all items with $k$ adoptions, we consider items with $k$ adoptions within about the same amount of time. Given the dominance of cumulative advantage, such a formulation would be better suited for future research in understanding how items become popular, as gains in accuracy will likely shed light on attributes of early adopters, items, and networks that affect their final popularity.
*k-t* problem formulation
-------------------------
We call this formulation Temporally Matched Balanced Classification, or a *k-t* formulation of the problem:
> **P2:** *Among items with exactly $k$ adoptions at the end of a fixed time period $t$, which ones would be higher than the median popularity at a later time $T$?*
To do this, for each dataset we filtered items to those that had exactly $k$ adoptions in $t$ days. We extracted features of these items as previously described, adding a new temporal feature for each day in $t$:
- $adoptions_i$: Number of adoptions on day $i$ of the early adopter period. [@szabo2010; @tsur2012; @pinto2013]
As before, we choose $k=5$ and $T=28$ days. For each dataset, we set $t$ to be the median time it took an item to reach five adoptions: $t=15$ for Goodreads, $t=7$ for Last.fm, and $t=1$ for Flickr. We exclude Twitter due to a lack of data when we filter for both $k$ and $t$. We again do 5-fold cross-validation, predicting if each item would be above or below the final median popularity after $T$ days.
Figure \[fig:hard-all-temporal\] shows the results. As we hoped, non-temporal features now provide most of the explanatory power in the full model. Further, comparing the all-temporal series with fixed $k$ and $t$ to the one with only fixed $k$ shows that the absolute accuracy of non-temporal features increases in this formulation. This suggests that de-emphasizing temporal features in prediction might in fact improve our understanding of other features that drive popularity.
Our understanding, however, is limited: even conditioning on a single temporal feature makes for a much harder problem, with the overall prediction accuracy below 65% for all datasets even when using all features. There is clearly much room for improvement.
![Percent accuracy for fixed $t$ & $k$ using all features and non-temporal features, and for fixed $k$ with all features and non-temporal features. $k=5$, $T=28$ days for all; $t=15$ days for Goodreads, $t=1$ days for Flickr, and $t=7$ days for Last.fm. Fixing $t$ reduces accuracy substantially compared to when $t$ is not fixed. As expected when controlling for time, non-temporal features now provide most of the explanatory power.[]{data-label="fig:hard-all-temporal"}](figures/fixed_t_k_temporal_vs_all.pdf){width="0.99\linewidth"}
Discussion and Conclusion
=========================
Using multiple problem formulations, we show that temporal features matter the most in predicting the popularity of items given data about initial adopters and our current ability to build explanatory features of those adopters and their networks. Using datasets from a variety of social networks, we show that temporal features are not only better at predicting popularity than all other features combined, but that they readily generalize to new contexts. When we discount temporal phenomena by removing temporal features or adjusting the problem formulation, accuracy decreases substantially.
From a practical point of view, these results provide empirical support for a promising approach where only temporal features are used to predict future popularity [@szabo2010; @zhao2015] because the drop in accuracy by casting aside non-temporal features is generally small. Maybe creative feature engineering is not worth the effort for the Balanced Classification task. This way of looking at the problem resonates a bit with the Netflix prize, where most of the learners that were folded into the winning model were never implemented in Netflix’s actual algorithm, in part because the cost of computing and managing those learners was not worth the incremental gains [@amatrain2012].
Although less valuable than temporal features, the non-temporal features examined so far do have some predictive power on their own. This might be useful when temporal information is unavailable: for example, for very new items [@borghol2012], or for external observers or datasets where timestamps are unavailable [@cosley2010]. Encouragingly, non-temporal features increase in accuracy a little on the *k-t* formulation compared to the fixed-$k$ balanced classification problem, suggesting that making time less salient might allow other factors to become more visible and modelable.
Using *k-t* models could also bend time to our advantage. Comparing the overall performance and predictive features in models with smaller versus larger $t$ might highlight item, adopter, and network characteristics that predict faster adoption (and eventual popularity). Another way to frame this intuition is that instead of predicting eventual popularity, we should try to predict initial adoption speed.
Deeper thinking about the context of sharing might also be useful. Algorithmic and interface factors, for instance, have been shown to create cumulative advantage effects; it would be interesting to look more deeply into how system features might influence adoption behaviors. Likewise, diffusion models tend to focus attention on sharers rather than receivers of information—but those receivers’ preferences, goals and attention budgets likely shape their adoption behaviors [@sharma2015-cscw]. Thus, consideration of audience-based features might be a way forward.
Most generally, we encourage research in this area to go beyond the low-hanging fruit of time. For building better theories of diffusion, maximizing accuracy with temporal information may act both as a crutch that makes the problem too easy, and as a blindfold that makes it hard to examine what drives those rapid adoptions that predict eventual popularity.
Acknowledgments
===============
This work was supported by the National Science Foundation under grants IIS 0910664 and IIS 1422484, and by a grant from Google for computational resources.
[^1]: Such featuring makes some items more salient than others and surely affects the final popularity of both featured and non-featured items; typically, formulations of the problem look at very small slices of early activity, which presumably minimizes these effects.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Within a phase field approach which takes the strain-induced elasticity into account, the kinetics of the coherent order-disorder transition is investigated for the specific case of $Al_3 Zr$ alloy. It is shown that a microstructure with cubic $L1_2$ precipitates appears as a transient state during the decomposition of a homogeneous disordered solid solution into a microstructure with tetragonal $DO_{23}$ precipitates embedded into a disordered matrix. At low enough temperature, favored by a weak internal stress, only $L1_2$ precipitates grow in the transient microstructure preceding nucleation of the $DO_{23}$ precipitates that occurs exclusively at the interface of the solid solution with the $L1_2$ precipitates.
Analysis of microstructures at nanoscopic scale shows a characteristic rod shape for the $DO_{23}$ precipitates due to the combination of their tetragonal symmetry and their large internal stress.
address: |
Laboratoire d’Etude des Microstructures\
ONERA - CNRS, BP72, 92322 Châtillon Cedex, France
author:
- 'Laurent Proville [^1] and Alphonse Finel'
title: 'Kinetics of coherent order-disorder transition in $Al_3 Zr$'
---
Introduction
============
Macroscopic properties of materials depend to a large extend on the microstructures they present at a mesoscopic scale. For the case of alloys the most efficient way to control the formation of microstructures is by phase transformations. Well known prototypes are the $\gamma-\gamma'$ superalloys, as $Ni-Al$ (see [@Ardell]). The microstructure of such binary alloys consists of stable ordered domains, dispersed in a disordered fcc matrix. The symmetry of the ordered domains is called $L1_2$: the atoms of the minority species are placed at the corner of the fcc motif and the atoms of the majority species occupies the center of the faces of the same motif.
The macroscopic properties are controlled by the size of precipitates, their spatial distribution, and their ability to resist to coarsening. In this context, we present a theoretical study of the dynamics of microstructures in an aluminum based alloy, namely $Al-Zr$.
In this system, for low enough concentration, the equilibrium ordered structure is the tetragonal $DO_{23}$ phase. Its motif is obtained from $L1_2$ structure with anti-phase boundaries in (100) directions with the period of $3$. In $Al-Zr$ the cubic $L1_2$ phase is known to be metastable at all temperature $T$. This is confirmed by ab-initio electronic calculations at $T=0 K$ [@Fontaine] and [@Freeman] which show that the energy difference between $L1_2$ and $DO_{23}$ is about $0.863 \times 10^8 J/m^3$ (or $9.1 meV/atom$), in favor of $DO_{23}$. However, the lattice misfit of $DO_{23}$ with respect to pure $Al$ is significantly larger than $L1_2$. Hence, the interplay between the chemical energy and, for coherent microstructure, the elastic energy may induce various precipitation processes.
The aim of the present work is to investigate the decomposition processes and to analyze the resulting microstructures at mesoscopic scale. In that aim, we use a phase field approach, where the incoherent chemical energy is represented by a Ginzburg-Landau free energy supplemented by a strain-induced elastic energy, in the form proposed by Khachaturyan [@turyan83; @turyan93; @Khacha00]. Using numerical simulations, we show that elastic stress favors the $L1_2$ order at low $Zr$ concentration and low temperature and in that range the kinetics generates $L1_2$ rather than $DO_{23}$ precipitates. If the coherent elastic interaction is not included, only $DO_{23}$ precipitates nucleate.
Our simulations prove that one may obtain metastable $L1_2$ phase in specific conditions that correspond to the experimental conditions leading to a $L1_2$ microstructure. The lifetime of that microstructure may be infinite compared to the simulation time, provided the temperature is not too high. Once the $L1_2$ precipitates have grown at sufficiently low temperature, the system may be brought to a higher temperature, where no $L1_2$ would nucleate if the system had been quenched directly to that temperature. If the already developed $L1_2$ microstructure is aged again at higher temperature, then the $DO_{23}$ precipitation starts and $DO_{23}$ precipitates nucleate exclusively along the interfaces between $L1_2$ inclusions and the matrix. Once $DO_{23}$ precipitates are formed, they grow and consume $L1_2$ domains. The latest transformation cannot be reversed with decreasing temperature.
In Sec.\[sec2\], the general principles of the kinetic model are presented. In Sec.\[sec3\], we outline how the microelasticity contribution can be included into the model. The implementation of the phase transition kinetics, the results obtained and their interpretation are given in Sec.\[sec4\]. Conclusions and perspectives are drawn in Sec.\[sec5\].
Ginzburg Landau Functional {#sec2}
==========================
Our mesoscopic method is a time-dependent Ginzburg-Landau kinetics driven by a functional with two parts: first a Ginzburg-Landau functional that includes the chemical interactions and the interface effects and second a strain-induced elastic energy.
The explicit form of the Ginzburg-Landau (GL) free energy is imposed by symmetry rules. The first step consists of identifying the long range order (LRO) parameters that represent the ordered phases we want to study. In the present situation, we should introduce the LRO parameters of $DO_{23}$ and $L1_2$. Nevertheless, to simplify the writing of the GL free energy, we choose to replace the $DO_{23}$ LRO parameters with the $DO_{22}$ ones as these phases are both tetragonal. What differs between them is only the periodicity of the anti-phase boundaries that is $2$ for $DO_{22}$ instead of $3$ for $DO_{23}$. As our aim is not to investigate the competition between the latter phases, our description only requires to capture the tetragonality of the possible stable ordered phase.
The $L1_2$ phase is simple: it consists in three independent parameters $\eta_1$, $\eta_2$, $\eta_3$ which correspond to the amplitudes of the three waves that contribute to $L1_2$. The waves correspond to the three vectors of the reciprocal space $K_1 =(100)$ , $K_2 =(010)$, $K_3 =(001)$, respectively. If the microstructure consists only in $L1_2$ domains embedded into a disordered fcc matrix, the local concentration $c(R)$ would be given by
$$\begin{aligned}
c(R)=c_0(R)+ \sum_{j=\{1,2,3\}}
{\eta_j}(R) \exp{(i 2 \pi K_j. R)} \end{aligned}$$
where the quantities $c_0(R)$ and ${\eta_j}(R)$ vary slowly in space. For a $DO_{22}$ structure, the probability $c(R)$ to find a atom $Zr$ at position $R$ can be written as follows:
$$\begin{aligned}
c(R)=c_0(R)+ \sum_{j=\{1,2,3\}}
{\eta_j}(R) \exp{(i 2 \pi K_j. R)} &+&
\gamma_j(R) \exp{(i 2 \pi Q_j . R)}+\nonumber\\
&& \gamma_j^*(R)
\exp{(-i 2 \pi Q_j . R)}\end{aligned}$$
where $\gamma_j^*$ is the complex conjugate of the amplitude parameter $\gamma_j$ and $Q_1 = (1/2\ 1\ 0)$, $Q_2=(0\ 1/2\ 1)$ and $Q_3=(1\ 0\ 1/2)$ blong to the reciprocal cubic lattice. This choice is not unique but is sufficient. For example, $(1/2\ 0\ 1)$ can be replaced by $(1/2\ 1\ 0)$); each index $j$ correspond to a possible orientation of the tetragonal transformation yielding $DO_{22}$. For each of the three orientational variants there are four translational variants. For example, for a perfect $DO_{22}$ phase with the orientation $j=1$, one has four translational variants defined by either $\gamma_1^*= \eta_1$, $\gamma_1^*= -\eta_1$ $\gamma_1^*=-i \eta_1$ $ \gamma_1^*= i\eta_1$. For simplicity we drop the complex variants. It does not affect the description as there are still two translation variants for each orientation of the $DO_{22}$ phase.
We note that the state $\gamma_j=0$ and $\eta_j \neq 0$ for $j=1,2,3$ leads to a density $c(R)$ that describes a $L1_2$ phase. A perfect $L1_2$ implies that the three (100)-type waves have the same amplitude. As $L1_2$ preserves the cubic symmetry there is no orientational variant but only four translational variants.
We now develop a uniform Landau functional as a function of $\{c_0,\eta_j,\gamma_j \}$. The terms of this analytical function $f_L(c_0,\eta,\gamma)=F_0 . {\hat f_L}$ are selected to fit with the symmetry of the fcc lattice, that is each term must be invariant under any operation of the fcc symmetry group. Formally there is no other rule to realize a polynomial expansion but the simplest form is probably the best. In practice, other conditions (see below) must be satisfied. We propose the following adimentional functional:
$$\begin{aligned}
{\hat f_L}&=& A_n (c_0-c_1)^n+\frac{A_2}{2} (c_2-c_0) \sum_i \eta_i^2
-A_3 \eta_1 \eta_2\eta_3
\nonumber\\
&&
+\frac{A_4}{4} \eta_i^4
+{K_3} \gamma_i^{2}\eta_i+\frac{K_4}{2} \sum_{j=i \pm 1}
\gamma_{i}^2 \eta_{j}^2\nonumber\\
&& +\frac{B_2}{2} (c_2-c_0) \gamma_i^2
- \frac{B_4}{4} \gamma_i^{4} +
\frac{B_{4}'}{2} \sum_{j=i \pm 1} \gamma_i^{2}\gamma_{j}^2
+\frac{B_6}{6} \gamma_i^{6} \label{func2}\\
\nonumber\end{aligned}$$
where order parameters indices are written modulo 3. The hat symbol points out the adimensional quantities: here ${\hat f_L}$ is an adimensional free energy density. The sets of coefficients $\{A_j \}$, $\{B_j \}$, $\{K_j \}$ and $\{c_1, c_2 \}$ should be function of the temperature but are chosen as constant parameters for simplicity. The first term in the right hand side of Eq.(\[func2\]) is the disorder contribution as it does not depend on the order parameters. The power $n$ of that term may have two values $n=n_+=2$ and $n=n_-=8$ depending whether $c_0> c_1$ or not. We introduce that property to adjust the topology of the Landau functional with both experimental and ab-initio measurements as described below. The continuity of the first and second derivatives of the Landau functional are preserved which is sufficient for the present case.
The next three terms with $A_2,A_3,A_4$ amplitudes are the contribution of the star $(1\ 0 \ 0)$. The four last terms associated with $B_2,B_{4},B_{4}',B_6$ are the contribution of the concentration waves $\{Q_i\}_i$. The expansion to the sixth order is required to obtain the linear stability of both $DO_{22}$ and $L1_2$ phases for the same range of concentration. The $\{K_4\}$ and $\{B_{4}'\}$ coefficients couple concentration waves with different orientations. They control the amplitude of the potential barrier between the minima that correspond to ordered phases. The $K_3$ term is the amplitude of the coupling between the waves that belong to the same orientational $DO_{22}$ variant, $(K_i ,Q_i)$ and equivalents.
In fact, the precise form of the Landau functional term by term has not a direct influence on the mesoscopic microstructure providing that the functional is globally invariant with respect to the space group of the fcc lattice. The important ingredients are the excess of free-energy associated with interfaces and long-range elastic interactions between domains. Therefore, we keep the lowest order coupling term between the LRO parameters associated with $(K_i ,Q_i)$ , i. e. , terms of the form $\gamma_i^2 \eta_i$.
The energy scale is fixed by the $F_0$ coefficient. The parameters of the adimensional free-energy density ${\hat f_L}(c_0,\eta,\gamma)$ are adjusted to fit the required qualitative thermodynamical properties. In Fig.1 is plotted versus concentration $c_0$ the free energy density $F_0.{\hat f_L}$ minimized with respect to the LRO parameters. The three types of minima correspond to the disordered phase, the $DO_{23}$ (or $DO_{22}$) and $L1_2$ structures. The common tangent constructions determine the regions where ordered phases may coexist with the solid solution. The concentrations $c_{DO_{23},a }$ and $c_{ DO_{23},b}$ are the limit of the region where coexist both disordered solid solution with the $DO_{23}$ phase. Similar quantities can be determine for the coexistence of solid solution with $L1_2$ phase: $c_{L1_2,a }$ and $c_{L1_2 ,b}$. The concentration $c_{L1_2,a}$ and $c_{DO_{23},a}$ are the solubility limit of the $L1_2$ and $DO_{23}$ phases, respectively.
In a non-uniform system, the order parameters $\eta_i$ and $\gamma_i$ and the local concentration $c_0$ are spatially dependant. Within a phase-field approach, these parameters vary continuously thought the system. The energy excess due to the interfaces of precipitate, is expressed as a continuous Hamiltonian of the parameter fields and their first derivatives. To the lowest order, this leads to the following Ginzburg-Landau (GL) free energy density:
$$f_{GL} = F_0. [{\hat f_L}(c_0,\eta,\gamma)+
\{\lambda_{c} |\nabla c_0|^2 + \sum_{i} \lambda_{\eta}
\nabla^2 \eta_i
+\lambda_{\gamma} \nabla^2 \gamma_i \}] \label{GL}$$
where the $\lambda$ coefficients are the weight of the gradient terms. Total energy is given by $F_{GL}=\int f_{GL} dV$. For the numerical implementation, we introduce a discrete space which is a cubic sub-lattice with unit cell of linear size $d$. This length must be large enough to justify our continuous approach and define the scale of one pixel in our simulations. The total free energy can be expressed as a discrete sum $F_{GL}=d^3 \sum_{L} {f}_{GL} $ where $L$ represents the set of the sub-lattice nodes.
We now describe the physical requirement we use to adjust the Ginzburg-Landau functional. First, $c_{L1_2 a}$ and $c_{DO_{23} a}$ (see Fig.1) are the solubility limit of both $L1_2$ and $DO_{23}$ if each phase is supposed to coexist alone with solid solution. The $c_{DO_{22} a}$ is given by the uncoherent phase diagram and $c_{L1_2 a}$ has been estimated by the measure of the lattice parameter of the solid solution by X ray using Vegard’s law [@Zedalis-Fine-86]. For $Zr$ at temperature $T=425^o C$, $c_{DO_{22} a}=0.0308$ and $c_{L1_2 a}=0.0426$ at.$\%$ [@Zedalis-Fine-86].
Second, to analyze the competition between the metastable $L1_2$ and stable $DO_{23}$ structures during the precipitation and aging processes, the free energy difference between both phases is an important quantity. In order to estimate this difference we refer to the theoretical studies of the formation energies obtained with ab-initio electronic structure methods [@Fontaine; @Freeman; @Meschter; @Watson]. In [@Fontaine], it is found that ground-state of $Al_3Zr$ is indeed $DO_{23}$, and its structure is stabilized by the relaxation of the atomic positions inside the elementary motif. The difference of energy between $L1_2$ and $DO_{23}$ is found to be $\delta_{L1_2}=0.863\times 10^8 J/m^3$. In term of a phase field approach, this corresponds to an uncoherent energy at zero temperature. We assume that for low enough temperature, the free energy difference $\delta_{L1_2}$ do not vary strongly. Therefore, the ab-initio quantity $\delta_{L1_2}$ is used to fix the scale $F_0$ of the free energy (Eq.\[GL\]) through the relation $F_0= {\delta}_{L1_2}/ {\hat \delta_{L1_2}}$ where ${\hat \delta_{L1_2}}$ is the energy difference between the corresponding $L1_2$ and $DO_{23}$ minima of the adimentional free energy $\hat f_L$ which has been minimized taking into account the conservation of the $c_0$ concentration parameter that yields the common tangent construction (see Fig.1).
Finally, another very important feature of the GL functional is the excess of energy of the interface between solid solution and precipitates, noted $I_o$ with $o=\{L1_2, DO_{23} \}$. These quantities play a role in nucleation and growth process. The interface energy of ordered precipitates in the solid solution can be measured experimentally at very low supersaturation. The Lifshitz-Slyozov-Wagner theory [@LSW] gives the interface energy as a function of the diffusion coefficient which is physically measured. Unfortunately the few we found in literature (see [@Zedalis-Fine-86]) about such a measurement is not satisfactory as the interface energy measured for $L1_2$ precipitates $I_{L1_2}$ is hundred times larger than the usual values. Thus we choose to estimate $I_{L1_2}$ to a value of $10 mJ/m^2$ which is the order of magnitude of interface energies measured in aluminum compounds. As we did not find measurement in literature concerning interface energy of $DO_{23}$ precipitates $I_{DO_{23}}$, we choose for $I_{DO_{23}}$ a similar value to that of $I_{L1_2}$ because there is no physical reason these two quantities to differ by an order of magnitude. We adjust the interface energies of the ordered phases $\{o\}$ of the adimensional GL functional ${\hat I}_o$ such as $F_0 d {\hat I}_o=I_o$. It implies ${\hat I}_{DO_{23}}/{\hat I}_{L1_2}= {I}_{DO_{23}}/{ I}_{L1_2}$. With some difficulties, we managed to adjust the GL functional such that both ${\hat I_{DO_{23}}}$ and ${\hat I_{L1_2}}$ have the same order of magnitude. The $F_0$ being fixed by the second criteriion stated previously, it imposes $d= I_o/{\hat I}_o. {\hat \delta_{L1_2}}/{\delta}_{L1_2}$ so we can define the scale of our simulation. In order to investigate the nanometer scale we choose $d = 1 nm$, then the $\hat{f}$ functional must verify $ {\hat I_o} / {\hat \delta}_{L1_2} \approx 0.26$. The only way we found to satisfy the previous criteria is to introduce the non-symmetric power $n_{\pm}$ for the term $(c-c_1)^{n_\pm}$ in the GL functional. Finally we obtain $c_{DO_{22} a}=0.0308 \%$, $c_{L1_2 a}=0.5 \%$ and $F_0. d. {\hat I_{L1_2}}= 8 mJ/m^2$ , $F_0. d. {\hat I_{DO_{22}}}=9 mJ/m^2$. The pixel of our simulation represents a cube of size $d = 0.5 nm $.
Microelasticity Contribution to Free Energy {#sec3}
===========================================
As described in [@Khacha96], the elastic energy $E_{el}$ is calculated assumsing that the local strain $(\epsilon_{kl})$ induces a relaxation that is calculated by setting a small volume $dV$ of the bulk to the mechanical equilibrium. It is supposed that the time needed to reach the mechanical equilibrium is negligible compared to the typical diffusive time of the ordering process. The key point of the phase field theory for alloys is that the stress free strain tensor can be expressed as a function of the local LRO parameters and local concentration $c_0 (R)$.
The geometrical operation to transform a cubic unit cell of the solid solution with lattice parameter $ \overline a $ into the cubic unit cell of the $L1_2$ phase with lattice parameter $ a_{L1_2} $ is given by the tensor:
$$\epsilon^{L1_2}_{kl}= \delta_{kl} \
(a_{L1_2}-{\overline a})/{\overline a}$$
where $\delta_{kl}$ is the unity tensor. The lattice parameter $ a_{L1_2}$ has been measured for a perfect $L1_2$ phase [@Zedalis-Fine-86] i. e. , with stoichiometry $0.25$ at. $\%$ $Zr$ and $\overline a $ is extrapolated from the lattice parameter $a_0$ of pure aluminum $Al$, using Vegard’s law.
We note $ a_{DO_{23}}$ and $b_{DO_{23}}$ the lattice parameters of the tetragonal $DO_{23}$ phase at stoichiometry $0.25$ at. $\%$ $Zr$. The geometrical operation to transform a cubic unit cell of the solid solution into the tetragonal elementary cell of the $DO_{23}$ phase is given by the tensor $\epsilon_{kl}^{DO_{23}}(p)$ if the orientational variant corresponds to the association of the $(K_p,Q_p)$ waves with $p=1,2\ or\ 3$: the cell is dilatated in either $(100)$, $(010)$ or $(001)$ direction. Here we choose as an example the orientational variant associated to $(K_1,Q_1)$ and the cubic unit cell is dilated in the direction $(100)$ so as $b_{DO_{23}}>{\overline a} > a_{DO_{23}}$. $$\begin{aligned}
\epsilon^{DO_{23}}(1)= \frac{b_{DO_{23}}-{\overline a}}{{\overline a}}
\left( \begin{array}{ccc}
1 & 0 & 0 \\
0 & t & 0 \\
0 & 0 & t \\
\end{array} \right) \label{strainDO23}
\end{aligned}$$ where $t=(a_{DO_{23}}-{\overline a})/(b_{DO_{23}}-{\overline a})$. Tensors $\epsilon_{kl}^{DO_{23}}(2)$ and $\epsilon_{kl}^{DO_{23}}(3)$ derive from $\epsilon_{kl}^{DO_{23}}(1)$ by permutation of the diagonal coefficients. For the general expression of the local strain, we propose the following form (see [@Khacha00]): $$\epsilon_{kl}^{0}(R)= \epsilon_{kl}^{00}(0)\ \psi_0(R) +
\sum_{p=1}^{3} \epsilon_{kl}^{00}(p)\ \psi_p(R) \label{epsil}$$ where $\psi_p(R)=(\gamma_p(R))^2$, and $\psi_0(R) = (c_0(R) - \overline c)$. The tensor coefficients $\epsilon_{kl}^{00}(p)$ are chosen in a such way that $\epsilon_{kl}^{0}(R)=\epsilon_{kl}^{L1_2}$ if a relaxed $L1_2$ inclusion is at the position $R$ and $\epsilon_{kl}^{0}(R)=\epsilon_{kl}^{DO_{23}(p)}$ if a $DO_{23}$ inclusion is in the same position with the orientational variant $p$. The Eq.\[epsil\] is rewritten in a compact form: $ \epsilon^{0}(R)=
\sum_{p=0}^{3} \epsilon^{00}(p)\ \psi_p(R)$ where the $p$ indice varies now from $0$ to $3$. The strain-induced elastic energy can be computed following the model presented in [@Khacha96] which gives: $$E_{el}=\frac{1}{2}
\sum_{p,q} \int [B_{pq}
] {\tilde \psi_p}^*{\tilde \psi_q} dK^3 \label{Eel}$$ where ${\tilde \psi_p}$ is the Fourier transform of the function $\psi_p$. There $$B_{p,q}(K)=\lambda_{ijkl}\ { \epsilon_{kl}^{00}(p)} \
{\epsilon_{ij}^{00}(q)}- b_{pq}(K)$$ and $b_{pq}(K)=\sigma_{i,j}^{00}(p)\ K_i \ G_{j k} \ K_l \ \sigma_{kl}^{00}(q)$ where $\lambda_{ijkl}$ is the elastic tensor and $\sigma_{ij}^{00}(p)=\lambda_{ijkl} \epsilon_{kl}^{00}(p)$. The tensor $G_{jk}$ is the elastic Green function. The tensor $\lambda_{ijkl}$ is assumed to be homogeneous in space and the simulations are realized with elastic coefficients of aluminum [@Lazarus].
Kinetics of the phase transformation {#sec4}
====================================
The total energy is given by the sum $F=F_{GL}+E_{el}$ (Eqs.(\[GL\]) and (\[Eel\])). At mesoscopic scale, the kinetics of the phase transition is well described by a phase field method (see [@Lebowitz]). In the context of a phase field approach, the local composition $c_0$ is a conservative order parameter and thus its time evolution is driven by the Cahn-Hilliard equation: $$\frac{\partial c_0(R ,t)}{\partial t}=
L_c \bigtriangleup \frac{\delta F }{\delta c_0(R ,t)} + \upsilon^{c}(R,t) \label{KinC}$$ and for the non-conservative LRO parameters the kinetic is given by $$\frac{\partial \eta_j }{\partial t}=-L_{\eta}
\frac{\delta F}{\delta \eta_j } + \upsilon^{\eta_j}(R,t)\label{KinEta}$$ $$\frac{\partial \gamma_j}{\partial t}=-L_{\gamma}
\frac{\delta F}{\delta \gamma_j} + \upsilon^{\gamma_j}(R,t)\label{KinGam}$$
where $\upsilon$’s are stochastic terms. To simulate thermal fluctuations, it is useful to introduce the Langevin noise which consists in assuming a white space-time noise for the stochastic terms and no cross-correlation between each other. Numerically, the random functions $\upsilon$’s are implemented with a gaussian probability density [@NR].
The set of Eqs.\[(\[KinC\])-(\[KinGam\])\] is the so called Time-Dependant Ginzburg-Landau equation [@Khacha00] and they can be derived from the microscopic Önsager equation with respect to the occupation probability of the solute atoms [@turyan83]. Numerically, for simplicity, we choose $L_{\gamma}=L_{\eta}=L_c=1$. So in order to deduce the approximative time unit of our simulations, our numerical time must be divided by the numerical value of the diffusion coefficient. The Langevin noise provides a very primitive description of the thermal fluctuations. All though it is the simplest and the less controversal way to implement the temperature in the simulation.
Equations \[(\[KinC\])-(\[KinGam\])\] are integrated on a system that is invariant along $z$-axis to save computation time. The initial state of the system is a uniform solid solution at concentration $c_{L1_2,a}<c_0={\overline c}< 0.25$ and all the set of LRO parameters are set to zero value, which represents an unstable disordered phase. The initial time of our simulation can be considered as an instantaneous quenching of the material.
To represent the microstructures we choose to color the $DO_{23}$ precipitates with either blue or red depending on their translational variant. The four translational variants of $L1_2$ are allocated to four other paler colors. The disordered phase is colored with black. The gray scale is sufficient to distinguish the different types of precipitates if the translational variants are forgotten. On Figs.\[(\[fig4a\])-(\[fig4c\])\], we present the dynamics of the phase transition from disordered solid solution to a microstructure with ordered precipitates embedded into a disordered matrix. The different sequences are realized for different average compositions $\overline c$ and different temperatures $T$ .
First, one notes the specific rod shape of the $DO_{23}$ precipitates. The tetragonal symmetry combined with a large misfit $(b_{DO_{23}}-{\overline a})/{\overline a}$ which involves a large intrinsic stress is well known to induce such a pattern [@Khacha96]. On the very last picture of Fig.2 the facets of the $DO_{23}$ inclusions correspond to the habit planes with orientation of around $20$ degrees with respect to the (1 0 0) directions. For any couple of external variables, namely the temperature $T$ and the composition $ \overline c$ the late stage of the kinetics is a microstructure which contains exclusively the $DO_{23}$ precipitates embedded in a disordered matrix. Nevertheless $L1_{2}$ structures may appear in the early regime of the dynamics (see Fig.3). As the $L1_{2}$ ordered precipitates involve a weak misfit $(a_{L1_2}-{\overline a})/{\overline a}$ compared to $(b_{DO_{23}}-{\overline a})/{\overline a}$, their shape is spherical. These spherical inclusions nucleate only at low temperature $T$ and low $Zr$ concentration $\overline c$. At low enough values of $T$ and for saturation $\overline c$ close enough from the solubility limit, the microstructure contains exclusively the $L1_{2}$ precipitates inside the solid solution. Since the $L1_{2}$ phase is metastable, any grain of this phase should not resist to thermal fluctuations and thus no $L1_{2}$ precipitates should grow in the microstructure. It is actually what is observed if the elastic energy $E_{el}$ is neglected in our simulations. However, in the limit of low $Zr$ saturation, the kinetics of the transition drives the system to a transient $L1_{2}$ microstructure which is favored by its intrinsic stress which is weaker than that of $DO_{23}$. If the temperature $T$ is not too high, i. e., $T<T_{L1_2} \approx 500 K$ (see Fig.2), the $L1_{2}$ precipitates can even grow by consuming the very few solute atoms contained in the disordered matrix. This implies that the solid solution becomes poor in solute and therefore the nucleation of $DO_{23}$ precipitates is no longer possible in the disoredered matrix. In fact, if temperature is lower than $T_{L1_2}$, the $L1_{2}$ microstructure may survive for a time much longer than the computation time. Nevertheless, a gradual increase of the temperature reveals the process of nucleation of the $DO_{23}$ precipitates (see Fig.2). A remarkable result is that nucleation of the stable phase occurs at the interface of the $L1_2$ precipitates with the solid solution where the local concentration $c_0$ is high enough. One note that the preferential growth of $DO_{23}$ precipitates occurs at temperature higher than $T_{L1_2}$ that demonstrates the robusness of the metastable $L1_{2}$ microstructure with respect to thermal fluctuations.
Similar simulations at larger saturation $\overline c$ show that both the $DO_{23}$ and $L1_{2}$ precipitates nucleate and may coexist in the microstructure (see Fig.4). In that range of $\overline c$, the $DO_{23}$ inclusions can nucleate at places different from $L1_{2}$ precipitates because the solute atoms have not been consumed by the growth of $L1_{2}$ inclusions. Once saturation of the solid solution is locally dried up, the $DO_{23}$ precipitates grow to the expense of the $L1_{2}$ grains via the solute diffusion throught the matrix. Then the persistent $L1_{2}$ precipitates are localized relatively far from the $DO_{23}$ grain. For a given temperature, if the average concentration $\overline c$ is increased again, the grains with different phases nucleate in neighboring regions of the supersaturated solid solution and the $DO_{23}$ inclusions absorbe the $Zr$ matter of $L1_2$ inclusions.
One remark on Figs.\[(2)-(4)\] that the orientational variant of $DO_{23}$ which is combination of waves $(K_3,Q_3)$ is inhibited. It is because the precipitates with such variant cannot relax their elastic energy because of z-invariance. Qualitatively it is not a problem as there are still two orientational variants for the $DO_{23}$ phase.
Conclusion and Perspective {#sec5}
==========================
The present paper treats of the specific case of the interplay between the $L1_2$ metastable phase and the $DO_{23}$ ordered ground-state during the order-disorder transition in $Al_3 Zr$ alloy. It is proved that for a sufficiently low temperature and weak solute saturation the metastable phase nucleates before the stable phase. It is the result of the dynamics of the phase transition that is deeply influenced by the microelasticity induced by the strain of precipitates. The metastable ordered phase with a weak internal stress may be favored with respect to the ordered ground-state which induces a much larger strain. The strain-induced elasticity may play a role at the early regime of the dynamics of the phase transition. Furthermore, depending on external variables, temperature and composition, we found different kinetics for the vanishing of the $L1_2$ precipitates. At low temperature and low saturation of solute, the $DO_{23}$ precipitates grow preferentially at the interfaces of $L1_2$ inclusions with solid solution. At higher temperature or equivalently higher $Zr$ saturation, the $DO_{23}$ precipitates nucleate into the solid solution. Once $DO_{23}$ precipitates have nucleated, they grow at expense of the $L1_2$ structure that disappears. Actually, such phenomena have not yet been observed experimentally. With this respect, the phase field method can be considered as a predictive method, though experimental confirmation is now required. On that trail, microscope analysis are programmed in the Laboratoire d’Etude des Microstructures (ONERA).
It is well known experimentally that the $L1_2$ microstructure has better mechanical properties than $DO_{23}$. As the $L1_2$ is metastable, the degradation of the mechanical properties with increasing temperature cannot be avoid. Nevertheless our results allow to hope that it is possible to increase the robusness of the $L1_2$ microstructure playing with elastic interaction. We expect our study will contribute to the understanding of precipitates formation and to improve the control of the alloys synthesis. To that purpose, the phase field method we use, can be extended for other alloys with similar phase transition as $Ti_3 Al$ or $Pd_3 V$.
Finally our calculations are only valid for coherent sample at the nanometer scale. It is possible to investigate non-coherent effect introducing dislocations in the phase-field method as it is discribed in both [@Khatcha-2000] [@Alphonse-2000]. It is of great interest to perform such simulation in the case of $Al_{3}Zr $ where discontinuous precipitations and dislocations modify strongly the precipitation process [@Nes; @Ryum] and play an important role in macroscopic properties of materials.
[99]{}
Kamara, A.B., A.J. Ardell and C.N.J. Wagner, [*Metall. and Mater. Trans. A*]{}, [**27A**]{}, 2888 (1996).
A.G. Khachaturyan, [*Theory of Structural Transformations in Solids*]{}, Ed. Wiley-Interscience (1983).
L.Q. Chen and A.G. Khachaturyan, [*Phys. Rev.Lett.*]{}, [**70**]{}, 1477 (1993).
A.G. Khachaturyan, in [*Phase Transformations and Evolution in Materials*]{}, Eds. P.E.A. Turchi and A. Gonis (2000), p.3.
M.S. Zedalis and M.E. Fine, [*Metallurgical Transactions A*]{}, [**17**]{}, 2187 (1986).
C. Amador, J. Hoyt, B. Chakoumakos and D. de Fontaine, [*Phys. Rev. Lett.*]{}, [**74**]{}, 4955 (1995).
Jian-hua Xu and A.J. Freeman, [*Phys. Rev. B*]{}, [**40**]{}, 11927 (1989); [**41**]{}, 12553 (1990).
A.E. Carlsson and P.J. Meschter, [*J. Mater. Res.*]{}, [**4**]{}, 1060 (1989).
M. Alatalo, M. Weinert, and R. E. Watson, [*Physical Review B*]{}, [**57**]{}, 2009 (1998).
I.M. Lifshitz and V.V. Slyozov, [*J. Phys. Chem. Solids*]{}, [**19**]{} 35 (1961);\
C. Wagner, [*Z. Elektrochem*]{} [**65**]{} 581 (1961).
Yunzhi Wang, L.Q. Chen and A.G. Khachaturyan, in [*Computer in Material Science*]{}, vol. [**308**]{} NATO ASI Series, Eds. H.O. Kirchner, L.P. Kubin and V. Pontikis (1996), p.325.
D. Lazarus, [*Phys. Rev.*]{}, [**76**]{}, 545 (1949).
J.D. Ganton, M.M. San, and P.S. Mahni, [*Phase Transitions and Critical Phenomena*]{} Academic Press, vol. 8, Eds. C. Domb and J.L. Lebowitz (1983), p.269.
$http://www.ulib.org/webRoot/Books/Numerical_Recipes/bookf90.html$.
Y. U. Wang, Y. M. Jin, A. M. Cuitiño, and A. G. Khachaturyan, [*Nanoscale Phase Field Microelasticity Theory of Dislocations: Model and 3D Simulations* ]{}, submitted.
A. Finel, [*Phase Field Theory of Dislocations* ]{}, submitted.
E. Nes, [*Acta Metallurgica*]{} [**20**]{}, 499 (1972);\
E. Nes and H. Billdal, [*Acta Metallurgica*]{} [**25**]{}, 1039 (1977).
N. Ryum, [*Acta Metallurgica*]{} [**17**]{}, 269 (1969).
[^1]: Present address: Groupe de Physique des Solides, UMR 7588 - CNRS, Universités Paris 6 et Paris 7, 2 pl. Jussieu 75251, Paris Cedex 05, France
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Certain quantum states are well-known to be particularly fragile in the presence of decoherence, as illustrated by Schrodinger’s famous gedanken cat experiment. It has been better appreciated more recently that quantum states can be characterized in a hierarchy of quantum quantities such entanglement, quantum correlations, and quantum coherence. It has been conjectured that each of these quantities have various degrees of fragility in the presence of decoherence. Here we experimentally confirm this conjecture by preparing tripartite photonic states and subjecting them to controlled amounts of dephasing. When the dephasing is applied to all the qubits, we find that the entanglement is the most fragile quantity, followed by the quantum coherence, then mutual information. This is in agreement with the widely held expectation that multipartite quantum correlations are a highly fragile manifestation of quantumness. We also perform dephasing on one out of the three qubits on star and $ W \bar{W} $ states. Here the distribution of the correlations and coherence in the state becomes more important in relation to the dephasing location.'
author:
- Huan Cao
- Chandrashekar Radhakrishnan
- Ming Su
- 'Md. Manirul Ali'
- Chao Zhang
- 'Y.F. Huang'
- Tim Byrnes
- Chuangfeng Li
- 'Guang-Can Guo'
title: Fragility of quantum correlations and coherence in a multipartite photonic system
---
[^1]
[^2]
Introduction
============
One of the main challenges in the development of quantum technologies is how to overcome decoherence [@steane1998quantum; @dowling2003quantum; @ladd2010quantum]. Quantum systems tend to couple very easily to their external environment, losing their quantum nature, reducing to a classical state [@breuer2002theory; @zurek2003decoherence]. It is however also well-known that the time scale for which a quantum state decoheres is very much a state-dependent process. For example, superpositions of macroscopically distinct states, such as Schrodinger cat states $ |0 \rangle^{\otimes N} + |1 \rangle^{\otimes N} $, where $ N $ is the number of qubits, collapse exponentially faster in comparison to a product state of qubits $ (|0 \rangle + |1 \rangle)^{\otimes N} $. The fragility (or conversely the robustness) of quantum states have been studied in numerous studies [@frowis2018macroscopic; @yu2002phonon; @shimizu2002stability; @janzing2000fragility]. The fragility of quantum states has been discussed in connection to measures of defining the macroscopicity of quantum superpositions [@frowis2018macroscopic; @frowis2012measures; @dur2002effective]. The fragility of particular quantum states can be considered the flip-side of the enhanced sensitivity of such states, the classic example being NOON states, which are fundamental in the field of quantum metrology [@boto2000quantum; @dowling2008quantum; @giovannetti2011advances].
Meanwhile, quantum information theory has provided numerous tools in order to better understand the nature of quantum states. Various quantifiers for strength of Bell correlations [@bell1964einstein; @einstein1935can], EPR steering [@wiseman2007steering], entanglement [@horodecki2009quantum], and quantum correlations [@ollivier2001quantum; @henderson2001classical] have been proposed, each characterizing different aspects of quantum states. For example, entanglement is strictly defined as any state that is not writeable in separable form, whereas quantum correlations arise when it is impossible to disturb a quantum state with local projective measurements [@ollivier2001quantum]. Recently another quantifier, quantum coherence, has attracted attention as another way of characterizing quantum states [@baumgratz2014quantifying]. Unlike quantum correlations that require at least bipartite systems to exist, quantum coherence can occur on a single system, and is a measure of the degree of superposition [@radhakrishnan2016distribution; @tan2016unified]. These quantifiers form a hierarchical structure, where quantities higher in the hierarchy possess attributes non-zero values of lower quantites [@adesso2016measures; @ma2019operational]. For example, a system possessing entanglement necessarily possesses quantum correlations and coherence, but does not necessarily show Bell correlations or steering. In particular, a unified theory connecting various types of quantum correlations was proposed by Modi, Vedral, Williamson, and co-workers in Ref. [@modi2010unified]. Giorgi and Zambrini extended this approach to include various types of coherence in Ref. [@giorgi2018hallmarking]. Various quantum technological tasks rely on different properties of quantum states, hence one of the major aims of quantum information theory is to understand the operational capability of these different resources [@chitambar2019quantum; @winter2016operational; @brandao2015reversible; @radhakrishnan2017quantum; @chitambar2016relating; @horodecki2013quantumness; @chitambar2016critical]. How these resources behave in a dynamical context has been a focus of several works [@xu2010experimental; @xu2010experimental2; @xu2013experimental; @bernardes2015experimental], motivated by the presence of environmental decoherence in quantum technological systems.
In this study, we experimentally show the effect of the different quantum correlations and coherences of a tripartite photonic system under the influence of a one and three qubit dephasing environment. We measure the six quantities: (1) entanglement, (2) total coherence, (3) global coherence, (4) local coherence, (5) mutual information, and (6) classical correlations and measure their decay dynamics under dephasing. The fragility of these quantities under dephasing is investigated by measuring the decay rate, which can quantify the fragility of the quantity under question. We note that investigations on the transient dynamics of entanglement and quantum discord have been performed in Refs. [@bellomo2007non; @eberly2007end; @lopez2008sudden; @yu2009sudden; @xu2010experimental; @xu2010experimental2; @xu2013experimental; @bernardes2015experimental; @mazzola2010sudden; @werlang2009robustness; @maziero2009classical]. Particularly in Ref. [@xu2010experimental; @xu2010experimental2; @xu2013experimental; @bernardes2015experimental] an experimental verification of the decay dynamics has been examined. In our work we focus on studying the [*comparative*]{} dephasing dynamics of different quantum properties using relative entropy measures. To observe the decay dynamics of multipartite quantum states, we generate the $ W \bar{W} $ and star states, which contain correlations and coherences at all levels. Such states are uniquely suited for examining multiple quantum properties simultaneously.
{width="\linewidth"}
Photonic state generation
=========================
$W \bar{W}$ and Star states
---------------------------
In this study we generate and study the dynamics of two quantum states under dephasing. The first state is the $W \bar{W}$ state defined as $$\begin{aligned}
|W \bar{W} \rangle &= \frac{1}{\sqrt{2}} \left( |W \rangle + |\bar{W} \rangle \right) \\
|W \rangle & = \frac{1}{\sqrt{3}} \left( |001 \rangle + |010 \rangle + |100 \rangle \right) \\
| \bar{W} \rangle &= \frac{1}{\sqrt{3}} \left( |110 \rangle + |101 \rangle + |011 \rangle \right) \end{aligned}$$ The $W \bar{W}$ state is an equal superposition of a standard $ W $ state and its spin-flipped version, the $ \bar{W}$ state. This type of state is chosen because it has quantum coherence at the single qubit, bipartite and tripartite levels, as well as bipartite and tripartite quantum correlations. Such a state is a good testbed for studying quantum correlations distributed at different levels. The presence of different types of correlations is one of the reasons that $W$ states are robust under local decoherence [@dur2000three].
The second state we investigate is the star state defined as $$\begin{aligned}
|S \rangle &= \frac{1}{2} \left(|000 \rangle + |100 \rangle + |101 \rangle + |111 \rangle \right).
\label{starstatedef} \end{aligned}$$ Like the $W \bar{W}$ state, the star state also has coherence and correlations distributed at all possible levels. However, the correlations are present in an asymmetric way for a star state, in contrast to the $W \bar{W}$ state which is symmetric for all qubits. The entanglement structure for the star state takes a form $ A \Leftrightarrow C \Leftrightarrow B $, where we have labeled the three qubits as $ ABC $ in (\[starstatedef\]) from left to right. For example, if qubits $ A $ or $ B $ are traced out, entanglement is present in the remaining qubits. However, if qubit $ C $ is traced out, the remaining qubits are left in a separable state. We thus call qubit $C$ is the central qubit, and qubits $ A $ and $ B $ the peripheral qubits. The star state is a very simple example of a graph state [@plesch2003entangled] which in multipartite cases are useful for quantum error correction [@anders2006fast]. More details on the distribution of correlations and coherence in the $W \bar{W}$ and star states are given in the Supplementary Material.
Experimental Preparation
------------------------
To experimentally realize the above states, polarization encoded photonic qubits are used, where the horizontal (H) and vertical (V) polarizations are encoded as the two levels $|0 \rangle$ and $|1 \rangle$ respectively. The detailed procedure of preparing these quantum states is shown in Fig. \[fig1\]. In our experiment we investigate the dynamics of various correlations and coherence in a tripartite quantum system which is under the influence of an external phase damping environment, realized by passing the photonic states through birefringent quartz crystals of different thicknesses. We perform two types of dephasing, where all three photons are dephased by a crystal of the same thickness, and another where only one of the photons is dephased. The dephasing on only one of the photons allows for a partial dephasing of the system, where some quantum property is retained even after complete dephasing.
The experimental set up to prepare the $W \bar{W}$ is shown in Fig. 1(a). Two pairs of down converted photons are simultaneously generated through a higher order emission of spontaneous down conversion (SPDC) process. These four photons are collected by a single mode fiber and then fed into a polarizing beam splitter (PBS) where they overlap and become indistinguishable in the spatial mode. The spectral selection is realized by inserting a 3nm interference filter after the PBS. The four photons are separated by three non-polarizing beam splitters (BS). The post selected four-fold coincidence count certifies the generation of four photon Dicke state with two excitations $|D_{4}^{2} \rangle = (|0011 \rangle + |0101 \rangle + |1001 \rangle + |0110 \rangle + |1010 \rangle + |1100 \rangle) /\sqrt{6}$. The $W \bar{W}$ state is generated from the Dicke state by projecting one of the qubits into the $(|0 \rangle + |1 \rangle)/\sqrt{2}$ basis.
The star state generation scheme is shown in Fig. 1(b). Two non-maximally entangled bipartite states $|\psi \rangle = \cos \theta |01 \rangle + \sin \theta |10 \rangle$ with the ratio $\cos^{2} \theta$:$\sin^{2} \theta$ $=6.8554$, are required to prepare the star state. These polarization entangled states are generated using a sandwiched geometry beam-like type II BBO entanglement resource. Such an entanglement resource was first devised by Zhang and co-workers in Ref. [@zhang2015experimental] and was later used in Ref. [@wang2016experimental] to realize ten photon entanglement. Applying single qubit unitary operators on each qubit, the state $|\psi \rangle$ is transformed to $(|00 \rangle + |10 \rangle + |11 \rangle)/\sqrt{3}$. The transformed states are fed into the PBS to overlap them and the Hong-Ou-Mandel interference visibility is enhanced using a 2nm band pass filter. The second one of the four qubit quantum states generated through this process is projected in the $(|0 \rangle + |1 \rangle)/\sqrt{2}$ basis. By exchanging the qubits $3$ and $4$ in the resulting quantum state, the star states are obtained.
Example reference state
--------------------------------------------------------------------- ----------------------------------- --------------------------------------------------------------------------- --------------------------------------------------------------------------------------
$ \sum_j p_j \rho^A_j \otimes \rho^B_j \otimes \rho^C_j $ $ E = \displaystyle{\min_{\sigma \in \mathcal{S}}} \; S(\rho \| \sigma)$
$\rho_d = \sum_{j} \langle j | \rho | j \rangle |j \rangle \langle j | $ $C = \displaystyle{\min_{\sigma \in \mathcal{I}}} \; S(\rho \| \sigma) $
Incoherent states $ \mathcal{I} $ $\pi_d(\rho) = \rho^A_d \otimes \rho^B_d \otimes \rho^C_d $ $ C_{L} = \displaystyle{\min_{ \pi(\rho) \in \mathcal{I}}} S(\pi(\rho) \| \sigma ) $
$\underset{\text{(Total correlations)}}{\text{Mutual Information}}$ $ \pi(\rho) = \rho^A \otimes \rho^B \otimes \rho^C $ $T= \displaystyle{\min_{\sigma \in \mathcal{P}}} \; S(\rho \| \sigma) $
$ \pi(\rho_d) $ $ K = \displaystyle{\min_{\sigma \in \mathcal{P}}} S(\rho_d \| \sigma ) $
$ \pi_d(\rho) = \pi(\rho_d) $ $M = \displaystyle{\min_{\sigma \in \bar{\mathcal{I}}}} \; S(\rho \| \sigma) $
Measures of correlations and coherence
======================================
We measure the correlations and coherence using the unified distance-based approach of Ref. [@modi2010unified]. The basic idea of any distance-based approach to quantify a quantum observable is as follows. First the set of all states do not have the relevant quantity is defined, and are called the reference states. For example, for entanglement, the reference states are the set of all separable states. Then to quantify the quantum property, one uses a suitable distance measure to find the distance to the closest reference state by minimization. In our case, we choose the distance measure to be relative entropy $$S(\rho \| \sigma) = \hbox{Tr} (\rho \ln \rho - \rho \ln \sigma),
\label{REdensitymatrices}$$ which is a popular choice due to its simplicity of computation and well-known properties [@vedral2002role]. The six quantities that we calculate are defined as below and summarized in Table \[table1\].
[*Entanglement:*]{} The entanglement is quantified as the minimum distance to the set of all separable states [@vedral1997quantifying; @vedral1998entanglement]. We perform a minimization procedure to separable states taking the form $ \sum_j p_j \rho^A_j \otimes \rho^B_j \otimes \rho^C_j $, where $ p_j $ is a probability and $ \rho^{A,B,C}_j $ are density matrices on subsystems $ A, B, C $.
[*Coherence:*]{} The total quantum coherence [@baumgratz2014quantifying] is defined as the distance to the closest incoherent state, which take the form $ \sum_{j}p_j |j \rangle \langle j | $, where $ |j \rangle $ are in the basis $ \{ |0 \rangle, |1 \rangle \} $ for $ A, B, C $. It has been shown that for the relative entropy, the closest incoherent state to a state $ \rho $ takes coefficients $ p_j = \langle j | \rho | j \rangle $, hence the minimization does not need to be explicitly performed [@baumgratz2014quantifying] and $$C(\rho) = \min_{\sigma \in \mathcal{I}} S(\rho \| \sigma) = S(\rho \| \rho_d ) = S(\rho_{d}) - S(\rho).
\label{totalcoherence}$$ Here we defined $ \rho_d $ as the matrix $ \rho $ with all off-diagonal terms set to zero in the basis $ | j \rangle $.
[*Local and global coherence:*]{} Quantum coherence can originate from both coherence which is localized on subsystems, or due to coherence due to a collective property of the whole system [@radhakrishnan2016distribution; @tan2016unified]. The former is called local coherence and is found by first breaking all the correlations between the subsystems. In a similar way to total coherence, the closest incoherent state is found by taking the diagonal form $$\begin{aligned}
C_{L}(\rho) & = \min_{\sigma \in \mathcal{I}} S(\pi(\rho) \| \sigma ) = S(\pi(\rho) \| \pi_d (\rho) ) ,\end{aligned}$$ where $ \pi_d (\rho) $ is the matrix $ \pi(\rho) $ but with all off-diagonal elements set to zero in the basis $ | j \rangle $. The coherence attributed to the collective nature of the system is called the global coherence and is defined as the difference of the total and local coherence $$\begin{aligned}
C_{G} (\rho) & = C(\rho) - C_{L}(\rho) . \end{aligned}$$
[*Mutual Information:*]{} Mutual information measures the total amount of correlations, including both quantum and classical parts [@modi2010unified]. The set of uncorrelated states takes the form of a product state $ \sigma^A \otimes \sigma^B \otimes \sigma^C $. It has been shown in Ref. [@modi2010unified] that for relative entropy the closest product state is the product state $ \pi (\rho) = \rho^A \otimes \rho^B \otimes \rho^C $ consisting of the reduced density matrices on each subsystem $ \rho^{A,B,C} $. Hence we can write $$T(\rho) = \min_{\sigma \in \mathcal{P}} S( \rho \| \sigma ) = S( \rho \| \pi(\rho)) \equiv S(\pi(\rho)) - S(\rho) .
\label{mutualinformation}$$ The total correlations as measured by the mutual information $T$ and the total quantum coherence $C$ are not completely independent quantities. Hence there is a common region of quantumness in a system which is measured by both these quantities. This region of overlap is the amount of global coherence in the system which arises due to quantum correlations between the qubits.
[*Classical correlations:*]{} For local coherence, first the correlations between the subsystems are broken, then the remaining coherence is measured. The reverse ordering can equally be performed, where first the coherence is removed from the system, then the remaining correlations are measured. The state with no coherence is $ \rho_d $, which can only contain classical correlations because it is a diagonal density matrix [@giorgi2018hallmarking]. In the same way as mutual information, the closest uncorrelated state is its corresponding product state, $$K(\rho) = \min_{\sigma \in \mathcal{P}} S(\rho_d \| \sigma ) = S(\rho_{d} \| \pi(\rho_{d})).
\label{classicalcorrelations}$$
[*Hookup:*]{} The reference state for total coherence $ C $ is $ \rho_d $, which is a state which has no coherence, but potentially classical correlations. Meanwhile, the reference state for the mutual information $ T $ is $ \pi (\rho) $, which has no correlations but potentially coherence. One can define a quantity with a reference state that has no correlations and no coherence. This was called the “hookup” in Ref. [@giorgi2018hallmarking] and can be evaluated to be $$M(\rho) = C(\rho) + K(\rho) = T(\rho) + C_{L} (\rho).
\label{hookup}$$ A detailed overview on the various correlations and coherence is given in the Supplementary Material.
Density matrix evolution under dephasing
========================================
Tomography reconstruction of states
-----------------------------------
Fig. \[tomography\] shows the tomographic reconstructions of the star and $ W \bar{W} $ states with various amounts of dephasing. For the case that the dephasing is applied to all the photons, the density matrix approaches its diagonal form as expected for larger values of $ \ell $, the thickness of the quartz plate. The case where dephasing is only applied to one of the photons, off-diagonal terms remain since the state is only partially dephased. This is due to the nature of the star and $ W \bar{W} $ states that are used which contain other types of coherence other than completely tripartite coherence (such as in a GHZ state). The tomographically reconstructed density matrix is compared to the theoretically calculated density matrix according to a dephasing channel for each qubit defined as $$\begin{aligned}
\rho \rightarrow (1- p(\ell)) \rho + p(\ell) \sigma_z \rho \sigma_z ,
\label{dephasingch}\end{aligned}$$ where $p(\ell) = [1-\exp(-\Gamma \ell^{2})]/2$ (see Supplementary Material). We obtain fidelities of the state with dephasing better than $ 93 \% $ for all dephasing values.
Decay of correlations and coherence with dephasing on all qubits
----------------------------------------------------------------
Using the tomographically reconstructed density matrices we calculate the various quantities summarized in Table \[table1\]. First we discuss the dephasing dynamics of the correlations in a $W \bar{W}$ state, as shown in Fig. \[fig3\](a) and (b). We observe that all quantities decay to zero for large dephasing, except the mutual information $T$ and classical correlations $K$, which saturate to finite values. This is due to the dephasing removing all coherence from the system, such that the state $$\begin{aligned}
\rho_d &=& \frac{1}{6} \Big( |001\rangle \langle 001| + |010\rangle \langle 010| + |100\rangle \langle 100| \nonumber \\
& & + |110\rangle \langle 110| + |101\rangle \langle 101|+ |011\rangle \langle 011| \Big)\end{aligned}$$ is progressively approached. This is a classically correlated state and hence the mutual information only contains classical correlations $ T = K $ as observed, and all other quantum properties decay to zero. In Fig. \[fig3\](b), we see that the global coherence starts at a larger value than the local coherence, but the global coherence decays faster than the local coherence. This is an indication of the greater robustness of the local coherence in the presence of dephasing than global coherence.
To examine this point in more detail, we plot the decay rates for the various quantities in Fig. \[fig3\](c). Due to the Gaussian nature of the dephasing channel (\[dephasingch\]), we expect the quantum properties to also approximately follow a Gaussian form $ \propto \exp( - \Gamma \ell^{2} ) $, hence the decay rate is the negative gradient on a semilog plot with $ \ell^2 $. Of all the quantum properties the fastest decay is for entanglement. The next fastest decay rate is displayed by global quantum coherence, followed by the total coherence. The very slow decay of mutual information is because it is composed of both quantum correlations and classical correlations. While quantum correlations decay due to the environment, the classical correlations remains unchanged, since the dephasing acts in the classical basis $ |0 \rangle, |1 \rangle $. Likewise, local coherence can be seen to decay more slowly than the total coherence. These results generally show that the quantities that are related to collective effects, such as entanglement and global coherence tend to decay at a faster than classical or local quantities. The star state generally shows similar behavior, as seen in Fig. \[fig3\](d) and (e). Here again the mutual information and classical correlations saturate towards a non-zero value, according to the classical correlations in the state $$\begin{aligned}
\rho_d = \frac{1}{4} \left( |000\rangle \langle 000| + |100\rangle \langle 100| + |101\rangle \langle101| + |111\rangle \langle 111| \right) . \end{aligned}$$ All other quantities decay to zero, in a similar way to the $ W \bar{W} $ state. The total coherence is less in the star state due to the smaller number of terms in the superposition. Nevertheless, as seen by evaluating the decay rates in Fig. \[fig3\](f), the entanglement shows the greatest rate of decrease, followed by the global and total coherences. The mutual information and local coherences decay with the slowest rates, similar to the $ W \bar{W} $ state. Thus despite the rather different structure of the states, a consistent picture emerges once the decay rates are examined.
{width="\linewidth"}
{width="\linewidth"}
Decay of correlations and coherence with one qubit dephasing
------------------------------------------------------------
One way of understanding the faster decay of the collective quantities such as entanglement and global coherence is that they are exposed to the dephasing effects from multiple qubits. This is in contrast to quantities that are localized on each qubit, such as local coherence, which can only affect one qubit at a time. In this picture, if the dephasing is only applied to one qubit, then we might expect that the rates for all quantities will be more similar. To test this hypothesis, we also perform dephasing on one qubit and investigate its effect on the various quantities as before.
The decay of various quantities for the $W \bar{W}$ due to dephasing is shown in Fig. \[fig4\](a) and (b). Due to the symmetric nature of the state, dephasing any one of the three qubits leads to the same result, hence in our case qubit $B$ is dephased. In this case all quantities saturate to a non-zero value, which is characteristic of the $W \bar{W}$ state. As is well-known, dephasing of a $W$ state only partially removes the entanglement from the system, and the remaining qubits are partially entangled. This means that both quantum correlations and coherence are preserved in the system. Due to the quantum correlations that are preserved in this case, we observe that the amount of correlations and coherence are always larger than the amount of classical correlations, in contrast to the three qubit dephasing case.
The entanglement structure of the state plays a more important role in the case of star states, as seen in Fig. \[fig4\](d) and (e). For the star state we show the effects of dephasing on the central qubit $ C $. In this case we observe the entanglement decaying to zero for large dephasing, as expected from the discussion surrounding Eq. (\[starstatedef\]). For dephasing on a peripheral qubit, we find that the entanglement does not decay to zero, in a similar way to the $ W \bar{W} $ state (see the Supplementary Material). Other quantities saturate to non-zero values, with the steady state value of the global coherence being higher than the amount of classical correlations in the system. This is in contrast to the entanglement and local coherence in which the steady state value is lower than the classical correlations. We note that compared to the other quantities the local coherence exhibits very minimal evolution due to dephasing.
Fig. \[fig4\](c) and (f) shows a comparison of the decay rates of the various quantities, which appears as the negative gradient on the semilog plot. We find that the ordering of the decay rates do not occur in a consistent order as before. For the $ W \bar{W} $ state, we find that all quantities generally decay with a similar rate, with the global coherence giving the largest value. On the other hand, for the star state, we clearly see that the entanglement decays at the fastest rate, in a similar way to the three qubit dephasing case. We attribute this to the different structure of entanglement that is present in the two states. For the $ W \bar{W} $ state, all the qubits can be considered “peripheral” qubits, since the dephasing only causes partial loss of entanglement. In the case of dephasing the central qubit of the star state, the destruction of entanglement is very effective, since it is a central qubit for the entanglement. Thus in this case we observe that the structure of the quantum correlations greatly affect the fragility of the state.
Summary and Conclusions
=======================
The effects of dephasing on quantum correlations and coherence was experimentally studied on photonic $W \bar{W}$ and star states, with one and three qubit dephasing. Such states have coherence and correlations of all types in a tripartite system. Using a Gaussian dephasing model, we are able to extract the effective decay rates for each state and each type of dephasing, as shown in Figs. \[fig3\](c) and (f) and \[fig4\](c) and (f). In the case that dephasing is applied on all the qubits, a consistent picture emerges, despite the different nature of the states. Here we find that $$\begin{aligned}
\Gamma(E) > \Gamma(C_G) > \Gamma(C) > \Gamma(C_L) > \Gamma(T) > \Gamma(K) ,
\label{robustnessordering}\end{aligned}$$ i.e. the dephasing rates occur in the order of entanglement, global coherence, total coherence, local coherence, mutual information, and classical correlations. We thus see a clear hierarchy in the decay rate of the various quantum properties, where the collective quantities decay at a faster rate than local and classical quantities. This can be understood as the result of collective quantities being affected by all the channels of dephasing, but local quantities only being affected by its local dephaser. In this way we verify the conjecture that collective quantities are more fragile than the local quantities, when local decoherence is applied on the whole system. For the case that only one qubit is dephased, the rates of decay depend more on the structure of the quantum state. In the case of dephasing the central qubit of a star state, we again recover the entanglement is the most fragile quantity. However, in the case of dephasing a peripheral qubit where entanglement can be retained in the strong dephasing limit, the rate of decay is much lower. Similar results were obtained in different models of dephasing theoretically [@radhakrishnan2019time].
The hierarchy between the different measurable quantities in (\[robustnessordering\]) can be viewed as a robustness hierarchy in terms of quantum properties as follows $$\hbox{NLQC} \prec \hbox{TQC} \prec \hbox{LS} .$$ In the above equation NLQC, TQC and LS stand for nonlocal quantum correlation, total quantum correlation and local superposition respectively and the notation $A \prec B$ denotes that $A$ decays faster than $B$. The NLQC is unique to quantum systems and TQC (both nonlocal and local quantum correlations) are inter-qubit correlations distributed between the qubits. LS is the superposition between the levels of a qubit and hence is a intra-qubit property which is localized within a qubit. Hence we find that the inter-qubit quantum properties which are spread out between the qubits are more likely to decay much faster when compared to the intra-qubit quantum properties which are relatively more robust. This suggest that in quantum information theoretic tasks it would be advantageous to use intra-qubit quantum properties as resources as they can be preserved over longer time intervals. By converting between local coherence to global coherence only when it is needed [@wu2018experimental], this could be used as a strategy for preserving coherence to longer times.
We note that in our approach the classical correlations are constant throughout the entire process of evolution. This is contrast to the theoretical results observed in Refs. [@maziero2009classical; @maziero2010system] and were subsequently experimentally examined [@xu2010experimental2]. The difference here originates from the different notions of classicality as defined by quantum discord and quantum coherence. In quantum discord, a state is classical correlated if there exists a local measurement and a conditioned measurement, in any basis, which does not disturb the quantum state [@ollivier2001quantum; @radhakrishnan2019quantum]. It is therefore a quantity that is invariant under local basis transformations. In contrast, coherence is a basis dependent quantity [@radhakrishnan2019basis]. The classical nature of the state is with respect to a particular basis choice, in our case the $ |0\rangle, |1 \rangle $ basis. Here our notion of classical correlations is in this fixed basis choice, and the dephasing removes coherence in this basis. This means that the classical correlations are always unchanged under this evolution. In the case of Refs. [@maziero2009classical; @maziero2010system; @xu2010experimental2], classical correlations can be dynamic because of the local basis optimization that is performed in evaluating the discord. In our view, these results are not inconsistent, but arise from different notions of classicality. In our approach, there is a preferred classical basis $ |0\rangle, |1 \rangle $, which is natural to consider since this is the basis that dephasing occurs in the system.
Another observation that can be made from Fig. \[fig3\] and \[fig4\] is that the amount of total quantum coherence is always higher than the entanglement present in the system. This is because the coherence originates due to nonlocal quantum correlations, local quantum correlations and local superpositions, whereas the entanglement arises only due to the nonlocal quantum correlations. This enables us to verify the theorem $E(\rho) \leq C(\rho)$ in Ref. [@streltsov2015measuring] in a dynamical scenario, when they are both measured using the same contractive distance. This relationship between entanglement and coherence was proved in Ref. [@streltsov2015measuring] under the condition that both these quantities are measured using the same contractive distance. In our work we also use the same contractive distance (relative entropy) and also verify that the relation holds under dephasing dynamics as well.
Our work demonstrates that various quantum information quantities can be used to effectively characterize quantum systems. These can be extended to larger quantum systems, where more dramatic phase transition phenomena can be observed [@radhakrishnan2017quantum]. Adding decoherence and observing the dynamics can be a direct quantifier for the fragility of various quantities. Since particular quantities are more relevant for a given quantum information task, this general method may find also practical uses in the context of applications to quantum technology.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is supported by the Shanghai Research Challenge Fund; New York University Global Seed Grants for Collaborative Research; National Natural Science Foundation of China (61571301,D1210036A); the NSFC Research Fund for International Young Scientists (11650110425,11850410426); NYU-ECNU Institute of Physics at NYU Shanghai; the Science and Technology Commission of Shanghai Municipality (17ZR1443600); the China Science and Technology Exchange Center (NGA-16-001); and the NSFC-RFBR Collaborative grant (81811530112).
[59]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, , , , , , ****, ().
, , , ** (, ).
, ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , , , , , ****, ().
, , , , , ****, ().
, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, **** ().
, ****, ().
, , , , , , , ****, ().
, , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , , ****, ().
, , , ****, ().
, ****, ().
, , , , , ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , , , , ****, ().
, , , , , , , , , , , ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, , , , , ****, ().
, , , , , , , , , , ****, ().
, , , , , ****, ().
, , , ().
, , , , , ****, ().
, , , , , ****, ().
[^1]: These authors contributed equally
[^2]: These authors contributed equally
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Josephson current between two one-dimensional nanowires with proximity induced $p$-wave superconducting pairing is calculated in the presence of Rashba spin-orbit interaction, in-plane and normal magnetic fields. We show that Andreev retro-tunneling is realized by means of three channels. The main contribution to the Josephson current gives a scattering in a conventional particle-hole channel, when an electron-like quasiparticle reflects to a hole-like quasiparticle with opposite spin yielding a current which depends only on the order parameters’ phase differences $\varphi$ and oscillates with $4\pi$ period. Second anomalous particle-hole channel, corresponding to the Andreev reflection of an incident electron-like quasiparticle to an hole-like quasiparticle with the same spin orientation, survives only in the presence of the in-plane magnetic field. The contribution of this channel to the Josephson current oscillates with $4\pi$ period not only with $\varphi$ but also with orientational angle of the in-plane magnetic field $\theta$ resulting in a magneto-Josephson effect. Third anomalous particle-particle channel, which represents a reflection of an electron-like (hole-like) quasiparticle to a electron-like (hole-like) quasiparticle with opposite spin-orientation, oscillates only with the in-plane magnetic field orientation angle $\theta$. We present a detailed theoretical analysis of both DC and AC Josephson effects in such a system showing contributions from all these channels and discuss experiments which can test our theory.'
author:
- 'E. Nakhmedov$^{1,2}$, B. D. Suleymanli$^3$, O. Z. Alekperov$^{1}$, F. Tatardar$^{1,4}$, H. Mammadov$^2$, A. A. Konovko$^{5}$, A. M. Saletsky$^{5}$, Yu. M. Shukrinov$^{6,7}$, K. Sengupta$^{8}$, and B. Tanatar$^{9}$'
title: 'Josephson current between two $p$-wave superconducting nanowires in the presence of Rashba spin-orbit interaction and Zeeman magnetic fields'
---
Introduction {#intro}
============
A key model for realization of Majorana fermion (MF) in a condensed matter is a spinless $p$-wave superconductor (SC) [@kitaev01; @kitaev03]. Majorana zero modes are excitations at zero energy which are typically localized at interface of the topological-non-topological phases and spatially separated from one another. They emerge as electrically neutral fermions indistinguishable from their antiparticles in subgap quasiparticle excitation spectrum of a topological superconductor (TSC). In $s$-wave superconductors, the quasiparticle excitations at the top of the supercoducting gap are indeed equal coherent superpositions of electrons and holes with opposite spins, and thus electrically neutral. Nevertheless, such a superposition of fermionic quasiparticles is not self-conjugate due to existence of spin. Therefore, MF can not appear in $s$-wave SC, and it is expected to occur in an effectively spinless $p$-wave SC [@kitaev01]. In $p$-wave SC model odd number of Majoranas reside at each end of the superconductor. However, electrons in conventional materials have spin-half particles; thus, the notion of a spinless SC does not seem immediately relevant to real physical systems.
Recently it was suggested that a topological (spinless $p$-wave) superconductivity can be effectively realized either in a spin-polarized normal metal or in a semiconductor nanowire with strong spin-orbit coupling under Zeeman magnetic field proximity-coupled to a conventional spin-singlet ($s$-wave) bulk SC [@lsd10; @oro10]. Spin-orbit interactions split the energy spectrum shifting the energy branches along the momentum axis. In contrast, an in-plane Zeeman magnetic field shifts the energy of up- and down-spin electrons opening a gap in the spectrum at zero momentum. When the chemical potential is located in the gap the upper spin-subband becomes empty, and the system is transformed to an effectively “spinless” electron model. However, owing to the spin-orbit interactions-induced rotation of the spins at the opposite Fermi points, the proximity-induced $s$-wave interaction opens a pairing gap in the spectrum. The resulting state is closely related to a spinless $p$-wave superconductor. One of the virtues of this model is that the proximitized nanowire can be driven into a topological phase by tuning the magnetic field or the chemical potential (Fig. \[Fig1\]). The emergence of Majorana zero modes at a certain critical value of a control parameter is necessarily accompanied by the closing of the bulk gap [@kitaev01], which corresponds to a topological quantum phase transition (that is, a quantum phase transition between topologically trivial and non-trivial states).
The topological phase in these wires is stable with respect to small perturbations (such as disorder) as long as they do not cause the bulk gap in the spectrum to collapse. The ability to realize the topological phase depends on the effective spin-orbit energy $E_{so}$, the proximity-induced gap $\Delta$ and the effective Zeeman energy $V_Z$ in the heterostructure. Note that highly controlled zero-energy quasi-particles Majoranas, produced in thin wire, can be utilized as a quantum information carrier qubit in quantum computer technology. MFs are exotic non-Abelian fermions obeying non-Abelian braiding statistics [@rg00; @ivanov01; @aoro11]. This unique property makes MF ideal for fault-tolerant quantum computation. These MFs typically arise at defect sites as Abrikosov vortex cores in bulk SCs, at interface of dielectric/TSC or normal metal/TSC, or at edges of TSC as localized excitations, and are topologically protected against any local perturbations. Essentially, two key points for the emergence of MFs are presented here: spin-orbit coupling (SOC) and superconducting proximity effect. A neutral excitation in a superconductor has a special property owing to the inherent particle-hole symmetry of the material: it is bound to zero energy, so that there is no cost to occupy such a state. One dimensional (1D) topological superconductor supports a non-local fermionic mode comprising two Majorana zero modes localized at opposite ends of the chain and are separated by a distance that can be much larger than the superconducting coherence length. An odd number of Majorana zero modes emerges per wire end; this is consistent with the fact that an even number of Majorana modes can pair up and locally form an Andreev state (which is a conventional fermionic state).
Recent investigations have shown [@gr01; @ljl14; @rm15; @yw16] that a proximity of semiconductor with $s$-wave superconductor induces not only $s$-wave but also $p$-wave superconducting pairing in the semiconductor. The pairing symmetry of a BCS-type two dimensional (2D) superconductor without inversion symmetry when the twofold degeneracy of the electron energy spectrum is lifted by SOC has been studied [@gr01; @edelstein89; @faks04] to be a mixture of singlet and triplet symmetries. Recently Reeg and Maslov have shown [@rm15] by directly solving the fully quantum-mechanical Gor’kov equations that spin-triplet superconducting correlations are induced by Rashba SOC in both 1D and 2D proximity junctions via the proximity effect. Furthermore, the induced triplet component in 1D was shown to vanish when integrated over the momentum; this result is in agreement with Ref.. The induced triplet amplitude in 2D was found to have an odd-frequency component that is isotropic in momentum.
In this work we study Josephson junction ($JJ$) of two superconductors with $p$-wave pairing of spinfull electrons separated by a $\delta$-function like insulator potential. $JJ$ consisting of $p$-wave superconductors in both sides of the junction has been studied [@ksy04] for a $\delta$-like insulator potential in the absence of SOC and Zeeman magnetic field. Since the superconductor is in the topological phase, a fractional oscillation of Josephson current was obtained in Ref. \[\]. The emergence of Majorana zero mode results in exotic Josephson effects [@kitaev01; @fk09; @jpar11; @jpar13; @nats17], when the current flowing between two topological superconductors in the junction oscillates with a fractional periodicity $4 \pi$ instead of $2\pi$ periodicity in a conventional Josephson junction. Additionally, “spin Josephson current” [@nb04; @asano06; @brydon09; @bat11] may flow across the junction, which is shown [@jpar11; @jpar13] also to be $4\pi$ periodic in the field orientations as a manifestation of the Majorana modes.
Josephson current in a junction of two $s$-wave symmetric superconductors has been studied by us in our previous work [@nats17]. In this work we study the Josephson current through a junction of two superconductors with $p$-wave pairing of spinfull electrons. This problem has been usually studied for spinless model [@kitaev01], although many aspects of spinfull $p$-wave symmetric $JJ$ have been investigated [@zclw12; @kecs14; @ehm16; @bs19; @mkc19] in the literatures. Furthermore, in this paper we want to understand how do SOC and external magnetic fields change Josephson current in p-type $JJ$ separated by a $\delta$-potential insulator. Similar effects have studied in Ref.for a junction consisting of $s$-wave SCs separated by $\delta$-potential thin insulator. Note that, the Josephson current in the case of proximity induced both $s$- and $p$-wave pairings in $JJ$, as argued in several recent papers [@gr01; @ljl14; @rm15; @yw16], can be found by summing up a corresponding result of Ref. \[\] with that presented in this paper. We show in this paper that the Andreev tunneling occurs in three channels, and clarify the origin of these channels. Two additional channels seem to vanish with [*in-plane*]{} magnetic field. All contributions to the Josphson current oscillate with fractional periodicity either with phase difference of the SC order parameters $\varphi$ or with tilded angle $\theta$ between the in-plane magnetic field and JJ. Simultaneous action of SOC and Zeeman magnetic fields results in openning of a forbidden gap in dependence of Josephson current on the phases at some definite values of SOC and magnetic fields.
The paper is organized as follow. In the next Section formulation of the problem is presented. In Section III Andreev bound state energy is calculated for different values of the external parameters as SOC constant, Zeenam energies for the magnetic fields normal to the junction $h$ and the tilted magnetic field ${\bf B}$ lying in the junction plane and forming an angle $\theta$ with SC wire. Section IV describes Josephson current as well as magneto-Josephson effect in the junction. $ac$-Josephson current and effects of SOC and magnetic field on the Shapiro step in the $JJ$ of p-wave superconductors are studied in the last Section V.
Model and formulation of the problem {#sec2}
====================================
We consider a junction of two $1D$ nanowires of proximity induced $p_x$-wave pairing symmetry superconductivity, having the effective pairing potentials $\Delta_L$ and $\Delta_R$ on the left(L)- and right (R)-side of an insulating potential barrier separated two superconductors, in the presence of Rashba spin-orbital interaction and external Zeeman magnetic fields. Hamiltonian for such a system reads $$\hat{H}= \hat{H}_{SC} + \hat{H}_R, \label{H}$$ where $\hat{H}_{SC}$ is Hamiltonian of the nanowire in the presence of external magnetic fields and $\hat H_R$ represents Rashba SOI. The former term is given by $$\begin{aligned}
\hat{H}_{SC} &=& \int dx \Bigg\{ \sum_{\sigma, \sigma'}
\psi_{\sigma}^{\dag}(x) \Big[[\xi_{\hat k}+U(x)] \sigma_0 + h
\sigma_z \nonumber\\
&& + B \{[\sigma_x \cos \theta_1 + \sigma_y \sin \theta_1] \theta(-x) +
[\sigma_x \cos \theta_2 \nonumber\\
&& + \sigma_y \sin \theta_2]
\theta(x) \} \Big] \psi_{\sigma'}(x) \nonumber\\
&& + \left[\Delta_L \theta(-x) + \Delta_R \theta(x)\right]
\psi_{\uparrow}^{\dag}(x) \psi_{\downarrow}^{\dag}(x)+ {\rm h.c.}
\Bigg\}, \label{H-sc}\end{aligned}$$ where $\xi_{\hat k} = \epsilon
\left(\frac{\hbar}{i}\frac{\partial}{\partial x}\right) -\epsilon_F$ denotes the electron kinetic energy as measured from the Fermi level $\epsilon_F$, $\psi_{\sigma}(x)$ is the electron annihilation operator, $h$ and ${\bf B}$ are external Zeeman magnetic fields in $z$ direction and in the [*x-y*]{} plane respectively, $\theta(x)$ is the Heaviside step function, and $\sigma_{x,y,z}$ and $\sigma_0$ denote Pauli and identity matrices respectively in spin space. Note that the magnetic field ${\bf B}$ forms an angle $\theta$ with wire which can be tuned externally. In what follows, we choose ${\bf B}$ in the left side of the junction to be aligned along the wire ($\theta_L=0$) while in the right side it is chosen to make an angle $\theta$ with it ($\theta_R= \theta$). In Eq. (\[H-sc\]), the pairing potential $\Delta_R$ in the right of the junction is chosen to have a phase difference $\varphi$ compared to its left counterpart: $\Delta_R = |\Delta| \exp(i\varphi)$ and $\Delta_L = |\Delta|$. The potential $U(x)=U_0 \delta (x)$, located at $x=0$, represents the barrier potential between two superconductors. The Hamiltonian of Rashba SOI can be written as $$\begin{aligned}
\hat{H}_{R}= \sum_{\sigma, \sigma'} \int dx \psi_{\sigma}^{\dag}(x) \alpha \left[v_x \sigma_z\right]\psi_{\sigma'}(x),
\label{rashba-o}\end{aligned}$$ where $\alpha$ is the strength of Rashba SOI which is chosen to be the same for both wires.
The order parameter $\Delta_{\alpha, \beta}^t (\hat{\bf k})$ for triplet pairing with $S=1$ can be presented as [@leggett75; @ms98] $$\Delta_{\alpha, \beta}^t (\hat{\bf k})= \Delta {\bf d}({\bf k})(i {\bf \sigma} \sigma_y)_{\alpha, \beta}$$ where ${\bf d}({\bf k})$ is odd function of ${\bf k}$, ${\bf d}({\bf k})=-{\bf d}(-{\bf k})$, and can be expanded over spherical harmonics $$d_{\alpha}({\bf k})=\sum_{m=-l}^{l}b_{l,m}^{\alpha}Y_{l,m}(\hat{\bf k}).
\label{d}$$ The quantum number $l$ in Eq. (\[d\]) takes odd values $1, 3, \dots$ corresponding to states of $p, f, \dots$- pairings. The coefficient $b_{l,m}^{\alpha}$ can be identified as $\alpha$ component of the superconducting order parameter with given $l$ and $m$. For a simple case of $p$-wave pairing for $l=1$ and $m=-1, 0,1,$ $d_{\alpha}({\bf k})$ can be expressed with the appropriate expressions for the spherical harmonics $Y_{1,-1} \propto (\hat{k}_x-i\hat{k}_y)$, $Y_{1,0} \propto \hat{k}_z$, and $Y_{1,1} \propto (\hat{k}_x+i\hat{k}_y)$, as, $$d_{\alpha}({\bf k})=A_{\alpha, i}\hat{k}_i.$$ Thereby, the $p_x$-wave symmetric order parameter $\Delta_{a, \sigma}(x, k_x)$ on the $a=R, L$-side of a junction between two superconductors aligned along the $x$-axis can be expressed as, $$\Delta_{a, \sigma}(x, k_x)=\Delta_a \frac{k_x}{k_F}.
\label{px}$$
It is advantageous to use a four component field operator for bulk superconductor at the right ($a=R$) and left ($a=L$) side of the junction as, $$\begin{aligned}
\Psi^{\dag}_a(x)=\left(\psi_{a,\uparrow, +}^{\dag}(x), \psi_{a,
\downarrow,+}^{\dag}(x), \psi_{a, \downarrow, -} (x), \psi_{a,
\uparrow, -}(x)\right) \label{op1}\end{aligned}$$ Here the third subscript of the annihilation operator (which we shall designate henceforth as $b$) labels the right- ($b=+$) and the left-moving $(b=-$) quasiparticles respectively while the index $a=R,L$ denotes either right ($R=-$) or left ($L=+$) superconductor. In terms of the field operator given by Eq. (\[op1\]), the Hamiltonian (Eq. (\[H\])) can be written as $\hat{H}= \sum_{a=R,L}
\int dx \Psi^{\dag}_a(x) \mathcal{H}_a \Psi_a(x)$ using the Pauli matrices $\sigma_i$ in spin- and $\tau_i$ in particle-hole spaces. From Eqs. (\[H\]) and (\[H-sc\]), we find $$\begin{aligned}
&& \mathcal{H}_R=\xi_{k, b} \tau_z \sigma_0 + h \tau_0 \sigma_z -i k
\alpha \tau_z \sigma_z \label{H2} \\
&& + B \tau_z \left(\sigma_x \cos \theta + \sigma_y \sin \theta \right)
+|\Delta| (\tau_x \cos \varphi - \tau_y \sin \varphi) \sigma_z,
\nonumber\end{aligned}$$ and ${\mathcal H}_L= {\mathcal H}_R (\theta=0;\varphi=0)$. In Eq.(\[H2\]), the energy spectrum of the electrons are linearized around the positive and negative Fermi momenta leading to $\xi_{k,
b}= b v_F\left(-i \frac{\partial}{\partial x} - k_F\right)$, where $v_F$ is the Fermi energy. Note that the Hamiltonian $\mathcal
{H}_{R,L}$ acquires a magnetism-superconductivity duality [@nab08; @jpar13] in the absence of the kinetic term, implying that it becomes invariant under the transformation $\{\Delta,
\epsilon_F, \varphi, \tau_i \} \to \{B, h, - \theta, \sigma_i \}$. The existence of a magneto-Josephson effect in a topological insulator is known to be a result of this duality [@jpar13]. We shall see that for the system we study, the magneto-Josephson effect takes place even in the presence of the additional quadratic kinetic energy term of the electrons.
The energy spectrum of a quasi-particle in a ’bulk’ quasi-1D superconductor is determined from the expression ${\rm Det} |\mathcal{H}_R - E|=0$, yielding
$$\left(E^2 -v_F^2k^2 + \alpha^2 k^2 -h^2-B^2 -|\Delta|^2 \right)^2 +4\left(Ekv_F + h \alpha k \right)^2 -
4 |\Delta|^2 \left(v_F^2 k^2 + B^2 + h^2 \right)=0.
\label{Eo}$$
This equation does not yield a simple analytic expression for $E$, while it contains a linear in energy term, which is a result of an alignment of ${\bf h}$ and the effective magnetic field of the SOI $\propto \alpha k$. The linear in $E$ term vanishes for either $h=0$ or $\alpha =0$, and Eq. (\[Eo\]) turns to quadratic equation for $E^2$, which gives two symmetric dispersion branches for quasi-particles and quasi-holes. Note that square of the momentum $k^2_{\pm}$, where the subscripts indices indicate the spin branches, can be obtained from Eq. (\[Eo\]). The evident expression for $k^2_{\pm}$ is presented by Eq. (\[k2\]) in Appendix. Equation (\[Eo\]) is strongly simplified for different limiting cases, and yields the following expressions for the energy dispersion,
$$E=
\begin{cases}
\pm \sqrt{|\Delta|^2 -v_F^2 k^2} & \text{if} \qquad \alpha = B =h=0,
%\label{E1}
%\\
%sh \pm \sqrt{|\Delta|^2-(v_F + s \alpha )^2 k^2} & \text{if} \qquad B=0, h \neq 0, \alpha \neq 0,
%\label{E2}
\\
s\left[\sqrt{|\Delta|^2 - v_F^2 k^2} \pm \sqrt{ B^2 + h^2}\right] & \text{if} \qquad \alpha =0, B \neq 0, h \neq 0,
%\label{E3}
\\
s\sqrt{B^2+|\Delta|^2-(v_F^2+\alpha^2)k^2 \pm 2\sqrt{v_F^2\alpha^2k^4+|\Delta|^2B^2-v_F^2k^2B^2}} & \text{if} \qquad
h=0, \alpha \neq 0, B \neq 0,
\label{E4}
\end{cases}$$
where $s=\pm$ indicates the particle- and hole-branches of the spectrum. The energy levels of the Bogolyubov-de Gennes (BdG) quasi-particles lie in the gap, symmetric to the Fermi level. SOI and/or magnetic field $h$ split both electron and hole levels due to Rashba ’momentum-shifting’ and/or Zeeman effect. The ’Fermi points’ around $+k_F$ and $-k_F$ are split also due to these effects. At the same time, the magnetic field makes the energy dispersion asymmetric. Note that in our case, all energies are measured from the Fermi energy; thus the condition for realization of a topological non-trivial superconducting gapped phase is $|\Delta|^2 \ge B^2+h^2$, [@jpar13]. Indeed, zero energy mode ($E=0$) at the center of the Brillouin zone ($k=0$) appears according to Eq. (\[Eo\]) under the condition $|\Delta|^2-B^2 - h^2=0$.
BdG equations for an isolated ’bulk’ superconductor in the case of an infinitely high potential between the right (R)- and left (L) parts ($a=R, L$) of the barrier is written as $$\mathcal{H}_a \mathbf{\eta}_a(x)= E \mathbf{\eta}_a(x), \quad a= R,
L\label{Sch}$$ where the four-component vector $\eta_a(x)=\left(\eta_{a,\uparrow, +}^{\dag}(x), \eta_{a,
\downarrow,+}^{\dag}(x), \eta_{a, \downarrow, -} (x), \eta_{a,
\uparrow, -}(x)\right)$ denotes the BdG wave function. In order to get the explicit expressions for the wave functions $\eta_{a,\sigma,b}$ and $\eta_{a,{\bar \sigma},b}^{\ast}$ we write Eq. (\[Sch\]) for finite value of the external parameters ${\bf B}$, $h$ and $\alpha$ as $$\begin{aligned}
&& (E + i a b v_F k + iab \alpha k - h)\eta_{a, \uparrow, b} -
Be^{-i\theta} \eta_{a, \downarrow, b} \nonumber\\
&&- \Delta_a \eta^{\ast}_{a, \downarrow, \bar{b}}= 0 \label{Sch1}\\
&&(E + i a b v_F k - i ab \alpha k + h)\eta_{a, \downarrow, b} -
Be^{i \theta} \eta_{a, \uparrow, b} \nonumber\\
&& + \Delta_a \eta^{\ast}_{a, \uparrow, \bar{b}} = 0 \label{Sch2}\\
&& (E - i a b v_F k - i ab \alpha k - h)\eta^{\ast}_{a, \downarrow,
\bar{b}}+ Be^{-i \theta} \eta^{\ast}_{a, \uparrow, \bar{b}} \nonumber\\
&&- \Delta_a^{\ast} \eta_{a, \uparrow, b} = 0 \label{Sch3}\\
&& (E - i a b v_F k + i a b \alpha k + h)\eta^{\ast}_{a, \uparrow,
\bar{b}} + Be^{i \theta}
\eta^{\ast}_{a, \downarrow, \bar{b}} \nonumber\\
&&+\Delta_a^{\ast} \eta_{a, \downarrow, b}= 0. \label{Sch4}\end{aligned}$$ In order to understand the features of Eqs. (\[Sch1\])-(\[Sch4\]) one considers several asymptotic cases. In the absence of the in-plane magnetic field ${\bf B}=0$, these equations link a particle wave function $\eta_{a, \sigma, b}$ with the hole one $\eta^{\ast}_{a, \bar{\sigma}, \bar{b}}$ of an opposite spin-polarized and opposite direction-moved quasiparticle state [@ksy04] and vice-versa, provided that the system is in a superconducting phase, $\Delta_a \neq 0$. Instead, in the absence of a superconducting phase, $\Delta_a=0$, these equations link a particle (hole) wave function $\eta_{a, \sigma, b}$ with the particle (hole) wave function with opposite spin-polarized quasiparticle state $\eta_{a, \bar{\sigma}, b}$ moving in the same direction provided that ${\bf B} \neq 0$. However, in the presence of in-plane magnetic field ${\bf B} \neq 0$ in the superconduction phase $\Delta_a \neq 0$, Eqs. (\[Sch1\])-(\[Sch4\]) connect a particle wave function $\eta_{a, \sigma, b}$ (a hole wave function $\eta^{\ast}_{a, \sigma, \bar{b}}$) with hole (particle) wave functions with the same $\eta^{\ast}_{a, \sigma, \bar{b}}$ ( $\eta_{a, \sigma, b}$) and opposite $\eta^{\ast}_{a, \bar{\sigma}, \bar{b}}$ ( $\eta_{a, \bar{\sigma}, b}$) spin-polarized quasiparticle states moving in the direction opposite to the particle (hole) one. Eqs. (\[Sch1\])-(\[Sch4\]) allow us to calculate all possible ratios $\eta^{\ast}_{a, \sigma, b}/\eta_{a,{\bar
\sigma},{\bar b}}$, $\eta^{\ast}_{a, \sigma, b}/\eta_{a,\sigma,
{\bar b}}$, and $\eta_{a, \sigma,b}/\eta_{a,{\bar \sigma}, b}$, $\eta^{\ast}_{a, \sigma,b}/\eta^{\ast}_{a,{\bar \sigma}, b}$. Furthermore, we note that only the ratio $\eta^{\ast}_{a, \sigma,
b}/\eta_{a,{\bar \sigma},{\bar b}}$ is non-zero for ${\bf B}=0$, which corresponds to the conventional Andreev reflection at the boundary of a superconductor with normal metal or insulator [@ksy04]. Eqs. (\[Sch1\])-(\[Sch4\]) provide the following expressions for these ratios for arbitrary values of the parameters $\alpha$, ${\bf B}$, $h$, $$\begin{aligned}
\hspace{-10mm}&&\frac{\eta^{\ast}_{a, \uparrow, {\bar b}}}{\eta_{a, \downarrow, b}}= -\frac{1}{\Delta_{a, b}}\Bigg\{E+iabkv_F-iab\alpha k+h-
\nonumber\\
\hspace{-10mm}&&\frac{2B^2(E - iab\alpha k)}{(E-h)^2+(kv_F+\alpha k)^2 + B^2-\Delta_{a, b}^2}\Bigg\}
\label{ratup-down}\\
\hspace{-10mm}&&\frac{\eta^{\ast}_{a, \uparrow, {\bar b}}}{\eta_{a, \uparrow, b}}=\frac{B e^{i \theta}}{\Delta_{a, b}}\Bigg\{1-\nonumber\\
\hspace{-10mm}&&\frac{2(E+iab\alpha k)(E+iabkv_F-iab\alpha k + h)}{(E+h)^2+(kv_F-\alpha k)^2 + B^2-\Delta_{a, b}^2}\Bigg\},
\label{ratup-up}\\
\hspace{-10mm}&&\frac{\eta_{a, \uparrow, b}}{\eta_{a, \downarrow, b}}= \frac{2B (E-i ab\alpha k)e^{-i \theta}}{(E-h)^2+(kv_F+k \alpha)^2 +
B^2 -\Delta_{a, b}^2};
\label{ratup-down2}\end{aligned}$$ where $\Delta_{a, b}=b \Delta_a$ according to Eq. (\[px\]), and $\Delta_R=|\Delta| \exp(i \varphi)$ and $\Delta_L=|\Delta|$. Note that the expressions for $\frac{\eta^{\ast}_{a, \downarrow, {\bar b}}}{\eta_{a, \uparrow, b}}$, $\frac{\eta^{\ast}_{a, \downarrow, {\bar b}}}{\eta_{a, \downarrow, b}}$, and $\frac{\eta_{a, \downarrow, b}}{\eta_{a, \uparrow, b}}$ can be obtained respectively from Eqs. (\[ratup-down\]), (\[ratup-up\]), and (\[ratup-down2\]) by replacing $\theta \to - \theta$, $\alpha \to - \alpha$, $h \to -h$, and by reversing the total sign of these expressions. According to these expressions the reflection channels, determined by the ratios $\frac{\eta^{\ast}_{a, \downarrow, {\bar b}}}{\eta_{a, \downarrow, b}}$, and $\frac{\eta_{a, \downarrow, b}}{\eta_{a, \uparrow, b}}$, vanish with in-plane magnetic field ${\bf B}$.
We note from Eqs. (\[Sch1\])-(\[Sch4\]) that the dependencies of these equations on $\varphi$ and $\theta$ are completely removed by transforming the wave functions as $$\begin{aligned}
\eta^{\ast}_a(x) &\to& \left(e^{-i(\varphi-
\theta)/2}\eta_{a,\uparrow, \bar{b}}^{\ast}(x), ~e^{-i(\varphi
+\theta/2)} \eta_{a,
\downarrow,\bar{b}}^{\ast}(x), \right. \nonumber\\
&& \left.~ e^{i(\varphi+ \theta)/2}\eta_{a, \downarrow, b} (x),
~e^{i(\varphi- \theta)/2} \eta_{a, \uparrow, b}(x)\right).\end{aligned}$$ In the transformed basis one has $$\begin{aligned}
\frac{\eta^{\ast}_{a, \uparrow, b}}{\eta_{a,\uparrow,{\bar b}}} \to
e^{-i(\varphi - \theta)} \frac{\eta^{\ast}_{a, \uparrow,
b}}{\eta_{a,\uparrow, {\bar b}}}, \quad \frac{\eta^{\ast}_{a,
\downarrow, b}}{\eta_{a,\downarrow, {\bar b}}} \to e^{-i(\varphi +
\theta)} \frac{\eta^{\ast}_{a, \downarrow,
b}}{\eta_{a,\downarrow, {\bar b}}} \label{dagup-up} \\
\frac{\eta^{\ast}_{a, \uparrow, b}}{\eta_{a,\downarrow, {\bar b}}}
\to e^{-i\varphi} \frac{\eta^{\ast}_{a, \uparrow,
b}}{\eta_{a,\downarrow, {\bar b}}}, \quad \frac{\eta^{\ast}_{a,
\downarrow, b}}{\eta_{a,\uparrow, {\bar b}}} \to e^{-i\varphi}
\frac{\eta^{\ast}_{a, \downarrow,
b}}{\eta_{a,\uparrow, {\bar b}}} \label{dagup-down} \\
\frac{\eta^{\ast}_{a, \uparrow, b}}{\eta_{a,\downarrow, b}^{\ast}}
\to e^{i\theta} \frac{\eta^{\ast}_{a, \uparrow,
b}}{\eta_{a,\downarrow, b}^{\ast}}, \quad \frac{\eta_{a, \uparrow,
b}}{\eta_{a,\downarrow, b}} \to e^{-i\theta} \frac{\eta_{a, \uparrow,
b}}{\eta_{a,\downarrow, b}}. \label{up-up}\end{aligned}$$ The different ratios that appear in the left hand-side of Eqs.(\[dagup-up\])-(\[up-up\]) can be understood on follows. The ratio $\eta^{\ast}_{a, \sigma, b}/\eta_{a,{\bar \sigma},{\bar b}}$ corresponds to the amplitude of conventional Andreev reflection channel which constitutes reflection of an electron-like quasiparticle to a hole-like quasiparticle with opposite spin on a N-S interface. In contrast, the ratio $\eta^{\ast}_{a, \sigma,
b}/\eta_{a,\sigma, {\bar b}}$ which is finite only in the presence of SOC and/or magnetic field, represents amplitude of Andreev reflection channel where the electron-like quasiparticle incident on the interface is reflected to a hole-like quasiparticle state with the same spin orientation. Finally, the ratio $\eta_{a,
\sigma,b}/\eta_{a,{\bar \sigma},b}$ represents a usual reflection channel of an electron-like quasiparticle on the boundary without creation of a Cooper pair in a superconducting part of the junction. We note that the ratio of wave functions in Eq. (\[dagup-up\]) depend on both $\varphi$ and $\theta$ while those in Eqs. (\[dagup-down\]) and (\[up-up\]) depend on either $\varphi$ or $\theta$. This suggests that the ratios (\[dagup-up\]) and (\[dagup-down\]) are responsible for the dependence of observable parameters on the order parameter phase difference $\varphi$, whereas the ratios (\[dagup-up\]) and (\[up-up\]) are responsible for the dependence on the magnetic field orientation angle $\theta$.
In what follows, we shall look for the localized subgap Andreev bound states with $\epsilon(k)< |\Delta|$ for the Josephson junction of two nanowires described by Eq.(\[H\]).\
**Andreev bound states, Josephson and magneto-Josephson effects** {#sec3}
=================================================================
In order to obtain a solution for the Andreev bound states for the junction described by Eq. (\[H-sc\]) one follows the method used in Ref. \[\]. The energy spectrum of an electron is splitted in the presence of Rashba SOC and/or Zeeman magnetic field, so that the Fermi level crosses the dispersion curve at four points, corresponding to right-mover $k_{F+}$, $k_{F-}$ and left-mover $-k_{F+}$, $-k_{F-}$ particles with oppositely polarized spin-states (see, Fig. \[Fig1\]), and $k_{F+}-k_{F-} \to 0$, $k_{F+}, k_{F-} \to k_F$ as $\alpha, B, h \to 0$. Furthermore, a condensation of the electron pairs in a superconducting state opens a gap around the Fermi level as is shown in Fig.\[Fig1\]b. We neglect here a difference between $k_F$ and $k_{F+}$, $k_{F-}$, and take $k_{F+} \approx k_{F-} \equiv k_F$. We assume that a transition occurs between the states with the same chirality. In order to obtain the wave function for $L$ and $R$ superconductors we superpose the wave functions for the left ($-$) and right ($+$) moving BdG quasiparticles correspondingly around the Fermi levels $k_{F}$, and $-k_{F}$ with arbitrary coefficients,
$$\begin{aligned}
\mathbf{\eta}_{a}(x)= e^{{\rm sgn}(a) k x} \left\{e^{i k_F x} A_a
\left(\begin{array}{ccc}
\eta_{a,\uparrow, +}(k) \\
\eta_{a, \downarrow, -}(k)\\
\eta_{a,\downarrow, -}^{\ast}(k) \\
\eta_{a, \uparrow, +}^{\ast}(k)
\end{array} \right)
+e^{-i k_F x} B_a \left(\begin{array}{ccc}
\eta_{a,\uparrow, -}(k) \\
\eta_{a, \downarrow,+}(k) \\
\eta_{a,\downarrow, +}^{\ast}(k)\\
\eta_{a, \uparrow, -}^{\ast}(k)
\end{array} \right)\right\}
%\end{displaymath}
\label{wave}\end{aligned}$$
where ${\rm sgn} (a)= + (-)$ for $a=L (R)$. SOC and magnetic field remove the spin degeneracy in a quasi-particle ($\eta_{a,b}$) and a quasi-hole ($\eta^{\ast}_{a, b}$) wave functions, and thereby split the wave functions written for the conventional superconductors [@ksy04] as is shown in the above given expression.
Andreev bound state energies are obtained by imposing the usual boundary conditions on each component of these wave functions $\eta_a(x)$. For a barrier modeled by the delta function potential $U(x)= U_0 \delta (x)$, the boundary conditions are provided at the merging point $x=0$ of two superconductors as, $$\begin{aligned}
\mathbf{\eta}_L(0)= \mathbf{\eta}_R(0),\qquad \qquad
\partial_x \mathbf{\eta}_R -
\partial_x \mathbf{\eta}_L= k_F Z \mathbf{\eta} (0),
\label{bc}\end{aligned}$$ where $Z=2mU_0/\hbar^2k_F$ and the transmission coefficient $D$ is expressed through $Z$ as $D=4/(Z^2 +4)$.
By choosing a pair of the wave functions from Eq. (\[wave\]) and substituting they into the boundary conditions (\[bc\]) one gets four linear homogeneous equations. The energy of the Andreev bound state for a transmission of the barrier through a particular channel is obtained from the determinant of these linear homogeneous equations. Selection, e.g. the second and fourth equations of the wave function (\[wave\]) under the boundary conditions (\[bc\]) yields the following expression for the determinant, $${\tilde F}_{\uparrow \downarrow}^{\ast}=
\frac{1}{D}\eta_{+,\downarrow,+}
\eta_{-,\downarrow,-}\eta_{+,\downarrow,-}\eta_{-,\downarrow,+}
F^{\ast}_{\uparrow \downarrow}, \label{Fup-down0}$$ where $$\begin{aligned}
F_{\uparrow \downarrow}^{\ast}=\left[\frac{\eta^{\ast}_{-,
\uparrow,-}}{\eta_{-,\downarrow,+}} - \frac{\eta^{\ast}_{+,
\uparrow,-}}{\eta_{+,\downarrow,+}}\right]
\left[\frac{\eta^{\ast}_{+, \uparrow,+}}{\eta_{+,\downarrow,-}}
-\frac{\eta^{\ast}_{-, \uparrow,+}}{\eta_{-,\downarrow,-}}\right]-\nonumber\\
(1-D) \left[\frac{\eta^{\ast}_{+,
\uparrow,-}}{\eta_{+,\downarrow,+}} -\frac{\eta^{\ast}_{-,
\uparrow,+}}{\eta_{-,\downarrow,-}}\right]
\left[\frac{\eta^{\ast}_{-, \uparrow,-}}{\eta_{-,\downarrow,+}}
-\frac{\eta^{\ast}_{+, \uparrow,+}}{\eta_{+,\downarrow,-}}\right].
\label{Fup-down1}\end{aligned}$$ Using Eq. (\[ratup-down\]) in this expression one gets an explicit expression for $F_{\uparrow \downarrow}^{\ast}$ $$\begin{aligned}
&&\hspace{-5mm}F_{\uparrow \downarrow}^{\ast}=-\frac{4e^{-i\varphi}}{|\Delta|^2M^2_-}\Bigg\{ \left[(E+h) M_-+2B^2 E \right]^2 -
D \cos^2\frac{\varphi}{2} \times \nonumber\\
&& \hspace{-5mm}\left[ \left[(E+h)M_--2B^2E\right]^2 + \left[k(v_F-\alpha) M_-
+2 B^2 \alpha k\right]^2 \right] \Bigg\},
\label{Fup-down}\end{aligned}$$ where $$M_{\pm}=(E\pm h)^2 + (v_F \mp \alpha)^2 k^2 +B^2 -|\Delta|^2.
\label{M}$$ The expression for $F_{\downarrow \uparrow}^{\ast}$ is obtained from Eq. (\[Fup-down\]) by replacing $\alpha \to - \alpha$, $h \to - h$ and $M_- \to M_+$, where $M_+(\alpha, h, B, \Delta)=M_-(-\alpha, -h, B, \Delta)$. Solution of the equation $F_{\uparrow \downarrow}^{\ast}(E_{\uparrow \downarrow})=0$ for energy, where $F_{\uparrow \downarrow}^{\ast}$ is given by Eq. (\[Fup-down\]), yields a contribution to the Andreev overlap energy $E_{\uparrow \downarrow}$ in the particle-hole channel.
Now we choose other pair, the first and fourth wave functions of (\[wave\]), and substitute they into the boundary conditions (\[bc\]). The determinant of four linear homogeneous equations yields the following expression to find the Andreev quasi-particle energy in the anomalous particle-hole channel, where the transition occurs between the spin states with the same chirality, $${\tilde F}_{\sigma \sigma}^{\ast}(k) = \frac{1}{D}\eta_{+,\sigma,+}(k) \eta_{-,\sigma,-}(k) \eta_{+,\sigma,-}(k)
\eta_{-,\sigma,+}(k) F^{\ast}_{\sigma \sigma}(k),
\label{feqss}$$ where $F^{\ast}_{\uparrow \uparrow}(k)$ ($F^{\ast}_{\downarrow \downarrow}(k)$) is obtained from Eq. (\[Fup-down1\]) by replacing all spin-down (all spin-up) with spin-up (spin-down). The evident expression for $F^{\ast}_{\uparrow \uparrow}(k)$ is obtained by using the ratio (\[ratup-up\]), which reads as, $$\begin{aligned}
&&F_{\uparrow \uparrow}^{\ast}=\frac{16 B^2 e^{-i(\phi- \theta)}}{|\Delta|^2M^2_+}\Bigg\{ \left(E k v_F+\alpha h k\right)^2 -\nonumber\\
&&|\Delta|^2 \left(E^2+ \alpha^2 k^2 \right)\left[1-D\cos^2\frac{(\phi- \theta)}{2}\right]\Bigg\}.
\label{Fup-up}\end{aligned}$$ The expression for $F^{\ast}_{\downarrow \downarrow}(k)$ can be obtained from Eq. (\[Fup-up\]) by replacing $\alpha \to - \alpha$, $h \to - h$, $\theta \to - \theta$, and $M_+ \to M_-$. The general feature of the Andreev quasi-particle energy in the anomalous particle-hole channel $E_{\uparrow \uparrow}$ with the same spin orientation is that it takes non-zero values only in the presence of in-plane magnetic field ${\bf B}$. Therefore, it depends on the angle $\theta$ between the junction and in-plane magnetic field. Oscillation of the Josephson current with $\theta$ yields a fractional magneto-Josephson effect.
Choice of the first and second equations of the wave function (\[wave\]) under the boundary conditions (\[bc\]) yields the following expression to determine the Andreev bound state energy in the anomalous particle-particle or hole-hole channel, $${\tilde F}_{\uparrow \downarrow}(k) =
\frac{1}{D}\eta_{-,\downarrow,-}(k) \eta_{+,\downarrow,+}(k) \eta_{+,\downarrow,-}(k) \eta_{-,\downarrow,+}(k)
F_{\uparrow \downarrow}(k),
\label{FA-up-down}$$ where $F_{\uparrow \downarrow}(k)$ is written as $$\begin{aligned}
F_{\uparrow \downarrow}=\left[\frac{\eta_{+,
\uparrow,+}}{\eta_{+,\downarrow,+}} - \frac{\eta_{-,
\uparrow,+}}{\eta_{-,\downarrow,+}}\right]
\left[\frac{\eta_{-, \uparrow,-}}{\eta_{-,\downarrow,-}}
-\frac{\eta_{+, \uparrow,-}}{\eta_{+,\downarrow,-}}\right]-\nonumber \\
(1-D) \left[\frac{\eta_{+,
\uparrow,+}}{\eta_{+,\downarrow,+}} -\frac{\eta_{-,
\uparrow,-}}{\eta_{-,\downarrow,-}}\right]
\left[\frac{\eta_{-, \uparrow,+}}{\eta_{-,\downarrow,+}}
-\frac{\eta_{+, \uparrow,-}}{\eta_{+,\downarrow,-}}\right].
\label{energyA-up-down}\end{aligned}$$ The evident expression for $F_{\uparrow \downarrow}$ can be obtained by substituting Eq. (\[ratup-down2\]) into Eq. (\[energyA-up-down\]), which yields, $$\hspace{-2mm} F_{\uparrow \downarrow}=-\frac{16 B^2 e^{-i\theta}}{M^2_-}\Big\{\alpha^2 k^2-
\left(E^2+ \alpha^2 k^2\right)D\sin^2\frac{\theta}{2}\Big\}.
\label{FAup-down}$$ Note that the expression for $F_{\downarrow \uparrow}(k)$ can be obtained from Eq. (\[energyA-up-down\]) by replacing $\alpha \to - \alpha$, $h \to - h$, $\theta \to - \theta$, and $M_- \to M_+$. The main feature of the Andreev bound state energy in the anomalous particle-particle channel is that it survives only in the presence of the in-plane magnetic field ${\bf B}$ and the spin-orbit interaction $\alpha$. $E'_{\uparrow \downarrow}$ and $E'_{\downarrow \uparrow}$ vanish in the absence of one of the factors either ${\bf B}$ or $\alpha$, and they depend on the angle $\theta$ between the in-plane magnetic field orientation and the junction, contributing to the fractional magneto-Josephson effect.
Andreev bound state energies and Josephson current, corresponding to different tunneling channels, demonstrate completely different oscillation. The conditions $F_{\uparrow \downarrow}^{\ast}(E_{\uparrow \downarrow})=0$ and $F_{\downarrow \uparrow}^{\ast}(E_{\downarrow \uparrow})=0$ with Eq. (\[Fup-down\]) for $F_{\uparrow \downarrow}^{\ast}(E)$ provide contributions to the Andreev bound state energy in the particle-hole channel, which oscillates fractionally with the order parameters’ phases difference $\varphi$. Additional contributions to the energy come from the conditions $F_{\uparrow \uparrow}^{\ast} (E_{\uparrow \uparrow})=0$ and $F_{\downarrow \downarrow}^{\ast} (E_{\downarrow \downarrow})=0$ with $F_{\uparrow \uparrow}^{\ast}$ given by Eq. (\[Fup-up\]), which arise only in the presence of an in-plane magnetic field ${\bf B}$ and oscillate not only with $\varphi$ but also with $\theta$. Contribution to the magneto-Josephson effect gives apart from the anomalous particle-hole channel also the anomalous particle-particle channel under the conditions $F_{\uparrow \downarrow}(E'_{\uparrow \downarrow})=0$ and $F_{\downarrow \uparrow}(E'_{\downarrow \uparrow})=0$, where the evident expression for $F_{\uparrow \downarrow}$ is given by Eq. (\[FAup-down\]). Furthermore, the contribution coming from the anomalous particle-particle channel vanishes not only at ${\bf B}=0$ but also in the absence of the spin-orbit interaction, $\alpha =0$.
The total Andreev bound state energy $E$ is obtained by finding overlap energies for each channel from the equations $F_{\uparrow \downarrow}^{\ast}(E_{\uparrow \downarrow})=0$, $F_{\downarrow \uparrow}^{\ast}(E_{\downarrow \uparrow})=0$, $F_{\uparrow \uparrow}^{\ast}(E_{\uparrow \uparrow})=0$, $F_{\downarrow \downarrow}^{\ast}(E_{\downarrow \downarrow})=0$, and $F_{\uparrow \downarrow}(E'_{\uparrow \downarrow})=0$, $F_{\downarrow \uparrow}(E'_{\downarrow \uparrow})=0$, and summing up of all the coupling energies $E(\varphi, \theta)=E_{\uparrow \downarrow}(\varphi)+E_{\downarrow \uparrow}(\varphi)+
E_{\uparrow \uparrow}(\varphi, \theta) + E_{\downarrow \downarrow}(\varphi, \theta) + E'_{\uparrow \downarrow}(\theta) +
E'_{\downarrow \uparrow}(\theta)$ in each reflection channel. Below we calculate the Andreev bound state energies for several asymptotic cases.
Andreev bound state energy in the absence of in-plane magnetic field, ${\bf B}=0$.
----------------------------------------------------------------------------------
Contribution to the Andreev bound state energy in the absence of [*in-plane*]{} magnetic field ${\bf B}=0$ comes only from the particle-hole scattering channel, determined by scattering amplitude Eq. (\[ratup-down\]), and all other channels vanish under this condition. The evident expression for the bound state energy in the particle-hole channel is obtained from the equation $F_{\uparrow \downarrow}^{\ast}(E_{\uparrow \downarrow})=0$, where $F_{\uparrow \downarrow}^{\ast}$ is given by Eq. (\[Fup-down\]). The general expression for $E_{\uparrow \downarrow}$ when all the external parameters take non-zero values, $\alpha \neq 0$, ${\bf B} \neq 0$ and $h \neq 0$, can be obtained from the expression (\[AEup-down\]) in Appendix. By putting ${\bf B}=0$ in this equation and replacing $k^2$ according to Eq. (\[k2\]) in Appendix one gets the following equation after routine calculations,
$$\begin{aligned}
&&\left[(E_{\uparrow \downarrow}+h)^2\left(1-D \cos^2\frac{\varphi}{2}\right)-\left(\frac{v_F-\alpha}{v_F+\alpha}\right) (E^2_{\uparrow \downarrow}
-h^2 -\Delta^2)D \cos^2 \frac{\varphi}{2}\right]^2 +\nonumber\\
&&\frac{4D\cos^2\frac{\varphi}{2}}{(v_F +\alpha)^2} \left\{ (E_{\uparrow \downarrow} +h)^2 \left(1-D \cos^2\frac{\varphi}{2}\right)
\left[(E_{\uparrow \downarrow}v_F+\alpha h)^2 - \Delta^2 v_F^2\right] -
(v_F -\alpha)^2 h^2 \Delta^2 D \cos^2\frac{\varphi}{2}\right\}=0.
\label{E-B=0}\end{aligned}$$
This equation is fourth order in $E_{\uparrow \downarrow}$ equation, and it can be in principle solved analytically.
Equation (\[E-B=0\]) yields exact analytical solutions for $E_{\uparrow \downarrow}$ in several asymptotic cases.
This equation is further simplified for , yielding $$E_{\pm} \equiv E_{\uparrow \downarrow}= \pm E_0 \equiv \pm |\Delta| \sqrt{D} \cos \frac{\varphi}{2},
\label{ksy}$$ which reproduces the well-known result [@ksy04] for the Andreev bound state energy of $JJ$ with $p$-wave superconductors in the absence of magnetic field and spin-orbit interactions. This expression provides the energy spectrum of quasi-electron and quasi-hole excitations, symmetrically located around the Fermi level in the gap.
, Rashba spin-orbit interaction splits both quasi-electron and quasi-hole spectra, and Eq. (\[E-B=0\]) yields four solutions for the bound state energy, $$E=E_{\uparrow \downarrow}= s\frac{\sqrt{D} |\Delta| \frac{(v_F \pm \alpha)}{(v_F+\alpha)} \cos \frac{\varphi}{2}}
{\sqrt{1- D \cos^2\frac{\varphi}{2}\frac{(v_F+\alpha)^2 - (v_F \pm \alpha)^2}{(v_F +\alpha)^2}}},
\label{E-B=01}$$ where $s = \pm$ assigns the electron and hole branches of the spectrum. Two solutions of this expression coincides with Eq. (\[ksy\]), and do not depend on the strength of Rashba spin-orbit interaction. Nevertheless other two solutions depend on $\alpha$. The expression for Andreev’s bound state energy $E_{\downarrow \uparrow}$, as mentioned above, is obtained by replacement of $\alpha \to -\alpha$ in the expression (\[E-B=01\]) written for $E_{\uparrow \downarrow}$. Figs.\[B0-alpha\]a, b and Figs.\[B0-alpha\]c, d depict the dependence of $E= E_{\uparrow \downarrow}$ and $E_{\downarrow \uparrow}$ respectively on $\varphi$ for two different values of $\alpha$ when $\alpha=0.2$ and $\alpha=0.5$. According to the figures, two branches of Andreev’s bound state energy, drawn by blue and dashed curves in Figs.\[B0-alpha\] do not depend on $\alpha$. Nevertheless, other two branches of $E_{\uparrow \downarrow}$ (of $E_{\downarrow \uparrow}$) decrease (increase) with increasing the strength of Rashba SOC. Note here that the parameters in all figures are given in a dimensionless form as $E \to E/\Delta$, $k \to (kv_F)/\Delta$, $h \to h/\Delta$, ${\bf B} \to {\bf B}/\Delta$, and $\alpha \to \alpha/v_F$. In this limiting case, the quasi-electron and quasi-hole spectra are again symmetrically located around the Fermi level.
Equation (\[E-B=0\]) yields the following expression for the Andreev bound state energy in this limit, $$\begin{aligned}
E_{\uparrow \downarrow}^+ &=& -h \left[1-2 D
\cos^2\frac{\varphi}{2}\right] \nonumber\\
&& + s \sqrt{D} \cos\frac{\varphi}{2}
\sqrt{\Delta^2-4h^2 + 4D h^2 \cos^2 \frac{\varphi}{2}} \nonumber\\
E_{\uparrow \downarrow}^- &=& -h + s \sqrt{D} \sqrt{\Delta^2-4h^2}
\cos\frac{\varphi}{2} \label{E-B=02}\end{aligned}$$ Andreev bound states are split again due to Zeeman effect. The dependence of $E_{\uparrow \downarrow}^{\pm} \equiv E_{\uparrow
\downarrow}$ on $\varphi$ is depicted in Figs. \[B0-h\] a, b for two different values of $h$. Note that contribution to Andreev bound state energy $E_{\downarrow \uparrow}$, found from the condition of $F^{\ast}_{\downarrow \uparrow}=0$, can be obtained again by replacing $h \to -h$, $\alpha \to - \alpha$, and $\theta \to -
\theta$ in Eqs. (\[ksy\]), (\[E-B=01\]), and (\[E-B=02\]). The dependence of $E_{\downarrow \uparrow}$ on $\varphi$ for different values of $h$ is depicted in Figs. \[B0-h\] c and d for completeness. Two branches of solution (\[E-B=02\]) differ from those given by Eq. (\[ksy\]) by shifting only the particle and hole pairs $E_{\uparrow \downarrow}$ ($E_{\downarrow \uparrow}$) to the value of $-h$ ( $+h$), without changing their oscillation characteristics (see, Figs. \[B0-h\] a, b and c, d). As it is seen clearly from Figs. \[B0-h\] a, b (Figs. \[B0-h\] c, d) the magnetic field reduces considerably the amplitude of the fractional oscillation for other two solutions of $E_{\uparrow \downarrow}$ ($E_{\downarrow \uparrow}$), at the same time shifts down (up) asymmetrically the quasi-particle and quasi-hole spectra. One of the quasi-hole (quasi-particle) branch of $E_{\uparrow \downarrow}$ ($E_{\downarrow \uparrow}$) is pushed off from the gap at higher magnetic field when $h>h_c=0.46 \Delta$.
is calculated numerically according to Eq. (\[E-B=0\]) writing this equation in the dimensionless parameters such as $\tilde{E}= E/\Delta$, $\tilde{k}=(kv_F)/\Delta$, $\tilde{h}=h/\Delta$, $\tilde{{\bf B}}={\bf B}/\Delta$, and $\tilde{\alpha}=\alpha/v_F$. Fig. \[B0-alpha-h\] shows the dependence of $E_{\uparrow \downarrow}$ on $\varphi$ for $D=0.3$ and $\tilde{\alpha}=0.4$ with different values of $\tilde{h}$, $\tilde{h}=0.3; 0.5001; 0.506$; and $0.5185$ (the parameters in all figures are given without tilde). One of the quasi-electron and quasi-hole pair of the spectrum, depicted by solid (blue and red) lines in Fig. \[B0-alpha-h\], shifts down with increasing the magnetic field $h$ without changing the form and amplitude. The amplitude of the other quasi-electron and quasi-hole branch of $E_{\uparrow \downarrow}$ (drawn by dashed blue and red curves) decreases, and the form of the curves is deformed with increasing the magnetic field $h$. At $\tilde{h} > \tilde{h}_g=0.5001$ a forbidden gap appears in the spectrum, i.e. as it is seen in Fig. \[B0-alpha-h\]c the quasi-electron and quasi-hole states disappear for some values of the order parameter phase difference $\varphi$. The quasi-particle and quasi-hole states, shown by dashed (blue and red) curves in Fig. \[B0-alpha-h\] vanishes by further increasing of the magnetic field at $\tilde{h} > \tilde{h}_c=0.51921$.
For completeness, $E_{\downarrow \uparrow}$ vs. $\varphi$ dependence is calculated also for $h=0.3; 0.5001; 0.506$; and $0.5185$ under the condition of ${\bf B}=0$, $\alpha=0.4$ and $D=0.3$, which is depicted in Fig. \[B0-h-alpha\]. As it is expected, $E_{\downarrow \uparrow}$ behaves like $E_{\uparrow \downarrow}$, i.e. the magnetic field shifts up one of the quasi-partice and quasi-hole pair, drawn by solid blue and red curves in Fig. \[B0-h-alpha\] without changing the amplitude and form. The other pair, presented by dashed blue and red curves Fig. \[B0-h-alpha\] deforms and amplitude decreases with increasing the magnetic field $h$. For $h>0.5001$ a forbidden gap is opened (see, Fig. \[B0-h-alpha\] b) in the spectrum. This branch (dashed curves in Fig. \[B0-h-alpha\] c, d) squeezes and disappears for $h>h_c=0.91124$.
Andreev bound state energy in the presence of [***in-plane***]{} magnetic field ${\bf B} \neq 0$
------------------------------------------------------------------------------------------------
In the presence of the in-plane magnetic field ${\bf B}$ all three channels described by Eqs. (\[ratup-down\])- (\[up-up\]) give contributions to Andreev bound state energy $E$. The expressions for general dependencies of $E_{\uparrow \downarrow}$, $E_{\uparrow \uparrow}$ and $E'_{\uparrow \downarrow}$ on $\alpha,~ {\bf B}, ~ h$ can be obtained from the Equations (\[AEup-down\]), (\[AEup-up2\]) and (\[FAup-down1\]) presented in Appendix after replacement of $k^2$ by $E$ according to Eq. (\[k2\]). These equations can be solved analytically for energy in several asymptotic cases. Note that main contribution to Andreev bound state energy still gives the conventional particle-hole channel.
In this limiting case an interference between SOC-induced effective magnetic field and $h$ vanishes, and hence the energy spectrum depends on the modulus of total magnetic field according to Eq. (\[E4\]) as $H= \sqrt{{\bf B}^2+ h^2}$. The expression (\[AEup-down\]) for $E_{\uparrow \downarrow}$ is strongly simplified in this limiting case, and substitution of $k^2$ from Eq. (\[E4\]) into this expression yields, $$E_{\uparrow \downarrow}^{\pm}= \pm \sqrt{B^2+h^2} +s \Delta \sqrt{D}
\cos \frac{\varphi}{2}, \label{Eup-down1}$$ where $s = \pm$.
The Andreev bound state energy in the anomalous particle-hole channel $E_{\uparrow \uparrow}$ can be found in this limiting case from the general expression given by Eq. (\[AEup-up3\]) yielding, $$E_{\uparrow \uparrow}^{\pm}= \pm \sqrt{B^2+h^2} +s \Delta \sqrt{D}
\cos \frac{\varphi-\theta}{2}. \label{Eup-up1}$$ This expression differs from that given by (\[Eup-down1\]) for $E_{\uparrow \downarrow}$ by dependence of cosine function not only on $\varphi$ but also on $\theta$.
Contribution to Andreev bound state energy from the third anomalous particle-particle channel vanishes, as is seen from Eq. (\[FAup-down2\]), in this limiting case. So, one can state that the third channel survives and gives a contribution to the bound state energy only in the presence of SOC ($\alpha \neq 0$) and in-plane magnetic field (${\bf B} \neq 0$) in the system. An absence at least one of these factors destroys this channel.
The total Andreev bound state energy in this limiting case contains (\[Eup-down1\]) and (\[Eup-up1\]), and also the energies $E_{\downarrow \uparrow}$ and $E_{\downarrow \downarrow}$, obtained from (\[Eup-down1\]) and (\[Eup-up1\]) by replacements $\alpha \to - \alpha$, $h \to -h$, $\theta \to - \theta$, $$\begin{aligned}
&&E= \pm 4 \sqrt{B^2+h^2} + s \Delta \sqrt{D} \Bigg\{2 \cos \frac{\varphi}{2} +\nonumber\\
&& \cos \frac{\varphi-\theta}{2} + \cos \frac{\varphi+\theta}{2} \Bigg\}.
\label{E1}\end{aligned}$$
. The expression $F_{\uparrow \downarrow}^{\ast}=0$ can be simplified for $h=0$ and $\alpha \neq 0$, ${\bf B} \neq 0$. Routine calculations yield the following expression to determine $E_{\uparrow \downarrow}$, $$\begin{aligned}
&& E^2_{\uparrow \downarrow} \Big\{k^2\left[v_F \alpha (B^2 +\Delta^2 -E^2_{\uparrow \downarrow}) -v_F^2B^2\right]-\nonumber\\
&&\frac{v_F \alpha}{(v_F +\alpha)^2}(E^2_{\uparrow \downarrow}-B^2-\Delta^2)^2+\Delta^2 B^2\Big\}-\nonumber\\
&& D \Delta^2 \cos^2 \frac{\varphi}{2}\Big\{k^2\alpha \left[v_F(\Delta^2-E^2_{\uparrow \downarrow})-B^2 (v_F- \alpha)\right]-\nonumber\\
&& \frac{v_F \alpha}{(v_F+\alpha)^2}\left[(E^2_{\uparrow \downarrow}+\Delta^2-B^2)^2-4E^2_{\uparrow \downarrow}\Delta^2\right]+\nonumber\\
&& \left(\frac{v_F-\alpha}{v_F+\alpha} \right)^2E^2_{\uparrow \downarrow}B^2\Big\}=0
\label{h=0}\end{aligned}$$ One can put the expression for $k^2_{\pm}$ from (\[k2\]) and solve numerically this equation for $E_{\uparrow \downarrow}$. Fig. \[h0-alpha-B\] shows the dependence of $E_{\uparrow \downarrow}$ on different values of the in-plane magnetic field $B$ for particular value of $\alpha=0.3$ and $h=0$. One quasi-particle and quasi-hole pair in the spectrum, depicted in blue (dashed lines) is enlarged and is partially pushed off from the gap with increasing the in-plane magnetic field $B$ at $B>B_c \approx 0.55$. On the other hand the pair, depicted in red in Fig. \[h0-alpha-B\], is narrowed with increasing $B$, and the gap is opened in the spectrum at $B>B_c \approx 0.95$. Further increase in $B$ makes this branch of the spectrum again regular at $B>1.063$.
The dependence of $E_{\uparrow \uparrow}$ on $\varphi$ in the anomalous particle-hole channel for non-zero values of the external parameters ${\bf B}$ and $\alpha$ but for $h=0$ is depicted in Fig. \[Up-alpha-B\] for $\alpha=0.2$, $h=0$, $D=0.3$ (a) $B=0.2$ and (b) $B=0.5$. The quasi-electron (quasi-hole) dispersion at $h=0$ is shifted to higher (lower) values with increasing ${\bf B}$ and/or $\alpha$ without changing the shape and symmetry of the energy spectrum.
The third anomalous particle-particle channel gives a contribution $E'$ to the Andreev bound state energy, hence to fractional magneto-Josephson effect provided that both parameters ${\bf B}$ and $\alpha$ take non-zero values. Contribution to $E'$ now is calculated according to Eqs. (\[FAup-down1\]) and (\[k2\]). The numerical calculations for $E'$ dependence on $\theta$ for the case of $h=0$ when $B$ takes different values is presented in Fig. \[F3-h0\]. According to Fig. \[F3-h0\] one of the quasi-particle and quasi-hole branch drawn by solid blue and red curves enlarges with in-plane magnetic field ${\bf B}$, nevertheless the (particle-hole) symmetry is preserved for all curves. The other branch of the spectrum drawn by dashed red and blue curves squees and disappears (see, Fig. \[F3-h0\]c) when $B$ approachs unity. For higher values of $B$ the forbidden gap (shown in Fig. \[F3-h0\]d) appears in the spectrum.
out-of-plane magnetic field $h$ destroys a particle-hole symmetry in the spectrum. Dependence of $E_{\uparrow \downarrow}$ and $E_{\downarrow \uparrow}$ on $\varphi$ in the particle-hole channel is depicted in Figs. \[Fup-downBah\] and \[Fdown-upBah\] for finite $B=0.4$ and different values of h, $h=0.2$, $h=0.4$, $h=0.6$, and $h=0.8$. The magnetic field seemly does not change the amplitude and structure one of the quasi-particle and quasi-hole energy pair, drawn by blue and dashed curves in Figs. \[Fup-downBah\] and \[Fdown-upBah\]. These bound state energies are shifted along the energy axis only. Instead, the magnetic fields strongly change other quasi-particle and quasi-hole pair, presented by red and solid curves in Figs. \[Fup-downBah\] and \[Fdown-upBah\]. This pair of the bound state energy is reduced in amplitude with increasing the magnetic field. At $h>h_c=0.623$ a forbidden gap appears in the spectrum, and $h_c$ increases with increasing $B$.
Numerical calculation of the Andreev bound state energy $E_{\uparrow \uparrow}$ in the anomalous particle-hole channel is shown in Fig.\[FUpUpBah\]. The dashed (red and blue) curves, corresponding to spin-up branches of the bound energy, move away each other with increasing the magnetic fields. Instead the solid (blue and red) curves, corresponding to the spin-down branch’s of the spectrum, is slightly narrowed with increasing the magnetic field. Andreev bound state energy $E'_{\uparrow \downarrow}$ in anomalous particle-particle channel is calculated for non-zero values of $\alpha$ and magnetic fields ${\bf B}$ and $h$, the result of which is presented in Fig. \[F3-Bah\]. The solid (blue and red) curves, corresponding to spin-up branch of the spectrum in the figure inclreases in amplitude with increasing the magnetic fields up to values $\sqrt{h^2+B^2}=1$. Instead the dashed (blue and red) curves, corresponding to spin-down branches of the spectrum, are narrowed and disappear at $\sqrt{h^2+B^2}=1$. A gap is opened in the spectrum with further increase in the magnetic fields.
Equilibrium Josephson current and spin current {#sec4}
==============================================
Josephson current carried by a quasi-particle state $a$ at zero temperature is $$J_a=\frac{2e}{\hbar}\frac{\partial E_a}{\partial \varphi}
\label{current1}$$ The current flowing thought the quasi-particle and quasi-hole states $J_{\pm}$ in the simplest case of ${\bf B} = h =0$ but $\alpha \neq 0$ can be obtained from the tunneling energy given by Eq. (\[E-B=01\]). For $J_+$, when the Andreev bound state energy becomes $$E_{\uparrow \downarrow}^+=s\sqrt{D}\Delta \cos \frac{\varphi}{2}
\label{Eksy1}$$ with $s=\pm$ assigning the quasi-particle and quasi-hole pair, one gets, $$J_{s}^+=-s \frac{e \Delta \sqrt{D}}{\hbar} \sin \frac{\varphi}{2}.
\label{ksy1}$$ For other particle-hole pair of the bound state energy Eq. (\[E-B=01\]) $$E_{\uparrow \downarrow}^-=s\sqrt{D} \Delta \frac{\frac{v_F-\alpha}{v_F+\alpha}}
{\sqrt{1-D\frac{4v_F \alpha}{(v_F+\alpha)^2}\cos^2\frac{\varphi}{2}}}\cos \frac{\varphi}{2}
\label{Eksy2}$$ Josephson energy $J_s^-$ reads as $$J_s^-=-s\frac{e\sqrt{D}\Delta}{\hbar}\frac{\frac{v_F-\alpha}{v_F +\alpha}}{\left[1-D\frac{4v_F\alpha}{(v_F+\alpha)^2}
\cos^2 \frac{\varphi}{2}\right]^{3/2}}\sin \frac{\varphi}{2}.
\label{ksy2}$$
In thermodynamic equilibrium at temperature $T$ the total contribution of the Andreev bound states to the Josephson current can be calculated according to the expression $$J=\frac{2e}{\hbar}\sum_{n=\pm} \frac{\partial E_{\uparrow \downarrow}^n}{\partial \varphi}~f(E_{\uparrow \downarrow}^n)=
- \frac{2e}{\hbar}\sum_{n= \pm}\frac{\partial E_{\uparrow \downarrow}^n}{\partial \varphi}\tanh\left(\frac{E_{\uparrow \downarrow}^n}{2k_BT}\right),
\label{current}$$ where the expression of $E_{\uparrow \downarrow}^{\pm}$ are presented by Eqs. (\[Eksy1\]) and (\[Eksy2\]). Taking into accout the expressions for the energies one gets for Josephson current in the simplest case when $B=h=0$ and $\alpha \neq 0$ $$\begin{aligned}
\hspace{-5mm}&&J=\frac{e\sqrt{D} \Delta}{\hbar}\sin \frac{\varphi}{2} \Bigg\{\tanh \frac{\sqrt{D}\Delta \cos \frac{\varphi}{2}}{2k_BT} +\nonumber\\
\hspace{-7mm}&&\frac{\frac{v_F-\alpha}{v_F+\alpha}}{\left[1-\frac{4v_F \alpha D \cos^2\frac{\varphi}{2}}{(v_F+\alpha)^2} \right]^{3/2}}
\tanh \frac{\sqrt{D} \Delta \frac{v_F -\alpha}{v_F+\alpha} \cos \frac{\varphi}{2}}
{2k_BT \sqrt{1-\frac{4v_F\alpha D \cos^2\frac{\varphi}{2}}{(v_F+\alpha)^2}}}\Bigg\}.\end{aligned}$$ Josephson current in this case will depend on the Rashba SOC coefficient $\alpha$, which can be experimentally determined.
Josephson current in the case of $B=\alpha=0$ and $h \neq 0$ can be calculated by using the expressions \[E-B=02\] for Andreev bound state energies $E^{\pm}_{\uparrow \downarrow}$. In this case the magnetic field $h$ makes asymmetric the bound energy. For simplicity we calculate here the total Josephson current $$I(h, T)=I_{\uparrow \downarrow}+I_{\downarrow \uparrow}$$ which correspon to bound state energies $E^{\pm}_{\uparrow \downarrow}$ and $E^{\pm}_{\downarrow \uparrow}$ respectively. The routine calculations yield
$$\begin{aligned}
I(h, T)=\frac{2e}{\hbar}\left[hD\sin \varphi +\frac{\sqrt{D}}{2}\sin \frac{\varphi}{2}\frac{{\mathbb D}^2 +
4h^2D\cos^2 \frac{\varphi}{2}}{{\mathbb D}} \right] \tanh \frac{-h\left(1-2D\cos^2\frac{\varphi}{2}\right)+
\sqrt{D} {\mathbb D}\cos \frac{\varphi}{2}}{2k_BT}-\nonumber\\
\frac{2e}{\hbar}\left[hD\sin \varphi - \frac{\sqrt{D}}{2}\sin \frac{\varphi}{2} \frac{{\mathbb D}^2 +
4h^2D\cos^2 \frac{\varphi}{2}}{{\mathbb D}} \right] \tanh \frac{h\left(1-2D\cos^2\frac{\varphi}{2}\right)+
\sqrt{D} {\mathbb D}\cos \frac{\varphi}{2}}{2k_BT}+\nonumber\\
\frac{e}{\hbar}\Delta \sqrt{D} \sin \frac{\varphi}{2}\left[\tanh \frac{-h +\Delta \sqrt{D} \cos \frac{\varphi}{2}}{2k_BT} +
\tanh \frac{h +\Delta \sqrt{D} \cos \frac{\varphi}{2}}{2k_BT} \right],
\label{current-h}\end{aligned}$$
where ${\mathbb D}=\sqrt{\Delta^2 -4h^2\left(1-D\cos^2 \frac{\varphi}{2}\right)}$. In two limiting case this expression is simplified. At $D \to 0$ and $h>k_BT$ Eq. (\[current-h\]) yields $$I(h, T) =-\frac{4e}{\hbar} h D \sin \varphi.$$ In the opposite case, when $D \to 0$ and $h< k_BT$ one gets, $$I=\frac{e}{\hbar}D \frac{\Delta^2 -2h^2}{k_BT} \sin \varphi.$$
The Josephson current is given as a partial derivative of the system’s energy with respect to the superconducting phase $\varphi$ as $J_Q=\frac{2e}{\hbar}\frac{\partial \langle H \rangle}{\partial \varphi}$, where $H$ is the system’s Hamiltonian. In the case of topological insulator edges, the spin currents arise as the exact duals of the Josephson current, $J_S =\frac{\partial \langle H \rangle}{\partial \theta}$. We define $\theta$ as the angle between the wire and the Zeeman field, which is also exact dual to the superconducting phase $\varphi$. Thereby, spin Josephson currents $j_S$ are equivalent to torques [@bb09] (driven partly by the Majoranas) that the wire domains apply on the external magnets. Our calculations allow us to find the spin current. Indeed, Andreev bound state energies in anomalous particle-hole channel $E_{\uparrow \uparrow}(\varphi, \theta)$, $E_{\downarrow \downarrow}(\varphi, \theta)$ and in anomalous particle-particle channel $E'_{\uparrow \downarrow}(\theta)$, $E'_{\downarrow \uparrow}(\theta)$ give contribution to the spin-Josephson current, which oscillate with $4\pi$ periodicity.
AC Josephson Effect {#sec5}
===================
In this section, we compute the AC Josephson effect for the tunnel junctions mentioned above. If there is the voltage in Josephson junction $V(t)=V_0+ V_1 \cos{\omega t}$, then from Josephson relation $\dot{\phi}=2eV/\hbar$ we get $$\begin{aligned}
\phi(t) &=& \varphi_0 + \omega_J t + \alpha_1 \sin{\omega t},
\label{phaseeq1}\end{aligned}$$ where $\omega_J= 2eV_0/\hbar$ and $\alpha_1= 2eV_1/(\hbar \omega)$. To obtain the AC Josephson current for such a voltage-biased JJ, we use the following procedure. We consider a JJ with phase difference $\phi$ and having Andreev bound state energies $E_n [\phi,\theta;
\alpha, h, B]$. The Josephson current at $T=0$ can be obtained from these bound states by using $ I_J = (2e/\hbar) \sum_n \partial E_n
/\partial \phi \, \theta(-E_n)$. One can then obtain the AC Josephson current by the replacement $$\begin{aligned}
I_J^{\rm AC} = I_J[\phi \to \phi(t)] \label{acjos1}\end{aligned}$$ For example, for pristine $p-$ wave JJs with $\alpha=h=B=0$, where $E= \pm \Delta \sqrt{D} \cos(\phi/2)$ according to Eq. (\[ksy\]), this procedure leads to $$\begin{aligned}
I_{J1}^{\rm AC} &=& \frac{I_0}{2} \sqrt{D} \sum_n J_n(\alpha_1/2)
\sin[(\varphi_0 +(\omega_J-2n \omega)t)/2] \nonumber\\ \label{ac1}\end{aligned}$$ where $I_0= 2 e \Delta/\hbar$ and we have used the identity $\exp[i
a \sin \theta]= \sum_n J_n[a] \exp[i n \theta]$. Eq. (\[ac1\]) reflects the fact that for $p$-wave junction one has Shapiro steps at $\omega_J/\omega_n = 2n$ for integer $n$; the odd Shapiro steps are absent. The width of the step corresponding to $n=n_0$ is given by $$\begin{aligned}
\Delta I_1 &=& \left|I_0 \sqrt{D} J_{n_0}(\alpha_1/2)
\right|\label{shap1}\end{aligned}$$ where we have used the fact that the maximum width of the step occurs at $\varphi_0 = \varphi_0^{m} = \pi$. The dependence of the step width on $\alpha_1$ for different values of $n_0=0, 1, 2, 3$ is presented in Fig. \[53\].
Next, we apply this procedure for the case where $B=h=0$ but $\alpha
\ne 0$. The Andreev bound state energy is given by Eq.(\[E-B=01\]) and consists of four branches, i.e. each electron and hole branch is split into two states. One of these split states, corresponding to the $+$ sign (Eq. \[E-B=01\]) are independent of $\alpha$. For these two states, the AC Josephson current can be easily shown to be given by Eq. (\[ac1\]); the corresponding Shapiro step width is given by Eq. (\[shap1\]). In contrast, the energy dispersion of the other two branches with designated by $-$ sign (Eq. \[E-B=01\]) depend on the ratio $\alpha/v_F$ and can be rewritten as $$\begin{aligned}
E_-^s &=& \frac{s \Delta_0 \beta \sqrt{D}
\cos(\phi/2)}{\sqrt{1-D(1-\beta^2) \cos^2(\phi/2)}} \label{balpha1}\end{aligned}$$ where $\beta= (1-\alpha/v_F)/(1+\alpha/v_F)$. We note that these branches do not contribute to the Josephson current if $\alpha=v_F$. For $\alpha < v_F$ the contribution from the $s=-$ branch to the current is given by,
$$\begin{aligned}
I_{J2}^{\rm AC} &=& \frac{I_0 \beta \sqrt{D}}{2} \frac{ \sum_n
J_n(\alpha_1/2) \sin[(\varphi_0 +(\omega_J-2n \omega)t)/2]}{\left
\{1 - D_0 - D_0 \sum_n J_n(\alpha_1) \cos[\varphi_0 +(\omega_J-n
\omega)t]\right\}^{3/2}} \label{ac2}\end{aligned}$$
where $D_0= D(1-\beta^2)/2$. The Shapiro step width corresponding to $n=n_0$ is given by $$\begin{aligned}
\Delta I_2 &=& I_0 \sqrt{D} \beta \left|\frac{J_{n_0}[\alpha_1/2]
\sin(\phi_0^m/2)}{\left\{1-D_0 -D_0 J_{n_0}(\alpha_1) \cos(\phi_0^m)
\right\}^{3/2}} \right| \nonumber\\ \label{shap2}\end{aligned}$$ In order to determine $\phi_0^m$, we need to find the value of $\varphi_0$ which maximizes the Shapiro step width. This can be computed easily from Eq. (\[ac2\]) by maximizing the current after setting $\omega_J=2 n_0 \omega$. This procedure yields $$\begin{aligned}
\phi_0^m &=& 2 \arcsin \left[\sqrt{(1-2D_0)/(4 D_0)}\right] \quad
{\rm if} \,
\, D_0 > 1/6 \nonumber\\
&=& \pi \quad {\rm otherwise} \label{maxang1}\end{aligned}$$ The dependence of the Shapiro step width on $\alpha_1$ (Eq. (\[shap2\])) is plotted, using Eq. (\[maxang1\]) in Fig. (\[56\]) for $D_0<1/6$ (Fig. \[56\](a)) and $D_0>1/6$ (Fig.\[56\](b)). We note that Eqs. (\[shap2\]) and (\[maxang1\]) allow one to obtain the strength of the spin-orbit coupling in these JJs from the Shapiro step width.
Next we consider the case where $\alpha=B=0$ and $h \ne 0$. Here the energy dispersion corresponds to four branches as can be seen from Eq. (\[E-B=02\]). We first consider the branches corresponding to $E_{\uparrow \downarrow}^-$. Here we note that when $E_{\uparrow
\downarrow}^-(s=\pm)<0$, the Josephson currents from these two branches cancel each other. Similarly, if $E_{\uparrow
\downarrow}^-(s=\pm)> 0$, none of the branches contribute at $T=0$. Thus these branches contribute to the Josephson current for $\sqrt{D} \sqrt{\Delta^2-4 h^2} |\cos(\varphi/2)|> h$; the presence/absence of Josephson currents from these branches can therefore be used to estimate $D$ in these junctions, provided $\Delta$ and $h$ are known. The Josephson current from these branches and the corresponding Shapiro step widths (when the above-mentioned condition is satisfied) are given by Eq.(\[ac1\]) and (\[shap1\]) respectively. In contrast, the contribution from the branches corresponding to $E_{\uparrow
\downarrow}^+(s)$ are given by
$$\begin{aligned}
I_{J3}^{\rm AC}/I_0 &=& -\frac{2e}{\hbar} \left[D h \sin \phi(t) +
\sum_{s=\pm} \frac{s \sqrt{D} (\Delta^2-4h^2 +4Dh(1+\cos\phi(t)))
\sin [\phi(t)/2]}{\left[ \Delta^2-4 h^2 + 2hD (1+
\cos(\phi(t)/2))\right]^{1/2}} \theta(-E_{\uparrow \downarrow}^-(s))
\right] \label{ac3}\end{aligned}$$
We note that when both $E^-_{\uparrow \downarrow}(s=\pm)<0$, the Josephson current is purely $2 \pi$ periodic and is given by $$\begin{aligned}
I_{J4}^{\rm AC} &=& -\frac{I_0 D h}{\Delta} \sum_n J_n(\alpha_1)
\sin[\varphi_0 + (\omega_J-n \omega)t] \label{ac4}\end{aligned}$$ In this case one has both odd and even Shapiro steps with the step width $\Delta I_4[n_0] = |I_0 D h J_{n_0}(\alpha_1)/\Delta|$. However, for $E_{\uparrow \downarrow}^-(s=+) >0$ and $E_{\uparrow
\downarrow}^-(s=-) < 0$, one has
$$\begin{aligned}
I_{J5}^{\rm AC} &=& I_{J4}^{\rm AC} + \frac{\sqrt{D} [\Delta^2-4h^2
+4Dh +4Dh \sum_n J_n(\alpha_1)\cos(\phi_0+(\omega_J-n \omega)t)]
\sum_n J_n(\alpha_1/2) \sin[(\phi_0+(\omega_J-2n
\omega)t)/2]}{\left[ \Delta^2-4 h^2 + 2hD +2hD \sum_n J_n(\alpha_1)
\cos(\phi_0 +(\omega_J-n \omega)t)\right]^{1/2}} \nonumber\\
\label{ac5}\end{aligned}$$
We note that in this case the JJ will have both $2 \pi$ periodic and $4 \pi$ periodic components. The corresponding Shapiro step width is given for $n=n_0$
$$\begin{aligned}
\Delta I_5 &=& \Delta I_4[n_0] \quad n_0= 2m_0+1 \label{shap5} \\
&=& \Delta I_4[n_0] \left|\sin[\varphi_0^m]\right| +
\left|\frac{\sqrt{D} [\Delta^2-4h^2 +4Dh +4Dh J_{n_0}(\alpha_1)\cos
\phi_0 ] J_{n_0}(\alpha_1/2)\sin \phi_0/2}{\left[ \Delta^2-4 h^2 +
2hD +2hD J_{n_0}(\alpha_1) \cos\phi_0\right]^{1/2}} \right| \quad
n_0=2m_0 \nonumber\end{aligned}$$
where $m_0$ is an integer and $\varphi_0^m$ denotes the value of $\varphi_0$ for which the stepwidth is maximum. This value needs to be numerically determined for the present case by minimizing Eq.(\[ac5\]) at $\omega_J= 2 n_0 \omega$ with respect to $\varphi_0$.
Finally, we treat the case $B,h \ne 0$ and $\alpha=0$. Here the Andreev bound states are given by Eq. (\[Eup-down1\]) and (\[Eup-up1\]). For $E\equiv E_{\uparrow \downarrow}^{\pm}(s)$ (Eq. (\[Eup-down1\])), there are four branches. For $E=E^+_{\uparrow \downarrow}$ both the branches are above the Fermi energy if $\sqrt{B^2 +h^2} \ge \Delta \sqrt{D}$. In this case, there is no Josephson current contribution from these branches. Similarly, for $E=E^-_{\uparrow \downarrow}$, the same condition results in both the branches being below the Fermi level. In this case, their contribution to the Josephson currents cancel each other. Thus the contribution to the Josephson current from $E_{\uparrow \downarrow}$ occurs only when $\sqrt{B^2+h^2} < \Delta \sqrt{D}$. However, even in this case, the contribution to the Josephson currents from the positive ($E^+$) and negative ($E^-$) branches cancel each other and the net Josephson current vanishes. A similar results can be easily deduced for $E= E_{\uparrow \uparrow}^{\pm}$ (Eq. (\[Eup-up1\])) and $E_{\downarrow \uparrow}$ (Eq. (\[E1\])) branches.
![The dependence of the Shapiro step width for $p$-wave JJs at $\alpha = {\bf B}=h=0$ on the amplitude of the $ac$-voltage $\alpha_1= 2eV_1/(\hbar \omega)$ at (a) $n_0=0$, (b) $n_0=1$, (c) $n_0=2$, and (d) $n_0=3$. $D$ is chosen to be $D=0.5$.[]{data-label="53"}](fig13a.eps "fig:"){width="4.7cm"}![The dependence of the Shapiro step width for $p$-wave JJs at $\alpha = {\bf B}=h=0$ on the amplitude of the $ac$-voltage $\alpha_1= 2eV_1/(\hbar \omega)$ at (a) $n_0=0$, (b) $n_0=1$, (c) $n_0=2$, and (d) $n_0=3$. $D$ is chosen to be $D=0.5$.[]{data-label="53"}](fig13b.eps "fig:"){width="4.7cm"}\
![The dependence of the Shapiro step width for $p$-wave JJs at $\alpha = {\bf B}=h=0$ on the amplitude of the $ac$-voltage $\alpha_1= 2eV_1/(\hbar \omega)$ at (a) $n_0=0$, (b) $n_0=1$, (c) $n_0=2$, and (d) $n_0=3$. $D$ is chosen to be $D=0.5$.[]{data-label="53"}](fig13c.eps "fig:"){width="4.7cm"}![The dependence of the Shapiro step width for $p$-wave JJs at $\alpha = {\bf B}=h=0$ on the amplitude of the $ac$-voltage $\alpha_1= 2eV_1/(\hbar \omega)$ at (a) $n_0=0$, (b) $n_0=1$, (c) $n_0=2$, and (d) $n_0=3$. $D$ is chosen to be $D=0.5$.[]{data-label="53"}](fig13d.eps "fig:"){width="4.7cm"}
Conclusion {#sec6}
==========
In this paper we study the Josephson current between 1D nanowires of $p$-wave superconductors separated by an insulating barrier in the presence of Rashba SOI and the magnetic fields ${\bf B}$ and $h$. The presence of the SOI and Zeeman magnetic fields enlarges the standard two-component $BdG$ system equations to four-component system equations (\[Sch1\])- (\[Sch4\]) for new BdG wave vector $\eta_a(x)=\left(\eta_{a,\uparrow, +}^{\dag}(x), \eta_{a,
\downarrow,+}^{\dag}(x), \eta_{a, \downarrow, -} (x),
\eta_{a,\uparrow, -}(x)\right)$. The BdG equations (\[Sch1\])-(\[Sch4\]) coincide with the standard BdG equations in the absence of the in-plane magnetic field ${\bf B}$, which provide only one relation $\eta^{\ast}_{a, \sigma, b}/\eta_{a,
{\bar \sigma}, {\bar b}}$ between quasi-particle and quasi-hole states, where $\sigma = \uparrow, \downarrow$ and ${\bar
\sigma}=\downarrow, \uparrow$. Instead, the BdG equations (\[Sch1\])-(\[Sch4\]) in the presence of ${\it in-plane}$ magnetic field ${\bf B}$ and Rashba SOI provide three relations $\eta^{\ast}_{a, \sigma, b}/\eta_{a,{\bar \sigma},{\bar b}}$, $\eta^{\ast}_{a, \sigma, b}/\eta_{a,\sigma, {\bar b}}$, and $\eta_{a, \sigma,b}/\eta_{a,{\bar \sigma}, b}$ between the quasi-particle and quasi-hole states, corresponding to new Andreev scattering channels. We studied in this paper all these scattering channels in detail by generalizing the method of Ref. for study of Josephson junction with $\delta$-function insulator between two $p$-wave superconductors to systems with SOI and Zeeman fields. We have shown in this paper that $\pi$-state is realized in Josephson junction with $p$-wave superconductors. Moreover, we have demonstrated the existence of magneto-Josephson effect in these systems. We note that although the existence of the magneto-Josephson effect in a topological superconductor has been predicted recently [@jpar13; @kss12; @pjpa13], the question of whether this effect is observable in superconducting junctions with $p$-wave superconductors and the presence of SOI was not addressed before. We have predicted in the paper new Andreev-type tunneling channels for quasi-particles and quasi-holes which are responsible to the magneto-Josephson effect.
In conclusion, we have studied Josephson effect in a junction between two $p$-wave 1D nanowires in the presence of SOI and Zeeman fields. We have analyzed the Josephson current in these junctions and provided analytical expressions of the Andreev bound states in several limiting cases. We have also demonstrated the presence of magneto-Josephson effect in these junctions. Our heoretical predictions are shown to be verifiable by straightforward experiments on these systems.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors kindly acknowledge the Scientific Fund of State Oil Company of Azerbaijan Republic (SOCAR) for financial support of 2019-2020 grant entitled ’Study of the impurity and correlation effects in graphene, fullerene and other topological nanostructures’. The reported study was partially funded by the RFBR research projects 18-02-00318 and 18-52-45011-IND. Part of the numerical calculations were made in the framework of the RSF project 18-71-10095. KS thanks DST, India for support through project INT/RUS/RFBR/P-314.
Andreev bound state energies at ${\bf B} \neq 0$
================================================
Andreev bound state energies are obtained from the conditions $F_{\uparrow \downarrow}^{\ast}(E_{\uparrow \downarrow})=0$, $F_{\uparrow \uparrow}^{\ast}(E_{\uparrow \uparrow})=0$, $F_{\uparrow \downarrow}({\tilde E}_{\uparrow \downarrow})=0$ corresponding to three channels, and also from the conditions, obtained by interchanging the spin orientations as $F_{\downarrow \uparrow}^{\ast}(E_{\downarrow \uparrow})=0$, $F_{\downarrow \downarrow}^{\ast}(E_{\downarrow \downarrow})=0$, $F_{\downarrow \uparrow}({\tilde E}_{\downarrow \uparrow})=0$.
Andreev bound state energy $E_{\uparrow \downarrow}$ is obtained from the condition $F_{\uparrow \downarrow}^{\ast}(E_{\uparrow \downarrow})=0$ with the expression (\[Fup-down\]) yielding
$$\begin{aligned}
&&\left\{(E_{\uparrow \downarrow}+h)\left[(E_{\uparrow \downarrow}-h)^2 +k^2(v_F+\alpha)^2+B^2-\Delta^2\right]^2-
2B^2E_{\uparrow \downarrow} \right\}^2\left(1-D\cos^2\frac{\varphi}{2}\right)-\nonumber\\
&&\left\{k(v_f-\alpha)\left[(E_{\uparrow \downarrow}-h)^2+k^2(v_F+\alpha)^2+B^2-\Delta^2\right]+2kB^2\alpha\right\}^2
D\cos^2\frac{\varphi}{2}=0.
\label{AEup-down}\end{aligned}$$
This expressin depends apart from the parameters $\alpha$, ${\bf B}$, and $h$ also on the momentum $k^2$. Expression for $k^2_{\pm}$, obtained from the energy spectrum (\[Eo\]), reads $$\begin{aligned}
&&k^2_{\pm}(v_F^2-\alpha^2)^2=(E^2-h^2-B^2-\Delta^2)(v_F^2-\alpha^2)-2(Ev_F+\alpha h)^2+2\Delta^2v_F^2 \pm
\Big\{\big[(E^2-h^2-B^2-\Delta^2)(v_F^2-\alpha^2)-\nonumber\\
&& 2(Ev_F+\alpha h)^2+2\Delta^2v_F^2\big]^2-
\left[(E^2-h^2-B^2-\Delta^2)^2-4\Delta^2(h^2+B^2)\right](v_F^2-\alpha^2)^2 \Big\}^{1/2}.
\label{k2}\end{aligned}$$ Elimination of $k^2$, by substituting it from (\[k2\]) into Eq. (\[AEup-down\]), yields a general expression to find $E_{\uparrow \downarrow}$, which is not easy to solve exactly.
The condition $F_{\uparrow \uparrow}^{\ast}(E_{\uparrow \uparrow})=0$ with the expression (\[Fup-up\]) yields $$k^2(Ev_F+\alpha h)^2-\Delta^2(E^2_{\uparrow \uparrow}+\alpha^2 k^2)+\Delta^2 D(E^2_{\uparrow \uparrow}+\alpha^2 k^2)
\cos^2 \frac{\varphi -\theta}{2}=0.
\label{AEup-up2}$$ Routine calculations, after substitution of $k^2$ from Eq. (\[k2\]) to this equation, result in $$\begin{aligned}
&&\left\{(v_F^2- \alpha^2)E^2_{\uparrow \uparrow}\Delta^2\left(1-D\cos^2\frac{\varphi - \theta}{2}\right)-(E^2_{\uparrow \uparrow}
+\Delta^2-B^2-h^2) \left[(E_{\uparrow \uparrow}v_F+\alpha h)^2- \alpha^2 \Delta^2\left(1-D\cos^2 \frac{\varphi -
\theta}{2}\right)\right]\right\}^2-\nonumber\\
&&4E^2_{\uparrow \uparrow} \Delta^2 D\cos^2\frac{\varphi - \theta}{2} (E_{\uparrow \uparrow}v_F +\alpha h)^2
\left[(E_{\uparrow \uparrow} v_F +\alpha h)^2-\alpha^2 \Delta^2 \left(1-D \cos^2 \frac{\varphi - \theta}{2}\right)\right]=0
\label{AEup-up3}\end{aligned}$$ This equation can be solved numerically for a general case when $\alpha,~ {\bf B},~ h \neq 0$. The expression for $F_{\downarrow \downarrow}^{\ast}(E_{\downarrow \downarrow})=0$ is obtained from Eq. (\[AEup-up3\]) by interchanging $\alpha \to -\alpha$, $h \to - h$, and $\theta \to - \theta$.
The bound state energy in the particle-particle channel with opposite spin orientations is determined from the condition $F_{\uparrow \downarrow}(E'_{\uparrow \downarrow})=0$, which can be written by using the expression (\[FAup-down\]) for $F_{\downarrow \uparrow}$ as, $$\alpha^2 k^2- \left(E^{' 2}_{\uparrow \downarrow}+\alpha^2 k^2\right)D\sin^2\frac{\theta}{2}=0.
\label{FAup-down1}$$ Substituting $k^2$ from Eq. (\[k2\]) to this equation yields the equation to determine $E'_{\uparrow \downarrow}$, $$\begin{aligned}
&&\alpha^4 \left[ (E_{\uparrow \downarrow}^{' 2}+ \Delta^2 -B^2-h^2)^2-4 \Delta^2 E^{' 2}_{\uparrow \downarrow} \right]+2\alpha^2 D \sin^2
\frac{\theta}{2} \Big\{v_F^2 E^{' 4}_{\uparrow \downarrow} + (v_F^2-\alpha^2)\Delta^2 E^{' 2}_{\uparrow \downarrow}+
(v_F^2+\alpha^2)B^2 E^{' 2}_{\uparrow \downarrow}+\nonumber\\
&& (v_F^2+3\alpha^2) h^2 E^{' 2}_{\uparrow \downarrow}-4v_F\alpha h E^{' 3}+\alpha^2(\Delta^2-B^2-h^2)^2 \Big\}+D^2\sin^4\frac{\theta}{2}
\Big\{ v_F^2(v_F^2-4\alpha^2)E^{' 4}_{\uparrow \downarrow}+\nonumber\\
&& 2\alpha^2 E^{' 2}_{\uparrow \downarrow}\left[v_F^2\Delta^2-v_F^2B^2-(v_F^2+2\alpha^2)h^2\right]-8\alpha^3 v_FhE^{' 3}_{\uparrow \downarrow}+
\alpha^4 (\Delta^2-B^2-h^2)^2\Big\}=0.
\label{FAup-down2}\end{aligned}$$ Note that the condition $F_{\downarrow \uparrow}(E'_{\downarrow \uparrow})=0$ provides for $E'_{\downarrow \uparrow}$ exactly the same expression as (\[FAup-down2\]), i. e. $E'_{\downarrow \uparrow}=E'_{\uparrow \downarrow}$. Indeed, a reality of this fact can be tested according to the rule that interchanging the spin directions is equivalent to the replacements of $\alpha \to - \alpha$, $h \to - h$, and $\theta \to - \theta$.
[99]{} A.Yu. Kitaev, Phys. Usp. [**44**]{}, 131 (2001). A. Yu. Kitaev, Annals Phys. [**303**]{}, 2 (2003). R. M. Lutchyn, J. D. Sau, and S. Das Sarma, Phys. Rev. Lett. [**105**]{}, 077001 (2010). Y. Oreg, G. Refael, and F. von Oppen, Phys. Rev. Lett. [**105**]{}, 177002 (2010). N. Read and D. Green, Phys. Rev. B [**61**]{}, 10267 (2000). D. A. Ivanov, Phys. Rev. Lett. [**86**]{}, 268 (2001). J. Alicea, Y. Oreg, G. Refael, F. von Oppen, and M. P. A. Fisher, Nat. Phys. [**7**]{}, 412 (2011). L. P. Gor’kov and E. I. Rashba, Phys. Rev. Lett. [**87**]{}, 037004 (2001). X. Liu, J. K. Jain, and C. X. Liu, Phys. Rev. Lett. [**113**]{}, 227002 (2014). C. R. Reeg and D. L. Maslov, Phys. Rev. B [**92**]{}, 134512 (2015). T. Yu and M. W. Wu, Phys. Rev.B [**93**]{}, 195308 (2016). V. M. Edelstein, Zh. Eksp. Teor. Fiz. 95, 2151 (1989) \[Sov. Phys. JETP 68, 1244 (1989)\]. P. A. Frigeri, D. F. Agterberg, A. Koga, and M. Sigrist, Phys. Rev. Lett. 92, 097001 (2004). H. -J. Kwon, K. Sengupta, and V. M. Yakovenko, Eur. Phys. J. B [**37**]{}, 349 (2004). L. Fu and C. L. Kane, Phys. Rev. B [**79**]{}, 161408 (2009). L. Jiang, D. Pekker, J. Alicea, G. Refael, Y. Oreg, and F. von Oppen, Phys. Rev. Lett. [**107**]{}, 236401 (2011). L. Jiang, D. Pekker, J. Alicea, G. Refael, Y. Oreg, A. Brataas, and F. von Oppen, Phys. Rev. B [**87**]{}, 075438 (2013); S. Jacobsen, I. Kulagina, and J. Linder, Sci. Rep. 6, 23926 (2016). E. Nakhmedov, O. Alekperov, F. Tatardar, Yu. M. Shukrinov, I. Rahmonov, and K. Sengupta, Phys. Rev. B [**96**]{}, 014519 (2017). F. S. Nogueira and K. H. Bennemann, Europhys. Lett. [**67**]{}, 620 (2004). Y. Asano, Phys. Rev. B [**74**]{}, 220501 (2006). P. M. R. Brydon, Phys. Rev. B [**80**]{}, 224520 (2009). P. M. R. Brydon, Y. Asano, and C. Timm, Phys. Rev. B [**83**]{}, 180504 (2011). H. Zhang, K. S. Chan, Z. Lin, and J. Wang, Phys. Rev. B [**85**]{} 024501 (2012). L. Klam, A. Epp, W. Chen, M. Sigrist, and D. Manske, Phys. Rev. B [**89**]{}, 174505 (2014). L. Elster, M. Houzet, and J. S. Meyer, Phys. Rev. B [**93**]{}, 104519 (2016) A. G. Bauer and B. Sothmann, Phys. Rev. B [**99**]{}, 214508 (2019). M. T. Mercaldo, P. Kotetes, and M. Cuoco, Phys. Rev. B [**100**]{}, 104519 (2019). A. J. Leggett, Rev. Mod. Phys. [**47**]{}, 331 (1975). V. P. Mineev and K. V. Samochin, [*Introduction to the theory of unconventioal superconductivity*]{}, Moscow Engeneering-Technical Institute Press, Moscow 1998, p. 144, ISBN 5-89155-024-5. J. Nilsson, A. R. Akhmerov, and C. W. J. Beenakker, Phys. Rev. Lett. [**101**]{}, 120403 (2008). T. Birol and P. W. Brouwer, Phys. Rev. B [**80**]{}, 014434 (2009). P. Kotetes, G. Schön, and A. Shnirman, J. Korean Phys. Soc. [**62**]{}, 1558 (2013). F. Pientika, L. Jiang, D. Pekker, J. Alicea, G. Refael, Y. Oreg, and F. von Oppen, New J. Physics [**15**]{}, 115001 (2013).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Since the 1980’s it has been becoming increasingly clear that the Solar System’s irregular satellites are collisionally evolved. The current populations are remnants of much more massive swarms that have been grinding away for billions of years. Here, we derive a general model for the collisional evolution of an irregular satellite swarm and apply it to the Solar System and extrasolar planets. The model uses a particle in a box formalism and considers implications for the size distribution, which allows a connection between irregular satellite populations and predicted levels in the resulting dust cloud. Our model reproduces the Solar System’s complement of observed irregulars well, and suggests that the competition between grain-grain collisions and Poynting-Robertson (PR) drag helps set the fate of the dust. In collision dominated swarms most dust is lost to interplanetary space or impacts the host planet, while PR dominated grains spiral in towards the planet through the domain of regular satellites. Because swarm collision rates decrease over time the main dust sink can change with time, and may help unravel the accretion history of synchronously rotating regular satellites that show brightness asymmetries, such as Callisto and Iapetus. Some level of dust must be present on AU scales around the Solar System’s giant planets if the irregular satellites are still grinding down, which we predict may be at detectable levels. We also use our model to predict whether dust produced by extrasolar circumplanetary swarms can be detected. Though designed with planets in mind, the coronagraphic instruments on JWST will have the ability to detect the dust generated by these swarms, which are most detectable around planets that orbit at many tens of AU from the youngest stars. Because the collisional decay of swarms is relatively insensitive to planet mass, swarms can be much brighter than their host planets and allow discovery of Neptune-mass planets that would otherwise remain invisible. This dust could have been detected by HST ACS coronagraphic observations, and in one case dust may have already been detected. The observations of the planet Fomalhaut b can be explained as scattered light from dust produced by the collisional decay of an irregular satellite swarm around a $\sim$10$M_\oplus$ planet. Such a swarm comprises about 5 Lunar masses worth of irregular satellites. Finally, we briefly consider what happens if Fomalhaut b passes through Fomalhaut’s main debris ring on a coplanar orbit, which allows the circumplanetary swarm to be replenished through collisions with ring planetesimals. This scenario, in which the planet is at least of order an Earth mass, may be ruled out by the narrow structure of the debris ring.'
author:
- |
G. M. Kennedy[^1] and M. C. Wyatt\
Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK
bibliography:
- 'ref.bib'
title: 'Collisional Evolution of Irregular Satellite Swarms: Detectable Dust around Solar System and Extrasolar Planets.'
---
circumstellar matter – stars: planetary systems: general – Solar System: planets and satellites
Introduction {#s:intro}
============
With a penchant for retrograde, barely-bound, high-eccentricity orbits and flatter-than-usual size distributions, the irregular satellites are one of the Solar System’s rebel populations. Their presumed capture into these unusual orbits around the Solar System’s giant planets has long been a puzzle. Because irregulars exist at all four outer planets, capture mechanisms specific to formation of gas giants like Jupiter and Saturn [e.g. @1977Icar...30..385H; @1979Icar...37..587P] are not general enough. Dynamical mechanisms, which do not rely on the existence or growth of large gaseous atmospheres, can be applied to gas and ice-giants alike and are therefore favoured [e.g. @1971Icar...15..186C]. @2007AJ....133.1962N recently proposed a different dynamical mechanism as part of a unified model of outer Solar System formation [@2005Natur.435..466G; @2005Natur.435..459T], where irregulars are captured during a period of instability via planet-planet interactions.
An apparent weakness of the planet-planet interaction mechanism is that the irregular satellites should share the same size distribution as Jupiter’s Trojan asteroids, which in this model were captured at the same time from the same population [@2005Natur.435..462M]. In fact the differences in the Trojan and irregular satellite size distributions are marked, with irregular satellites being much flatter for sizes larger than about 10km. To overcome this hurdle, @2010AJ....139..994B showed that while the size distribution was indeed *initially* steeper like the Trojans, 4.5Gyr or so of collisional evolution is sufficient to reduce primordial irregular satellite populations to a size distribution that matches those currently observed. This result implies that there were previously more irregular satellites and perhaps most significantly, that copious amounts of dust were produced during the depletion of these satellites.
The evidence for collisional evolution of irregular satellites has been mounting for some time. @1981Icar...48...39K showed that the four prograde irregulars known to orbit Jupiter at the time had a relatively short collisional lifetime. The advent of large-format CCD surveys since the turn of the century has seen a dramatic increase the number of irregulars and made further theoretical advances possible . @2003AJ....126..398N noted that irregular satellites around planets closer to the Sun have larger orbits (in Hill radii), suggesting that the lack of satellites closer to Jupiter is due to their erosion through collisions, which proceed at a faster rate closer to the planet. Using numerical integrations to derive average orbital elements, they also proved the existence of collisional families [see also @2001Natur.412..163G]. This latter discovery is particularly important, because the current collision rate amongst the irregular satellites is in some cases too low to explain their existence [@2003AJ....126..398N]. The inference is again that the number of irregular satellites, and thus their collision rate, was much greater in the past and that they decayed through collisions to the current level.
Coinciding with these theoretical advances was the discovery by direct imaging of Fomalhaut b [@2008Sci...322.1345K], an extrasolar planet predicted to exist based on the elliptical orbit of Fomalhaut’s circumstellar debris ring [@2005Natur.435.1067K; @2006MNRAS.372L..14Q]. The ring structure suggests that the planet is less than 3[$M_{\rm {Jup}}$]{}, though the planet could be much less massive [@2009ApJ...693..734C]. This discovery appears unrelated to irregular satellites, but the inability of Hubble, Keck, and Gemini photometry to pin down whether the planet looks like a planetary atmosphere or reflected starlight provides the link.
While the planets discovered to orbit HR 8799 appear to be consistent with $\sim$1000K substellar-mass objects [@2008Sci...322.1348M], multi-wavelength photometry of Fomalhaut b appears bluer than expected for a 200Myr old gas-giant planet. Specifically, Fomalhaut b has so far defied detection at wavelengths longer than 1[$\mu$m]{}, leading @2008Sci...322.1345K to suggest that the spectrum is actually starlight scattered from an optically thick circumplanetary disk of about 20 Jupiter radii. Though such a scenario is plausible, we argue that dust produced by a swarm of colliding irregular satellites is also a possibility.
Given that the Solar System’s irregular satellite complement decayed to its current state through collisions and the exciting possibility that Fomalhaut b may harbour the first circumplanetary dust seen outside the Solar System, the time seems right to consider whether such clouds of irregular satellites could be visible around extrasolar planets. In the following sections, we derive a simple model for the evolution of a circumplanetary satellite swarm and the all-important extrasolar observable—the dust. We compare our model with the Solar System irregulars, and comment on the fate and observability of dust. We then apply the model to circumplanetary swarms around extrasolar planets. Finally, we explore what kind of satellite swarm could exist around Fomalhaut b, and what constraints this proposed swarm puts on the planet mass.
Model of a Circumplanetary Swarm {#s:mod}
================================
The irregular satellite swarms described in this paper have not knowingly been detected around other planets that orbit other stars. Therefore, like the pre-1995 days of planet formation theory, we must take cues from the Solar System. However, based on experience gained from the surprising diversity of extrasolar planets, we should not assume that our irregular satellite complement is typical, or that extrasolar analogues should follow all the same rules.
Thankfully, some of the most important irregular satellite properties are dynamical and would have been discovered even if the Solar System had no irregulars. The main dynamical curiosity is their inclinations, which are all within about 60$^\circ$ of the ecliptic (but include retrograde orbits). This evacuation of near-polar orbits is due to Solar and planetary perturbations, which drive the eccentricities of highly inclined orbits to such large values that they either encounter regular satellites or leave the Hill sphere [@2002Icar..158..434C; @2003AJ....126..398N].
Another constraint comes from the stability of circumplanetary orbits. Although @2008AJ....136.2453S find that satellites out to a few Hill radii could survive the age of the Solar System around Uranus and Neptune, all currently known irregulars have orbits with semi-major axes $a_{\rm pl}$ less than half the Hill radius [$R_{\rm H}$]{} [ e.g. @2003Natur.423..261S; @2004Natur.430..865H; @2005AJ....129..518S] $$\label{eq:rh}
R_{\rm H} = a_{\rm pl} \left( M_{\rm pl} / 3\,M_\star \right)^{1/3} \, ,$$ where $M_{\rm pl}$ is the planet mass and $M_\star$ the stellar mass. On the sky, the giant planets’ Hill radii span several degrees. The exact stability limit has a small inclination dependence in that retrograde orbits are stable at larger distances than the widest stable prograde orbits [@2003AJ....126..398N].
In this section we outline a model for the collisional evolution of irregular satellite swarms. Because we want to make predictions of the only possible extrasolar observable—dust—we keep our model simple. There are many uncertainties in extrapolating a swarm of irregular satellites to a cloud of dust, such as the strength of satellites, the size distribution slope, and the minimum grain size, which at this stage make the development of a more complex collisional model largely unnecessary.
The next four subsections contain many equations that describe properties of a circumplanetary swarm. Readers looking for actual numbers may like to refer ahead to Table \[t:ssdust\], which shows estimates of some properties for the Solar System giant planets.
Collisional mass loss {#ss:decay}
---------------------
In a steady-state collisional cascade the mass within a given size range decreases as these objects are destroyed in collisions, but is replaced at the same rate by fragments created by destruction of larger objects. Mass is lost at the bottom end of cascade, usually by radiation forces that remove grains smaller than some minimum size. The evolution of the size distribution is therefore dictated by the collisional decay of the largest objects.
The result of such a collisional cascade would, in an ideal situation (cascade infinite in extent, strength independent of size), have a steady state size distribution with a well defined slope of $n(D) = K \, D^{2-3q}$ where $n(D)dD$ is the number of satellites between $D$ and $D+dD$ and $q=11/6$ [@1969JGR....74.2531D].
In fact the true distribution of circumstellar collisional cascades like the asteroid belt is expected to have different slopes in different size ranges due to the way strength depends on size [@2003Icar..164..334O]. Strength is typically described by the parameter [$Q_{\rm D}^\star$]{}, which is the kinetic collision energy per target mass needed to shatter *and disperse* the target, such that the largest remnant is half the mass of the original target. Such a collision is commonly termed “catastrophic.”
Small objects are held together by their own material strength, and grow weaker with increasing size due to the increased likelihood of the presence of a significant flaw. To quote @1994Icar..107...98B, “Subdivide this same rock into 100 equal pieces and 99 of them are now stronger than the original, owing to the simple fact that they do not contain the one weakest flaw.” Above some transition size ($D_{\rm t} \sim 0.1km$) bodies gain strength from self-gravity. The energy needed for catastrophic disruption now increases with size. Though objects may be shattered, extra energy is required to ensure the fragments have sufficient escape speeds and are no longer bound. Gravity also limits fracture propagation within the material, thus adding strength [e.g. @1999Icar..142....5B]. This behaviour is usually modelled using complex numerical codes, and parameterised by a power law for each of the strength and gravity regimes. In fact [$Q_{\rm D}^\star$]{} varies by about a factor 10 over the range of impact parameters and there are similar differences between strength laws derived by different studies [e.g. @1999Icar..142....5B; @2009ApJ...691L.133S]. Thus [$Q_{\rm D}^\star$]{} and the resulting size distribution are the most uncertain inputs for our model.
We set planetesimal strength with the @1999Icar..142....5B law for ice at 3km s$^{-1}$. For objects larger than $D_{\rm t} = 0.1$ (in km), $$\label{eq:qd}
Q_{\rm D}^\star = 0.1\, \rho \, D^{1.26} / f_{\rm Q}$$ in J kg$^{-1}$ where the mass density $\rho$ is in kg m$^{-3}$. The strength dependence of small objects ($\propto$$D^{-0.39}$) is only used in setting the size distribution of objects smaller than [$D_{\rm t}$]{}. Following @2010AJ....139..994B, we allow objects to be weaker than the @1999Icar..142....5B law by including the factor [$f_{\rm Q}$]{} [see also @2009Natur.460..364L]. The strength law is similar to the @2009ApJ...691L.133S strength law when $f_{\rm Q} =
8$. @2010AJ....139..994B reproduce the Solar System’s irregular satellite populations best when $f_{\rm Q} > 3$ so we set $f_{\rm Q} = 5$.
The size distribution of objects with such strength properties is expected to have a slope with $q_{\rm{s}}=1.9$ at $D<D_{\rm t}$ (in the strength-dominated regime) and $q_{\rm{g}}=1.7$ for $D>D_{\rm t}$ [in the gravity-dominated regime, @2003Icar..164..334O]. Although several wiggles are also expected in the distribution , here the size distribution is assumed to be continuous with the appropriate slopes ($q_{\rm{s}}$ or $q_{\rm{g}}$) between the smallest objects of size $D_{\rm{min}}$ (in [$\mu$m]{}) and the largest objects participating in the collisional cascade of size $D_{\rm{c}}$ (in km). As we show later, [$D_{\rm c}$]{} may be smaller than the largest object, which has size [$D_{\rm max}$]{}.
With this two phase size distribution, the conversion between the size distribution’s surface area ([$\sigma_{\rm tot}$]{} in AU$^2$) and mass ([$M_{\rm tot}$]{} in $M_\oplus$) is $$\label{eq:mtotgen}
M_{\rm tot} = 0.0025 \, \rho \, \sigma_{\rm tot} \, \frac{3q_{\rm s}-5}{6-3q_{\rm g}} \,
D_{\rm c}^{6-3q_{\rm g}} \, D_{\rm t}^{3q_{\rm g}-3q_{\rm s}} \, D_{\rm min}^{3q_{\rm s}-5}$$ where we assume $D_{\rm min} \ll D_{\rm t} \ll D_{\rm c}$. For $q_{\rm s} = 1.9$, $q_{\rm
g} = 1.7$, and $D_{\rm t} = 0.1$ (in km), this relation simplifies to $$\label{eq:mtot}
M_{\rm tot} = 3.9 \times 10^{-6}\, \rho\, \sigma_{\rm tot} \,D_{\rm c}^{0.9}\, D_{\rm
min}^{0.7} \, .$$ This mass only includes objects between [$D_{\rm min}$]{} and [$D_{\rm c}$]{} (not [$D_{\rm max}$]{}) because we use [$M_{\rm tot}$]{}below to calculate collision rates.
The collisional lifetime of satellites of size $D_{\rm{c}}$ can be calculated from the total mass using the particle-in-a-box approach. Here we follow @2010MNRAS.402..657W, who also take into account that objects of size [$D_{\rm c}$]{} can be destroyed in catastrophic collisions by those down to a size $X_{\rm{c}}\,D_{\rm c}$ (assumed to be $\gg 0.1$km) where $X_{\rm c} = \left( 2 Q_{\rm D}^\star / v_{\rm rel}^2
\right)^{1/3}$ and the collision velocity [$v_{\rm rel}$]{} is in m s$^{-1}$. The rate of catastrophic collisions is $$\label{eq:rcc}
R_{cc} = 8.4 \times 10^{-5} \, \frac{6-3q_{\rm g}}{3q_{\rm g}-5}\,\frac{ v_{\rm{rel}} \, C_1
X_{\rm c}^{C_2}\, M_{\rm tot} }
{ \rho \, D_{\rm c} \, V }$$ in years$^{-1}$, where $V$ is the volume occupied by the satellites in AU$^3$. The inverse of the collision rate is called the collision time $t_{\rm c}$. The assumption of $D_{\rm t} \ll X_{\rm{c}}\,D_{\rm c}$ means that only $q_{\rm g}$ is needed for the collision rate. To obtain equation (\[eq:rcc\]), it was necessary to integrate over the size distribution from $X_{\rm c}\,D_{\rm c}$ to [$D_{\rm c}$]{}, which yields a function called $G(q,X_{\rm c})$ by @2007ApJ...658..569W. In @2010MNRAS.402..657W the approximation $G \left( 11/6,X_c \right) \approx C_1 X_c^{C_2}$ with $C_1 = 0.2$ and $C_2
= -2.5$ is used (the limit of small $X_c$). Substituting these values and $q_{\rm g} =
11/6$ yields Equations (9) & (10) of @2010MNRAS.402..657W. Here, we find a numerical approximation for the function $G \left( q,X_c \right)$, which is within 10% for the more physically plausible range $0.01 < X_c < 0.75$ and $1.7 < q <2$, yielding $C_1 = 2.62(q-1.66)$ and $C_2 = 2.70(0.98-q)$. For $q_{\rm g} = 1.7$, $C_1 = 0.1$ and $C_2 = -1.9$.
Irregular satellites are assumed to orbit the planet at semi-major axes relative to the Hill radius in the range $\eta \pm d\eta/2$ (we use $d\eta = \eta/2$). The volume the satellites occupy is $V = 4 \pi \eta_{\rm{s}}^2 d\eta_{\rm{s}} r_{\rm{H}}^3 \times
0.866$, where the extra factor of $0.866$ accounts for the lack of near polar orbits.
The mean relative velocity of collisions is expected to be some fraction [$f_{\rm v_{\rm rel}}$]{} of the Keplerian velocity at $\eta$, which (in m s$^{-1}$) is $$\label{eq:vk}
v_{\rm k} = 516 \, M_{\rm pl}^{1/3} \, M_\star^{1/6} / ( \eta \, a_{\rm pl} )^{1/2} \, .$$ That fraction will depend on the eccentricities and inclinations of the satellite swarm. A simple estimate of the mean collision velocity comes from assuming circular and isotropic orbits, yielding $f_{\rm v_{\rm rel}} = 4/\pi$ and typical impact velocities of $\sim$0.5–3km/s for Solar System giant planets. These velocities are high enough that the impactor/target mass ratio for catastrophic collisions is small, so the energy lost in a collision is also small. Thus very little kinetic energy is lost in a typical collision and collisional damping is unimportant. Using the Monte-Carlo eccentric ring model of @2010MNRAS.402..657W, we find that the mean collision velocity is similar for a realistic orbital distribution, and somewhat lower when eccentricities are introduced. The values vary between about 0.9–1.3, so we adopt $f_{\rm v_{\rm rel}} =
4/\pi$ throughout. In fact orbits have a range of inclinations and eccentricities, and each collision has a different probability, which is itself a function of the relative velocity [@1994Icar..107..255B].
Substituting our approximation for $C_1$ and $C_2$ for $q_{\rm g} = 1.7$ yields the rate of catastrophic collisions $$\label{eq:rccdc}
R_{cc} = 1.3 \times 10^7 \,
\frac{ M_{\rm tot} \, M_\star^{1.38} \, f_{\rm v_{\rm rel}}^{2.27} }{ \, {Q_{\rm D}^\star}^{0.63}
\rho \, D_{\rm c} \, M_{\rm pl}^{0.24} \, (a_{\rm pl} \, \eta)^{4.13} }$$ in years$^{-1}$. As one expects, the rate is independent of $a_{\rm pl}$ for the same physical swarm (i.e doubling $a_{\rm pl}$ halves $\eta$). The factors that largely set the collision rate are $\eta$, [$D_{\rm c}$]{}, $a_{\rm pl}$, and [$f_{\rm v_{\rm rel}}$]{}. The strongest contributions are from $\eta$ and $a_{\rm pl}$, which set the cloud volume and space density. The rate also depends strongly on the mean collision velocity because this speed sets both the rate at which an object sweeps through space and the number of impactors that result in a catastrophic disruption. Greater collision velocities mean smaller impactors can destroy a given object, and smaller impactors are more numerous. Because [$Q_{\rm D}^\star$]{} also depends on [$D_{\rm c}$]{} (eq. \[eq:qd\]), the collision rate also depends strongly on [$D_{\rm c}$]{}. For the same total mass, larger [$D_{\rm c}$]{} means fewer large objects, which are also stronger.
Curiously, the least important parameter is the mass of the planet itself. This result arises because in equation (\[eq:rccdc\]), [$v_{\rm rel}$]{} is slightly less than linearly dependent on [$M_{\rm pl}$]{}, which nearly cancels with the linear dependence of volume on [$M_{\rm pl}$]{}. In simple terms, for fixed $\eta$ and [$M_{\rm tot}$]{} the increase in volume with planet mass works out to be slightly stronger than the increase in velocity, so the collision rate decreases slowly as the planet mass increases.
We validate our model with the eccentric ring model [@2010MNRAS.402..657W]. Our collision rate is within a factor of three for a range of eccentricities with the best agreement for high values, sufficient for our purposes here considering that much larger uncertainty lies with assumed material properties and the resulting size distribution.
Radiation forces on dust {#ss:rad}
------------------------
To derive the surface area in small grains (dominated by small objects) from the total mass (dominated by large objects) we need to know the size of the smallest grains that can survive in circumplanetary orbits. Two notable detections are micron-sized grains found orbiting at large (50–350[$R_{\rm {Jup}}$]{}) distances from Jupiter , and the large tenuous ring found orbiting Saturn [@2009Natur.461.1098V]. Both studies attribute material released by impacts from interplanetary grains as a likely source, though @2009Natur.461.1098V note that debris from irregular satellite collisions impacting Phoebe could also be the cause.
As with grains orbiting a star, the effect of radiation forces on dust characterised by $\beta = F_{\rm radiation}/F_{\rm gravity}$ (both due to the star) plays the most important role in setting the minimum size of grains that survive on circumplanetary orbits [@1979Icar...40....1B]. Other effects related to interaction with planetary magnetospheres play some part but are less important for the $\gtrsim$[$\mu$m]{} grain sizes and wide orbits considered here.
### Radiation pressure
Radiation pressure is the radial component of the force, which in contrast to circumstellar orbits causes the orbits of dust grains to evolve. While semi-major axes remain constant, eccentricities oscillate with a period equal to the planet’s orbital period, with a maximum that depends on $\beta$ and the grain orbit. The maximum $\beta$ of grains that survive in orbit around the planet with $e < 1$ have [@1979Icar...40....1B] $$\label{eq:betac}
\beta_{\rm c} = v / 3 v_\odot =
5.8 \times 10^{-3} M_{\rm pl}^{1/3} / ( M_\star^{1/3} \eta^{1/2} ) \, ,$$ where $v$ is the velocity of a grain as it orbits the planet, and $v_\odot$ is the velocity of the planet as it orbits the star. For typical planets and irregular satellite orbits $\beta_{\rm c}$ is much smaller than the blowout limit of 0.5 for stellocentric orbits.
Because $\beta$ for normal grains peaks where the star radiates most of its radiation, it might be possible for $\beta_{\rm c}$ to allow both large and very small grains to survive, with only grains in the peak being excluded. However, for sub-micron “astronomical silicate” grains, those on the small side of the peak, the smallest grains have $\beta \sim 0.11$ [e.g. @1994AREPS..22..553G] and is higher for more massive stars, so grains smaller than the wavelength of typical stellar radiation will usually be ejected.
To convert $\beta_{\rm c}$ into a minimum size, we use $D_{\rm min} = (1150/\rho
\beta_{\rm c})L_\star/M_\star$ in $\mu$m [e.g. @2008arXiv0807.1272W]. The minimum size is therefore $$\label{eq:dmin}
D_{\rm min} = 2 \times 10^5 \frac{ \eta^{1/2} \, L_\star }{ \rho \, M_{\rm pl}^{1/3} \, M_\star^{2/3}}$$ in [$\mu$m]{}. For typical parameters, [$D_{\rm min}$]{} is at least [$\mu$m]{}-size. For Jupiter and Neptune, Equation (\[eq:dmin\]) yields 12 and 23[$\mu$m]{}, an order of magnitude larger than the minimum (blowout) size for the same grains on circumsolar orbits.
### Poynting-Robertson drag
An alternative to grain removal by radiation pressure is orbital decay due to Poynting-Robertson (PR) drag. The decay timescale is similar to the heliocentric case $$\label{eq:tpr}
t_{\rm PR} = 530 \, a_{\rm pl}^2 / ( \beta \, M_\star )$$ in years [@1979Icar...40....1B]. However, as noted above $\beta$ for the smallest grains is typically much smaller than 0.5 and $t_{\rm PR}$ correspondingly longer.
For grains to spiral into the planet by PR drag they must avoid colliding with other grains first, which breaks them into smaller particles that are instead removed by radiation pressure. The competition between PR drag and collisions can be characterised by $\chi_{\rm PR} = t_{\rm PR} / t_{\rm col}$. @1999ApJ...527..918W showed that for the smallest grains this collision rate is roughly $t_{\rm per} r_{\rm dust}^2 / (4\,
\sigma_{\rm tot})$ for a flat disk with radial extent $r_{\rm dust} \pm r_{\rm dust}/4$ (where $t_{\rm per}$ is orbital period). Adapting this expression to an isotropic case results in a small change due to the greater cloud volume and faster collision velocities: $t_{\rm col,dust} = t_{\rm per} r_{\rm dust}^2 / (4 \, f_{v_{\rm rel}}
\sigma_{\rm tot})$, or using our parameters $$\label{eq:tcoldust}
t_{\rm col,dust} = 10^{-5} \, \frac{ \left( \eta \, a_{\rm pl} \right)^{7/2} M_{\rm
pl}^{2/3} } { M_\star^{7/6} \, \sigma_{\rm tot} }$$ in years. Expressed in terms of our basic parameters, the ratio is $$\label{eq:chipr}
\chi_{\rm PR} = 4 \times 10^{4} \, \frac{ \rho \, D_{\rm min} \, \sigma_{\rm tot} \, M_\star^{7/6} }
{ a_{\rm pl}^{3/2} \, \eta^{7/2} \, M_{\rm pl}^{2/3} \, L_\star }$$ When this ratio is larger than unity, grains suffer collisions before their orbits have time to decay due to PR drag, and are subsequently removed by radiation pressure (termed “collision dominated”). When this ratio drops below unity, grains spiral in towards the planet before they collide and may encounter any existing regular satellites as they do so (“PR dominated”). Substituting Equation (\[eq:dmin\]) for [$D_{\rm min}$]{} yields $$\label{eq:chipr2}
\chi_{\rm PR} = 8 \times 10^9 \, \frac{ \sigma_{\rm tot} \, M_\star^{1/2} }
{ a_{\rm pl}^{3/2} \, \eta^3 \, M_{\rm pl} } \, .$$ With only the Solar System example to go by (see §\[ss:ssdust\]), it is impossible to tell whether the “typical” satellite dust cloud will be collision or PR dominated. However, due to the amount of dust needed for an extrasolar irregular satellite swarm to be detectable ($\sigma_{\rm tot} \gtrsim 10^{-4}$AU$^2$), any observed extrasolar swarms will likely be collision dominated.
### Summary
In order to estimate the minimum grain size we have necessarily made a number of simplifications. We have used expressions for radiation forces assuming low inclination orbits akin to Saturn’s ring. In fact $\beta_{\rm c}$ varies with the orientation of the orbit relative to the Solar direction and is smallest for those with pericenter initially aligned with the Solar direction. Equation (\[eq:betac\]) assumes a coplanar orbit with pericenter perpendicular to the Solar direction and underestimates $\beta_{\rm c}$ for certain loss by a factor 2–3 [@1979Icar...40....1B]. We therefore overestimate the minimum grain size, and underestimate [$\sigma_{\rm tot}$]{}.
Our prescription for [$D_{\rm min}$]{} does not preclude detection of grains smaller than [$D_{\rm min}$]{}. Grains below the minimum size may complete many orbits before their eccentricity exceeds unity. Also, the minimum grain size decreases with $\eta$ so small bound eccentric grains with smaller planetocentric semi-major axes can be found anywhere within $2\eta$. On a detailed level, [$D_{\rm min}$]{} takes on a range of values and for the smallest grains with $\beta \sim \beta_{\rm c}$, the orbit averaging used in deriving Equation (\[eq:betac\]) breaks down. Using numerical simulations, @2002Icar..157..436K find that [$D_{\rm min}$]{} for Jupiter is $\sim$1[$\mu$m]{} (compared to our value of 12.4[$\mu$m]{}, see Table \[t:ssdust\]). Considering that our value is overestimated by the factor 2-3 noted above, the minimum size is probably 1 to a few [$\mu$m]{}, with differences in the assumed grain properties contributing some uncertainty.
While our simplifications are reasonable, they gloss over important aspects of grain dynamics. For example, @2002Icar..157..436K show circumplanetary dust clouds to have both size-dependent and pro/retrograde orbit sensitive structure. Because non-gravitational forces cause small objects to deviate from the orbits of their parent bodies, such forces lead to effects beyond the scope of our model. Therefore, future work will need to consider how grain orbital evolution affects both the physical appearance of the cloud, and the underlying size distribution. For example, in a typical collision dominated circumstellar disk the minimum (blowout) grain size is a single number, independent of radial distance. This fairly sharp truncation creates a wave in the size distribution due to the alternating lack and then over-abundance of projectiles that destroy larger targets . However, in a circumplanetary cloud the minimum grain size varies with circumplanetary distance (Eq. \[eq:dmin\]), which could lead to qualitatively different and spatially varying size distributions. In §\[ss:ssdisc\] we suggest that Jupiter’s swarm may be PR dominated, which can lead to further differences at the small end of the size distribution. In addition, other non-gravitational forces known to be important for circumstellar dynamics [e.g. Yarkovsky effect, @1979Icar...40....1B] may be important for the evolution of circumplanetary orbits and consequent collisional evolution.
Time evolution {#ss:evol}
--------------
We now turn to the time evolution of disk properties. Assuming that the size distribution remains fixed and that mass is lost by catastrophic disruption of the largest objects, the disk mass remaining as a function of time is found by solving [e.g. @2007ApJ...658..569W] $$\label{eq:dmdt}
\frac{dM_{\rm tot}}{dt} = -M_{\rm tot} \, R_{\rm cc} \, ,$$ which yields $$\label{eq:mtime}
M_{\rm tot}(t) = \frac{ M_{\rm tot}(0) }{ 1 + R_{\rm cc}(0) \, t }$$ where we use $t$ or 0 to indicate variables at a particular time where necessary. The initial cloud mass is $M_{\rm tot}(0)$ and $M_{\rm tot}(t) = M_{\rm tot}(0) / R_{\rm
cc}(0) \, t$ when $R_{\rm cc}(0)\,t \gg 1$. Equation (\[eq:mtot\]) shows that mass and surface area are linearly proportional for fixed [$D_{\rm c}$]{} and [$D_{\rm min}$]{} so the surface area of small grains decays in the same way. Because the collision rate depends on the remaining mass, after a few collision times the remaining mass is independent of the initial mass [@2007ApJ...658..569W].
### Where are the most massive swarms? {#ss:aopt}
There is an interesting interplay between the initial cloud mass and planet semi-major axis. For arbitrarily large initial masses, swarms at any distance and time are on the $1/t$ part of their evolution where the remaining mass is independent of the initial mass [called a “collision limited” disk by @2010MNRAS.401..867H]. Because these systems have a maximum remaining mass that depends on the collision time ($1/R_{\rm cc}$) without the mass term, the remaining mass increases strongly with planet semi-major axis. In fact, $t_{\rm c} \propto a_{\rm pl}^{4.13}$ (i.e. more strongly than $a_{\rm
pl}^2$), so collision limited swarms are brighter in scattered light for larger planet semi-major axes (for fixed planet mass, age, etc.).
Of course, the initial swarm mass cannot be arbitrarily large, so swarms around sufficiently distant planets will take some time to start colliding. Thus, for fixed initial mass, swarms around close planets will rapidly decay due to the short collision times, while more distant planets all have the same mass in satellites because the largest irregulars have not yet, or only just begun to collide. Of these more distant planets, the one whose swarm has just started to suffer collisions is the one that receives the most stellar insolation and is brightest in scattered light.
The semi-major axis of the planet hosting this swarm can easily be worked out from the collision rate in equation (\[eq:rccdc\]), because a swarm that has just started to collide has $t = 1/R_{\rm cc}$, therefore $$\label{eq:aopt}
a_{\rm opt} = 50 \, \frac{ M_\star^{0.33} \, f_{v_{\rm rel}}^{0.55} }
{ M_{\rm pl}^{0.06} \, {Q_{\rm D}^\star}^{0.15} \, \eta }
\left( \frac{ t \, M_{\rm tot}(0) }{ \rho \, D_{\rm c} } \right)^{0.24}$$ in AU. The “opt” subscript indicates that this planet lies at the optimum distance to be detected at “opt”-ical wavelengths. This prescription for the brightest swarm is complicated by the fact that faint objects are harder to detect close to host stars, an issue we return to in §\[ss:exo\]. The same concept of an optimal distance applies to thermal emission, but is somewhat different because the cloud temperature changes with planet semi-major axis.
### Stranding the largest objects {#ss:stranding}
If collisional evolution proceeds for long enough, the mass in the swarm will drop to the point where it is comparable with the mass contained in a single largest object. Around this time the evolution of the largest objects changes from being reasonably well described by our particle-in-a-box formalism, to a regime where individual collisions and cratering are important [@2010AJ....139..994B]. In this regime, the largest objects are less likely to be destroyed due to their small number and the decreased number of potential destructors. While our model cannot take cratering or stochastic collisions into account, we can approximate the evolution by assuming that when the number of largest objects drops too low, they lose their connection with the rest of the size distribution and become “stranded.” The size of the largest non-stranded object is of size [$D_{\rm c}$]{} by definition, which decreases over time and leaves a relatively flat size distribution of stranded objects between sizes [$D_{\rm c}$]{} and [$D_{\rm max}$]{}, whose evolution is halted due to a lack of would-be destructors.
We implement this simple approximation by assuming an object of size $D$ is stranded when the number of objects between sizes $D/2$ and $D$ $$\label{eq:nd2}
n(D/2 \rightarrow D,t) = \frac{ K \left( 2^{3q_{\rm g}-3} -1 \right)
( 10^3 \, D )^{3-3q_{\rm g}} }{ 3 q_{\rm g} - 3 }$$ drops to some number $N_{\rm str}$. This number sets the normalisation for the size distribution of stranded objects. This assumption sets the size distribution slope of stranded objects $q_{\rm str} = 1$ because $n(D/2 \rightarrow D,t)$ is independent of $D$ for $D_{\rm c} > D > D_{\rm max}$.
The first object is stranded at [$t_{\rm nleft}$]{}, which can be calculated from the initial number of objects in this size range $$\label{eq:tnl}
t_{\rm nleft} = \frac{ n(D_{\rm c}/2 \rightarrow D_{\rm c},t=0) }{ R_{\rm cc}(0) \, N_{\rm str} }$$ The difference between $t_{\rm col}$ and [$t_{\rm nleft}$]{} is simply a measure of the initial number $n$ of large objects, which take $n/N_{\rm str}$ collision times to become stranded. Like the remaining mass once collisions occur, this time is independent of the initial cloud mass.
At this point, [$D_{\rm c}$]{} and [$D_{\rm max}$]{} become distinct sizes, with [$D_{\rm max}$]{} remaining fixed and [$D_{\rm c}$]{} decreasing with time as smaller objects are stranded. The remaining planetesimal population decays at the collision rate for [$D_{\rm c}$]{} size and strength objects and the mass remaining in the size distribution below [$D_{\rm c}$]{}. Therefore, by substituting [$D_{\rm c}$]{} for [$M_{\rm tot}$]{}in equation (\[eq:dmdt\]), the evolution after [$t_{\rm nleft}$]{} obeys $$\label{eq:dcdt}
\frac{dD_{\rm c}^3}{dt} = - D_{\rm c}^3 \, R_{\rm cc}$$ where we have used $M_{\rm tot} \propto D_{\rm c}^3$ (for fixed [$n(D_{\rm c}/2 \rightarrow D_{\rm c})$]{}). The collision rate is for size [$D_{\rm c}$]{} objects, on which both [$M_{\rm tot}$]{} and [$Q_{\rm D}^\star$]{} depend, and varies as $R_{\rm cc}
\propto D_{\rm c}^{1.2}$ (eqs. \[eq:qd\] & \[eq:rccdc\]). Integrating equation (\[eq:dcdt\]) yields $$\label{eq:tnlevol}
D_{\rm c} = \frac{ D_{\rm max} }{ \left( 1 + 0.4 (t - t_{\rm nleft}) / t_{\rm nleft}
\right)^\alpha } \, .$$ where $\alpha = 1/1.2$. This evolution is illustrated in Figure \[fig:strandeg\], which shows the time evolution of the top section of the size distribution and [$\sigma_{\rm tot}$]{}. The size distribution initially decays straight down (i.e. $K$ decreases) with [$D_{\rm c}$]{} fixed. When there are only $N_{\rm str}$ largest objects left, [$D_{\rm c}$]{} begins to decay as dictated by Equation (\[eq:tnlevol\]). The size distribution then moves to the left (smaller [$D_{\rm c}$]{}), with both [$\sigma_{\rm tot}$]{} and [$M_{\rm tot}$]{} continuing to decay. The right panel shows that [$\sigma_{\rm tot}$]{} drops more quickly after the first objects are stranded, tending to $\sigma_{\rm tot} \propto
t^{-1.75}$ (using eqs. \[eq:mtot\] and \[eq:tnlevol\]). Such evolution is potentially interesting, as a dust cloud is accelerated towards being PR-dominated after stranding due to the stronger decrease in small grains with time. Entering the PR-dominated regime, the size distribution is effectively truncated at larger sizes, leading to an even faster decay of [$\sigma_{\rm tot}$]{} than shown in Figure \[fig:strandeg\] [@2003ApJ...598..626D].
This evolution is necessarily very simple because objects are only destroyed by catastrophic collisions in our model. However, at these late stages the mass released into the cascade by cratering may be as or more important [@2010AJ....139..994B]. Therefore, though our model is physically plausible, the actual evolution will depend on details such as the relative importance of catastrophic disruptions vs. cratering or differences between pro and retrograde populations. We treat $N_{\rm str}$ and $\alpha$ as free parameters when comparing our model with the Solar System irregulars in §\[ss:ss\].
Observable quantities {#ss:obs}
---------------------
Given the surface area in dust, we derive the flux density $F$ due to the cloud and planet from both scattered light and thermal emission. Because distance and surface area can be in different units (e.g. m, pc, AU), quantities in these equations have dimensions. We take Solar System planetary effective temperatures and radii from @2000asqu.book.....C. The stellar flux at the planet is $$F_\star = L_\star \, B_\nu(\lambda,T_\star) / ( 4 \, \sigma_{\rm K} \,
T_\star^4 \, a_{\rm pl}^2 )$$ where $\sigma_{\rm K}$ is the Stefan-Boltzmann constant. The scattered light seen from Earth is [e.g. @2002MNRAS.330..187C] $$F_{\rm scat} = F_\star \, A \, g \, Q / (\pi \, d_{\rm pl}^2)$$ where $A$ is either the projected area of the planet or [$\sigma_{\rm tot}$]{} for dust. The geometric albedo $Q$ is assumed to be 0.08 for dust, similar to both the values for Kuiper belt objects, Jovian trojans, and irregular satellites in the Solar System [@2008ssbn.book..161S; @2008DPS....40.6108M; @2009AJ....138..240F] and to that inferred for the Fomalhaut dust ring [@2005Natur.435.1067K]. For planets we use an albedo of 0.5. The phase function $g$ is set to unity for the Solar System because planets are on exterior orbits. For extrasolar planets and swarms we set $g=0.32$, the value for a Lambert sphere at maximum extension from the host star [e.g. @2002MNRAS.330..187C]. We use a blackbody estimate for the dust temperature $T_{\rm dust} = 278.3 \, L_\star^{1/4} / a_{\rm pl}^{1/2}$K. The distance $d_{\rm pl}$ is the distance to the system in the extrasolar case, or $a_{\rm pl}$ in the Solar System (i.e. roughly the distance for a Sun-Earth-Planet angle of 90$^\circ$ for outer planets). For thermal emission $$F_{\rm th} = B_\nu(\lambda,T) \, A / d_{\rm pl}^2 \, .$$ For dust we include a “greybody” decrease in emission of $210/\lambda$ beyond 210[$\mu$m]{}to account for inefficient emission from small grains . Though the actual spectrum depends on the grain properties and size distribution, this addition provides more realistic (sub-)mm flux densities than a plain blackbody.
Applications {#s:application}
============
In this section we apply our model to three different irregular satellite populations. We first compare our model evolution with the Solar System’s complement of irregulars and make predictions of the current levels of dust. We suggest that the fate of dust changes over time, with implications for regular satellites such as Callisto and Iapetus. We then apply our model to possible satellite clouds around extrasolar planets. Finally, we apply our model to Fomalhaut b.
Solar System Irregular Satellites {#ss:ss}
---------------------------------
In this section we compare our model with the Solar System’s irregular satellites. We model a swarm with initial mass $M_{\rm tot}(0) = 0.01\,M_{\rm moon}$. This mass ensures all swarms are collision limited and therefore does not influence the model. We set $D_{\rm c} = 150$km for Jupiter and Uranus and $D_{\rm c} = 250$km for Saturn and Neptune. We set the density similar to known irregular satellites $\rho = 1500$kg m$^{-3}$ . Though we assume $f_{v_{\rm rel}} = 4/\pi$ at each planet, this number varies somewhat depending on the specific orbital properties of each swarm. We use the average $\eta$ of known irregulars at each planet.
Figure \[fig:ssirreg\] shows a comparison of our model with the Solar System’s irregular satellites.[^2] We have combined the pro/retrograde satellites for this distribution. The simplicity of our model means that we cannot account for the different evolution of pro/retrograde populations, which appears to be important at Uranus [@2010AJ....139..994B]. Because all satellites at each planet are at a similar distance from Earth, these distributions are near-complete to the smallest observed size and need no correction [e.g. @2005AJ....129..518S].
We vary the normalisation $N_{\rm str}$ and rate of decay after stranding power-law index $\alpha$ to obtain a by-eye fit. Variation of these parameters over a fairly small range allows an excellent match for Jupiter, Saturn, and Uranus. Given the differences in the irregular populations at each planet and the simplicity of the prescription for stranding, some variation is expected. The first few columns of Table \[t:ssdust\] show the $\eta$, $t_{\rm col}$, and [$t_{\rm nleft}$]{} for each planet. The slower collisional evolution for more distant planets means that [$t_{\rm nleft}$]{} is longer and that the size distribution at the current epoch turns up (i.e is stranded) at larger sizes. The minimum known satellite size increases for more distant planets, making the model comparison less certain for Uranus and Neptune. We do not know if their size distributions are similar to Jupiter and relatively flat to $\sim$8km, or instead turn up at larger sizes like our model. If evolution at Uranus is similar to Jupiter and Saturn, our model predicts that the size distribution should turn up at a few tens of km due to the slower collisional evolution.
There are few Neptune irregulars to compare with, perhaps because many were depleted by Nereid [@2003AJ....126..398N] and/or Triton [@1989Sci...245..500G; @2005ApJ...626L.113C]. Nereid’s low orbital inclination (7$^\circ$ relative to the ecliptic) and relatively close prograde orbit ($\eta = 0.05$), as well as a colour and albedo similar to the Uranian satellites Umbriel and Oberon [@1997Icar..126..225B], mean that it may in fact have formed as a regular satellite of Neptune. For these reasons @2010AJ....139..994B did not model Neptunian irregulars.
Given these complications, it is perhaps unsurprising that our model for Neptune’s irregulars needs very different [$N_{\rm str}$]{} and $\alpha$ to the other giant planets, and is still a poor match. This difference is due to the slow collisional evolution, which predicts that [$t_{\rm nleft}$]{} is a sizeable fraction of the Solar System age. With stranding occurring at such a late time, the subsequent evolution must be very rapid to deplete the population to that currently observed. With values for [$N_{\rm str}$]{} and $\alpha$ more like the other three planets, our model predicts that Neptune would have two orders of magnitude more satellites (i.e. be at about the second to lowest curve). This discrepancy suggests that either the initial conditions for Neptune’s swarm were quite different to the other planets, or as already proposed, the irregulars were affected by Nereid and Triton.
Based on these comparisons we conclude that our model provides a reasonable description of the collisional evolution of an irregular satellite swarm.
### Current irregular dust levels {#ss:ssdust}
Given the ongoing collisional erosion of the irregular satellites, the presence of dust is inevitable. Here we make some estimates of the expected level of dust at each planet and relate them to a few relevant observations. Table \[t:ssdust\] shows estimates for each planet based on [$\sigma_{\rm tot}$]{} from the models of Figure \[fig:ssirreg\] at $t = 4.5
\times 10^9$years. That is, they are extrapolated using our size distribution and independent of the collisional model. We calculate the flux density and surface brightness at 1 and 100[$\mu$m]{} (distributed uniformly over a disk of radius 0.5[$R_{\rm H}$]{} for simplicity), which roughly correspond to peaks in scattered light and thermal emission respectively.
--------- ----------------- -------- --------------- --------------------- ------------------- ------------------------ ------------- --------------------- ---------------- -------------- ---------- ---------------- -------------- ----------
[$\sigma_{\rm tot}$]{}
[$R_{\rm H}$]{} $\eta$ $t_{\rm col}$ [$t_{\rm nleft}$]{} [$D_{\rm min}$]{} ($10^{-9}$ dens [$\chi_{\rm PR}$]{} $F_{\rm dust}$ $F_{\rm pl}$ $B$ $F_{\rm dust}$ $F_{\rm pl}$ $B$
Planet (AU) (Myr) (Myr) ([$\mu$m]{}) AU$^2$) (km$^{-3}$) (Jy) (Jy) (MJy/sr) (Jy) (Jy) (MJy/sr)
Jupiter 0.35 0.4 2.5 170 12 1.5 0.9 0.05 12.0 37000 0.0034 960 450000 0.26
Saturn 0.43 0.3 17.0 740 16 7.5 1.5 0.8 5.7 2400 0.0035 840 60000 0.52
Uranus 0.46 0.2 14.0 1100 24 16.0 1.1 13.0 0.8 28 0.0016 210 950 0.45
Neptune 0.77 0.2 250.0 3500 23 36.0 0.6 13.0 0.3 4 0.0005 99 350 0.19
--------- ----------------- -------- --------------- --------------------- ------------------- ------------------------ ------------- --------------------- ---------------- -------------- ---------- ---------------- -------------- ----------
The predicted surface brightness levels are much fainter than the background in the ecliptic, but expected to vary on a similar scale (i.e. degrees). For comparison, the Zodiacal background at 1.25 and 100[$\mu$m]{} is about 0.4 and 9MJy/sr respectively [@1998ApJ...508...44K]. Detection of these clouds therefore requires accurate subtraction of this (and the galactic and cosmic) background.
Of course dust that originates from irregular satellites has been detected already (see also §\[ss:ssdisc\] below). The largest of Saturn’s rings is probably fed by material generated when inter/circumplanetary grains impact the irregular satellite Phoebe [@2009Natur.461.1098V]. At 24[$\mu$m]{}, this ring has a surface brightness of $\sim$$0.4$MJy sr$^{-1}$, which is less than 1% of the ($\sim$70MJy/sr) zodiacal background. For comparison, our Saturn swarm model has $\sim$0.2MJy sr$^{-1}$ when spread over a 0.5[$R_{\rm H}$]{} radius disk.
To take a more detailed look at Saturn, Figure \[fig:saturnpic\] shows an example of what such a cloud might look like from Earth at 24[$\mu$m]{}. In creating this image we have assumed that the dust follows the same orbits as the parent bodies, which have semi-major axes distributed between 0.04–0.16AU, and $e = 0.1$–0.6 based on Figure 7 of @2007AJ....133.1962N. The image would therefore look the same at any wavelength, though the total flux changes. The total flux in this 24[$\mu$m]{} image is 320 Jy, which yields a peak surface brightness of around 1.25 MJy/sr. Therefore, the surface brightness predictions in Table \[t:ssdust\] are a factor $\sim$6 too low if grains follow their parent body orbits. The peak level is about three times brighter than Saturn’s Phoebe ring.
Though the cloud is spread over a much larger region of sky, for it to have evaded detection thus far, particularly in the Phoebe ring observations, it appears that our prediction is at least a few times too high. However, @2009Natur.461.1098V note an apparent trend of increasing surface brightness towards the Phoebe ring, which could be due to dust from irregular satellites. The narrow vertical range of their Figure 3 does not constrain our prediction, because the Phoebe ring could be sitting on top of a larger background.
Several uncertainties with our model may explain this apparently high prediction for the level of dust at Saturn, the most likely being that small departures from our assumed size distribution can lead to large differences in the predicted surface area in dust. In addition, we have assumed that grains follow the orbits of their parent bodies, but @2002Icar..157..436K show that the detailed cloud structure is more complex. Finally, Table \[t:ssdust\] shows that grains around Saturn may be in the PR-dominated regime, which can alter the cloud structure and reduce the dust level.
Despite our apparent over-prediction of the cloud surface brightness, detection of such a cloud is difficult, not least because it may be unexpected. The large extent means achieving sufficient coverage (ideally the entire Hill sphere) is expensive at IR wavelengths (i.e. using space telescopes). Poorly characterised dust bands will also hinder background subtraction. At longer wavelengths, such as those covered by *Herschel* SPIRE, achieving decent coverage is easier, but the background comprises a combination of Zodiacal and galactic light. Care would be needed to ensure good background subtraction that minimises elongation effects.
Now looking at Jupiter, from Galileo dust detection data @2002Icar..157..436K derive a constant dust number density of $\sim$10km$^{-3}$ between 50 and 350[$R_{\rm {Jup}}$]{}. report detection of micron-size particles (and note the lack of detections of smaller particles) by the Galileo dust detector at a distance of approximately 350[$R_{\rm {Jup}}$]{}. Such large planetocentric distances are prime irregular satellite territory, being about half the Hill radius. These grains have previously been explained as ejecta from impacts of interplanetary dust on irregular satellites, which produce about the same level of dust as detected by Galileo [@2002Icar..157..436K]. The predicted space density of dust for Jupiter shown in Table \[t:ssdust\] (again within 0.5[$R_{\rm H}$]{}) is at a lower level to the Galileo detections and may not have contributed to the measurements. However, the space density is an extremely strong function of [$D_{\rm min}$]{}. For example, if the minimum grain size in our model were 1[$\mu$m]{} the predicted space density at Jupiter would be $\sim$100km$^{-3}$. Therefore, Galileo observations constrain either the grain size in our model to be larger than a few microns, or the dust level to be lower than predicted if the minimum grain size is 1[$\mu$m]{}.
### Fate of Dust {#ss:ssdisc}
The best way to probe the small end of the irregular satellite size distribution is to detect the dust cloud directly. However, as noted above such an observation is difficult and there are complementary ways to detect irregular dust. One of the most interesting signatures exists on the surfaces of some regular satellites. The orbits of the smallest grains are strongly affected by radiation forces and may end up on a regular satellite. However, because this deposition is only one of several possible fates a grain may meet, an understanding of which is more likely (and when) is needed to make a strong connection between irregular dust and regular satellite surfaces. Our model is too simple to model irregular satellite evolution in the Solar System at a detailed level. However, we offer some order of magnitude arguments concerning the fate of dust, which highlight questions that should be asked by more detailed studies.
Like a circumstellar disk, a young circumplanetary swarm will most likely be collision dominated. At this stage grains are lost to the planet and interplanetary space. When the mass has been sufficiently depleted it becomes PR dominated and grains spiral in toward the planet. Grains destined to impact the planet, or pass nearby, may meet a third fate and be swept up by a regular satellites. Though only a small fraction of mass may be lost this way, it is important due to the visible effect of leading/trailing asymmetries on some tidally locked Solar System regular satellites, such as Callisto and Iapetus. These asymmetries are thought to arise from the higher accretion rate of retrograde dust by the leading hemisphere [e.g. @1996ASPC..104..179B].
Though most grains in collision dominated swarms either leave the Hill sphere or impact the planet due to radiation pressure, nearly all must pass through the regular satellite domain to do so. Grains destined to hit the planet or leave the Hill sphere do not reach their maximum eccentricity immediately, but instead make a number of pericenter passages first. The smallest grains with $\eta \sim 0.1$–0.5 only complete a few, to a few tens of orbits before their eccentricities exceed unity (i.e. $t_{\rm per,grain} \sim
\eta^{3/2} \, t_{\rm per,planet}$) and are therefore unlikely to encounter regular satellites before removal.
Grains in the small size range where $e$ grows large enough to pass through the regular satellite region but not high enough to hit the planet or leave the Hill sphere have the best chance of colliding with regular satellites. However, this collision time must be shorter than the time for grains to collide with themselves. Taking Saturn as an example, at $\eta=0.3$ with $\sigma_{\rm tot} = 7.5 \times 10^{-9}$AU$^2$ (Table \[t:ssdust\]) the current time for collisions between the smallest grains is predicted to be $\sim$10$^6$years (eq. \[eq:tcoldust\]). In the past the level of dust was much higher, and the grain-grain collision time correspondingly shorter.
Continuing with the Saturn example, grains with $0.83 > e > 0.997$ traverse the region between Iapetus’ orbit and Saturn’s surface near pericenter. This range corresponds to grains of sizes $16 > D > 25$[$\mu$m]{}. The orbital period $t_{\rm per}$ is about 3years for these grains. Grains on coplanar orbits with Iapetus in this size range spend no more than 40% of their lifetimes in Iapetus-crossing orbits. Iapetus occupies about 1/20000th of its orbital torus, so coplanar grains have a collision time of $\sim$$10^4$ orbits, or $\sim$$10^4$ years. This time is shorter than the grain-grain collision time, so coplanar grains in this size range appear likely to impact Iapetus.
However, grains on inclined orbits (i.e. the majority) only have a chance of hitting Iapetus if they cross Iapetus’ plane at precisely the right radial distance. Iapetus is $\sim$1500km across, compared to a semi-major axis of $3.5 \times 10^6$km, so the chances of a plane-crossing particle encountering Iapetus *orbit* are roughly 1/2000. The chance of an inclined grain impacting Iapetus at each plane crossing is $\sim$10$^{-7}$ per orbit, or an impact time of $\sim$10$^{7}$ years. The Iapetus collision time is therefore longer than the grain-grain collision time, so grains are more likely fragmented to smaller sizes first. Taking these numbers as representative, it appears that in collision dominated swarms only a small fraction of grains, those within a small size range and on coplanar orbits, will impact regular satellites.
In contrast to the highly variable eccentricities of radiation pressure affected grains, PR drag causes grain orbits to collapse slowly. The chance of impacting a regular satellite is larger than for the more distant and eccentric radiation pressure induced orbits because grains orbit the planet more often as their semi-major axes shrink. For example, in the case of grains released from Phoebe, collision with Iapetus is nearly guaranteed, with escapees destined to hit Hyperion or Titan [e.g. @1996ASPC..104..179B].
Table \[t:ssdust\] shows that dust from irregular satellites at Jupiter and Saturn is near or in the regime where PR drag dominates, but dust at Uranus and Neptune is not. If grains in PR dominated swarms are more likely to impact regular satellites, this finding is consistent with the brightness asymmetries seen on Callisto and Iapetus, which are less marked (but still present) on the Uranian satellites.
While we suggest that the Jupiter and Saturn swarms are PR dominated now, they were collision dominated in the past. The possibility that collision dominated swarms do not coat regular satellites as efficiently as PR dominated ones therefore has implications for the interpretation of the brightness asymmetries. If the mass deposited on regular satellites is some fraction of the total mass lost, this fraction will increase as the swarm changes from collision to PR dominated. This change will affect the accretion history and more mass may be accreted at late stages if the difference in accretion efficiency is large enough.
Such interpretations may be too simple. For example, Iapetus’ asymmetry is likely due to accretion of grains from Saturn’s Phoebe ring, so is due to grains released from an individual irregular satellite rather than the cloud in general. In summary, further study of the fate of irregular satellite debris and potential observables requires consideration of the competition between collisions, PR drag, and radiation pressure.
Irregular Satellite Clouds around Exoplanets {#ss:exo}
--------------------------------------------
While the dust produced by Solar System irregular satellites is at very low levels, these clouds were many times brighter at earlier times. In this section we explore the prospects for discovery of young extrasolar circumplanetary swarms. Like circumstellar debris disks, dust can be discovered in scattered light at optical wavelengths, and thermal emission at IR wavelengths. In the case of thermal emission, one may need to distinguish a planetary atmosphere from dust at a similar temperature via spectral features.
Because there is little *a priori* reason to choose any particular system configuration, we first model a particular system and then show which configurations are detectable in a more general sense. Clearly, to be detectable with current or near-future technology these systems must possess more dust than predicted for the Solar System’s irregulars. Further, because the first objects discovered where sensitivity is an issue tend to be the most extreme (e.g. hot Jupiters), we do not restrict the initial swarm masses. While the initial conditions used by @2010AJ....139..994B based on the @2007AJ....133.1962N simulations were at most a small fraction of the Moon’s mass, extrasolar systems may have much more massive planetesimal belts from which to capture satellites, and mechanisms disfavoured for the Solar System may also operate. We discuss capture mechanisms further in §\[ss:exodisc\].
In this section we first use the example of a Jupiter-mass planet orbiting an A5 star at 10pc. We choose this spectral type of star simply because the only planets directly imaged to date orbit A-stars . However, more massive stars do not necessarily produce brighter clouds because their high luminosity and mass increases [$D_{\rm min}$]{} and changes the collision rate. For the evolution of the planet, we use the non-irradiated Jupiter-mass ($Z=0.02$) model from . We set $M_{\rm tot}(0) = 10\,M_{\rm moon}$, and $D_{\rm c} =
250$km. Another choice is when the satellite swarm is captured (or otherwise formed) and evolution begins. For simplicity, we assume swarms start evolving at $t=0$.
Figure \[fig:egpspec\] shows the evolution of scattered light and thermal emission from both planet and dust cloud when the planet orbits at 50AU. The Figure is drawn to highlight where the spectrum of each component moves over time (which is downwards in this plot). In scattered light, the dust cloud is initially much brighter than the planet, but decreases significantly as the satellite swarm collides and the dust level subsides. The planet shrinks somewhat as it cools, but in scattered light is nearly constant. At early times the planet’s thermal emission peaks slightly short of 10[$\mu$m]{}, but moves to longer wavelengths over time. Thermal emission from the dust beyond 20[$\mu$m]{}is brighter than the planet at all times. They become comparable at mm wavelengths at late times, which shows the importance of including grain emission inefficiencies.
\
This evolution therefore highlights wavelengths where irregular satellite clouds may be bright enough to be detected, both in absolute terms and relative to thermal and scattered emission from the planet. Unsurprisingly these lie at the thermal and stellar peaks. The key to detection lies with rejection of starlight, the same issue faced by those looking for planets or circumstellar debris disks [e.g. @2007prpl.conf..915B; @2010PASP..122..162B]. This rejection is characterised by the star/planet contrast ratio that a particular instrument can detect, which is usually a function of angular separation between the two. While Figure \[fig:egpspec\] shows that the youngest dust clouds may be bright enough for detection by the *Herschel* Photodetector Array Camera and Spectrometer (PACS) at 100[$\mu$m]{}, the 7” resolution means that resolving these systems is difficult. We therefore focus on optical and near/mid-IR wavelengths. We model the detectability of satellite swarms using simple approximations to published contrast ratios for several instruments.
For the *Hubble Space Telescope Advanced Camera for Surveys* (HST ACS) the contrast is based on actual roll-subtracted coronagraphic observations [@Krist2006] and we set an absolute detection limit of 0.1$\mu$Jy. We use predicted *James Webb Space Telescope* (JWST) Near IR Camera (NIRCAM) 4.4[$\mu$m]{} and Mid IR Imager (MIRI) 11.4 contrast ratios from @2010PASP..122..162B. For MIRI at 23[$\mu$m]{}, we use the same contrast as at 11.4[$\mu$m]{} for the same $\lambda/D$ [i.e. at twice the separation, @2004EAS....12..195B]. The other difference from the three shorter MIRI coronagraph wavelengths is that only a Lyot stop is offered at 23[$\mu$m]{} so the inner working angle is larger. We set absolute detection limits of 68nJy, 2.5$\mu$Jy, and 50$\mu$Jy for 4.4, 11.4, and 23[$\mu$m]{} respectively. These sensitivities are based on a 5$\sigma$ detection in the difference of two 1hour exposures [@2010PASP..122..162B]. The contrast ratios are shown in Figure \[fig:egpcontr\].
Figure \[fig:egpcoron\] shows detectability contours for the Jupiter-mass planet at a range of possible planetary semi-major axes from our A5 star at 10pc. For each instrument there are two curves; one for the swarm (black curve) and one for the planet (grey curve). Swarms and planets in the space to the left of curves are detectable. The optimum planet semi-major axis for detection in scattered light for our chosen parameters (eq. \[eq:aopt\]) is drawn as a dotted line.
Looking at the detection limits for each instrument individually, HST ACS detects scattered light from swarms across a wide range of the parameter space. As expected from equation (\[eq:aopt\]), these swarms lie many tens of AU from the star. Detectable swarms are at larger semi-major axes at later times as swarms around planets on closer orbits deplete. As expected from Figure \[fig:egpspec\], Jupiter-mass planets are hard to detect with ACS at any separation (so there is no grey HST curve on Fig. \[fig:egpcoron\]).
\
\
JWST instruments are well suited to planet detection by design so planets and swarms are detectable. The detection space for satellite swarms with NIRCAM covers semi-major axes greater than 70-120AU until a Gyr. Planets are detectable for a shorter time but to much closer separations. The difference is because the swarms are detected in scattered light, and planets from their thermal emission. At 11[$\mu$m]{} MIRI does not detect the satellite swarms, but is ideal for detecting cooling planets, whose SEDs initially peak in or near this region. The detection region therefore covers even more age and separation space than at 4.4[$\mu$m]{}.
At 23[$\mu$m]{} MIRI detects the cooling planet, but also sees the Wein side of the dust cloud for systems up to $3 \times 10^8$years old. At this wavelength the visibility space for MIRI lies closer than suggested by [$a_{\rm opt}$]{} because the dust clouds are hotter. With the exception of the vertical parts of curves (most notably at 23[$\mu$m]{}), detectability in this example is set by contrast rather than absolute sensitivity.
Therefore, optical wavelengths appear to be the best place to detect extrasolar irregular satellite swarms around Jupiter-mass planets. Though detectable, at IR wavelengths the swarms are fainter than their host planets. Discovery then involves finding whether a (presumably previously known) planet has an optical or IR excess.
Over the parameter space in Figure \[fig:egpcoron\] we find that the largest excess ratio ($F_{\rm swarm}/F_\star$) is about 2% at 100[$\mu$m]{} and 3% at 160[$\mu$m]{} for a young swarm at 70–80AU. These excesses would not be detectable given current uncertainties with *Herschel* PACS photometry and planet atmospheres. Given the resolution, a sufficiently bright swarm detected this way would be classed as a circumstellar disk unless it were resolved at shorter wavelengths or orbiting a nearby star. Such a limitation may be overcome by high resolution facilities such as ALMA.
The predictions of Figure \[fig:egpcoron\] for all instruments can be extended in several directions. For the same star moved to 5pc, swarms and planets at the same semi-major axis are more detectable because they lie at larger angular separations (where the instrument contrast ratio is better). Alternatively, a swarm around a planet at 10” separation around a star at 10pc (planet at 100AU) is about equally detectable at optical/near-IR wavelengths when the star is at 5pc (planet at 50AU), but more easily detectable at mid-IR wavelengths because the dust is hotter. For the same range of planet semi-major axes shown in Figure \[fig:egpcoron\], the same swarms become much harder to detect beyond a few tens of parsecs. The difficulty arises due to the poorer contrast at smaller angular separations. If swarms are detectable beyond a few tens of parsecs, they must have of order an Earth-mass or more in irregular satellites.
As the stellar mass decreases the detection space at optical/near-IR wavelengths increases due to the much lower stellar luminosity and consequently smaller minimum grain size. However, the clouds become undetectable in the mid-IR because the Wien side of their spectrum is too faint. The planets themselves become somewhat more detectable at IR wavelengths because their thermal emission is the same but the star is fainter (and we have assumed all planetary luminosity is intrinsic).
Decreasing the planet mass can make for undetectable planets that host detectable swarms. The weak Hill radius dependence on planet mass means that collisional evolution is only somewhat slower (eq. \[eq:rccdc\]). In addition, the minimum grain size only depends weakly on planet mass (eq. \[eq:dmin\]). To highlight the relative unimportance of planet mass on the evolution of satellite swarms, the lower panel of Figure \[fig:egpcoron\] shows detectability contours for a $20\,M_\oplus$ planet. The much fainter planet can only be detected at the earliest times with MIRI at 11[$\mu$m]{}. Despite the 16 times decease in planet mass, the HST detection space is not much smaller than for the Jupiter-mass planet in the upper panel and still covers a wide range of orbits and ages. The detection space for NIRCAM at 4.4[$\mu$m]{} is extremely small. At 23[$\mu$m]{}, MIRI can still detect young swarms orbiting planets at 25 to 80AU from the star. This Figure therefore shows that relatively low-mass planets that would otherwise be invisible can be detected thanks to the luminosity of their irregular satellite swarms. As we argue below for the specific case of Fomalhaut b, these swarms could be misidentified as more massive planets in the first instance.
### Discussion {#ss:exodisc}
The ability of planets to capture irregular satellite swarms and reside at tens of AU sets the likelihood that any will be detected outside the Solar System. The wide range of planet-masses about which swarms should be detectable means that how capture mechanisms and efficiencies change with planet mass is important. In particular, our prediction that swarms may be visible around relatively low-mass planets relies on the ability of these planets to capture swarms.
Giants forming by core accretion may capture irregulars passing through their primordial envelopes [e.g. @1979Icar...37..587P], or by “pull-down” during a phase of rapid growth [e.g. @1977Icar...30..385H]. Giants that form by gravitational instability may have analogous capture mechanisms. suggest that the similarity of the irregular populations at each planet argues against gas giant specific capture processes for the Solar System. However, @2010AJ....139..994B note that because the size distributions are a result of the collisional evolution, they cannot be used to constrain the capture mechanism.
Any planet may capture satellites during three and $n$-body interactions [e.g. @1971Icar...15..186C; @2006Natur.441..192A]. Planets may capture satellites as they themselves interact in the presence of a planetesimal disk. Irregulars captured earlier by gas drag and pull-down are supplanted by those captured during planet-planet interactions. Such a scenario has been proposed for the origin of the Solar System’s irregular satellites within the context of the Nice model [@2007AJ....133.1962N]. The strength of this model is that it results in similar populations at each planet. In fact, in most model runs @2007AJ....133.1962N find that Jupiter captures the least irregular satellites, because it undergoes the fewest encounters with other planets. The prospects for lower mass planets harbouring swarms of irregulars are therefore good.
Because we predict that swarms are most detectable around planets with large semi-major axes, the probability of detection depends on the ability of planets to form on, or move to such orbits. The recent discovery of planets around HR 8799 [@2008Sci...322.1348M] and Fomalhaut [@2008Sci...322.1345K] on such wide orbits is encouraging, not only due to their very existence, but because both stars also harbour planetesimal disks [@1984BAAS...16..483A; @1986PASP...98..685S]. Planets that originate on closer orbits must either scatter or migrate to such distances. If these systems have planetesimal disks, satellite capture during the scattering and migration process is likely. In such a scenario, the swarm’s radial extent ($\eta$) cannot be much larger than half the Hill radius at the planetary semi-major axis where the swarm was captured (see also §\[ss:fombdisc\]).
The presence or absence of irregular swarms within the context of discovered planetary systems should provide information about planet formation and evolution. For example, the @2007AJ....133.1962N scenario requires the apparently specialised circumstances of a set of giant planets that are destabilised and interact in the presence of a planetesimal disk, but appears to be a robust way to capture irregulars. Such a scenario suggests that swarms may be discovered around planets in multiple systems that harbour debris disks (such as HR8799), especially those where scattering rather than migration has occurred. On the other hand, if capture of irregulars is associated with major planetesimal depletion events [as proposed by @2007AJ....133.1962N], systems with irregular swarms may tend *not* to have visible debris disks. If primordial irregulars are formed in all planetary systems, the properties of swarm-harbouring systems may be much more general. Therefore, it will be interesting to study the expected detection outcomes for a range of capture and formation mechanisms.
As noted previously, there are effects beyond the scope of our model that can affect our predictions. More theoretical effort should be made to understand which are important and how they affect our conclusions, in particular the effect of radiation forces on dust. Efforts should be made to detect dust clouds around Solar System giant planets, not only for their intrinsic interest, but to provide an empirical conversion between the mass in irregulars and the surface area in dust to calibrate our model predictions.
Fomalhaut b {#ss:fomb}
-----------
An interesting application of our model is the planet orbiting the nearby star Fomalhaut, which also harbours a narrow ring-like circumstellar debris disk [@2008Sci...322.1345K]. The planet was predicted to exist based on the elliptical shape of the debris ring [@2005Natur.435.1067K]. Using the sharpness of the inner edge of the dust ring @2006MNRAS.372L..14Q estimated the planet’s orbit and a range of possible masses between Neptune’s and Saturn’s. In fact the structure of the ring continues to provide the most stringent constraints on the planet’s mass, with more recent modelling providing an estimate of $<$3[$M_{\rm {Jup}}$]{} if this planet is the “sole sculptor” of the disk [@2009ApJ...693..734C].
@2009ApJ...693..734C note that Fomalhaut b need not be the only sculptor of the belt. This point is in part made because the derived orbital velocity for Fomalhaut b is marginally inconsistent with that expected if the disk and planet are apsidally aligned. With Fomalhaut b only observed at two epochs, this inconsistency is a minor issue at present, but highlights the possibility that another planet may be partly responsible for the offset and truncation of the debris disk, and Fomalhaut b may be much less massive.
Another reason the mass is so poorly constrained at present is the lack of information about the planet’s spectrum [@2008Sci...322.1345K]. Thus far it has only been detected in two bands (0.6 and 0.8$\mu$m), the first of which shows a factor 3 change in brightness between measurements at two epochs separated by 2 years. Neglecting the potential temporal evolution, these observations and the non-detections in other wavebands more closely resemble reflected starlight than thermal emission from a planet.
These observations lead @2008Sci...322.1345K to suggest that the planet hosts a circumplanetary ring akin to Saturn’s, which would extend to at least 20 Jupiter radii for an assumed albedo of 0.4 to recover the observed fluxes. If the emission is indeed scattered light from dust then the planet mass could be much less than 3[$M_{\rm {Jup}}$]{} . In this section we suggest that the observed spectrum could indeed be reflected starlight, scattered off the dusty debris from a circumplanetary swarm of irregular satellites.
A convenient parameter space for this problem is the cloud semi-major axis in units of Hill radii ($\eta$) vs. the planet mass ($M_{\rm pl}$). The observed quantities constrain $\eta$, because the planet is not resolved and the disk size is limited by optical depth. The planet mass will be constrained because a circumplanetary swarm that evolves like our model will not survive for the age of Fomalhaut over all parts of this parameter space. That is, for a swarm orbiting a $\sim$Jupiter-mass planet to be unresolved it would have small $\eta$, where it would be rapidly depleted to undetectable levels.
### Observational Constraints
The scattered light model in @2008Sci...322.1345K requires a total cross-sectional area of dust: $$\label{eq:stot}
\sigma_{\rm{tot}}= 5.8 \times 10^{-4} (0.08/Q)(1/\cos{i}) \, {\rm AU}^2 \, ,$$ where we assume $Q = 0.08$ (see §\[ss:obs\]) and $\cos{i}$ is the factor that would be required if the geometry was that of a flat ring inclined $i$ to our line-of-sight, assumed to be $\sim$1 as the model of §\[s:mod\] is optically thin.
The light from Fomalhaut b looks unresolved so the dust must be confined to a region smaller than the Hubble PSF FWHM of 0.53AU. This region has an area of 0.22AU$^2$ so the geometrical optical depth of the dust could be as low as $2.6 \times 10^{-3}$. The resolution constraint means that $\eta r_{\rm{H}} \lesssim 0.53$ so $\eta <
0.6/M_{\rm{pl}}^{1/3}$, where we have allowed the dust cloud to be a factor of two larger in extent to account for the concentration of brightness closer to the planet seen in Figure \[fig:saturnpic\]. As noted in §\[s:mod\], another constraint on $\eta$ comes from the stability of circumplanetary orbits, $\eta < 0.5$.
Assuming that the dust is uniformly projected on the sky across $\pi (\eta
r_{\rm{H}})^2$, its geometrical optical depth is $\tau = \sigma_{\rm{tot}} / ( \pi (\eta
r_{\rm{H}})^2 )$. Since $\tau < 1$ the dust must be located at $\eta > 0.015 / M_{\rm
pl}^{1/3}$, which sets a lower limit on planet mass of $26 \times 10^{-6}M_\oplus$ (i.e., 6 times less massive than Ceres and 80 times less massive than Pluto).
The above constraints, and the dynamical constraint of $M_{\rm pl} \lesssim 3$[$M_{\rm {Jup}}$]{} from @2009ApJ...693..734C, are summarised by solid lines in Figure \[fig:fomb\]. The satellite cloud may reside anywhere in the region enclosed by the solid lines, which spans over seven orders of magnitude in planet mass, and more than two in semi-major axis. These constraints are independent of our model (except perhaps the 0.5[$R_{\rm H}$]{}stability limit), and apply to any circumplanetary dust population. The $\sim$20[$R_{\rm {Jup}}$]{}disk proposed by @2008Sci...322.1345K would lie in the lower right region of the allowed space in Figure \[fig:fomb\].
Our aim is now to narrow this parameter space further. Because the collision rate depends strongly on $\eta$, satellite clouds appear more likely to survive for the age of Fomalhaut around planets much less massive than Jupiter.
Currently, the typically quoted age of Fomalhaut is 200Myr [@1997ApJ...475..313B]. However, recent work suggests that the age may be closer to 400Myr (E. Mamajek, *priv. comm.*). We adopt 200Myr as the age, and note below the (small) difference an older age makes to our model.
### Collisional constraints
To find where satellite clouds survive for the age of Fomalhaut, we set the collision time ($1/R_{\rm cc}$) to 200Myr (i.e. it is collision limited). It remains to substitute appropriate equations and estimates for parameters to reduce Equation (\[eq:rccdc\]) to $\eta$ as just a function of $M_{\rm pl}$. We use $M_\star =
2\,M_\odot$, $L_\star = 21\,L_\odot$, $\rho = 1500$kg m$^{-3}$, $a_{\rm pl} = 115$AU, and $D_{\rm c} = 500$km. For comparison with Table \[t:ssdust\], the Hill radii for Earth, Neptune, and Jupiter-mass planets at this distance are 0.9, 2.3, and 6.2AU respectively. We need to convert the observed [$\sigma_{\rm tot}$]{} into [$M_{\rm tot}$]{}, for which we use equation (\[eq:mtot\]). Because the effect of radiation pressure in setting [$D_{\rm min}$]{}varies over the parameter space, we also need equation (\[eq:dmin\]), which gives the mass as $M_{\rm tot} = 300 \, \sigma_{\rm tot} \eta^{0.35} / M_{\rm pl}^{0.23}$. The mass in satellites is therefore about 0.06$M_\oplus$, or 5$M_{\rm moon}$ for $\eta = 0.2$ and $M_{\rm pl} = 10\,M_\oplus$, about 100-1000 times more massive than the initial conditions in the @2010AJ....139..994B models, and similar to §\[ss:exo\]. Such a low planet mass and higher stellar mass and luminosity result in $D_{\rm min} \sim 300$[$\mu$m]{}, much higher than in the Solar System.
Substituting these parameters into equation (\[eq:rccdc\]) yields a locus for a collision limited satellite swarm around Fomalhaut b that reproduces the observed [$\sigma_{\rm tot}$]{}. This locus $\eta = 0.27 / M_{\rm pl}^{0.12}$ is shown in Figure \[fig:fomb\] as a dotted line. Because [$D_{\rm c}$]{} may be smaller than our assumed 500km, but based on the Solar System is unlikely to be significantly larger, the space above this line is also allowed. Thus, the collision rate in concert with the resolution limit constrains the planet to have a maximum mass of about 100$M_\oplus$. The minimum mass is lower than the mass in satellites, and therefore physically unreasonable.
There is a further complication due to the large semi-major axis of Fomalhaut b, which results in low collision velocities $v_{\rm rel} = 68 M_{\rm pl}^{1/3} / \sqrt{\eta}$ (in m s$^{-1}$). A rough estimate of the largest object that can be destroyed by another of the same size (and will therefore participate in the collisional evolution) can be derived from equation (\[eq:qd\]) and $X_{\rm c} = 1$ (§\[ss:decay\]), giving $D_{\rm c} \approx v_{\rm rel}^{1.6}/28$ (in km). This estimate yields another collision rate from equation (\[eq:rcc\]) that applies when [$D_{\rm c}$]{} is set by collision velocities $$R_{\rm cc} = 114 \, \frac{ M_{\rm tot} \, M_\star }
{ D_{\rm c}^{0.37} \, \rho \, ( \eta \, a_{\rm pl} )^3 \, M_{\rm pl} }$$ in years$^{-1}$. This equation yields another locus in the $\eta$ vs. $M_{\rm pl}$ parameter space where the collision rate equals the system age of $\eta = 0.67 / M_{\rm
pl}^{0.46}$. This limit is shown in Figure \[fig:fomb\] as a dashed line. Though it would appear that the dotted line allows swarms with $D_{\rm c} = 500$km to survive for planet masses right down to $10^{-3}\,M_\oplus$, satellites this large cannot be destroyed at the low collision velocities around such low mass planets. The dashed line is a lower limit because [$D_{\rm c}$]{} could be smaller than the smallest object that can be destroyed. This limit therefore constrains the mass of Fomalhaut b to be more than a few Earth masses.
These two loci combined with the previous constraints map out a region of parameter space in Figure \[fig:fomb\] where a satellite swarm could survive for 200Myr around Fomalhaut b. The planet is $\sim$2–100 Earth masses and the swarm lies at 0.1-0.4 Hill radii. The swarm mass is of order a few Lunar masses (but varies with planet mass due to the changing [$D_{\rm min}$]{}, see above). The model predicts that a cloud of the observed luminosity could survive around a more massive planet, but this possibility is excluded because Fomalhaut b is unresolved. If Fomalhaut is 400Myr old, the allowed parameter space is pushed even closer to the resolution limit, but does not change Figure \[fig:fomb\] significantly.
Fomalhaut b is marked in the lower panel of Figure \[fig:egpcoron\] (based only on age and orbit). As expected it lies within the region where swarms can be detected around a 20$M_\oplus$ planet with HST and JWST NIRCAM.
Of course there is considerable uncertainty in both our model and the parameters it uses, but based on evidence that satellite swarms exist around Solar System planets, it at least suggests that circumplanetary swarms should be considered a possibility around extra solar planets. In the particular case of Fomalhaut b, the question of planet vs. cloud can resolved if new observations show Fomalhaut b looks like a planetary atmosphere. In the event that scattered light cannot be ruled out, both circumplanetary rings [e.g. @2008Sci...322.1345K] and swarms remain plausible options.[^3]
### Discussion {#ss:fombdisc}
While our model provides a possible explanation of Fomalhaut b’s apparently blue spectrum, the provenance of such a configuration is unclear. Any scenario faces the difficulty of explaining how Fomalhaut b came to be at such a large distance from Fomalhaut itself. In a core accretion scenario, the planet presumably originates somewhere much closer to the star [e.g. @2008ApJ...673..502K], and somehow scatters or migrates to its current location. Because the Hill radius expands as the planet moves outward, the satellite cloud would need to be captured while the planet was orbiting at least $\sim$20AU from the star for it to reside at $\eta > 0.1$ now. Capture of satellites might happen as Fomalhaut b scatters off other planets or migrates through a planetesimal disk on the way to its current location. The gas drag and pull-down mechanisms are unlikely to operate because our predicted mass of Fomalhaut b is insufficient for it to have a significant gaseous envelope.
Our predicted mass for Fomalhaut b is similar to or less than the mass of the main debris ring itself, which is $\sim$3–300$M_\oplus$ [e.g. @2002MNRAS.334..589W; @2009ApJ...693..734C]. If the planet and ring masses are similar, it is unlikely that Fomalhaut b is responsible for truncating and imposing eccentricity on ring particles. Even if such a low-mass planet could reproduce the debris ring structure, the relatively small chaotic zone width may require that the planet lie closer to the ring than observed.
These issues do not necessarily pose a major problem for our model. @2009ApJ...693..734C emphasise that Fomalhaut b may not actually be responsible for sculpting Fomalhaut’s debris ring, in part because the planet’s orbit is mildly inconsistent with the expected apsidal alignment of planet and ring particle orbits. However, an as yet undiscovered object massive enough to sculpt the debris ring may have more serious implications for the stability of Fomalhaut b.
### An alternative model
The current poor constraint on Fomalhaut b’s orbit allows for other interesting possibilities. For example, its orbit may pass through the circumstellar ring. When the planet is within the ring, planetesimal impacts will generate a surrounding cloud of regolith, mantle, and planetesimal fragments. The material may free-fall back to the planet as in the model of @2002MNRAS.334..589W, or in a picture more like the formation of the Earth-Moon system or Kuiper belt binaries some fraction of the fragments can remain in bound orbits. If Fomalhaut b spends a non-neglible fraction of its orbit within the ring, possible if the planet orbit and ring are coplanar, this scenario provides a mechanism by which a population of objects bound to the planet may be built up and replenished, thus avoiding the requirement that the swarm must survive for the age of Fomalhaut to be observed. This way, even a small amount of mass launched into orbit each time the planet passes through the ring will build up over time, eventually reaching an equilibrium state where the bound mass is limited by collisions. Finding this equilibrium mass is then simply a matter of estimating the rate at which mass is added to the swarm and balancing it with the depletion rate. Because we expect the new satellites to be launched into near-isotropic orbits (i.e. to look like an irregular satellite swarm), the depletion rate is calculated from the model developed in §\[s:mod\].
The rate at which new satellites are added to the swarm is uncertain, so we adopt the approach taken by @2002MNRAS.334..589W; an impacting planetesimal can at most launch its own mass of regolith and fragments to a significant distance from the planet. For impacts by large planetesimals these collisions may be more akin to the “graze and capture” like scenario suggested for Pluto-Charon and Haumea’s satellites [e.g. @2005Sci...307..546C; @2010ApJ...714.1789L]. Not all planetesimals accreted by the planet will launch their own mass in regolith and fragments away from the planet, nor will all launched material end up in orbit. We therefore expect that the accretion rate must be at least a few orders of magnitude higher than the loss rate for the existence of a cloud to be feasible.
The gravitationally focused mass accretion rate can be calculated from $$\label{eq:macc}
\dot{M}_{\rm acc} =
\frac{M_{\rm ring}}{2\,\pi\,r_{\rm ring}^2\,dr_{\rm ring}\,i_{\rm ring}} \,
\pi R_{\rm pl}^2 \, \left( 1 + \frac{v_{\rm esc}^2}{v_{\rm rel}^2} \right) \, v_{\rm rel}$$ We assume $r_{\rm ring} = 141$AU, $dr_{\rm ring} = 25$AU, and $i_{\rm ring} =
0.026$rad [using ring parameters from @2005Natur.435.1067K]. We assume the largest ring planetesimal is 500km in diameter, which corresponds to a ring mass of about 75$M_\oplus$ [@2005Natur.435.1067K], which is intermediate to the two cases considered in @2002MNRAS.334..589W. The accretion rate of a 10$M_\oplus$ planet with mass density 5000kg m$^{-3}$ is about $10^{-10} M_\oplus/$year. Because the planet cannot spend all its time within the ring, we (arbitrarily) decrease the accretion rate by a factor of two. The accretion rate would be much lower for a non-coplanar orbit, on which the planet would spend much less time (if any) within the ring. For a 10$M_\oplus$ planet with mass density 5000kg/m$^3$, and $v_{\rm rel} = 0.1\,v_{\rm
K}$, Equation (\[eq:macc\]) reduces to $2 \times 10^{-10} M_{\rm
pl}^{4/3}\,M_\oplus/$yr.
The mass loss rate is simply the current mass times the collision rate, which for a swarm in collisional equilibrium is also the inverse of the age (equation \[eq:dmdt\]). More generally, using Equations (\[eq:qd\]), (\[eq:mtot\]), and (\[eq:rccdc\]), the mass loss rate is $$\label{dmdtstot}
\dot{M}_{\rm loss} = 0.002 \,
\left[ \frac{\rho^{0.37} \, D_{\rm min}^{1.4} \, M_\star^{1.38} \, f_{v_{\rm rel}}^{2.27}}
{M_{\rm pl}^{0.24} \, \left(a_{\rm pl} \, \eta\right)^{4.13}} \right] \,
\sigma_{\rm tot}^2$$ in $M_\oplus /$ year. For our purposes here, the mass loss rate is simply proportional to the square of the observed $\sigma_{\rm tot}$. That is, more mass is lost and at a faster rate for higher [$\sigma_{\rm tot}$]{} or [$M_{\rm tot}$]{}. All parameters in the large parenthesis depend on the particular model assumptions. The rate is independent of the maximum planetesimal size [$D_{\rm c}$]{}. This independence can be viewed as due to our assumption of a self-similar size distribution, where the mass lost is the same in each logarithmic size bin. If the small end of the size distribution is fixed, larger [$D_{\rm c}$]{} means more mass is available to be depleted ([$M_{\rm tot}$]{}), but collisions are less frequent and the largest objects are stronger.
Evaluating Equation (\[dmdtstot\]) for Fomalhaut b yields $\dot{M}_{\rm loss} = 5.5
\times 10^{-12} / \left( \eta^{3.4} \, M_{\rm pl}^{0.7} \right)$. Using the system age and a planet with a collision limited swarm of mass $0.06\,M_\oplus$ from the previous section yields $3 \times 10^{-10}\,M_\oplus/$year, in agreement with the general expression for these parameters.
Using these two expressions for the mass accretion and loss rates and assuming an efficiency parameter $f_{\rm acc}$, a swarm will replenished as fast as it decays when $f_{\rm acc} \, \dot{M}_{\rm acc} = \dot{M}_{\rm loss}$. Figure \[fig:cre\] shows where this equation is satisfied for different $f_{\rm acc}$. The line where the planet mass is too large to be consistent with the observed debris ring structure is omitted because a planet that passes through ring will perturb it in a different way. This line would move to lower planet masses, and constrain this model. Another constraint on the upper planet mass would come from the mass above which a significant atmosphere is present, meaning that ring planetesimals are engulfed, rather than launch regolith and fragments. However, an atmosphere under periodic bombardment would be more prone to thermal escape due to heating.
The location of the lines in Figure \[fig:cre\] suggest that replenishing a swarm from ring planetesimals is difficult. Even if a tenth of the accreted mass is launched into orbit the planet must be more than an Earth mass, which seems likely to significantly perturb the main ring on a timescale less than 200Myr. One way to lower the planet mass but maintain the accretion rate is if the planet is a binary. The cross section for an interaction with a passing planetesimal scales with the binary semi-major axis and is therefore much larger than the physical size of either object [@1975AJ.....80..809H; @1989AJ.....98..217L]. Such an interaction does not guarantee a collision, but allows a greater chance of one for the same (total) planet mass. This may allow a binary to launch sufficient material into a swarm, yet also have a low enough mass to minimise dynamical perturbations to Fomalhaut’s debris ring.
Summary and Conclusions {#s:conc}
=======================
The purpose of this paper is to derive a simple but general model for collisional evolution of irregular satellites and apply it to planets in the Solar System and elsewhere. Our model uses the particle-in-a-box formalism to describe the collision rate of the largest objects in a satellite swarm. A model size distribution allows the mass evolution to be converted to the surface area in dust and flux densities. Though the model can be developed further in many ways, a comparison with the Solar System irregular satellites shows that our model is reasonable.
Application of our model to the Solar System suggests that the bulk of grains may be lost to the planet or interplanetary space when the cloud is collision dominated. It may be that deposition on regular satellites only becomes important at later times when PR drag starts to dominate grain orbital evolution. Some level of dust must be present if the irregular satellites are still grinding down, which we suggest may be at detectable levels at any of the Solar System’s outer planets.
A remarkable feature of irregular satellite swarm evolution is its relative insensitivity to planet mass. Swarms are nearly as bright and last nearly as long around Neptune-mass planets as around Jupiter-mass planets. In contrast to extrasolar planets, we find that satellite swarms are most visible at optical wavelengths around planets that orbit at many tens of AU from their parent stars. Lower mass planets without gaseous envelopes cannot capture swarms by gas drag and pull-down, but dynamical mechanisms can still operate. There is an optimum distance for detection, which arises because swarms around planets on close orbits decay too rapidly and swarms around planets on distant orbits take too long to collide.
We propose a plausible model for a satellite swarm around Fomalhaut b. This model provides an alternative explanation for the spectrum of the planet being consistent with reflected starlight. In this picture Fomalhaut b is predicted to be $\sim$1–100 Earth masses, further illustrating that planet mass is relatively unimportant for the evolution of irregular satellite swarms. The allowed parameter space for the model lies very close to the resolution limit, so observations at higher resolution (${\rm FWHM} < 0.53$AU or 69mas) should test our model. However, the order-of-magnitude spirit of our model and large uncertainties in model parameters that affect our allowed parameter space, such as satellite size, composition and strength, mean that the best way to test our model in this particular instance is to ascertain whether Fomalhaut b has the spectrum expected from a planet.
We briefly outline another possibility for a swarm around Fomalhaut b based on speculation that it could pass through Fomalhaut’s large debris ring. In this case, a swarm of objects is maintained by regolith and fragments launched by planetesimal impacts. The $\sim$Earth mass planet required by such a model may be too high to avoid large (and unobserved) perturbations to the main ring. This issue may be somewhat alleviated if the planet is actually a binary, which would enhance the accretion rate and allow a lower total planet mass.
The prospects for detecting dust created by irregular satellite collisions, both in the Solar System and around planets orbiting other stars appear good. Along with further characterisation of Fomalhaut b, the four populations of irregulars on our doorstep seem a good place to start.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to the Isaac Newton Institute for Mathematical Sciences in Cambridge where part of this work was carried out during the Dynamics of Discs and Planets research programme. We would also like to thank Paul Kalas for helpful comments and sharing preliminary observations of Fomalhaut b, Eric Mamajek for sharing his results on Fomalhaut’s age, and the referee Bill Bottke for a constructive review.
[^1]: Email: <[email protected]>
[^2]: Taken from <http://www.dtm.ciw.edu/users/sheppard/satellites/> in March 2010
[^3]: Patient observers may someday find photometric phase variations that would be caused by an optically thick circumplanetary disk . However, other methods will likely become available before any variation can be found due to Fomalhaut b’s $\sim$10$^3$ year period.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We analyze the influence of the characteristics of the STM-tip (applied voltage, tip radius) on the electroluminescence spectra from an STM-tip-induced quantum dot taking into account the many-body effects. We find that positions of electroluminescence peaks, attributed to the electron-hole recombination in the quantum dot, are very sensitive to the shape and size of the confinement potential as determined by the tip radius and the applied voltage. A critical value of the tip radius is found, at which the luminescence peak positions as a function of the tip radius manifest a transition from decreasing behavior for smaller radii to increasing behavior for larger radii. We find that this critical value of the tip radius is related to the confinement in the lateral and normal direction.'
author:
- 'M. D. Croitoru$^{a}$, V. N. Gladilin$^{a}$, V. M. Fomin$^{a,b}$, J. T. Devreese$^{b}$[^1]'
- 'M. Kemerink, P. M. Koenraad, J. H. Wolter'
title: 'Influence of the Characteristics of the STM-tip on the Electroluminescence Spectra'
---
Introduction
============
Scanning-tunnelling microscopy techniques have developed into an important tool for studying semiconductor nanostructures in recent years. The scanning-tunnelling microscope (STM) covers a wide field of applications. An example of such an application is the so-called tip-induced quantum dot, a special and unique type of quantum dot that can only be studied using a STM [@wenderot; @Dombrowski; @Yamauchi; @Kemerink]. When a negative bias is applied between the metallic STM-tip and the semiconductor sample, the electric field extends into the structure, and a hole accumulation layer is formed at the surface under the tip. Such an accumulation layer effectively screens the electric field. For sufficiently small tip apex radii, quantization occurs both in the radial direction and along the axis normal to the surface of the structure. It should be noted that the tip-induced band bending confines only a single type of carriers, holes in our case. The electrons are repelled from the region under the tip. By putting a barrier just below the surface, injected electrons are confined to the surface layer and radiative recombination with holes in the dot becomes possible.
In this paper we analyze the influence of the characteristics of the confinement potential, determined by the STM-tip radius and the applied voltage, on the electroluminescence spectra of a tip-induced quantum dot in the GaAs layer of the GaAs/Al$_{0.25}$Ga$_{0.75}$As/In$_{0.25}$Ga$_{0.75}$As quantum well structure [@Kemerink].
Model
=====
In our model we consider a multilayer structure sketched in Fig. 1. The shape of the STM tip is modelled by a truncated cone because, when the tip made of a soft metal (like Pt in our case) is pressed to the semiconductor layer, it becomes flat. The STM-tip and the GaAs layer (cap layer) are separated from each other by an insulator layer of width $l_{1}$. The holes and electrons are confined in the cap layer of width $l_{2}$. The width of the Al$_{0.25}$Ga$_{0.75}$As barrier is $l_{3}$. The Al$_{0.25}$Ga$_{0.75}$As layer borders on the In$_{0.25}$Ga$_{0.75}$As quantum well. The contact area of the STM-tip with the structure is a circle of radius $R_{\mathrm{tip}}$. Owing to the axial symmetry of the system, cylindrical coordinates $r$, $\varphi,$ $z$ are used. The $z$-axis is directed along the symmetry axis of the tip, which is chosen parallel to the normal to the surface of the structure. While the hole motion is confined in all directions by the potential well, the electron motion is confined only along the $z$-axis.
In order to find energy spectra of holes, trapped in the quantum dot, we solve self-consistently the Poisson equation $$\frac{1}{r}\frac{\partial}{\partial r}\left( r\frac{\partial V\left(
r,z\right) }{\partial r}\right) +\frac{\partial^{2}V\left( r,z\right)
}{\partial z^{2}}=-\frac{\rho_{h}\left( r,z\right) }{\varepsilon
_{0}\varepsilon_{i}}, \label{Poisson}$$ where $\rho_{h}\left( r,z\right) \equiv en_{h}\left( r,z\right) $ is the hole charge density and $\varepsilon_{i}$ is the dielectric constant of the material, together with the Schrödinger equation governing the hole motion:$$\begin{gathered}
-\frac{\hbar^{2}}{2}\left( \frac{1}{m_{j}^{||}\left( z\right) }\frac{1}{r}\frac{\partial}{\partial r}\left( r\frac{\partial}{\partial r}\right)
+\frac{1}{m_{j}^{||}\left( z\right) }\frac{1}{r^{2}}\frac{\partial^{2}}{\partial\varphi^{2}}+\frac{\partial}{\partial z}\frac{1}{m_{j}^{\bot}\left(
z\right) }\frac{\partial}{\partial z}\right) \Psi_{j,s,n,m}(r,\phi,z)\\
+V(z,r)\Psi_{j,s,n,m}(r,\phi,z)=E_{j,s,n,m}\Psi_{j,s,n,m}(r,\phi,z),\nonumber\end{gathered}$$ $$m_{j}^{||,\bot}\left( z\right) =\left\{
\begin{array}
[c]{l}m_{j}^{||,\bot\left[ GaAs\right] },\;\mathrm{if}\;z\in\mathrm{GaAs}\\
m_{j}^{||,\bot\left[ Al_{0.25}Ga_{0.75}As\right] },\;\mathrm{if}\;z\in\mathrm{Al}_{0.25}\mathrm{Ga}_{0.75}\mathrm{As.}\end{array}
\right.$$ Here the index $j$ labels the hole band type: $j=1$ for a hole, which is heavy for the motion along the $z$-axis and light in the plane of the GaAs layer and $j=2$ for a hole, which is light for the motion along the $z$-axis and heavy in the plane of the GaAs layer. For the case of a quantum well, the light and heavy holes exactly decouple at $\mathbf{k}_{\parallel}=0$. Since the lateral size of the quantum dot is much larger than the GaAs unit cell, only states around $\mathbf{k}_{\parallel}=0$ contribute to the wavefunctions in the quantum dot. The effect of coupling between the valence bands is very weak in the close vicinity of the center of the Brillouin zone $\mathbf{k}_{\parallel
}=0$ (see, e.g., Refs. [@BN; @koen]). The index $s$ labels subbands due to the size quantization of the hole degree of freedom along the $z$-axis. The index $n$ is the radial quantum number and $m$ is the angular quantum number of the lateral (in the plane of the GaAs layer) motion of the hole. The kinetic energy of the hole motion along the $z$-axis is larger than the kinetic energy of the lateral motion. (Our calculations show that the charge of holes is concentrated in a relatively thin ($\sim3~\mathrm{nm}$) layer near the interface insulator/GaAs. The radial size of the smallest dot considered in this paper is $12~\mathrm{nm}$.) Hence the adiabatic approach can be used, assuming that the hole degree of freedom along the $z$-axis is “fast” while the lateral one is “slow”. Therefore, the hole wave function is represented in a product form: $\Psi_{j,s,n,m}(r,\phi,z)=\Psi_{j,s}^{\bot}(z;r)\Psi
_{j,s,n,m}^{||}(r)e^{im\phi},$ where $\Psi_{j,s}^{\bot}(z;r)$ and $\Psi_{j,s,n,m}^{||}(r)e^{im\phi}$ are the wave functions describing the motion along the $z$-axis and the lateral motion, respectively. The Schrödinger equation governing $\Psi_{j,s}^{\bot}(z;r)$ reads$$-\frac{\hbar^{2}}{2}\frac{\partial}{\partial z}\frac{1}{m_{j}^{\bot}\left(
z\right) }\frac{\partial}{\partial z}\Psi_{j,s}^{\bot}(z;r)+V(r,z)\Psi
_{j,s}^{\bot}(z;r)=E_{j,s}\left( r\right) \Psi_{j,s}^{\bot}(z;r),
\label{Sch}$$ where $V(z,r)=eU(z,r)+V_{\mathrm{bar}}^{h}(z)$ with $U(z,r)$, the electrostatic potential. $V_{\mathrm{bar}}^{h}(z)$ represents band offsets for a hole [@Har].
We solve the Schrödinger equation (\[Sch\]) using the method described in Ref. [@Bal; @my]. It yields the spectrum of the subband energies $E_{j,s}\left( r\right) $ and wavefunctions $\Psi_{j,s}^{\bot}(z;r)$. Each of the energies $E_{j,s}\left( r\right) $ determines the top of a hole subband in the structure and plays the role of an adiabatic potential for the lateral motion.
Consequently, the equation, which describes the lateral motion, is as follows:$$-\frac{\hbar^{2}}{2m_{j}^{||}\left( z\right) }\left( \frac{1}{r}\frac{\partial}{\partial r}\left( r\frac{\partial}{\partial r}\right)
-\frac{m^{2}}{r^{2}}\right) \Psi_{j,s,n,m}^{||}(r)+E_{j,s}\left( r\right)
\Psi_{j,s,n,m}^{||}(r)=E_{j,s,n,m}\Psi_{j,s,n,m}^{||}(r). \label{Sch_L}$$ After solving this equation as described in Ref. [@my], we obtain the energy spectrum of the holes, confined in the STM-tip-induced quantum dot.
Given a particular number of holes in the quantum dot, we find the quasi-Fermi level $E_{F}^{h}$ for holes from the equation: $$N_{h}=\sum_{j=1,2}~\sum_{s,n,m}\frac{2}{\exp\left( \frac{E_{j,s,n,m}-E_{F}^{h}}{kT}\right) +1}.$$ Then, within the Hartree approximation scheme, the hole density may be extracted from $$n_{h}\left( r,z\right) =\sum_{j=1,2}~\sum_{s,n,m}\left\vert \Psi
_{j,s,n,m}(r,\phi,z)\right\vert ^{2}\frac{2}{\exp\left( \frac{E_{j,s,n,m}-E_{F}^{h}}{kT}\right) +1}.$$ This is the basic formula employed to invoke a self-consistent solution to the Schrödinger and Poisson equations.
In order to include in our scheme the exchange effects we use the local density approximation (LDA) as described in [@DFT; @Jones; @Kohn] $$E_{\mathrm{xc}}^{\mathrm{LDA}}\left[ n\right] =\left.
{\displaystyle\int}
\mathrm{d}\mathbf{r}~e_{\mathrm{xc}}^{\hom}\left( n_{0}\right) \right\vert
_{n_{0}\longrightarrow n\left( \mathbf{r}\right) },$$ where $e_{\mathrm{xc}}^{\hom}$ is the exchange energy of the homogeneous gas per particle.
The exchange energy of a hole gas was evaluated in Ref. [@Bobbert]. The result for exchange energy per hole is $$e_{\mathrm{xc}}^{\hom}\left( n\right) =\frac{E_{\mathrm{xc}}^{\mathrm{LDA}}}{N}=-~\frac{e^{2}}{4\pi\varepsilon_{0}\varepsilon}~\frac{3}{4\pi}\left(
3\pi^{2}n\right) ^{\frac{1}{3}}\zeta\left( w\right) ,$$ where the numerical function $\zeta\left( w\right) $ is given by$$\zeta\left( w\right) =\frac{w^{4}+3w^{3}+3w+1-\frac{3}{4}\left(
1-w^{2}\right) ^{2}\ln\left\vert \frac{1+w}{1-w}\right\vert +\frac{3}{4}\left( 1-w^{4}\right)
{\displaystyle\int\limits_{w}^{1}}
\frac{\mathrm{d}x}{x}\ln\left\vert \frac{1+x}{1-x}\right\vert }{4\left(
1+w^{3}\right) ^{4/3}}.$$ A parameter $w$ represents the ratio between the light- and heavy-hole Fermi wave vectors:$$w\equiv\frac{k_{\mathrm{Fl}}}{k_{\mathrm{Fh}}}=\sqrt{\frac{m_{l}}{m_{h}}.}$$ In the local density approximation, the exchange potential is related to the exchange energy $e_{\mathrm{xc}}^{\hom}\left( n\right) $ by [@Kohn]$$V_{\mathrm{xc}}\left( \mathbf{r}\right) =\frac{1}{e}\frac{\mathrm{d}}{\mathrm{d}n}\left[ n~e_{\mathrm{xc}}^{\hom}\left( n\right) \right]
\left( \mathbf{r}\right) .$$ The exchange potential is thus$$V_{\mathrm{xc}}\left( r,z\right) =-\frac{e}{4\pi\varepsilon_{0}\varepsilon_{\mathrm{GaAs}}}~\left( \frac{3}{\pi}\sum_{j=1,2}\sum
_{s,n,m}\left\vert \Psi_{j,s,n,m}(r,\phi,z)\right\vert ^{2}\right) ^{\frac
{1}{3}}\zeta\left( w\right) , \label{final}$$ This formula is employed to include exchange corrections within LDA in the self-consistent solution of the Schrödinger and Poisson equations. When $V(r,z)$ is substituted by $V(r,z)+V_{\mathrm{xc}}\left( r,z\right) $, we need a straightforward numerical solver as shown in Fig. 2. The iteration procedure, involved to calculate the hole states and the corresponding electrostatic potential is subdivided into four steps. In the first step, the electrostatic potential $V(r,z)$ in the whole structure is calculated in the absence of holes. Then the obtained potential is used as the initial guess in the iteration procedure. In the second step, the Schrödinger equation, describing the hole motion is solved. The obtained wave functions are used to calculate the hole density. In the third step, solving the Poisson equation with the charge density obtained in the previous step yields a new approximation for the electrostatic potential, which is involved in the next iteration. In the fourth step, we use the charge density obtained as a result of the second step to calculate the exchange potential Eq. (\[final\]), which is also used in the next iteration. At a given number of holes in the tip-induced quantum dot, the second, the third and the fourth steps are repeated consecutively. This sequence of steps continues till the maximal absolute value of the difference between the values of the electrostatic potential at consecutive iterations becomes less than a certain threshold ($\lesssim0.1~\mathrm{mV}$), which establishes the measure of the accuracy. When increasing the number of holes in the quantum dot, the electrostatic potential, obtained earlier for a smaller number of holes, is used as an initial guess for calculations. This procedure guarantees a continuous link between the solutions obtained at the consecutive numbers of holes.
Given the electrostatic potential formed by the hole charge, we solve the Schrödinger equation for electrons,$$\begin{gathered}
-\frac{\hbar^{2}}{2}\left( \frac{1}{m_{e}^{||}\left( z\right) }\frac{1}{r}\frac{\partial}{\partial r}\left( r\frac{\partial}{\partial r}\right)
+\frac{1}{m_{e}^{||}\left( z\right) }\frac{1}{r^{2}}\frac{\partial^{2}}{\partial\varphi^{2}}+\frac{\partial}{\partial z}\frac{1}{m_{e}^{\bot}\left(
z\right) }\frac{\partial}{\partial z}\right) \Psi_{s_{e},n_{e},m_{e}}^{e}(r,\phi,z)\\
+V^{e}(z,r)\Psi_{s_{e},n_{e},m_{e}}^{e}(r,\phi,z)=E_{s_{e},n_{e},m_{e}}^{e}\Psi_{s_{e},n_{e},m_{e}}^{e}(r,\phi,z),\end{gathered}$$ where $V^{e}(z,r)=-eU(z,r)+V_{\mathrm{bar}}^{e}(z)$, $U(z,r)$ is the electrostatic potential, and $V_{\mathrm{bar}}^{e}(z)$ describes band offsets for an electron [@Har], $$m_{e}^{||,\bot}\left( z\right) =\left\{
\begin{array}
[c]{l}m_{e}^{||,\bot\left[ GaAs\right] },\;\mathrm{if}\;z\in\mathrm{GaAs}\\
m_{e}^{||,\bot\left[ Al_{0.25}Ga_{0.75}As\right] },\;\mathrm{if}\;z\in\mathrm{Al}_{0.25}\mathrm{Ga}_{0.75}\mathrm{As.}\end{array}
\right.$$ The index $s_{e}$ labels subbands due to the size quantization of the electron motion in the direction normal to the surface of the structure. The index $n_{e}$ is the radial quantum number and $m_{e}$ is the angular quantum number of the lateral motion of the electron. The electron wave function is represented in a product form: $\Psi_{s_{e},n_{e},m_{e}}^{e}(r,\phi
,z)=\Psi_{s_{e}}^{\bot,e}(z;r)\Psi_{s_{e},n_{e},m_{e}}^{||,e}(r)e^{im_{e}\phi
}$, where $\Psi_{s_{e}}^{\bot,e}(z;r)$ and $\Psi_{s_{e},n_{e},m_{e}}^{||,e}(r)e^{im_{e}\phi}$ are the wave functions describing the motion along the $z$-axis and the lateral motion, respectively.
Our numerical approach to solve the Schrödinger equation, which governs the lateral degree of freedom of the electron, is described in Ref. [@my]. The set of energies $E_{s_{e},n_{e},m_{e}}^{e}$ and wave functions $\Psi_{s_{e},n_{e},m_{e}}^{e}(r,\phi,z)$ together with the set of hole energies $E_{j,s,n,m}$ and wave functions $\Psi_{j,s,n,m}(r,\phi,z)$ are used to describe the electron-hole radiative recombination. The intensity of electroluminescence at a frequency $\Omega$ in the quantum dot created by the STM tip is described by the following expression [@BN]:$$I\left( \hbar\Omega\right) \sim\sum_{j,s,n,m}~\sum_{s_{e},n_{e},m_{e}}f_{h}\left( j,s,n,m\right) f_{e}\left( s_{e},n_{e},m_{e}\right) P\left(
j,s,n,m,s_{e},n_{e},m_{e},\hbar\Omega\right) , \label{inten}$$ where $f_{h}\left( j,s,n,m\right) $ and $f_{e}\left( s_{e},n_{e},m_{e}\right) $ are, respectively, probabilities of finding holes and electrons in the corresponding quantum states, $P\left( j,s,n,m,s_{e},n_{e},m_{e},\hbar\Omega\right) $ is the probability of the electron-hole recombination.
We assume that the distribution functions $f_{h}\left( j,s,n,m\right) $ is described by the Fermi-Dirac function with the quasi-Fermi level $E_{F}^{h}$ :$$f_{h}\left( j,s,n,m\right) =\frac{1}{\exp\left( \frac{E_{j,s,n,m}-E_{F}^{h}}{kT}\right) +1}.$$ Following Ref. [@my], we assume that due to the radial accelerating field the non-equilibrium electron distribution is a strongly decreasing function of the angular quantum number$\ m_{e}$. Hence, the electron distribution function is approximated as $$f_{e}\left( s_{e},n_{e},m_{e}\right) \sim\delta_{m_{e}0}. \label{distr2}$$
Results and discussion
======================
The characteristics of the structure under consideration are as follows. The thickness of the vacuum barrier is $l_{1}=0.5~\mathrm{nm}$, while the thickness of the GaAs and Al$_{0.25}$Ga$_{0.75}$As layers are, correspondingly, $l_{2}=17~\mathrm{nm}$ and $l_{3}=46~\mathrm{nm}$. The respective dielectric constants are $\varepsilon_{3}=1.0$, $\varepsilon
_{2}=13.2$ and $\varepsilon_{3}=12.2$.
Figure 3 illustrates the position of the calculated peaks in the electroluminescence spectrum as a function of the voltage $V_{\mathrm{tip}}$ applied between the STM-tip and the semiconductor structure at $4.2~\mathrm{K}$. Clearly, the peak positions as a function of the bias between the sample and the tip reveal a red shift. This is explained as follows. An increase of the voltage shifts the bottom (top) of the dot (antidot) upwards nearly proportional to the applied voltage. Since the number of holes increases with the voltage, the width of the dot in the $z$-direction is reduced due to a stronger screening. Hence, an increase of $V_{\mathrm{tip}}$ leads to a deepening of the potential well for holes and to a rise of the barrier for electrons. This, in its turn, raises the energy of the electron states, which contribute to the recombination, with respect to the bottom of the conduction band in the absence of the applied voltage. The holes experience both the rise of the energy level due to the shift of the bottom of the dot and the lowering of the energy level due to a stronger confinement. The net effect of both trends is an upward shift of the hole energy levels. The upward shift of the hole energy levels with increasing $V_{\mathrm{tip}}$ is larger than that for electrons. Thus, these shifts result in a *red shift* in the electroluminescence spectrum (see also Ref. [@my]).
In Fig. 4 we present the peak positions for the spectrum shown in the inset of Fig. 3 as a function of the radius of the STM-tip. When the radius of the STM-tip $R_{\mathrm{tip}}>12\ \mathrm{nm}$, one notices that the enlargement of the radius $R_{\mathrm{tip}}$ leads to a rise of the transition energies (blue shift). This is understood on the basis of the fact that an enlargement of the contact area shifts on average the conduction and valence bands upwards, resulting in a rising potential barrier for electrons and a slight deepening of the potential well for holes. Consequently, the recombining electron has to occupy a higher energy level in order to have a considerable overlap integral with a hole wave function. Our calculations show that the upward shift of the hole levels is smaller than the increase of the energy of the recombining electron. So, a change of the contact area influences both hole and electron levels, but due to the weak confinement for holes in the radial direction, the shifts of the electron and the hole levels jointly lead to a *blue shift* of the transition energy (this behavior of the transition energy as a function of the tip radius has been analyzed in Ref. [@my]).
As distinct from that type of behavior, for $R_{\mathrm{tip}}<12\ \mathrm{nm}$ an increase of the contact area between the STM-tip and the semiconductor sample results in a *red shift* of the transition energy. The explanation of this phenomenon is as follows. When the radius of the STM-tip $R_{\mathrm{tip}}$ is much smaller than the thickness of the GaAs layer, the electrostatic field from the STM-tip hardly extends into this layer, as shown schematically in Fig. 5(a). Hence, the band bending in the GaAs layer is negligible. In this case the energies of the transitions between hole and electron levels are practically equal to those in the case without any external field. An enlargement of the contact area increases the extension of the electrostatic field into the structure, as shown in Fig. 5(b). This effect is similar to that of an increase of the applied voltage. In both cases transition energies become smaller (red shift). However, under the further enlargement of the contact area the electrostatic field within this area extends almost homogeneously over the whole semiconductor layer, see Fig. 5(c). As the luminescence originates in this case from the edge of the STM-tip-induced quantum dot, there occurs a smaller rate of decrease of the transition energies with increasing radius. Finally, when the radius $R_{\mathrm{tip}}>12\ \mathrm{nm}$, the effects giving rise to the blue shift, described in the previous paragraph, prevail.
Thus, we have shown a strong effect of the confinement potential, as determined by the STM-tip radius and the applied voltage, on the electroluminescence peak positions. Namely, while an increase of the voltage on the STM-tip results in a *red shift* of the electroluminescence peaks [@my], an increase of the contact area between the STM-tip and the semiconductor structure results in a *non-monotonous* variation of peak positions. Below a certain critical value of the tip radius, the transition energies decrease with increasing radius, while above the critical radius of the tip, the transition energies are an increasing function of radius. These results demonstrate vast possibilities of tuning optical characteristics of the STM-tip induced quantum dot by varying parameters of the STM-tip.
This work has been supported by GOA BOF UA 2000, IUAP, FWO-V projects Nos. G.0274.01, G.0435.03, WOG WO.025.99 (Belgium) and the European Commission GROWTH Programme, NANOMAT project, contract No. G5RD-CT-2001-00545. The research of M.K. has been made possible by a fellowship of the Royal Netherlands Academy of Arts and Sciences.
[999999]{}
Permanent address: Department of Theoretical Physics, State University of Moldova, str. A. Mateevici 60, MD-2009 Kishinev, Republic of Moldova.
Also at: Technische Universiteit Eindhoven, P. B. 513, 5600 MB Eindhoven, The Netherlands.
M. Wenderoth, M. A. Rosentreter, K. J. Engel, A. J. Heinrich, M. A. Schneider, and R. G. Ulbrich, Europhys. Lett. **45**, 579 (1999).
R. Dombrowski, C. Steinebach, C. Wittneven, M. Morgenstern, and R. Wiesendanger, Phys. Rev. B **59**, 8043 (1999).
M. Yamauchi, T. Inoshita, and H. Sakaki, Appl. Phys. Lett. **74**, 1582 (1999).
M. Kemerink, K. Sauthoff, P. M. Koenraad, J. W. Gerritsen, H. van Kempen, V. M. Fomin, J. H. Wolter, and J. T. Devreese, Appl. Phys. A. **72**, S239 (2001).
G. A. Bastard, *Wave Mechanics Applied to Semiconductor Heterostructures* (Les Editions de Physique, Paris, 1988).
M. Kemerink, P. M. Koenraad, and J. H. Wolter, Phys. Rev. B **54**, 10644 (1996).
P. Harrison, *Quantum wells, wires and dots. Theoretical and computational physics* (John Willey & Sons, Chichester, 2000).
E. P. Pokatilov, V. M. Fomin, S. N. Balaban, V. N. Gladilin, S. N. Klimin, J. T. Devreese, W. Magnus, W. Schoenmaker, N. Collaert, M. Van Rossum, and K. De Meyer, J. Appl. Phys. **85**, 6625 (1999).
M. Croitoru, V. N. Gladilin, V. M. Fomin, J. T. Devreese, M. Kemerink, P. M. Koenraad, K. Sauthoff, J. H. Wolter, Phys. Rev. B **68**, 195307 (2003).
R. M. Dreizer, E. K. U. Gross, *Density Functional Theory. An Approach to the Quantum Mamy-Body Problem* (Springer, Berlin, 1990).
R. O. Jones, O. Gunnarsson, Rev. Mod. Phys. **61**, 689 (1989).
W. Kohn and L. J. Sham, Phys. Rev. **140**, A1133 (1965).
P. A. Bobbert, H. Wieldraaijer, R. van der Weide, M. Kemerink, P. M. Koenraad, J. H. Wolter, Phys. Rev B **56**, 3664 (1996).
**FIGURE CAPTIONS**
1. Scheme of the tip/insulator/GaAs/Al$_{0.25}$Ga$_{0.75}$As/In$_{0.25}$Ga$_{0.75}$As quantum well structure. The shown hole density distribution is calculated for the STM-tip voltage $-3.1\ \mathrm{V}$ and the radius of the tip $20~\mathrm{nm}$.
2. Flow diagram of the numerical calculations.
3. Positions of peaks in the electroluminescence spectra, calculated for the STM-tip-induced quantum dot at $4.2~\mathrm{K}$, $R_{\mathrm{tip}}=12~\mathrm{nm}$, versus bias $V_{\mathrm{tip}}$. Inset: Typical electroluminescence spectrum calculated for the voltage on the STM-tip $V_{\mathrm{tip}}=-3.1~\mathrm{V}$ and the STM-tip radius $R_{\mathrm{tip}}=12~\mathrm{nm}$.
4. Positions of the electroluminescence peaks from Fig. 3 as a function of the STM-tip radius at the voltage on the STM-tip $V_{\mathrm{tip}}=-3.1~\mathrm{V}$.
5. Sketch of the electrostatic potential profile for the STM-tip induced quantum dots with small (a), intermediate (b) and large (c) radius. Darker regions correspond to a larger absolute value of the electrostatic potential.
![M. D. Croitoru, V. N. Gladilin, V. M. Fomin, J. T. Devreese, M. Kemerink, P. M. Koenraad, J. H. Wolter[]{data-label="fig1"}](fig1.eps)
![M. D. Croitoru, V. N. Gladilin, V. M. Fomin, J. T. Devreese, M. Kemerink, P. M. Koenraad, J. H. Wolter[]{data-label="fig2"}](fig2.eps)
![M. D. Croitoru, V. N. Gladilin, V. M. Fomin, J. T. Devreese, M. Kemerink, P. M. Koenraad, J. H. Wolter[]{data-label="fig3"}](fig3.eps)
![M. D. Croitoru, V. N. Gladilin, V. M. Fomin, J. T. Devreese, M. Kemerink, P. M. Koenraad, J. H. Wolter[]{data-label="fig4"}](fig4.eps)
![M. D. Croitoru, V. N. Gladilin, V. M. Fomin, J. T. Devreese, M. Kemerink, P. M. Koenraad, J. H. Wolter[]{data-label="fig5"}](fig5.eps)
[^1]: E-mail: [email protected]; tel.: +323 820 2460; fax: +323 820 2245
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We demonstrate that the spectrum rearrangement can be considered as a precursor and a basis for the metal-insulator transition observed in graphene dosed with hydrogen atoms. The Anderson-type transition is attributed to the Fermi level’s entering into the quasigap, which develop in the vicinity of the impurity resonance energy due to the anomalous spectrum rearrangement. Theoretical results for the Lifshitz impurity model are in a reasonable agreement with available experimental data.'
author:
- 'Yuriy V. Skrypnyk'
- 'Vadim M. Loktev'
title: 'Metal-insulator transition in hydrogenated graphene as manifestation of quasiparticle spectrum rearrangement of anomalous type'
---
introduction
============
Since the existence of Dirac quasiparticles has been proved for graphene, one of the most intriguing issues of its physics is the possibility of their localization by whichever imperfection that can appear in the honeycomb lattice.[@castr; @beena; @altsh; @gorba; @chang; @bardas] Moreover, It should not be overlooked that the effect of disorder on the transport of quasiparticles is sensitive to the nature of inhomogeneity, be it caused by short-range or long-range defects, neutral or charged impurities, adsorbed or interstitial atoms or molecules, vacancies, spatial distortions or structural irregularities, including ripples on the graphene sheet and other long-wave random modulations. All these different types of imperfections necessitate corresponding dedicated studies, which can not be accomplished by resorting to a single impurity model. Since these imperfections are rather dissimilar both in their character and action on basic properties of graphene, respective theoretical models should be diverse as well.
Early experiments on graphene-based devices, which were engineered like a commonplace field effect transistors, revealed that the sample conductivity never drops below a certain *minimal* value.[@natur] This fact, indeed, considerably reduced audacious expectations that corresponding devices are capable to serve as next-generation electronic switches. Nevertheless (or, maybe, exactly because of this), the minimal conductivity existence has produced quite a stir, and its origin has been relentlessly debated. Distinctive features of charge carriers, which were shown to obey the linear dispersion, constituted the core of this discussion. Uniqueness of the electron subsystem in graphene were pushed to its limits so much that former physics of semiconductors were sometimes categorically declared being utterly unsuitable for this uncommon material. It has been speculated that massless, according to their Dirac dispersion, charge carriers can not be localized by any degree of disorder caused by lattice imperfections or impurity centers. The presumed impossibility to localize Dirac excitations were directly linked to the minimal conductivity phenomenon by a simple reasoning. Since the product of the wave vector modulus and the mean free path is confined from below for propagating states according to the Ioffe-Regel criterion, the conventional Drude expression immediately yields the minimum value of $\sim e^{2}/h$ for each conducting channel. This argument, along with available experimental results, has led to an opinion that the minimum conductivity value has a *universal* character for graphene, and is expressed only through fundamental constants.
In the mean time, there appeared occasional theoretical studies, which did not deny a possibility to localize charge carriers in graphene, and argued for the mobility edge appearance under certain circumstances, either for a disorder of a general type,[@alein; @mirli] or for a specific impurity model, in particular, for substitutional defects,[@naumi] and for the Anderson model of disorder.[@amini; @roche; @feng] Later, it has become apparent that the minimal conductivity in graphene does depend on the sample quality and noticeably varies with the impurity concentration.[@fuhre; @mucci; @peres; @sarma; @ugart] Furthermore, not so long ago, a metal-insulator transition (MIT) in graphene dosed with atomic, which is essential, hydrogen has been convincingly observed. The MIT has been manifested by an increase of the room temperature resistance in about four orders of magnitude.[@roten] The transition has been reported for the grown on SiC surface graphene with a low hydrogen coverage, namely, around $0.1\%$. The presence of the mobility gap has been also reported for the fluorinated and ion bombarded graphene.[@savch; @ionir]
As well, corresponding data of ARPES measurements indicated the disappearance of quasiparticle excitations in the system. The MIT of the Anderson type comes about when the Fermi level finds itself inside a domain of *localized* states. Since the graphene in the actual experiment was pristine before the hydrogen deposition and possessed sounding metallic properties with conventional Fermi-liquid behavior, it is tempting to anticipate that some amount of deposited hydrogen is sufficient to open up a reasonably wide region filled with localized states, or a *quasigap*, in the spectrum. In essence, this draft description of the leading to the MIT process corresponds to the well-known phenomenon of *spectrum rearrangement*,[@ivlpo; @lifsh] which deals with decisive modifications in the elementary excitation spectrum upon increasing the impurity concentration in the system. When the quasigap is located in the vicinity of the Fermi level position, the MIT can take place in spite of the rather low hydrogen concentration. Below we are going to link together the predicted spectrum rearrangement in graphene with point defects and the observed MIT under the hydrogen dosing.[@roten] The spectrum rearrangement takes place when a single impurity induces a local or a resonance level in the spectrum. It has been demonstrated recently that graphene almost inevitably contains traces of impurity centers, which are capable in producing resonance states.[@nov]
Criticality and types of spectrum rearrangement
===============================================
Let us discuss some general spectral properties of impure systems and assume for this purpose that inside an isolated host band in a hypothetical $d$-dimensional system a single impurity center accounts for a resonance state with the energy $\varepsilon_{r}>0$, which is measured against the band edge and has been made dimensionless by the corresponding bandwidth. The dispersion relation for this host band is supposed to be of the form $\varepsilon(\bm{k})\sim k^{p}$. Then, the spatial behavior of the host Green’s function at a given energy inside the band, apart from any other details, should be governed by oscillations with the characteristic radius $r\sim\varepsilon^{-1/p}$, which is expressed in units of the appropriate lattice constant. At the energy $\varepsilon_{r}$, these oscillations determine the effective radius $r_{im}\sim\varepsilon_{r}^{-1/p}$ of the impurity state. Through a closeness of the resonance energy to the band edge, the effective radius of the impurity state may far exceed the lattice period. As the impurity concentration is increased, the individual impurity states are becoming more tightly packed. It can be anticipated that a significant spatial overlap between these states is capable in provoking qualitative changes in the spectrum. The average distance between impurities $\bar{r}\sim c^{-1/d}$, where $c$ is the (relative) impurity concentration, is gradually shrinking with increasing $c$, and at some stage of the process becomes of the order of $r_{im}$. This general condition defines the critical concentration $c_{SR}\sim\varepsilon_{r}^{d/p}$ of the spectrum rearrangement in the corresponding impure system. It seems justifiable to reiterate here that the value of $c_{SR}$ can be fairly low due to the smallness of the energy $\varepsilon _{r}$.
Graphene is, evidently, a two-dimensional system, so that $d=2$, and features the linear dispersion of charge carriers, which implies that $p=1$. Consequently, the critical concentration of spectrum rearrangement in the impure graphene is expected to be $\sim\varepsilon_{r}^{2}$, where energy is counted from the Dirac point, at which the valence band and the conduction band coincide. In a case of the conventional dispersion with $p=2$, the identical in appearance relation between $\varepsilon_{r}$ and $c_{SR}$ is inherent in four-dimensional systems. In this sense graphene, as regards the spectrum rearrangement, can be formally viewed as a conventional system of *increased* spatial dimensionality. This fact further accentuates the uniqueness of graphene, since the spectrum rearrangement has not been studied for such systems.
Let us return back to experimental data contained in Ref. . The comparison of angle-integrated spectrum for clean and hydrogen dosed graphene suggests that the dimensional resonance energy $E_{r}$ lies somewhere around $0.25$ eV lover than the initial Fermi level, while the Dirac point position $E_{D}$ in the undoped crystal is about $0.45$ eV below it. A bit more careful analysis of the spectrum rearrangement in impure graphene, [@skrlo] which is based on the expansion of the self-energy into a cluster series,[@ivlpo] brings about the following expression for $c_{SR}$, which, for a convenience, is rewritten here in terms of the hydrogen coverage, $$n_{SR}\sim n_{C}\frac{(E_{r}-E_{D})^{2}}{W^{2}}, \label{critcov}$$ where $$n_{C}=\frac{4}{\sqrt{3}a^{2}}\approx3.8\times10^{15}\textrm{cm}^{-2}$$ is the areal density of carbon atoms in graphene, $a=0.246$ nm – the lattice constant of graphene, $n_{SR}$ – the critical coverage of hydrogen atoms, $W\approx 6.3$ eV – the bandwidth parameter, and the logarithmic correction is omitted because of the relatively large resonance energy. Substitution of the guessed value for the resonance energy results in $n_{SR}$ that is reasonably close to the reported MIT critical hydrogen coverage of $3.8\times10^{12}$ cm$^{-2}$.[@roten] As a matter of fact, the critical coverage of the spectrum rearrangement $n_{SR}$, should not be identical to the experimentally obtained hydrogen coverage required for the MIT, which will be discussed later.
It should be also noted that there are two main types of the spectrum rearrangement: the *cross* one, which is usually accompanied by a sharp single impurity resonance, and the *anomalous* one. The first type of the spectrum rearrangement results in a quasiparticle dispersion that looks like a familiar hybridization between the host branch and a dispersionless branch at the impurity resonance energy. In this case, two new branches are separated by a gap, which widens with an increase in the impurity concentration. Consequently, two different wave vectors correspond to the same energy in this split spectrum. However, this feature has not been detected in the ARPES measurements on the hydrogenated graphene. The second type of the spectrum rearrangement, which is usually encountered in low-dimensional systems, is of a more diffused nature and frequently corresponds to a considerably smeared single impurity resonance. This anomalous spectrum rearrangement is characterized by the opening of the quasigap, which is filled with localized states and separates two non-overlapping branches of extended states exhibiting a renormalized dispersion.
Interplay between spectrum rearrangement and Anderson transition
================================================================
The anomalous type of the spectrum rearrangement has been shown to unfold in graphene for impurities described by the Lifshitz model.[@skrlo] Within this model, identical impurity centers are supposed to be randomly distributed on sites of the host lattice. Each impurity is only allowed to change the on-site energy at its location in the corresponding tight-binding Hamiltonian. It is well-known that the electron subsystem in graphene encompasses practically free $\pi$-electrons, which number is equal to the number of lattice sites. As soon as an additional hydrogen atom lands on graphene, its uncoupled s-electron immediately enters a chemical bond with one of the $\pi$-electrons. That should make the latter one localized almost completely at the impurity site. In the first approximation, one can assume that the resulting perturbation comes only from the appearance of strong attracting potential on the occupied carbon atom. Thus, the disordered system is described by following operators, $$\bm{H}=\bm{H}_{0}+\bm{H}_{im},\qquad\bm{H}_{im}=V_{L}\sum\limits_{\bm{n}\alpha}\eta_{\bm{n}\alpha}^{\phantom{\dagger}}
c_{\bm{n}\alpha}^{\dagger}c_{\bm{n}\alpha}^{\phantom{\dagger}},
\label{ham}$$ where $$\bm{H}_{0}=t\sum\limits_{\bm{n}\alpha,\bm{m}\beta}c_{\mathbf{n}\alpha}^{\dagger}c_{\mathbf{m}\beta}^{\phantom{\dagger}}
\label{ham0}$$ is the host Hamiltonian, in which summation is restricted to nearest neighbors, $\bm{n}$ runs over lattice cells, indices $\alpha$ and $\beta$ enumerate sublattices, $c_{\bm{n}\alpha}^{\dagger}$ and $c_{\bm{n}\alpha}^{\phantom{\dagger}}$ are the creation and annihilation operators at the corresponding lattice site, $t\approx2.7$ eV – is the tight-binding transfer integral for the $\pi$ bands in graphene, the variable $\eta_{\mathbf{n}\alpha}$ takes the value of $1$ with the probability $n_{H}/n_{C}$ or the value of $0$ with the probability $1-n_{H}/n_{C}$, where $n_{H}$ is the hydrogen coverage, and $V_{L}<0$ is the difference in the on-site potential at the defect position.
Within this model, the local density of states (LDOS) at the lattice site occupied by a hydrogen reads, $$\rho_{im}(E)=-\frac{1}{\pi}\operatorname{Im}\frac{g_{0}(E)}{1-V_{L}g_{0}(E)},
\label{ldos}$$ where $g_{0}(E)$ is the diagonal element of the host Green’s function $$\bm{g}=(E-\bm{H}_{0})^{-1}.$$ In the vicinity of the Dirac point in the spectrum, namely, within the window stretching up to around $0.75$ eV to each side of it, this diagonal element can be easily approximated, $$g_{0}(E)\approx\frac{E-E_{D}}{W^{2}}\ln\left(\frac{E-E_{D}}{W}\right)^{2}-i\pi\frac{|E-E_{D}|}{W^{2}},
\label{apr}$$ where $$W=\sqrt{\pi\sqrt{3}}t$$ is the same bandwidth parameter as in (\[critcov\]). Substitution of this approximation to Eq. (\[ldos\]) yields $$\rho_{im}(E)\approx\frac{|\mathcal{E}|}{W[(1-2\frac{V_{L}\mathcal{E}}{W}\ln|\mathcal{E}|)^{2}+(\pi\frac{V_{L}\mathcal{E}}{W})^{2}]},
\label{ldosappr}$$ where $$\mathcal{E}=(E-E_{D})/{W}.$$ The LDOS at the impurity site is shown in Fig. \[f1\] for $V_{L}=-25$ eV. As follows from the Figure, the resonance energy $E_{r}$ is located approximately $0.2$ eV above the Dirac point and thus corresponds to the experimentally observed peak. While the required change in the on-site potential is quite substantial, the large perturbation magnitude should be considered, first of all, as a result of the chosen simplified impurity model. Unless another is explicitly specified, the perturbation value of $-25$ eV will be used below for all subsequent estimations.
![Local density of states at the impurity site for $V_{L}=-25$ eV.[]{data-label="f1"}](f1){width="96.00000%"}
When a certain amount of deposited hydrogen atoms is taken into account, the negative potentials induced by them on the occupied lattice sites impede the electron movement. For the Lifshitz impurity model, which has been described above, the course of the spectrum rearrangement has been studied both analytically and numerically.[@skrlo; @persk] The analysis has shown that states are subject to localization near the resonance energy, where the impurity scattering is the strongest. At some coverage, a quasigap occupied by localized states opens around the energy $E_{r}$. With increasing impurity concentration, the quasigap gradually broadens and the degree of localization inside it rises. This quasigap expansion and the localization enhancement are not symmetric about the energy $E_{r}$ and are more expressed above the resonance (for an attractive potential). Finally, the quasigap consumes all the space between $E_{r}$ and $E_{D}$ in the spectrum, which corresponds to the onset of the spectrum rearrangement by the definition. In other words, the spectrum rearrangement does follow the anomalous scenario.
In the experiment discussed, dosed hydrogen atoms act not only as scattering centers, but also as acceptors. Thus, their presence inevitably lowers the Fermi level of the system. Therefore, while the hydrogen coverage is increasing, the Fermi level and the upper boundary of the quasigap, where the localization of states is most pronounced, are moving toward each other. Sooner or later, the Fermi level should appear inside the quasigap, which will cause the MIT of the Anderson type. Since the energy $E_{r}$ is positioned somewhere in-between $E_{D}$ and the bare Fermi level $E^{(bare)}_{F}$, the critical coverage of the spectrum rearrangement should be close but not identical to the critical coverage of the MIT.
The acceptor effect of the deposited hydrogen atoms is also implicitly contained in the Lifshitz model. Strong attracting potential $V_{L}<0$ produces a deep local level below the conduction band. Because of its remoteness from the continuous spectrum, the corresponding narrow impurity band should hold almost $n_{H}N_{C}/(2n_{C})$ energy levels, where $N_{C}$ is the total amount of carbon atoms in the system. Thus, taking into account the spin degeneracy, this impurity band consumes nearly one electron per deposited hydrogen atom. However, apart from these qualitative considerations, the adopted impurity model allows to address the question on the Fermi level position more precisely. Consider the (normalized) total number of states with energies that are less than a specified one, $$N(E)=\int_{-\infty}^{E}\rho(E)\,\mathrm{d}E,$$ where $\rho(E)$ is the density of states. In a disordered system this quantity can be expanded in powers of the impurity concentration,[@lifsh] $$N(E)=N_{0}(E)+\frac{2 n_{H}}{\pi n_{C}}\arg(1-V_{L}g_{0}(E))+\dots,$$ where the presence of two sublattices in graphene and the Lifshitz impurity model are taken into account, and $N_{0}(E)$ is the total number of states in the host system. Keeping in mind that the deep impurity band is modeling the acceptor effect of impurities, it is natural to demand the conservation of the number of occupied states in the system with varying the amount of dosed hydrogen. This yields a kind of a balance condition, $$\operatorname{sgn}\left(\frac{E_{F}-E_{D}}{W}\right)\frac{(E_{F}-E_{D})^{2}}{W^{2}}
+\frac{2n_{H}}{\pi n_{C}}\arg(1-V_{L}g_{0}(E_{F}))=\frac{(E^{(bare)}_{F}-E_{D})^{2}}{W^{2}},
\label{ef}$$ where the constant term is dropped from the both sides of the equation, and Eq. (\[apr\]) is used to obtain $N_{0}(E)$ for the clean graphene. From here and on this condition will be employed to determine the Fermi level position at a given dosing level. It is worth mentioning that the Dirac point position (which should not be confused with its value for the unperturbed system $E_{D}$) is gradually shifted with increasing the impurity concentration by approximately $n_{H}V_{L}/n_{C}$. Since the Dirac point and the Fermi level are both moving in one direction, namely to lower energies, the distance between them does not shorten remarkably.
Conduction band spectrum
========================
For $n_{H}<n_{SR}$, the presence of attracting impurities does not distort significantly the quasiparticle dispersion in the valence band. In contrast, impurities noticeably affect the conduction band. The self-energy $\bm{\Sigma}$, which enters the Dyson equation $$\bm{G}=\bm{g}+\bm{g}\bm{\Sigma}\bm{G}$$ for the averaged over impurity distributions Green’s function of the disordered system $$\bm{G}=\biglb<(E-\bm{H})^{-1}\bigrb>$$ can not be attributed only to the impurity scattering in the actual experiment of Ref. , because the linewidth of the ARPES data is not negligible even at the absence of impurities. To deal with this issue, we assume that all other sources of its broadening, except the controlled amount of hydrogen atoms, can be roughly described by a constant dumping term, $$\bm{\Sigma}\approx\bm{\Sigma}_{im}-i\Gamma\bm{I},\qquad \Gamma>0,
\label{sg}$$ where $\bm{\Sigma_{im}}$ is those part of the self-energy, which results from the impurity scattering, and $\bm{I}$ is the identity matrix. A comparison of the density of states obtained by the numerical simulation with the one resulted from the standard average T-matrix method showed that at $n_{H}<n_{SR}$ this approximation works rather well in the vicinity of the impurity resonance.[@persk] Within this single-site approximation, the impurity-induced part of the self-energy is diagonal in lattice sites and sublattices for point defects, $$\bm{\Sigma}_{im}\approx\frac{\frac{n_{H}}{n_{C}}V_{L}}{1-(1-\frac{n_{H}}{n_{C}})V_{L}g_{0}(E)}\bm{I}.
\label{tm}$$ Thus, taking into account both assumptions, the total self-energy is also diagonal, $$\bm{\Sigma}\approx\Sigma(E)\bm{I}.$$ According to the conventional expression for the spectral function, $$A(E,\bm{k})\approx-\frac{1}{\pi}\operatorname{Im}\frac{1}{E-\Sigma(E)-v_{F}k},
\label{sf}$$ where $$v_{F}=\frac{\sqrt{3}at}{2}$$ is the Fermi velocity, the wave vector $\bm{k}$ is counted from the corresponding Dirac point, and only one branch, which is manifested in the actual experiment, is retained, it is straightforward to put down corresponding expressions for the inverse width of the momentum distribution curve at its half-height, $$L(E)\approx -\frac{v_{F}}{2\operatorname{Im}\Sigma(E)},$$ and for the inverse of its median $$k^{-1}(E)\approx\frac{v_{F}}{E-\Sigma(E)}.$$ Both this quantities are taken at the Fermi level, $L_{F}\equiv
L(E_{F})$ and $k_{F}^{-1}\equiv 1/k(E_{F})$, and depicted in Fig. \[f2\] in their dependence on the hydrogen coverage. The Fermi level position is calculated from Eq. (\[ef\]), and the external damping $\Gamma$ is chosen to be $0.03$ eV.
![The mean free path $L_{F}$ (solid line) and the inverse wave vector $k_{F}^{-1}$ (dashed line) at the Fermi level *vs* the hydrogen coverage.[]{data-label="f2"}](f2){width="96.00000%"}
The ARPES linewidth analysis given in Fig. 3(c) of Ref. can be compared with Fig. \[f2\]. Since the inverse momentum width $L(E)$ corresponds to the magnitude of the mean free path of charge carriers, the Ioffe-Regel criterion[@ioffe] is violated for states at the Fermi level, *i.e.* $k_{F}L_{F}=1$, at somewhat higher hydrogen coverage, namely, around $9.5\times 10^{12}$ cm$^{-2}$, than in the experiment – $6.5\times 10^{12}$ cm$^{-2}$. This discrepancy, for the most part, arise from the noticeably reduced wave vectors in the conduction band of the sample containing no hydrogen. The reduction is evident in comparison with the dispersion in the valence band, or with the idealized graphene (\[ham0\]). The experimentally obtained quasiparticle dispersion looks already distinctly distorted before any hydrogen deposition. The distortion can be attributed to the possible presence of unidentified impurities in the system.[@roten] To a certain extent, this issue can be crudely addressed by a corresponding adjustment of the hopping integral $t$ in the Hamiltonian. Results of such quick guesstimate are shown in Fig. \[f3\].
![The mean free path $L_{F}$ (solid line) and the inverse wave vector $k_{F}^{-1}$ (dashed line) at the Fermi level *vs* the hydrogen coverage for $t=3.5$ eV, $V_{L}=-35$ eV, and $\Gamma=0.04$ eV.[]{data-label="f3"}](f3){width="96.00000%"}
The deliberate adjustment of the transfer integral $t$ pushes $k_{F}^{-1}$ to where it should be at the zero hydrogen coverage. Correspondingly, the Ioffe-Regel criterion is violated at lower hydrogen coverage matching the experimental data. However, such a forthright measure will be suitable only in the Fermi level vicinity. It is more consistent to work out a proper model of the host system that is capable in simulating its experimentally obtained spectrum. Anyhow, the resulting adjustment of the zero coverage Fermi wave vector moves the entire curve for the inverse Fermi vector upwards and, thus, diminish the impurity concentration, at which the Ioffe-Regel criterion ceases to hold.
It should be mentioned that, according to the routine of the spectrum rearrangement analysis, one should expect the Fermi level to enter inside a domain of localized states at $k_{F}L_{F}\sim\sqrt{3}/2$. This will happen for the idealized host spectrum at the hydrogen coverage around $10.5\times 10^{12}$ cm$^{-2}$, which is a little bit higher than the coverage corresponding to $k_{F}L_{F}=1$. With further increase in the impurity concentration, the average T-matrix approximation (\[tm\]) becomes non-valid because of the increased scattering on impurity clusters, and the approach outlined above is only suitable for signaling a strong localization at the Fermi level. On the other side, the experimentally detected sharp increase in the sample resistance, see Fig. 3(b) of the Ref. , also succeeds the Ioffe-Regel criterion violation.
Within the Kubo approach the zero temperature conductivity can be written as follows,[@minca] $$\begin{gathered}
\tilde{\sigma}_{cond}=\left(\frac{e^{2}}{h}\right)\sigma_{cond}, \\
\sigma_{cond}=\frac{2}{\pi}\left[1+(\cot\varphi_{F}+\tan\varphi_{F})\Bigl(\frac{\pi}{2}-\varphi_{F}\Bigr)\right], \label{kub}\end{gathered}$$ where $$\varphi_{F}=\arg(E_{F}-\Sigma(E_{F})).$$ The dependence of the dimensionless conductivity $\sigma_{cond}$ on the hydrogen coverage is plotted in Fig. \[f4\].
![Logarithm of the dimensionless zero temperature conductivity *vs* the hydrogen coverage.[]{data-label="f4"}](f4){width="96.00000%"}
It is evident from Fig. \[f4\] that the conductivity is falling down nearly exponentially with increasing the impurity concentration. Indeed, the Kubo formula and the average T-matrix approximation are becoming not so reliable as approaching the mobility edge. Thus, the zero temperature conductivity can not be satisfactorily described by Eqs. (\[tm\]) and (\[kub\]) close to the MIT.
The concentration dependence of the energy distribution curve at the Fermi wave vector $k_{F}$ can be easily re-created for the Lifshitz impurity model with the help of Eqs. (\[ef\]), (\[sg\]), (\[tm\]), and (\[sf\]). The corresponding magnitude, $A(E,k_{F})$, is given in Fig. \[f5\] as a function of the binding energy $E-E_{F}$ and the hydrogen coverage $n_{H}$.
![Evolution of the energy distribution curve at $k_{F}$ with increasing the hydrogen coverage.[]{data-label="f5"}](f5){width="96.00000%"}
It is clearly seen from the Figure that the energy width of the quasiparticle peak at $k_{F}$ is gradually increasing with increasing the hydrogen coverage. Close to the critical impurity concentration for the MIT, the width of the energy distribution curve occupies nearly all conduction band below the Fermi level, which indicates the presence of a developed quasigap in the spectrum. This broadening translates into the complete breakdown of the quasiparticle picture near the MIT. The characteristic curve shape consisting of two competing peaks reproduces well the experimentally observable one, which can be seen in Fig. 2(c) of Ref. . At that, the peak, which is developing with increasing the hydrogen coverage, is connected with the impurity resonance.
Finally, we render the density plots of the spectral function (\[sf\]) at hydrogen coverages that are close to the critical one for the MIT in order to reproduce Fig. 1(d) of Ref. . In Fig. \[f6\] the Fermi level is about to enter the spectral region near the resonance energy, where the quasigap is forming due to strong impurity scattering. In Fig. \[f7\] the Fermi level is already inside the developed quasigap. The pattern of Fig. \[f7\] suggests the apparent upturn of the dispersion as approaching the Fermi level of the system, which has been noticed in the experiment. It is worth mentioning here that point defects does not cause a uniform broadening of the spectral function, on the contrary, the impurity-induced broadening is most pronounced in the vicinity of the resonance energy $E_{r}$ in this case.
![Density plot of the spectral function at $n_{H}=5\times 10^{12}$ cm$^{-2}$.[]{data-label="f6"}](f6){width="96.00000%"}
![Density plot of the spectral function at $n_{H}=10\times 10^{12}$ cm$^{-2}$.[]{data-label="f7"}](f7){width="96.00000%"}
Conclusion
==========
To summarize, we have established that point defects are introducing a new length parameter into the impure graphene. This length parameter results from the presence of a single-impurity resonance and shows up as the effective radius of a single-impurity state. When the average distance between impurities decreases up to this effective radius, the quasiparticle spectrum undergoes the cardinal rearrangement. The spectrum rearrangement is manifested in the opening of the quasigap around the impurity resonance energy in the spectrum. The quasigap progressively widens with increasing the impurity concentration. The Fermi level moves due to the doping effect of impurities, as well as confining the quasigap mobility edges are moving due to its expansion. If the Fermi level crosses one of the mobility edges and enters the quasigap, where states are localized, the MIT of the Anderson type takes place in the disordered system.
We have presented arguments, which confirm that the above scenario can be considered as a sounding candidate for the explanation of the experimentally observed MIT in graphen dosed with hydrogen. Even employing the simple Lifshitz model for the impurity centers, it appears possible to achieve a semi-quantitative interpretation of the experimental data. The MIT of the Anderson type in this case is prompted by the gradual lowering of the Fermi level due to the acceptor effect of the hydrogen atoms combined with the persistent raising of the mobility edge due to the ongoing spectrum rearrangement. Thus, the spectrum rearrangement acts as the main cause of the MIT, and, as a common phenomenon in semiconductors, provides the basis for understanding the physics of the process. Indeed, the oversimplified impurity model is not able to convey all the detail of the system’s behavior. To fulfill this task, more sophisticated impurity models are required. Among them the two-parametric s-d model looks like a more natural choice for the deposited hydrogen atom. However, more sophisticated impurity models should not change the general physics of the transition process, which is already captured by the Lifshitz impurity model.
Authors are grateful to Eli Rotenberg for reading the manuscript and inspiring comments.
This work was supported by the SCOPES grant No. IZ73Z0-128026 of Swiss NSF, the SIMTECH grant No. 246937 of the European FP7 program, the State Program “Nanotechnologies and Nanomaterials”, project No. 1.1.1.3, and by the Program of Fundamental Research of the Department of Physics and Astronomy of the National Academy of Sciences of Ukraine.
[99]{}
A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov, and A. K. Geim, Rev. Mod. Phys. [**81**]{}, 109 (2009).
J. H. Bardarson, J. Tworzydlo, P. W. Brouwer, and C. W. J. Beenakker, Phys. Rev. Lett. [**99**]{}, 106801 (2007).
K. Kechedzhi, E. McCann, V. I. Fal’ko, H. Suzuura, T. Ando, and B. L. Altshuler, Eur. Phys. J. Special Topics, [**148**]{}, 39 (2007).
F. V. Tikhonenko, A. A. Kozikov, A. K. Savchenko, and R. V. Gorbachev, Phys. Rev. Lett. [**103**]{}, 226801 (2009).
J. Bang, and K. J. Chang, Phys. Rev. [**81**]{}, 193412 (2010).
J. H. Bardarson, M. V. Medvedyeva, J. Tworzydlo, A. R. Akhmerov, and C. W. J. Beenakker, Phys. Rev. B [**81**]{}, 121414 (2010).
K. S. Novoselov, A. K. Geim, S. V. Morozov, D. Jiang, M. I. Katsnelson, I. V. Grigorieva, S. V. Dubonos, and A. A. Firsov, Nature [**438**]{}, 197 (2005).
I. L. Aleiner, and K. B. Efetov, Phys. Rev. Lett. [**97**]{}, 236801 (2006).
P. M. Ostrovsky, I. V. Gornyi, A. D. Mirlin, Phys. Rev. Lett. [**98**]{}, 256801 (2007).
G. G. Naumis, Phys. Rev. B 76, 153403 (2007).
M. Amini, S. A. Jafari, and F. Shahbazi, Europhysics Lett. [**87**]{}, 37002 (2009).
A. Lherbier, B. Biel, Y.-M. Niquet, and S. Roche, Phys. Rev. Lett. [**100**]{}, 036803 (2008).
Y. Song, H. Song, and S. Feng, arXiv:1007.3309 (unpublished).
J.-H. Chen, C. Jang, S. Adam, M. S. Fuhrer, E. D. Williams, and M. Ishigami, Nature Phys. [**4**]{}, 377 (2008).
E. R. Mucciolo, and C. H. Lewenkopf, J. Phys.: Condens. Matter [**22**]{}, 273201 (2010).
N. M. R. Peres, Rev. Mod. Phys. [**82**]{}, 2673 (2010).
S. Das Sarma, S. Adam, E. H. Hwang, and E. Rossi, arXiv:1003.4731 (unpublished).
V. Ugarte, V. Aji, and C. M. Varma, arXiv:1007.3533 (unpublished).
A. Bostwick, J. L. McChesney, K. V. Emtsev, T. Seyller, K. Horn, S. D. Kevan, and E. Rotenberg, Phys. Rev. Lett. [**103**]{}, 056404 (2009).
F. Withers, M. Dubois, and A. K. Savchenko, Phys. Rev. B [**82**]{}, 073403 (2010).
J.-H. Chen, W. G. Cullen, C. Jang, M. S. Fuhrer, and E. D. Williams, Phys. Rev. Lett. [**102**]{}, 236805 (2009).
Z. H. Ni1, L. A. Ponomarenko, R. R. Nair, R. Yang, S. Anissimova, I. V. Grigorieva, F. Schedin, Z. X. Shen, E. H. Hill, K. S. Novoselov, and A. K. Geim, Nano Lett., Article ASAP (2010).
M. A. Ivanov, V. M. Loktev, and Yu. G. Pogorelov, Phys. Rep. [**153**]{}, 209 (1987).
I. M. Lifshits, S. A. Gredeskul, and L. A. Pastur, *Introduction to the Theory of Disordered Systems*, Wiley, N.Y. (1988).
Yu. V. Skrypnyk, and V. M. Loktev, Phys. Rev. B [**73**]{}, 241402(R) (2006).
S. S. Pershoguba, Yu. V. Skrypnyk, and V. M. Loktev, Phys. Rev. B [**80**]{}, 214201 (2009).
A. F. Ioffe, and A. R. Regel, Prog. Semicond. [**4**]{}, 237 (1960).
H. Kumazaki, and D. S. Hirashima, J. Phys. Soc. Jpn. [**75**]{}, 053707 (2006).
|
{
"pile_set_name": "ArXiv"
}
|
jytex.tex =1200 =17truept =6truein=8.5truein =1 =1 =0
\#1[[*Exercise*]{}: ]{} \#1[[*Exercise$^*$*]{}: ]{} \#1[[*Exercise$^{**}$*]{}: ]{} \#1 \[\] \#1 =msbm10 =eufm10 =cmss10 =cmss8 \#1 \#1 \#1\#2 \#1/\#2[.1em .5ex-.1em/-.15em .25ex]{} \#1/\#2[.1em .5ex-.1em/-.15em .25ex]{} \#1\#2
\#1[[\#1]{}]{} \#1[[\#1]{}]{} \#1[[\#1]{}]{} \#1[\^[.18ex-.06em \#1]{}]{} /\#1[.18ex-.68em \#1]{}
\#1[1.8ex-17mu \#1]{} \#1[ 1.5ex-22.8mu \#1]{} \#1 \#1 \#1\#2 \#1[$\underline{\smash{\hbox{#1}}}$]{} \#1[$\underline{\vphantom{y}\hbox{#1}}$]{} \#1[\_[\#1]{}]{} \#1[\^[\#1]{}]{} \#1[\_[|\#1]{}]{} \#1[\^[|\#1]{}]{} \#1\#2[[[d \#1]{}]{}]{} \#1\#2[([[d \#1]{}]{})]{} \#1\#2[[\#1\#2]{}]{} \#1\#2\#3[[\^[\#1]{}]{}\_[\#2\#3]{}]{} \#1\#2\#3[[’\^[\#1]{}]{}\_[\#2\#3]{}]{} \#1\#2[[(\#1\#2)]{}]{} \#1\#2[[\_[\#1]{}]{}\^[\#2]{}]{} \#1\#2[[\^[\#1]{}]{}\_[\#2]{}]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1\#2 \#1\#2 \#1\#2\#3 \#1\#2\#3\#4[[Z|[\#1\#3]{}[\#2\#4]{}|]{}]{} \#1\#2\#3\#4[[Z’|[\#1\#3]{}[\#2\#4]{}|]{}]{} \#1[$^{#1}$]{} \#1[[[D]{}\_]{}\_[\#1]{}]{} \#1[[\_]{}\_[\#1]{}]{} \#1[[\_[\#1]{}]{}]{} \#1\#2[[\_[\#1]{}\^[\#2]{}]{}]{} \#1\#2[[\_[\#1]{}\^[\#2]{}]{}]{} \#1 \#1\#2\#3\#4\#5\#6[([\#1\#4]{}[\#2\#5]{}[\#3\#6]{})]{} \#1\#2\#3\#4\#5\#6[([\#1\#4]{}[\#2\#5]{}[\#3\#6]{})]{} j\#1\#2\#3\#4\#5\#6[ ]{} \#1\#2\#3\#4\#5\#6[{}]{} \#1\#2\#3\#4\#5\#6\#7\#8\#9[{}]{}
?
\#1\#2\#3[[*Ann. Phys.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Can. J. Phys.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Comm. Math. Phys.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Class. Quant. Grav.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*J. Chem. Phys.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Int. J. Mod. Phys.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*J. Math. Phys.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*J. Phys.*]{} [**A[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Mod. Phys. Lett.*]{} [**A[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Nucl. Phys.*]{} [**B[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Phys. Lett.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Phys. Rep.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Phys. Rev.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Phys. Rev.*]{} [**A[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Phys. Rev.*]{} [**B[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Phys. Rev.*]{} [**D[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Phys. Rev. Lett.*]{} [**\#1**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Prog. Theor. Phys.*]{} [**\#1**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Theor. Mat. Phys.*]{} [**\#1**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Rev. Mod. Phys.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Z. f. Naturf.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Z. f. Phys.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{}
\#1\#2\#3[[*Ann. Sci. École Norm. Sup. (Paris)*]{} [**[\#1]{}**]{} (\#2) \#3]{} \#1\#2\#3[[*Ann. Inst. H. Poincaré (Paris)*]{} [**[\#1]{}**]{} (\#2) \#3]{} \#1\#2\#3[[*Comptes Rend. Acad. Sci. (Paris)*]{} [**[\#1]{}**]{} (\#2) \#3]{} \#1\#2\#3[[*Proc. Roy. Soc.*]{} [**A[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Proc. Camb. Phil. Soc.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{}
\#1\#2\#3[[*Abh. Math. Sem. Ham.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Am. Math. Mon.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Ann. der Math.*]{} [**[\#1]{}**]{} ([\#2]{}) \#3]{} \#1\#2\#3[[*Can. J. Math.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Compos. Math.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Duke Math. J.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Invent. Math.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*J. Diff. Geom.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*J. Lond. Math. Soc.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Math. Ann.*]{} [**[\#1]{}**]{} ([\#2]{}) \#3]{} \#1\#2\#3[[*Math. Zeit.*]{} [**[\#1]{}**]{} ([\#2]{}) \#3]{} \#1\#2\#3[[*Proc. Am. Math. Soc.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Proc. Edin. Math. Soc.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Proc. Lond. Math. Soc.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Proc. Glasgow Math. Ass.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Quart. J. Math.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Quart. J. Pure and Appl. Math.*]{} [**[\#1]{}**]{} ([\#2]{}) \#3]{} \#1\#2\#3[[*Rend. Circ. Mat. Palermo*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Russ. Math. Surveys*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Topology*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Trans. Am. Math. Soc.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*Z.f.Math.*]{} [**[\#1]{}**]{} ([\#2]{}) \#3]{}
20truept 100truept 15truept 7truept 60truept 7truept
The specific form of the constant term in the asymptotic expansion of the heat-kernel on an axially-symmetric space with a codimension two fixed-point set of conical singularities is used to determine the conformal change of the effective action in four dimensions. Another derivation of the relevant coefficient is presented.
5truept 100truept
=0
**1. Introduction and geometry**
================================
In previous work \[\] we have used the transformation of the effective action, obtained by integrating the conformal anomaly in two dimensions, to relate the effective actions on regions of the two-sphere and plane. An extension to four dimensions is technically feasible. The relevant transformations when the manifolds have no boundary were given some time ago \[\], and, when a boundary exists, in \[.\] (See also \[\]). However, as a preliminary, it is necessary to extend the analysis to manifolds with conical, or vertex, singularities and this is the object of the present paper.
The general method depends upon knowing the constant term in the heat-kernel expansion, in this case the $C_2$ coefficient. When the manifold possesses a singular O(2) fixed-point set of simple conical type, the extra term in $C_2$ has been determined by Fursaev \[\]. In \[\] the expression was rederived and its conformal behaviour discussed. A further analysis is presented in a later section.
We enlarge on the geometry. Let ${\cal N}$ be a totally geodesic submanifold of $\man$, in particular a fixed-point set of codimension two. The global symmetry group is O(2), generated by the circular Killing vector $\pa/\pa\phi$, making $\man$ axially symmetric. Reperiodising the polar angle, $\phi$, from $2\pi$ to $\be$ turns $\man$ into $\man_\be$, a space with a simple conical singularity of ‘extent’ ${\cal N}$. Close to ${\cal N}$, $\man_\be$ approximates to the product ${\cal C}_\be\times{\cal N}$ where ${\cal C}_\be$ is a cone of angle $\be$.
As hyper-cylindrical coordinates of a point $P$ on $\man$ we take (i) the distance, $r$, from $P$ along the normal geodesic to its foot on ${\cal N}$, (ii) the coordinates, $y^a$, of this foot and (iii) the angle $\phi$, between the tangent to this geodesic, at $r=0$, and some fiducial normal vector (parallely propagated along ${\cal N}$). $\phi$ is the rotation angle ‘about’ ${\cal N}$ and, with $r$, makes up the coordinates of the normal space. Close to ${\cal N}$, $r$ and $\phi$ are the usual plane polar coordinates.
The metric of $\man$ is generally $g_{\mu\nu}dx^\mu dx^\nu$ and in cylindrical coordinates is taken to be $$ds^2= dr^2+f(r,y)d\phi^2+g_{ab}(y,r)dy^ady^b
\eql{metricm}$$where $f(r,y)$ is an even function of $r$ and tends to $r^2$ as $r\to0$. Also $g_{ab}(y,0)=h_{ab}(y)$, the metric on ${\cal N}$. The region $r\le b$ forms a tube, $U{\cal N}$, surrounding ${\cal N}$. If $g_{ab}(y,r)=h_{ab}(y)$, all the surfaces $r=$ const., $\phi=$ const. are totally geodesic.
**2. Heat-kernel coefficients**
===============================
The integrated heat-kernel expansion is written $$K_\be(t)\sim{1\over(4\pi t)^{d/2}}\sum_{n=0,1/2,\ldots}^\infty C_nt^n =
{1\over(4\pi t)^{d/2}}\sum_{n=0,1/2,\ldots}^\infty (A_n+F_n)t^n
\eql{hker}$$ where $A_n$ is the usual volume integral, over $\man_\be$, of a local, scalar density involving the curvature of $\man$. The $F_n$ are the due to the conical singularity and are integrals over ${\cal N}$. In particular, $F_2$ is given by $$F_2=\int_{\cal N}f_2\,h^{1/2}d^{d-2}y =\int_{\cal N} f_2\,d\,\vol_{\cal N}(y)
\eql{eff2}$$ with the integrand, $$f_2={\pi\over B}(B^2-1)f_{2,1}+{\pi\over360 B}(B^4-1)f_{2,2}
\eql{coeff2}$$where $B=2\pi/\be$ and $$\eqalign{
f_{2,1}=&\big({1\over6}-\xi\big)R+\la_1\big(\bka{\bf.}\bka
-2\tr(\bka{\bf.}\bka)\big)\cr
f_{2,2}=&\bigg(2R_{\mu\nu\rho\si}({\bf n}^{\mu}{\bf .n}^{\rho})({\bf n}^{\nu}
{\bf .n}^{\si})-R_{\mu\nu}{\bf n}^\nu{\bf .n}^\mu-{1\over2}\bka{\bf.}\bka\cr
&\hspace{*************************}
+\la_2\big(\bka{\bf.}\bka-2\tr(\bka{\bf.}\bka)\big)\bigg).\cr}
\eql{coeff22}$$ The constant coefficients, $\la_1$ and $\la_2$, of the conformally covariant combination $\big(\bka{\bf.}\bka-2\tr(\bka{\bf.}\bka)\big)$ are unknown. The ${\bf n}^\mu_i$ are the normals to ${\cal N}$ and we may take $n^\mu_i{n_\mu}_j=\de_{ij}$.
Although the extrinsic curvatures $\bka$ are zero for a fixed-point set, it is necessary to retain them when making [*general*]{} conformal transformations. A derivation of () is given in section 4.
**3. Conformal transformations and the effective action**
=========================================================
Under a Weyl rescaling, $g_{\mu\nu}\to e^{-2\om}g_{\mu\nu}$, $\man_\be\to{\overline\man_\be}$ and ${\cal N}\to\overline{\cal N}$. In order to preserve the topology, the transformation function $\om(r,\phi,y)$ must have period $\be$ in $\phi$. In general, the O(2) symmetry is destroyed by the rescaling. $\overline{\cal
N}$ is a submanifold of ${\overline\man_\be}$, but not a totally geodesic one. It has nonzero extrinsic curvatures.
When evaluated on ${\cal N}$, or equivalently on ${\overline{\cal N}}$, $\om$ becomes independent of the coordinates of the normal space so that, at the singularity, ${\overline\man_\be}\to{\overline{\cal C}_\be}\times{\overline{\cal N}}$ where ${\overline{\cal C}_\be}$ is a cone of angle $\be$ scaled by a factor depending on its position in ${\overline{\cal N}}$. We could say that ${\overline\man_\be}$ has a ‘squashed’ conical singularity.
We turn now to an evaluation of the change in the renormalised effective action, $W_\rR$, under a conformal transformation for conformal coupling, $\xi=1/6$.
The technique used in the present paper is that explained in \[\], involving the conformal transformation of the constant term in the heat-kernel expansion. The general formula is $$W_\rR[e^{-2\om}g]-W_\rR[g]=\lim_{d\to
d'}(4\pi)^{-d/2}{C^{(d)}_{d'/2}[e^{-2\om}g]-C^{(d)}_{d'/2}[g]\over d-d'},
\eql{effct}$$ where $d'$ is the dimension of $\man_\be$ and $d$ is an arbitrary dimension. We set $d'=4$.
This method, in contrast to that of integrating the conformal anomaly, requires an application of finite conformal transformations in $d$ dimensions. In the present instance there is little to choose between the two methods so far as effort goes.
The total coefficient $C_2$ contains the standard volume term $A_2$, which is dealt with in \[\]. In the present paper we are interested only in the effect of $F_2$.
The transformations needed are $$\eqalign{
&R_{\mu\nu}{\bf n}^\nu{\bf .n}^\mu\to
e^{2\om}\big(R_{\mu\nu}{\bf n}^\nu{\bf .n}^\mu+(d-2)\om_{\mu\nu}{\bf n}^\mu
{\bf .n}^\nu+2\De_2\om+2(d-2)\De_1\om\big)\cr
\noalign{\vskip 4truept}
&R_{\mu\nu\rho\si}({\bf n}^{\mu}{\bf .n}^{\rho})({\bf n}^{\nu}{\bf .n}^{\si})
\to e^{2\om}\big(R_{\mu\nu\rho\si}({\bf n}^{\mu}{\bf .n}^{\rho})
({\bf n}^{\nu}{\bf .n}^{\si})+
2\om_{\mu\nu}{\bf n}^\mu{\bf .n}^\nu+2\De_1\om\big),\cr}$$where $\om_{\mu\nu}=\om\cd{\mu\nu}+\om\od\mu\om\od\nu$, and $$\eqalign{
&\bka_{ab}\to e^{-\om}\big(\bka_{ab}+h_{ab}{\bf n}^\mu\om\od\mu\big),\cr
&\bka\to e^{\om}\big(\bka+(d-2){\bf n}^\mu\om\od\mu\big)\cr}$$ for codimension 2. Also, for reference, $h^{1/2}\to e^{(2-d)\om}h^{1/2}$.
The change in $h^{1/2}f_{2,2}$, for example, is $$\eqalign{
(4-d)\bigg(\om f_{2,2}+&2\om\uod{\mu}\om\od{\mu}+\om\cd{\mu\nu}{\bf n}^\mu
{\bf .n}^\nu+(1+d/2)\om\od\mu\om\od\nu{\bf n}^\mu{\bf .n}^\nu+\bka{\bf.}
{\bf n}^\mu\om\od\mu+\cr
&\la_2(d-2)\big(\om\od\mu\om\od\nu{\bf n}^\mu{\bf .n}^\nu+2\bka{\bf.}{\bf n}^
\mu\om\od\mu\big)\bigg)+O\big((4-d)^2\big)-2\widehat\De_2\om.\cr}$$ all multiplied by $h^{1/2}$. $\widehat\De_2$ is the Laplacian intrinsic to ${\cal N}$.
It is then easy to show that the change in the integral () is proportional to $(d-4)$ so that, from (), our final result can be written $$W_\rR[e^{-2\om}g]-W_\rR[g]= \int_{\cal N}\De w\,d\vol_{\cal N}(y)$$with $$\De w=-\om f_2+{2\pi\over B}\sum_{k=1}^2(B^k-1)\De w_k,
\eql{deltaw}$$ where, after setting $d=4$ and $\bka=0$, $$\De w_1=-{\la_1\over8\pi^2}\,\om\od\mu\om\od\nu{\bf n}^\mu{\bf .n}^\nu
\eql{ch1}$$ and $$\De w_2=-{1\over16\pi^2}\bigg(2\om\uod{\mu}\om\od{\mu}+
\om\cd{\mu\nu}{\bf n}^\mu{\bf .n}^\nu+
(2\la_2+3)\,\om\od\mu\om\od\nu{\bf n}^\mu{\bf .n}^\nu\bigg).
\eql{ch2}$$
The constants $\la_1$ and $\la_2$ remain undetermined. The derivation of the corresponding terms in the presence of a conventional boundary is somewhat complicated, involving either a direct solution of the differential equations or the cleaner, but still longish, functorial properties, \[\].
Note that, even if $\overline{\cal N}$ were an O(2) fixed-point subspace of $\overline{\man}_\be$, there would still be a contribution from the $-\bka{\bf.}\bka/2$ term in (). In this case $\om\od\mu{\bf n}^\mu$ is zero removing dependence on $\la_1$ and $\la_2$ in the final answer.
**4. The heat-kernel expansion**
================================
For completeness we present a derivation of () with (), similar to that in \[\]. The basic idea is that, close to ${\cal N}$, the heat-kernel on $\man_\be$ is approximated by that obtained by a process of [*reperiodisation*]{} from the heat-kernel on $\man$. This has been described and used in earlier work \[\] where the Green functions in some specific curved spaces, de Sitter and Schwarzschild, were considered.
We are interested in the integrated, diagonal $K_{\be}$ $$K_{\be}(t)=\int_{\man_\be}K_{\be}(y,{\bf r};y,{\bf r},t)$$ and its asymptotic expansion ().
Because $\man_\be$ is identical to $\man$ off ${\cal N}$, the local heat-kernel expansions of $K_\be$ and $K_{2\pi}$ will be the same in $\man_\be-{\cal N}$. It is therefore sufficient for the asymptotic expansion to write, thickening out ${\cal N}$ by setting $\man_\be=U{\cal N}\cup\overline {U{\cal N}}$, $$K_{\be}(t)=\int_{U{\cal N}}K_{\be}(y,{\bf r};y,{\bf r},t)+
\int_{\overline {U{\cal N}}}K_{2\pi}(y,{\bf r};y,{\bf r},t)
\eql{split1}$$ valid up to terms exponentially small as $t\to0$.
Because of the O(2) symmetry, the Laplacian heat-kernel $K_{2\pi}(y,r,\phi;y',r',\phi',t)$ depends on the polar angles through the difference $\phi-\phi'$ only, and the approximation in the narrow tube $U{\cal N}$ is conveniently written as a contour integral, $$K_{\be,\de}(\phi-\phi',t)\approx\int_{A}K_{2\pi}(\al,t)\,
P(\al-\phi+\phi';\be,\de)\,d\al
\eql{perint}$$ where $P(\al;\be,\de)$ is the reperiodising function, $$P(\al;\be,\de)={e^{\pi i\al(2\de-1)/\be}
\over2i\be\sin\big(\pi\al/\be\big)}.
\eql{perfunc}$$ As is our wont, a phase change $\de$, $0<\de\le1$, has been included. This will not be made use of here but allows one to discuss fluxes running along the singularity. The contour $A$ has two parts. In the upper half-plane it runs from $(\pi-\ep)+i\infty$ to $(-\pi+\ep)+i\infty$ and in the lower half-plane from $(-\pi+\ep)-i\infty$ to $(\pi-\ep)-i\infty$.
It is helpful to exhibit the arguments now,
$$K_{\be}(y,{\bf r};y',{\bf r}',t)\approx{1\over\be}\int_{A}K_{2\pi}(y,{\bf
r};y',
\brho',t){e^{2\pi i\al/\be}\over e^{2\pi i\al/\be}-e^{2\pi i(\phi-\phi')/\be}}
\,d\al\eql{Kbetacont}$$
$$K_{\be,\de}(y,{\bf r};y',{\bf r}',t)\approx{1\over2\be i}
\int_{A}\,K_{2\pi}(y,{\bf r};y',{\bf r}'_\al,t){e^{\pi i(\al-\phi+\phi')
(2\de-1)/\be}\over\sin\big(\pi(\al-\phi+\phi')/\be\big)}\,d\al.
\eql{perintd}$$ The ‘complex point’ ${\bf r}'_\al$ has polar coordinates $(r',\al+\phi)$.
We also give the expression for $\de=1$, no phase change, $$K_{\be}(y,{\bf r};y',{\bf r}',t)\approx{1\over2\be
i}\int_{A}\,K_{2\pi}(y,{\bf r};
y',{\bf r}'_\al,t)\cot\big({\pi(\al-\phi+\phi')\over\be}\big)\,d\al
\eql{perintg}$$ where the symmetry of the contour under reversal of $\al$ has been used and the fact that $K_{2\pi}(\phi,t)=K_{2\pi}(-\phi,t)$ by orientation arguments. (The metric () is unchanged under reversal of $\phi$. If the cone were spinning, these arguments would have to be revised.)
To isolate the effect of the singularities, the contour $A$ is deformed to a small loop around the origin plus two infinite ‘vertical’ curves, labelled $A'$, which, to avoid problems, are taken to skirt the origin. The small loop evaluates by the pole at $\al=0$ to $K_{2\pi}$ and so one has the split $$K_{\be}\approx K_{2\pi}+K'_{\be}
\eql{split}$$ where $K'_{\be}$ is the effect of the singularity and is given by a formula like () but now over the $A'$ contour.
Effecting the split (), and combining with (), one finds $$K_\be(t)\sim{\be\over2\pi}K_{2\pi}(t)+
{1\over2\be i}\int_{A'}\int_{U{\cal N}}\,K_{2\pi}(y,{\bf r};y,{\bf r}_\al,t)
\,\cot\big({\pi\al\over\be}\big)\,d\al
\eql{tracecont}$$where $K_{2\pi}(t)$ is the integrated kernel on the smooth manifold $\man$ and has the standard asymptotic expansion. It will not be considered further. The $\be/2\pi$ is a volume factor that reflects the O(2) symmetry. The second term is the effect of the singularity. We denote it by $K'_\be(t)$. The point $(y,{\bf r}_\al)$ is $(y,{\bf r})$ rotated through $\al$ about ${\cal N}$.
As explained in our earlier works, the contour $A'$ can be replaced by a small clockwise loop around the origin, and we will imagine this to have been done.
Donnelly \[\] has elucidated the asymptotic expansion of $$K_{2\pi}(\phi,t)=\int_{\man}\,K_{2\pi}(y,{\bf r};y,{\bf r}_\phi,t)=
{2\pi\over\be}\int_{\man_\be}\,K_{2\pi}(y,{\bf r};y,{\bf r}_\phi,t)
\eql{parhk}$$ and his results can be substituted directly into () for, as we see, the complex activity takes place in the normal space, the point $y$ of ${\cal N}$ being a spectator.
It is clear from the classic results of Minakshisundaram and Pleijel that, up to exponentially small terms, the integral in () gets its value from the fixed-point set ${\bf r}={\bf r}_\al$ from ${\cal N}$, and so, following Donnelly, for $\phi\ne0$, $$K_{2\pi}(\phi,t)\sim{1\over(4\pi t)^{(d-2)/2}}\sum_{n=0}^\infty t^n\int_
{\cal N}b_n(\phi,y)\,d\,\vol_{\cal N}(y).
\eql{Donn}$$ Substitution into (), after setting $\phi\to\al$, gives $$K'_\be(t)\sim{1\over(4\pi t)^{(d-2)/2}}{1\over4\pi i}
\sum_{n=0}^\infty t^n\int_{\cal N}
\int_{A'}
b_n(\al,y)\cot\big({\pi\al\over\be}\big)\,d\al\,d\,\vol_{\cal N}(y)
\eql{contoursum}$$ and we now concentrate on the contour integral part of this equation $$b_n(y)={1\over4\pi i}\int_{A'}
b_n(\al,y)\cot\big({\pi\al\over\be}\big)\,d\al.
\eql{bcoeffn}$$
If $S(\phi)$ is the linear O(2) (actually SO(2)) action on the normal fibre, the general form of the coefficients is, \[\], $$b_n(\phi,y)={1\over|\det(1-S)|}b_n'(\phi,y)
\eql{coeffk}$$where $b_n'(\phi,y)$ is an O$(d-2)\times$O(2) invariant polynomial in the components of $T\equiv(1-S)\invo$, the curvature of $\man$ and its covariant derivatives.
We indicate the origin of () and (), \[\]. The local Minakshisundaram-Pleijel parametrix expansion is $$K_{2\pi}(x;x',t)\sim{e^{-\Om(x,x')/2t}\over(4\pi t)^{d/2}}\sum_{n=0}^\infty
a_n(x,x')t^n
\eql{MandP}$$ where $\Om(x,x')$ is half the square of the geodesic distance between $x$ and $x'$. This is substituted into () and the integral divided into one over ${\cal N}$, with coordinates $y$, and one over the normal fibre, with coordinates ${\bf r}$. Because of the exponential cutoff the integral is restricted to the tubular neighbourhood $U{\cal N}$. Transforming the fibre coordinates to normal coordinates, $x^i$, based at $y$, $2\Om$ becomes $\tilde
x\widetilde{(1-S)}(1-S)x$. (This is actually a diagonal form.) The remainder of the integrand, including the volume form and the $a_n$ coefficients, is expanded about the point $(y,0)$ and the integrals over the $x^i$ extended to $\pm\infty$, again up to exponentially small errors. Standard Gaussian integrals, familiar from perturbation theory, then yield (). The nontrivial O(2) tensor dependence of $b'_n$ originates in the expansion of the volume factor. We remark that Donnelly’s general expression shows that there is always the contribution $a_n(y,y)$ to $b'_n(\phi,y)$.
In particular, Donnelly calculates for any codimension (we include the $\xi R$ coupling) $$\eqalign{
b_0'(\phi,y)&=1\cr
b_1'(\phi,y)&=\big({1\over6}-\xi\big)R+{1\over6}R_i^i
+{1\over3}
R_{ijkl}T^{ji}T^{lk}+{1\over3}R_{ijkl}T^{jk}T^{li}-R_{ij}T^{ik}T^{jk}\cr}
\eql{Donncoeff}$$ where $$R_{ijkl}=R_{\mu\nu\rho\si}n^\mu_i n^\nu_j n^\rho_k n^\si_l\quad{\rm and}\quad
R_{ij}=R_{\mu\rho\nu\si}h^{\rho\si}n^\mu_i n^\nu_j.
\eql{normcurv}$$ $h_{\mu\nu}=g_{\mu\nu}-{\bf n}_\mu{\bf.n}_\nu$ is equivalent to the metric on ${\cal N}$, therefore $$R_{ij}=R_{\mu\nu}n^\mu_i n^\nu_j-
R_{\mu\rho\nu\si}{\bf n}^\rho{\bf.n}^\si n^\mu_i n^\nu_j$$ and $$R_i^i=R_{\mu\nu}{\bf n}^\mu{\bf. n}^\nu-
R_{\mu\rho\nu\si}({\bf n}^\rho{\bf.n}^\si)( {\bf n}^\mu{\bf. n}^\nu).$$
For codimension two the calculation is simplified by noting that $$R_{ijkl}={1\over2}R^{(2)}\big(\de_{ik}\de_{jl}-\de_{il}\de_{jk}\big),\quad
{\rm where}\quad
R^{(2)}=R_{\mu\nu\rho\si}({\bf n}^{\mu}{\bf .n}^{\rho})({\bf n}^{\nu}
{\bf .n}^{\si})$$so $$R_{ij}=R_{\mu\nu}n^\mu_i n^\nu_j-{1\over2}R^{(2)}\de_{ij}.$$
Substituting into () gives $$b_1'(\phi,y)=\big({1\over6}-\xi\big)R+{1\over6}R_i^i
-R^{(2)}\tr T^2-{1\over6}R^{(2)}(\tr T)^2
+R_{\mu\nu}n^\mu_i n^\nu_j(T^2)^{ij}.
\eql{beeone}$$ With respect to normal coordinates, $$S=\left(\matrix{\cos\phi&\sin\phi\cr-\sin\phi&\cos\phi\cr}\right);\qquad
T={1\over2(1-\cos\phi)}\left(
\matrix{1-\cos\phi&\sin\phi\cr-\sin\phi&1-\cos\phi\cr}\right)$$and $$T^2=-{1\over2(1-\cos\phi)}\left(
\matrix{\cos\phi&-\sin\phi\cr\sin\phi&\cos\phi\cr}\right).$$ In terms of indices $$T_{ij}={1\over2}\big(\de_{ij}+{\sin\phi\over1-\cos\phi}\ep_{ij}\big)
={1\over2}\big(\de_{ij}+\cot(\phi/2)\ep_{ij}\big).
\eql{indices}$$ The characteristic equation is $T^2-T=-{\bf 1}\,\det T=-\widetilde T T$.
One sees from the symmetry of the matrices that, in (), $(T^2)^{ij}$ can be replaced by $\de^{ij}\,\tr T^2/2$. Furthermore $\tr T=1$ giving, $$b_1'(\phi,y)={1\over6}\bigg((1-6\xi)R+\big(1+6\tr T^2\big)\big(
R_{\mu\nu}{\bf n}^\mu{\bf. n}^\nu-2R^{(2)}\big)\bigg).
\eql{beeone2}$$
These results give the explicit dependence on the angle $\phi$. If $\phi$ is replaced by the complex angle $\al$, the expressions can be substituted into the contour integral (). Noting that $|\det(1-S(\phi))|=2(1-\cos\phi)\to2(1-\cos\al)$ and $\tr T^2=-1+1/(1-\cos\phi)$, from (), () and () we encounter the polynomials, \[\], $$P_k(\be,\de)={1\over\be i}\int_{A'}\,{1\over(1-\cos\al)^k}
{\cos(\pi\al(2\de-1)/\be)\over\sin\big(\pi\al/\be\big)}\,d\al.
\eql{perintf}$$
A routine residue calculation gives $$\eqalign{
P_1(\be,\de)&= {1\over3}(B^2-1)-2B^2\si,\cr
\noalign{\vskip 5truept}
P_2(\be,\de)&={1\over90}(B^2-1)(B^2+11)
-{1\over3}B^2\si\big(B^2\si+2\big)\cr
\noalign{\vskip 5truept}
P_3(\be,\de)&= {1\over3780}(B^2-1)(2B^4+23B^2+191),\cr
\noalign{\vskip 5truept}
&\hspace{**}-{1\over90}B^2\si\big(B^4\si(2\si+1)-15B^2\si-24\big),\cr
P_4(\be,\de)&={1\over85050}(B^2-1)(B^2+11)(3B^4+10B^2+227),\cr
\noalign{\vskip 5truept}
&\hspace{**}-{1\over5670}B^2\si\big(B^6(3\si^2+4\si+2)+28B^4\si(2\si+1)
+294B^2\si+432\big),\cr}
\eql{eyes}$$where $\si=\de(1-\de)$. In this paper $\si=0$.
Combining terms in () we obtain () and () for $\xi=0$, as promised.
**5. The general coefficient**
==============================
A typical term in the general coefficient $b_n'(\al,y)$ has the form $$M_{ij\ldots kl}T^{ij}\ldots T^{kl}
\eql{gencoeff}$$ where $M_{\cdots}$ is an appropriate combination of the curvature, $R_{\mu\nu\rho\si}$, its covariant derivatives and the normal vectors $n^\mu_i$. Using (), a parity argument, or the symmetry of the $\al$-integral, shows that there can only be an even number of $\ep$-symbols so () reduces to a series of contractions of $M$. From () it follows that each pair of $\ep$-symbols will increase the order of the polynomials, $P_k$, by one.From dimensions, the maximum number of pairs in $b'_n$ equals $n$, an example being $(R_{ijkl}T^{ij}T^{kl})^n$, and so the general form of () is $$b_{n-1}(y)={2\pi\over B}\sum_{k=1}^n P_k(B)G_{nk}
\eql{genform}$$ where the $G_{nk}$ are integrals over ${\cal N}$ of a local, scalar expression constructed from the curvature of $\man$, its covariant derivatives and the normals $n^\mu_i$. Form () was given by Fursaev \[\]. It allows one to set up a functorial method on the lines of Branson and Gilkey \[\] for the determination of the coefficients.
As an aside we note that in the case of the $b_1$ coefficient, it was not necessary to use the symmetry of the $\al$-integral. The reduction of () to () was purely algebraic, something that will not happen for the higher coefficients. For example, the possible combination $********$ does not seem to reduce to traces algebraically.
**6. Comments**
===============
Donnelly’s expression can also be applied to the special case where $\be$ is an integral part of $2\pi$, $\be=2\pi/q$, $q\in\oZ$. This would involve a preimage summation which can be effected to yield the expected answer, as mentioned by Fursaev \[\].
A natural extension is to higher codimensions. However a simple process similar to that of periodisation does not appear to exist. If we think of codimension two as corresponding to a dihedral angle (with sides identified) then codimension three corresponds to a trihedral corner, and the heat-kernel for such a domain is unknown except for special cases.
Regarding the conformal transformation, in order to apply the result () with confidence, it would be helpful to have an independent check. This would entail finding two conformally related spaces, with conformally related singular fixed-point sets, on which one could determine the effective action, or at least that part due to the singularities. A possibility is the Einstein universe with a cosmic string \[\]. This is conformal to $\oR^4$ with a string.
Another possibility is to check the heat-kernel coefficients () themselves. This can be done in any dimension since the coefficients are universal. However it necessitates finding a tractable space with a deformed conical submanifold one with nonzero extrinsic curvatures.
**Acknowledgements**
====================
I would like to thank D.V.Fursaev for helpful remarks.
5truept 1truept
1truept
\[Donnelly\][H.Donnelly 161.]{} \[Dow\][J.S.Dowker [*Effective actions on spherical domains*]{}, [*Comm.Math.Phys*]{}, in the press.]{} \[Dow4\][J.S.Dowker .]{} \[KCD\][G.Kennedy, R.Critchley and J.S.Dowker .]{} \[Fur2\][D.V.Fursaev [*Spectral geometry and one-loop divergences on manifolds with conical singularities*]{}, JINR preprint DSF-13/94, hep-th/9405143.]{} \[HandE\][S.W.Hawking and G.F.R.Ellis [*The large scale structure of space-time*]{} Cambridge University Press, 1973.]{} \[DandK\][J.S.Dowker and G.Kennedy .]{} \[ChandD\][Peter Chang and J.S.Dowker .]{} \[FandM\][D.V.Fursaev and G.Miele .]{} \[Dowkerccs\][J.S.Dowker .]{} \[BandH\][J.Brüning and E.Heintze .]{} \[Cheeger\][J.Cheeger .]{} \[DandS\][J.S.Dowker and J.P.Schofield .]{} \[Dow2\][J.S.Dowker [*Heat kernels on curved cones*]{}, hep-th/9606002, to be published.]{} \[Dow3\][J.S.Dowker .]{} \[BandO\][M.R.Brown and A.Ottewill .]{} \[conft\][M.R.Brown ; R.J.Riegert ; E.S.Fradkin and T.Tseytlin ; L.Bukhbinder, S.Odintsov and A.Shapiro ; L.Bukhbinder, V.P. Gusynin and P.I.Fomin [*Yad. Phys.*]{} [**44**]{} (1986) 828 \[Sov.J.Nucl.Phys.[**44**]{} (1986) 534\]; J.S.Dowker .]{} \[Dowcone\][J.S.Dowker .]{} \[DowRSS\][J.S.Dowker .]{} \[Dowcs1\][J.S.Dowker .]{} \[Dowcs\][J.S.Dowker [*Quantum field theory around conical defects*]{} in [*The formation and evolution of Cosmic strings*]{} edited by S.W.Hawking, G.W.Gibbons and T.Vachaspati, Cambridge University Press, Cambridge 1990.]{} \[BandG\][T.Branson and P.B.Gilkey [*Comm. Partial Diff. Equns.*]{} [**15**]{} (1990) 245.]{} \[Duff\][M.J.Duff, .]{} \[DandW\][A.Dettki and A.Wipf .]{} \[CdeV\][Y.Colin de Verdière [*Compos. Mat.*]{} [**27**]{} (1973) 159.]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The XENON1T experiment uses a time projection chamber (TPC) with liquid Xenon to search for Weakly Interacting Massive Particles (WIMPs), a proposed Dark Matter particle, via direct detection. As this experiment relies on capturing rare events, the focus is on achieving a high recall of WIMP events. Hence the ability to distinguish between WIMP and the background is extremely important. To accomplish this, we suggest using Convolutional Neural Networks (CNNs); a Machine Learning procedure mainly used in image recognition tasks. To explore this technique we use XENON collaboration open-source software to simulate the TPC graphical output of Dark Matter signals and main backgrounds. A CNN turns out to be a suitable tool for this purpose, as it can identify features in the images that differentiate the two types of events without the need to manipulate or remove data in order to focus on a particular region of the detector. We find that the CNN can distinguish between the dominant background events (ER) and 500 GeV WIMP events with an accuracy of 87.0% and a recall of 88.2%.'
author:
- 'Charanjit K. Khosa'
- Lucy Mars
- Joel Richards
- Veronica Sanz
title: Convolutional Neural Networks for Direct Detection of Dark Matter
---
Introduction
============
The mystery of Dark Matter is one of the main motivations to search for physics beyond the Standard Model of Particle Physics. The detection of Dark Matter interactions beyond gravitational would be a crucial step for both the Particle and Astroparticle Physics communities. Direct Detection experiments such as XENON1T search for instances where Dark Matter particles scatter with Standard Model (SM) particles and transfer energy to them inside a detector.
For particle Dark Matter direct detection experiments, the main building block is cryogenic material single-phase time projection chamber (TPC), dual-phase TPC, and bubble chambers. Among the leading direct detection experiments, TPC is used in DarkSide-50 [@Agnes:2018fwg], LUX [@Akerib:2016vxi], PandaX-II [@Cui:2017nnn] and XENON1T [@Aprile:2017aty]. In this work, we focus on the output of TPC as a part of the XENON1T experiment. The light signals recorded by the photomultiplier tubes (PMTs) due to the prompt scintillation and secondary scintillation are used to infer the type of interactions, namely to distinguish WIMP and background events. The dominant background sources are beta particles, neutrons and gamma-ray photons.
Typically, in order to achieve a large signal-to-background ratio in the data, one requires a substantial number of cuts to the data to be performed based on certain discrimination parameters. To squeeze as much signal as possible, it is crucial to improve on recall of anomalous signal events, and more effective limits to these discrimination parameters are sought in an attempt to improve the detection probability of such anomalous events.
Machine Learning (ML) provides a unique and flexible alternative to these rigid cut-based methods. In particular, Deep Learning models such as Convolutional Neural Networks (CNN) are able to heuristically learn patterns in highly-complex input space, leading to an ability to detect anomalous signals without the need to manipulate or remove as much data. In this paper, a novel, Deep Learning model is developed using a CNN which can discriminate between simulated background and WIMP waveform images from the XENON1T experiment.
In the field of Particle and Astrophysics, Deep Learning is showing promising results (see e.g. [@Carleo:2019ptp]). Convolution Neural Networks has also been found very efficient for simulating the Dark Matter in N-body simulations of the galaxies [@Zhang:2019ryt]. They also offer improved sensitivity for cosmological observations from weak lensing maps [@Fluri:2019qtp]. Machine learning shows promising reach for disentangling among collider Dark Matter searches [@Khosa:2019kxd], using substructure based Dark Matter probes for non-collider searches [@Brehmer:2019jyt; @Alexander:2019puy; @DiazRivero:2019hxf], and for cosmological Dark Matter [@cosmoDM].
This paper is organised as follows. In Section \[NoEvents\] we describe the XENON1T experiment, explaining how the time projection chamber is used to look for proposed WIMP events and the types of backgrounds. Section \[simulations\] is devoted to the process of simulating Dark Matter and electronic recoil background events using the open source data processing software created by the XENON collaboration. The architecture of the CNN used is explained in section \[cnn\], as well as the training and testing procedure for the model. In the last section \[conclusions\], we discuss the results.
WIMP and Background Events at XENON1T\[NoEvents\]
=================================================
The XENON experiment is an underground research facility, operated at the Laboratori Nazionali del Gran Sasso (LNGS) in Italy. Starting in 2006, the XENON10 experiment used a 25 kg (14 kg target mass) Xenon dual phase time projections chamber (XeTPC) to search for WIMPs [@Angle:2007uj]. This was followed by the XENON100 experiment in 2008, a liquid Xenon TPC (LXeTPC) containing 62 kg target mass of LXe (161 kg total) [@Aprile:2011dd]. The most recent experiment is the XENON1T experiment, a 3.2 tonne LXeTPC with a fiducial volume of roughly 2 tonnes [@Aprile:2017aty]. XENON1T is designed for detecting nuclear recoils (NRs) from WIMP particles scattering off the Xe nuclei [@Aprile:2017iyp].
The TPC used in the XENON1T experiment was built roughly 97 cm long by 96 cm wide and contained (2004 $\pm$ 5) kg of LXe, with another 1200 kg of LXe used as shielding [@Aprile:2017iyp]. The TPC was then placed in a 10 m tall water tank with a diameter of 9.6 m, used to further shield it from any background radioactivity. A vital part of a TPC is the photomultiplier tubes. When a photon hits the photocathode in the PMT it produces electrons which are then accelerated by a high-voltage field. The number of electrons increases within a chain of dynodes by secondary emission. A total of 248 PMTs are used in the TPC to record signals. They are split into the top array (which contains 127 PMTs) and bottom array (121 PMTs), in order to maximise the scintillation-light collection efficiency [@Aprile:2017aty]. Fig. \[fig:Experiment\] shows the PMTs and the TPC used in the XENON1T experiment. The PMTs with the highest quantum efficiency were placed in the centre of the bottom array since that is where most of the scintillation light is collected. More information about the XENON1T TPC can be found in [@Aprile:2017aty].
If an incoming particle interacted with the liquid Xenon, it would produce a scintillation of light and ionization. The S1 signal is the light seen by the top and bottom PMTs (due to total internal reflection at the boundary). The electron charges that were released during ionisation then drift upwards towards the gaseous Xenon (GXe) due to the electric field between the cathode gate and anode. They are then extracted by a stronger extraction field, E$_{extraction}$, creating another larger scintillation light signal (S2) seen by the top PMTs. The position of the original interaction can be reconstructed in 3 dimensions by using the S2 signal pattern (giving the lateral position) and the difference in time between S1 and S2 (depth of interaction) [@Aprile:2017aty].
However, there are many other processes that can lead to the creation of a light signal within the XENON1T experiment. For example, in order to reduce cosmic rays reaching the Xenon, the experiment was carried out deep underground, under the Italian Apennines. To reduce natural background radioactivity the TPC was submerged in a water tank and made from shielding material. Furthermore, only events that occurred within the inner tonne of LXe were used, allowing for the rest of the LXe to be used as more shielding. This was possible due to the large mass density of LXe (almost 3000 kg/m$^3$), meaning particles, such as gammas, can only travel for a short distance before being stopped [@Pelssers].
Despite the shielding, there are six types of background events that can be detected within the search region. Table \[table:Background\] shows the expected number of each of these background events, as well as the expected number of events for a spin-dependent 500 GeV/$c^2$ WIMP with a cross-section of $10^{-45}$ cm$^2$, over the time period of the first science run of XENON1T (34.2 live-days). This WIMP benchmark is chosen among the allowed values by direct, indirect and collider Dark Matter searches [@WIMPbanchmark]. The expected number of events were calculated using [Laidbax]{} [@Laidbax] and the [blueice]{} Monte Carlo model [@Blueice]. These results agree with the XENON1T collaboration paper ([@Aprile:2017iyp]), however, they used a WIMP mass of 50 GeV/$c^2$ and a $10^{-46}$ cm$^2$ cross-section.
The particles in the detector release energy in the form of a nuclear or electronic recoil (NR or ER). This means the incoming particle will either scatter directly from the nucleus of the target atom, or it will interact with the electron cloud [@Pelssers]. Since we expect WIMPs to have a very low interaction probability (low cross-section), they should only cause nuclear recoils. Therefore, our ability to differentiate between NR and ER is very important when searching for WIMPs. Table \[table:Background\] shows that the number of expected WIMP events is much higher than for any of the NR background. Hence, in this work we have focused on comparing the ER background with WIMPs[^1].
Name Expected number of events
--------------------------------------- ---------------------------
Electronic recoils (ER) 61.879487
CNNS ($\nu$) 0.000901
Radiogenic neutrons 0.058570
Accidental coincidences (acc) 0.220000
Wall leakage (wall) 0.520000
Anomalous (anom) 0.090004
500 GeV/$c^2$, $10^{-45}$ cm$^2$ WIMP 35.029005
: Expected number of events for each type of background over the time period of the first science run of XENON1T (34.2 live-days) within the fiducial mass and a 500 GeV/$c^2$, $10^{-45}$ cm$^2$ WIMP (Generated using [Laidbax]{}).[]{data-label="table:Background"}
WIMP mostly causes the Nuclear Recoils (NR) but there are some backgrounds also which can have NR. One way the XENON1T experiment cut down on NR background was to only look at low energy, single-scatter nuclear recoils, since those are the type we expect WIMPs to create. To do this, only events that had at least one S2 signal above 200 single photoelectrons (PE) were included and the signal duration of S2 had to be consistent with the interaction depth (calculated from the drift time) [@Aalbers:2018mfc].
The XENON1T was designed as an ultra-low background experiment including high rejection of ER backgrounds. Even when set to 50% NR acceptance the XENON1T could detect WIMP-like events while also rejecting roughly 99.8% of all the ER events. The main source of the ER background is the beta decays of $^{85}$Kr and $^{214}$Pb, which causes a flat energy spectrum within the interested range [@Aalbers:2018mfc].
When reconstructing the energy deposits of particles interacting with the LXe, the S1 and S2 signals need to be corrected in case of any time or spatial dependent signal losses. The light signal, S1, is corrected for the (x,y,z)-dependent light collection efficiency in the TPC (cS1) [@Breur]. The charge signal, S2, is corrected for the time and depth dependent electron lifetime since electrons can be lost while drifting in the LXe if they attach to impurities within the Xenon (cS2) [@Breur]. In Fig. \[fig:csc\] we plot the distribution of these events for ER and WIMP interactions.
Event Simulation\[simulations\]
===============================
In the following we introduce the three types of images that were used during our analysis, and explain how we generated them using [PaX]{} (Processor for Analysing XENON) software, created by the XENON collaboration [@PAX] [^2].
At the first step, we generate the energy spectrum of the Dark Matter particle, using [wimprates]{} [@wimprate]. Given the proposed mass and cross-section of a WIMP, [wimprates]{} produces the differential rates of a WIMP-nucleus scattering in the standard halo model, within a liquid Xenon detector [@wimprate]. In this work, we used a WIMP mass of 500 $GeV^{-1}$ and a cross-section of $10^{-45} cm^{-2}$ for illustration. More details on how to set-up the simulation are given in App. \[setup\].
Next, [Laidbax]{} (Likelihood- And Interpolated Density Based Analysis for XENON) [@Laidbax] is used to convert the energy spectrum into a model that is compatible with the [PaX]{} [@PAX] software which we use to create the graphical output of the simulated WIMP and background events. [PaX]{} has an in-built simulator called [FaX]{} (Fake XENON) which requires an input [*csv*]{} file consisting of a set of variables for each interaction to construct the waveform of the event. These variables are: instruction (simply a number given to each interaction); recoil type (NR or ER); x-position, y-position and depth of the interaction (in cm); number of photons produced; number of electrons produced; time of the interaction (in ns).
The ER background energy spectrum we used came pre-built in [Laidbax]{}. Note, though, that the models used in [Laidbax]{} are not the official models approved by the XENON collaboration, which include several more uncertainties. For example, the ER background model is a polynomial ER fit [@Aalbers:2018mfc] and the yields for the NR background model are specified by the parameterisation of the global fit found in [@Lenardo:2014cva]. The [Laidbax]{} model produces a database of simulated interactions. The parameters of the interactions include; radial distance $r^2$, angle $\theta$, z-coordinate, number of photons produced, number of electrons produces, the cS1 and cS2 values (corrected S1 and S2 signals). As mentioned earlier, in Fig. \[fig:csc\], simulated WIMP and ER background events are shown. We can see the overlap between these two types of events. To convert the [Laidbax]{} model to the input file for FaX, a series of calculations and assumptions have to be made. For example, the polar coordinates need to be converted to cartesian coordinates. The time variable is not given by [Laidbax]{}, hence it is assumed that the actual time of the interaction is not relevant, as long as it occurred within a particular time range that an event is defined by.
Image Processing
----------------
The original graphical output of [PaX]{} [@PAX] is shown in Fig. \[fig:Event0\]. The images show the largest S1 and S2 peaks (top left), the hitpatterns for the top and bottom PMTs (top right), all the S1 and S2 peaks in the event (middle), and the PMTs that detected the signal (bottom).
In order to use the images as an input to CNN, we edited them to remove unnecessary features that are the same in every image so that the CNN can focus on features that are unique to a WIMP or background event. First, we looked at whether we could differentiate a WIMP event from the background just using the S1 and S2 hitpatterns. The images were edited to remove the text labels in order to increase effective learning. The final image produced is 800$\times$400 pixels and an example is shown in Fig. \[fig:Hit\]. Next, we looked at the largest S1 and S2 peak graphs. Fig. \[fig:Peak\] shows an example of the edited peaks for a WIMP event. Finally, we combined the hitpatterns and peak graphs into one “HitPeak" image of size 800$\times$800 (shown in Fig. \[fig:WIMP\]). The similar image for the ER is shown in Fig. \[fig:ER\]; comparing this to Fig \[fig:WIMP\] we can see that both events are visually very similar, hence the need for Machine Learning to differentiate them.
CNN Architecture\[cnn\]
=======================
This section discusses the details of a Convolutional Neural Network used in this work (python code can be found in the github repository [@lucygithub]).
The majority of Machine Learning problems involve a dataset ***X**=$\{$($y_{i}$,$x_{i}$), i=1,...,N$\}$*, a model *g(**$\theta$**)* with parameters ***$\theta$***, and a cost function *C(**X**,g(**$\theta$**))* (also known as the loss function). The cost function allows you to judge how well the model performs on the given dataset. A model is fitted by calculating the value of ***$\theta$*** that minimises the cost function.
The most common way to minimise the cost function is to use *Gradient Descent*; an algorithm that finds the local or global minima of a function. The parameters are adjusted in the direction where the gradient of the cost function is large and negative, and then the gradient is recalculated in the new position. After each iteration the model gradually converges towards a minimum (where any changes to the parameters will produce little or no change in the loss) resulting in the weights being optimised. Given the cost function *C($\theta_{i}$)*, it simultaneously updates for each *i = 0,...,n* until convergence is reached: $$\theta_i := \theta_i - \eta \nabla_{\theta} C(\theta_0,...,\theta_n)
\label{eq:Gradient}$$ The learning rate, $\eta$, controls how large each step is taken during gradient descent.
Neural networks (NN) contain multiple neurons[^3] stacked into hidden layers. The output of each layer then serves as the input for the following one. Each neuron takes a vector of inputs, **x**, and produces a scalar output *a$_i$*(**x**). The function *a$_i$* depends on the NN but it can be separated into a linear operation (which weighs the importance of the inputs) and a non-linear transformation (performed by an activation function). In this paper we used a Leaky ReLU (Rectified Linear Unit) activation function.
A NN calculates the gradient of the cost function using backpropagation. This algorithm contains a forward pass (going from the input layer to output layer), calculates the weighted inputs and activations for all the neurons, and then backpropagates the error (output to input layer), calculating the gradients.
A Convolutional Neural Network is a type of Neural Network Machine Learning algorithm that primarily takes images as its input and assigns weights and biases to different parts of the image. A CNN is comprised of many layers of different types, including *Convolutional Layers*, *Pooling Layers*, and *Fully-Connected (FC) Layers*. The convolutional layer is used to extract features from the input image by passing a filter (kernel) over the image matrix. To perform different operations on the image, such as edge detection or sharpening, different types of filters are used. The layer outputs the image matrix at a reduced volume, depending on the size of the filter. A nonlinear activation function is applied after each Convolutional layer.
Pooling layers are used to reduce the dimensions of the data by combining the output of a neuron cluster in one layer to a single neuron in the next. The FC layer connects the neurons to all the activations in the previous layer.
The final layer is non-linear classification layer and assigns decimal probabilities to each mutually exclusive class.
Before running the images through CNN, we scale down the image resolution. The original images had a resolution of 800$\times$400 for the individual Hit and Peak graphs and 800$\times$800 for the images showing both (HitPeak). A lot of computational power would be needed to run these images through the CNN due to the large number of parameters. We chose to scale the HitPeak images to 75x75 resolution and the separate Hit and Peak images to 100x50. Next, the images are converted into a single array, the shape of which depends on the image spatial resolution. Each pixel is one of 256 possible values (0-255) since the images are 8-bit, i.e. 256 = 2$^8$. Most Machine Learning algorithms perform better on small, floating point values instead of large pixel values. Hence, we scale the image pixels to between 0 and 1 by dividing by 255. Each image is labelled as signal or background since we are using supervised learning. The data is split into training and test sets. We separated 90$\%$ of the images into the training set with the other 10% into the test set. A further step is to create another subset from the training data, assigning 20% of the training set as the validation set. The validation set provides an unbiased evaluation of the models fit on the training data while also tuning the hyperparameters[^4]. The final model can then be used on the test dataset.
The CNN was created using the [Keras]{} Sequential model within the [Tensorflow]{} [@Tensorflow] environment to linearly stack the layers. The number of convolution layers and number of neurons are two of the hyperparameters we tested during this project (see Sections \[neurons\] and \[layers\]). Each convolution layer contains an L2 weight regularizer (“kernel\_regularizer") where the alpha value[^5] is 0.005. Regularization is used to prevent overfitting due to intrinsic noise.
Each convolutional layer also includes a LeakyReLU activation function and a pooling layer. Max pooling, with a pool size of 2x2, is used to reduce the spatial dimensions of the output volume. This means after each pooling layer the output volume is half that of the input volume.
The classification stage has the most parameters and so requires a “dropout" regularisation layer to prevent overfitting (with the dropout rate set to 25%). The data is flattened into a 1-dimensional array to connect the 2-dimensional convolutional and pool layers to the 1-dimension FC layers. The output layer contains a single neuron which is used to make predictions. A sigmoid activation function produces a probability output between 0 and 1.
Finally, the model is compiled using an [Adam]{} optimiser [@Kingma:2014vow] to minimise the cost function. [Adam]{} (derived from ‘adaptive moment estimation’) computes adaptive learning rates for each parameter. Binary cross entropy was used for the loss function (since we were completing a binary classification task). Fig. \[fig:Tensorboard\] shows the architecture of CNN, created using [Tensorboard]{} [^6]. Finally, the model is fitted with the defined parameters and the accuracy and loss for the test dataset are calculated.
Testing the CNN
---------------
We simulated 10,000 events for both WIMP and ER particles, making a total of 20,000 images. Most of the data (90%) was included in the training data set with the other 10% in the test set. The training set was then split again to include a validation set (14,400 training and 3600 validation images). The test accuracy and loss (value of the cost function) are given for each test.
The CNN assigns random weights and biases at the beginning of each iteration and so there will always be a slight variation in results after each run. Therefore, the model will give slightly different results in each test, even when using the same parameters. The CNN was run in a CPU [TensorFlow]{} environment.
In this project we compared the effect of using the three types of images as well as multiple different hyperparameters, including the number of neurons in a convolution layer, number of convolutional layers in the CNN, the batch size, and the number of epochs.
### Image Type\[Image\]
First, we compared how the type of image affects the CNN output. To do this we tested each of the three image types (Hit, Peak and HitPeak) using a one convolution layer CNN with 20 neurons over 40 epochs, with a batch size of 100. Fig. \[fig:Graphs\] compares the accuracy and loss for the training and validation datasets of the three images. Table \[table:Type\] shows the accuracy and loss for their test sets. We can see that the HitPeak images give far better results than the seperated images. Note, the HitPeak images have slightly more parameters (519,181) than the seperated images (471,181) due to the shape of the original figures. The rest of the tests discussed in this paper use the 75x75 HitPeak images.
Image Type Scale Parameters Test Acc Test Loss
------------ -------- ------------ ---------- -----------
Hit 100x50 471,181 0.820 0.433
Peak 100x50 471,181 0.827 0.452
HitPeak 75x75 519,181 0.865 0.395
: Test accuracy and loss for the three types of image; Hit, Peak and HitPeak.[]{data-label="table:Type"}
 
 
### Number of Neurons\[neurons\]
Next, we explore how the choice of neuron number affected the performance of the CNN. The number of neurons represents the number of filters that run in parallel for a given input, allowing the model to learn multiple features. To do this we used a CNN with one convolutional layer and set the neuron number to 5, 10, 20, 40 and 60. Fig. \[fig:neuron40\] shows the training and validation results for the different number of neurons for the HitPeak images over 40 epochs with a batch size of 100. Table \[table:Neuron\] shows the test accuracy and loss for each neuron number as well as the number of parameters in the CNN.
Neuron Number Parameters Test Accuracy Test Loss
--------------- ------------ --------------- -----------
5 129,826 0.838 0.411
10 259,611 0.847 0.404
20 519,181 0.865 0.395
40 1,038,321 0.857 0.406
60 1,557,461 0.841 0.427
: Test accuracy and loss for different neuron numbers in a CNN with one convolutional layer for HitPeak images.[]{data-label="table:Neuron"}
The results show that the test accuracy increases with increasing neuron number before peaking at 20 neurons (0.865) and then decreasing. This is due to the fact that while more neurons can express more complicated function, using too many neuron on a simple dataset can lead to overfitting on the test data.
### Number of Convolutional Layers\[layers\]
 
 
Having multiple convolutional layers allows a model to extract more complex features. The earlier layers (closer to the input image) learn the lower level features, with the complexity of the features increasing with each layer. To observe how the number of convolutional layers affect our model, CNNs with 1, 2 or 3 layers were used. Each layer contained 20 neurons since that was the optimal number found for one convolutional layer in the previous test.
The results, shown in Fig. \[fig:conv\] and Table \[table:Conv\], show that two convolutional layers give the best test accuracy (0.870) and loss (0.372), as well as the best training/validation accuracy and loss. Since our images are not very complex they do not have many features, and so a very deep CNN with three or more layers is not needed. This conclusion was also reached when increasing the number of neurons in each layer.
As the number of convolutional layers increases, the total number of parameters decreases (Table \[table:Conv\]). This is because there is a pooling layer associated with each convolutional layer which reduces the shape of the output volume. Hence, even though each convolution layer introduces new parameters, the number due to the output volume (when the data is flatten into a 1D array) will be much less than for a CNN with one convolutional layer.
Conv layers Parameters Test Acc Test Loss
------------- ------------ ---------- -----------
1 519,181 0.842 0.420
2 120,001 0.870 0.372
3 27,621 0.853 0.393
: Test accuracy and loss for 1, 2, or 3 convolutional layers for HitPeak images. There are 20 neurons in each convolution layers.[]{data-label="table:Conv"}
### Batch Size
We tested the 75$\times$75 HitPeak images over 40 epochs using the CNN with two convolutional layers (with 20 neurons each) using different batch sizes; 50, 100, 200, 500 and 700. Results are shown in Table \[table:Batch\] and Fig. \[fig:Batch\].
We can see that the smaller the batch size, the faster the model will converge to a “good" solution (steeper curve at the beginning), since the model will start learning before seeing all the data. However, this does not guarantee the model will converge to the best possible result (as it would when using the whole dataset). The best test accuracy (0.870) and loss (0.385) were found using a batch size of 100, hence this batch size was used in the future tests.
Batch Size Test Accuracy Test Loss
------------ --------------- -----------
50 0.841 0.405
100 0.869 0.385
200 0.854 0.387
500 0.850 0.396
700 0.847 0.413
: Test accuracy and loss for different batch sizes; 50, 100, 200, 500 and 700 for HitPeak images.[]{data-label="table:Batch"}
 
### Number of Epochs
One epoch occurs when the entire dataset in passed through the neural network (via backpropagation) once. The choice of epoch number can lead to overfitting or underfitting of training data depending on whether there are too many or too few epochs. The effect of epoch number is different for different datasets and depends on how diverse the data is (more diverse data requires more epochs so the CNN has time to learn all the data features). Table \[table:Epoch\] gives the test results for different epoch numbers (10, 20, 30, 40 and 50), while Fig. \[fig:Epoch\] shows the training and validation accuracy and loss. Each test was carried out using a CNN with two convolutional layers with 20 neurons each, and a batch size of 100.
Epoch Test Accuracy Test Loss
------- --------------- -----------
10 0.799 0.470
20 0.861 0.379
30 0.862 0.380
40 0.870 0.370
50 0.865 0.367
: Test accuracy and loss for different number of epochs (CNN has two convolutional layers with 20 neurons each and batch size of 100).[]{data-label="table:Epoch"}
In our tests we have found that a CNN with two convolutional layers (both containing 20 neurons) using the HitPeak images scaled to 75$\times$75 for 40 epochs and a batch size of 100, gives the best results overall. Finally, we test the CNN with all the optimal parameters. Fig. \[fig:final2\] shows the ROC Curve for the optimal CNN. Our final run gives an accuracy of 87.0% and loss of 36.7% for the HitPeak images. We also show the ROC curves for separate Hit and Peak images.
Conclusions and Outlook\[conclusions\]
======================================
In this work, we have introduced a new method of differentiating between proposed Dark Matter (WIMP) events and the background events (electronic recoils) in the XENON1T experiment using the graphical output of PaX. The CNN correctly identified a proposed WIMP event 87.0% of the time. When testing our CNN, we found that images which showed both the hitpattern and largest peak figures (HitPeak) gave the best accuracy, compared to images showing only one or the other.
We found that 20 neurons is the optimal number of neurons in a CNN with one convolutional layer (0.865 accuracy and 0.395 loss). The number of neurons in a convolution layer represents the number of filters used. The higher the neuron number, the higher the total number of trainable parameters. Since our images are not very complex, the CNN does not require many neurons. The best results were obtained when using a CNN with two convolution layers with 20 neurons in each (0.870 accuracy and 0.372 loss). This test was also carried out using different numbers of neurons in each layer (first layer having 20 neurons, then increasing by 20 each layer); the results followed the same pattern as using a constant number of neurons. The optimal batch size was found to be 100 (0.869 accuracy, 0.385 loss). The smaller batch sizes reached their maximum performance quicker than the larger ones since the model will start training before all the data is seen. If the batch size is too large it can lead to poor generalisation, resulting in lower test accuracy. Finally, 40 epochs gave the best results overall (0.870 accuracy and 0.370 loss). One could run CNN using a GPU (Graphics Processing Unit) which would be faster than the current CPU and would be able to use higher resolution images, possibly resulting in higher accuracy. This set-up could be used for the real Dark Matter and background models created by the XENON group to simulate more realistic events. The CNN can then be adapted and improved using this data, with the aim of running CNN using real data generated during the XENON1T experiment. If this stage is successful then the method could be used in the next XENON collaboration experiment; the XENONnT (an upgrade of the XENON1T). Experiments aiming for the detect Dark Matter via direct detection focus on the particular regions of the detector to improve the signal and background ratio. This method does not require any reduction in the data sets. It just tries to differentiate the events based on the PMTs signals. In fact same idea could be extended for other TPCs based detectors. Finally, this work could be further extended by the ML based models to identify the rare WIMP events among the many different type of background events, not just the dominant ER background.
acknowledgments
===============
C.K.K. wishes to acknowledge support from the Royal Society-SERB Newton International Fellowship (NF171488). The work of V.S. by the Science Technology and Facilities Council (STFC) under grant number $\text{ST/P000819/1}$. We would like to thank Michael Soughton and Adam Matthews for many fruitful discussions. We would also like to acknowledge Prof. Christopher Tunnell and Dr Jelle Aalbers for their help in explaining how we could use PaX, and the other open-source software they created, to carry out this work.\
Analysis set-up\[setup\]
========================
This section follows the Linux installation process suggested by the PAX github page [@PAX].
1. Install Python 3 and Anaconda libraries
2. Set up Anaconda libraries
`export PATH=\sim/anaconda3/bin:$PATH conda config –add channels http://conda.anaconda.org/NLeSC`
3. Add additional python packages
`conda install conda conda create -n pax python=3.4 root=6 numpy scipy=0.18.1 pyqt=4.11 matplotlib pandas cython h5py numba pip python-snappy pytables scikit-learn rootpy pymongo psutil jupyter dask root_pandas jpeg=8d isl=0.12.2 gmp=5.1.2 glibc=2.12.2 graphviz=2.38.0=4 gsl=1.16 linux-headers=2.6.32 mpc=1.0.1 mpfr=3.1.2 pcre=8.37 python-snappy=0.5 pyopenssl=0.15.1`
4. Activate the environment
`source activate pax `
5. Installing Pax
`git clone https://github.com/XENON1T/pax.git source activate pax cd pax python setup.py develop`
Simulate Data - Package Installation
====================================
- [@Laidbax] `git clone https://github.com/XENON1T/laidbax cd laidbax python setup.py develop cd ..`
- [@Blueice] `git clone https://github.com/JelleAalbers/blueice cd blueice python setup.py develop cd ..`
- [@wimprate]
`pip install wimprates`
- [@multihist] `pip install multihist `
Simulate Data - Running
=======================
The code to simulate and plot the data can be found in [@lucygithub].
- Run “`python Simulate.py`" (This python file is based on the laidbax tutorial notebook [@Laidbax]). This will save two csv files called ‘ERSIM.csv’ and ‘WIMPSIM.csv’ in the data folder ($\sim$/pax/pax/data)
- To generate the hitpattern and peak images run: “`paxer –config XENON1T Simulation –input \sim/pax/pax/data/WIMPSIM.csv –plot`"
- Replace ‘WIMPSIM’ with ‘ERSIM’ for ER plots.
Running the CNN
===============
To run the CNN create a conda environment containing TensorFlow:\
`conda create -n tensorflow_env tensorflow`\
`conda activate tensorflow_env`\
Then install “Keras" [@Keras], “Matplotlib" [@Matplotlib], “Scikit-learn" [@Sklearn], “Pandas" [@Pandas] and “Imageio" [@Imageio]:\
`conda install -c conda-forge keras`\
`conda install -c conda-forge matplotlib`\
`conda install scikit-learn`\
`conda install -c anaconda pandas`\
`conda install -c menpo imageio`
[99]{}
P. Agnes [*et al.*]{} \[DarkSide Collaboration\], Phys. Rev. D [**98**]{} (2018) no.10, 102006, \[arXiv:1802.07198 \[astro-ph.CO\]\]. D. S. Akerib [*et al.*]{} \[LUX Collaboration\], Phys. Rev. Lett. [**118**]{} (2017) no.2, 021303, \[arXiv:1608.07648 \[astro-ph.CO\]\]. X. Cui [*et al.*]{} \[PandaX-II Collaboration\], Phys. Rev. Lett. [**119**]{} (2017) no.18, 181302, \[arXiv:1708.06917 \[astro-ph.CO\]\]. E. Aprile [*et al.*]{} \[XENON Collaboration\], Eur. Phys. J. C [**77**]{} (2017) no.12, 881, \[arXiv:1708.07051 \[astro-ph.IM\]\]. G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, L. Vogt-Maranto and L. Zdeborová, arXiv:1903.10563 \[physics.comp-ph\].
X. Zhang, Y. Wang, W. Zhang, Y. Sun, S. He, G. Contardo, F. Villaescusa-Navarro and S. Ho, arXiv:1902.05965 \[astro-ph.CO\]. J. Fluri, T. Kacprzak, A. Lucchi, A. Refregier, A. Amara, T. Hofmann and A. Schneider, Phys. Rev. D [**100**]{} (2019) no.6, 063514, \[arXiv:1906.03156 \[astro-ph.CO\]\].
C. K. Khosa, V. Sanz and M. Soughton, arXiv:1910.06058 \[hep-ph\]. J. Brehmer, S. Mishra-Sharma, J. Hermans, G. Louppe and K. Cranmer, arXiv:1909.02005 \[astro-ph.CO\].
S. Alexander, S. Gleyzer, E. McDonough, M. W. Toomey and E. Usai, arXiv:1909.07346 \[astro-ph.CO\].
A. Diaz Rivero and C. Dvorkin, arXiv:1910.00015 \[astro-ph.CO\].
C. Escamilla-Rivera, M. A. C. Quintero and S. Capozziello, arXiv:1910.02788 \[astro-ph.CO\].
J. Angle [*et al.*]{} \[XENON Collaboration\], Phys. Rev. Lett. [**100**]{} (2008) 021303, \[arXiv:0706.0039 \[astro-ph\]\]. E. Aprile [*et al.*]{} \[XENON100 Collaboration\], Astropart. Phys. [**35**]{} (2012) 573, \[arXiv:1107.2155 \[astro-ph.IM\]\]. E. Aprile [*et al.*]{} \[XENON Collaboration\], Phys. Rev. Lett. [**119**]{} (2017) no.18, 181301, \[arXiv:1705.06655 \[astro-ph.CO\]\]. B. Pelssers, *Position reconstruction and data quality in Xenon*, Master’s thesis, Universiteit Utrecht (2015).
Laidbax Github repository, XENON Collaboration https://github.com/XENON1T/laidbax
J. Aalbers, Blueice Github repositroy, https://github.com/JelleAalbers/blueice
J. Aalbers, *Dark matter search with XENON1T*, PhD thesis, University of Amsterdam (2018). P. A. Breur, *Backgrounds in XENON1T*, PhD thesis, University of Amsterdam (2019).
PaX Github repository, XENON Collaboration, https://github.com/XENON1T/pax.
U. Simola, B. Pelssers, D. Barge, J. Conrad and J. Corander, JINST [**14**]{} (2019) no.03, P03004, \[arXiv:1810.09930 \[astro-ph.IM\]\].
J. Aalbers, B. Pelssers and K. Morå, *wimprates: v0.3.0* (2019). http://doi.org/10.5281/zenodo.3345959.
B. Lenardo, K. Kazkaz, A. Manalaysay, J. Mock, M. Szydagis and M. Tripathi, IEEE Trans. Nucl. Sci. [**62**]{} (2015) no.6, 3387, \[arXiv:1412.4417 \[astro-ph.IM\]\]. L. Mars, Github repository for this paper, https://github.com/LucyMars/SearchForDarkMatter
M. Abadi, et al., “TensorFlow: Large-scale Machine Learning on heterogeneous distributed systems" arXiv:1603.04467 (2016).
D. P. Kingma and J. Ba, arXiv:1412.6980 \[cs.LG\]. J. Aalbers, Multihist github repository https://github.com/JelleAalbers/multihist
F. Chollet, Keras GitHub repository, https://github.com/fchollet/keras (2015).
J. Hunter, “Matplotlib: A 2D Graphics Environment", *Computing in Science and Engineering*, **9**, 90-95 (2007).
F. Pedregosa, et al., “Scikit-learn: Machine learning in Python", *Journal of Machine Learning Research*, **12**, pp. 2825-2830 (2011).
W. McKinney, “Data Structures for Statistical Computing in Python", *Proceedings of the 9th Python in Science Conference*, 51-56 (2010).
A. Klein, et al., imageio/imageio: V2.6.0 (Version v2.6.0). Zenodo. http://doi.org/10.5281/zenodo.3475011 (2019).
[^1]: Other background sources are discussed in Ref. [@Aalbers:2018mfc].
[^2]: Note that in [@Simola:2018ntn], Machine Learning was used for the accurate reconstruction of 2-D interaction in the TPC.
[^3]: A computational unit which performs a nonlinear transformation of its input.
[^4]: A parameter whose value is chosen before learning begins. Other parameters are derived during training.
[^5]: Transformation being applied to the coefficients in the weight matrix before being added to the total loss.
[^6]: TensorBoard is a browser based application that is used to visualise the architecture of the CNN and the scalar values, such as the classification accuracy.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Progress on six dimensional ionization muon cooling with relatively small rings of magnets is described. Lattices being explored include scaling sector cyclotrons with edge focusing and strong focusing, fixed field alternating gradient (FFAG) rings. Ionization cooling is provided by high pressure hydrogen gas which removes both transverse and longitudinal momentum. Lost longitudinal momentum is replaced using radio frequency (RF) cavities, giving a net transverse emittance reduction. The longer path length in the hydrogen of higher momentum muons decreases longitudinal emittance at the expense of transverse emittance. Thus emittance exchange allows these rings to cool in all six dimensions and not just transversely. Alternatively, if the RF is located after the ring, it may be possible to cool the muons by stopping them as they spiral adiabatically into a central swarm. As $p \to 0$, $\Delta{p} \to 0$. The resulting cooled muons can lead to an intense muon beam which could be a source for neutrino factories or muon colliders.'
address:
- |
Department of Physics and Astronomy, University of Mississippi–Oxford\
University, MS 38677, USA
- |
Department of Physics, University of California–Los Angeles\
Los Angeles, CA 90095, USA
- |
Department of Physics, University of California–Riverside\
Riverside, CA 92521, USA
- |
Brookhaven National Laboratory\
Upton, NY 11973, USA
author:
- 'D. J. SUMMERS,[^1] S. B. BRACKER, L. M. CREMALDI, and R. GODANG'
- 'D. B. CLINE and A. A. GARREN'
- 'G. G. HANSON and A. KLIER'
- 'S. A. KAHN, H. G. KIRK, and R. B. PALMER'
title: 6D IONIZATION MUON COOLING WITH TABLETOP RINGS
---
Introduction
============
Muons at rest have a 2.2 $\mu$s lifetime; cooling an ensemble of muons must be completed faster than that. Ionization cooling can help.$^1$ Random muon motion is removed by passage through a low Z material, such as hydrogen, and coherent motion is added with RF acceleration. Designs to cool muons in six dimensions using linear helical channels$^{\,2}$ at 100 MeV kinetic energies and using frictional cooling$^{\,3}$ at 1 keV kinetic energies are under investigation. A number of muon cooling rings have been simulated at various levels.$^4$ In a ring structure, the same magnets and RF cavities may be reused each time a muon orbits. Transverse cooling can naturally be exchanged for longitudinal cooling by allowing higher momentum muons to pass through more material. Thus rings can cool in all six dimensions. The rings reported here are the smallest to date, and are basically radial sector cyclotrons.
Small emittance bunches of cold muons are useful for a neutrino factory$^{\,5}$ and are required for a muon collider.$^6$ At a neutrino factory, accelerated muons are stored in a racetrack to produce neutrino beams ($\mu^- \to e^- \, {\overline{\nu}}_e \, \nu_{\mu}$ and $\mu^+ \to e^+ \, \nu_e \, {\overline{\nu}}_{\mu}$). Neutrino oscillations have been observed$^{\,7}$ and need more study. Further exploration at a neutrino factory could reveal CP violation in the lepton sector,$^{8}$ and will be particularly useful if the $\nu_e$ to $\nu_{\tau}$ coupling, $\theta_{13}$, is small.$^9$ A muon collider can do s-channel scans to split the $H^0$ and $A^0$ Higgs doublet.$^{10}$ Above the ILC’s 800 GeV there are a large array of supersymmetric particles that might be produced$^{\,11}$ and, if large extra dimensions exist, so could mini black holes.$^{12}$ Note that the energy resolution of a 4 TeV muon collider is not smeared by beamstrahlung.
Hydrogen Gas Filled Sector Cyclotrons with Internal RF
======================================================
A low field ring is being designed using ICOOL$^{\,13}$ for optimization and SYNCH$^{\,14}$ for lattices. It could be built to demonstrate six dimensional muon cooling. Relatively constant $\beta$ functions allow a continuous placement of the energy absorber. Hydrogen gas cools best and also prevents breakdown in the 201 MHz RF cavities.$^2$ Skew quadrupole magnets mix vertical and horizontal betatron oscillations. The parameters for the ring appear in Table 1. Simulated 6D cooling is shown in Fig. 2.
The equation describing transverse cooling (with energies in GeV) is: $$\frac{d\epsilon_n}{ds}\ =\
-\frac{1}{\beta^2} \frac{dE_{\mu}}{ds}\ \frac{\epsilon_n}{E_{\mu}}\ +
\ \frac{1}{\beta^3} \frac{\beta_{\perp} (0.014)^2}{2\ E_{\mu}m_{\mu}\ L_R},
\label{eq1}$$ where $\beta =$ v/c, $\epsilon_n$ is the normalized emittance, $\beta_{\perp}$ is the betatron function (focal length) at the absorber, $dE_{\mu}/ds$ is the energy loss, and $L_R$ is the radiation length of the material. The first term in this equation is the cooling term, and the second is the heating term due to multiple scattering. This heating term is minimized if $\beta_{\perp}$ is small (strong-focusing) and $L_R$ is large (a low-Z absorber). The equilibrium emittance is achieved when the heating and cooling terms balance.
A higher field ring with smaller values of $\beta_{\perp}$ can give a higher cooling merit factor. Simulations show that high RF gradients are required as noted in Table 1.
A scaling fixed field, alternating gradient (FFAG) ring$^{\,15}$ can allow the use of lower RF gradients and magnetic fields. For the FFAG ring in Table 1, a focusing parameter, $n = -(r/B) (dB/dr) = -0.6$, was used. Each sector has a focusing-defocusing-focusing (FDF) geometry with three magnets. There are no open slots between sectors, so the RF cavities would have to placed within the magnets.
Sector Anti–Cyclotrons with External RF
=======================================
An anti-cyclotron has been used to slow LEAR anti-protons at CERN.$^{16, 17}$ An annular quasipotential well, $U(r,z)$, is formed which ferries anti-protons towards the center of an azimuthally symmetric cyclotron. The radius of the annulus decreases with the decreasing angular momentum of the $\overline{p}$.
$$U(r,z) = V(r,z) - (1/(2 \eta r^2))\, (L_g / M + \eta \, r A_{\theta})^2,$$
where $\eta = e/M$ and $L_g = L_z - e\,r\,A_{\theta}$ is a generalized angular momentum. The radial well deepens with decreasing radius and the vertical well grows shallower (see Fig. 2 of Ref. 16). Particles must adiabatically spiral to the center. If dE/dx is too large, particles will not stay in the magnetic wells. The final $\overline{p}$ swarm has a radius of 1.5 cm, a height of 4 cm, and a kinetic energy of 2 keV. A long bunch train is coalesced into a single swarm, which is roughly the same diameter as the incoming beam. The spiral time is 20 $\mu$s with 0.3 mbar hydrogen and about 1 $\mu$s with 10 mbar hydrogen. Given the dependence of the cyclotron frequency on mass, $f = \omega / 2 \pi = q B / 2 \pi m$, the spiral time for a muon is nine time less than for a $\overline{p}$. The gas pressure in the center must be low, both to allow a particle to spiral all the way in before stopping, and to allow reasonable kicker voltages for extraction. An 80 ns electric kicker pulse rising to 500 V/cm in 20 ns is employed. The $\overline{p}$’s move 32 cm in 500 ns. Given that $F = ma$, muons will go farther.
The anti-cyclotron has now been moved from LEAR at CERN to PSI where it is being used to slow negative muons to kinetic energies of a few keV.$^{18}$ Three centimeter diameter beams with 30000 $\mu^-$/s below 50 keV and 0.8 cm diameter beams with 1000 $\mu^-$/s in the 3 to 6 keV kinetic energy range are output for use. A static electric field continuously ejects the muons. The energy absorber and the electrode consist of a single 30 $\mu$g/cm$^2$ formvar foil with a 3 nm nickel film produced by 30 minutes of sputtering.
A scaling sector anti-cyclotron could allow greater acceptance than the azimuthally symmetric anti-cyclotron now running at PSI. For a given $\int{\bf{B}}{\cdot}d\,{\bf\ell}$, the ratio of the fields in the hills and valleys can be adjusted to maximize acceptance. Only radial and not spiral sectors have been explored so far. The sector anti-cyclotron may be able to function as a damped harmonic oscillator to lower the amplitude of horizontal and vertical betatron motion as a bunch train of muons spirals into a single central swarm. With $10^{12}$ muons in a swarm, space charge is a concern. A conductor might be used for neutralization. Movement of $10^{12}$ electrons in 100 ns only requires 1.6 amps of current. Muons must spiral in fast enough to minimize decay loss, but must not stop before reaching the central swarm. So, the density of the absorber must decrease smoothly with radius. Radial $LiH$ wedges immersed in a gas or high to low pressure gases separated by beam pipes might meet this criteria. The sector cyclotron geometry must transform into an azimuthally symmetric magnetic bottle as the muons approach the central swarm. Otherwise, as shown by GEANT3 simulations, muons will escape though the valleys. In the transition region the field might roughly resemble a hexapole or octupole field as used in an Electron Cyclotron Resonance Ion Source (ECRIS).$^{19}$ Busch’s theorem (eqn. 3)$^{\,20}$ has the effect of increasing the emittance as muons leave a magnetic field. A half Tesla field and a 10 cm radius give an 8 MeV/c azimuthal kick. One might be able to use radial iron fins in the exit port to alleviate this effect or reverse and increase the magnitude of the magnetic field to capture the unwanted angular momentum in an absorber just after extraction. Using low fields with cylindrical swarms that have small diameters works for sure. An RF quadrupole is perhaps a natural choice for acceleration that would immediately follow the extraction electric kicker.
$$\dot{\phi} = [e / (2 \pi \, \gamma \, m \, r^2(s))] [\Phi(s) - \Phi_k], \quad
L_z = x p_y - y p_x = r^2 \gamma \, m \, \dot{\phi} = -e \, B \, r^2 / 2$$
In summary, progress on large acceptance tabletop muon rings is underway, including energy loss injection (see Fig. 4), 6D cooling, and electric kicker extraction. Many thanks to Juan Gallardo and Franz Kottmann for useful suggestions.
[99]{}
A. Skrinsky and V. Parkhomchuk, [*Sov. J. Part. Nucl.*]{} [**12**]{}, 223 (1981);\
D. Neuffer, [*Part. Accel.*]{} [**14**]{}, 75 (1983); [*Nucl. Instrum. Meth.*]{} [**A532**]{}, 26 (2004);\
G. Penn and J. S. Wurtele, [*Phys. Rev. Lett.*]{} [**85**]{}, 764 (2000);\
K. Kim and C. Wang [*Phys. Rev. Lett.*]{} [**85**]{}, 760 (2000); [*ibid.*]{} [**88**]{}, 184801 (2002);\
G. Franchetti, [*Phys. Rev. ST Accel. Beams*]{} [**4**]{}, 074001 (2001); G. Dugan, 104001. Y.Derbenev, R.Johnson, [*NIM*]{} [**A532**]{}, 470 (2004); [*AIP Conf. Proc.*]{} [**671**]{}, 328 (2003). R. Galea [*et al.,*]{} [*Nucl. Instrum. Meth.*]{} [**A524**]{}, 27 (2004); [*J. Phys.*]{} [**G29**]{}, 1653 (2003);\
A. Caldwell, [*J. Phys.*]{} [**G29**]{}, 1569 (2003); H. Abramowicz [*et al.,*]{} physics/0410017;\
D. Taqqu, [*AIP Conf. Proc.*]{} [**372**]{}, 301 (1996); M. Muhlbauer [*et al.,*]{} [*Nucl. Phys. Proc. Suppl.*]{} [**51A**]{}, 135 (1996); [*Hyperfine Interact.*]{} [**119**]{}, 309 (1995). V. I. Balbekov and A. Van Ginneken, [*AIP Conf. Proc.*]{} [**441**]{}, 310 (1998);\
J. S. Berg, R. C. Fernow, and R. B. Palmer, [*J. Phys.*]{} [**G29**]{}, 1657 (2003);\
R. B. Palmer, [*J. Phys.*]{} [**G29**]{}, 1577 (2003); [*Nucl. Instrum. Meth.*]{} [**A532**]{}, 255 (2004).
A. Blondel [*et al.,*]{} [*Nucl. Instrum. Meth.*]{} [**A451**]{} (2000) 102; R. Palmer [*et al.,*]{} [*ibid.,*]{} 265;\
D. Ayres [*et al.*]{}, physics/9911009; N. Holtkamp [*et al.*]{}, “A neutrino source based on a muon storage ring," Fermilab-Pub-00-108-E; S. Ozaki [*et al.,*]{} “Study II," BNL-52623. G. Budker, [*AIP Conf. Proc.*]{} [**352**]{}, 4 (1996); [**352**]{}, 5 (1996); A. Skrinsky, [**352**]{}, 6 (1996);\
D. Neuffer, [*AIP Conf. Proc.*]{} [**156**]{}, 201 (1987); [*Nucl. Instrum. Meth.*]{} [**A350**]{}, 27 (1996);\
R. Palmer [*et al.,*]{} [*Nucl. Phys. Proc. Suppl.*]{} [**51A**]{}, 61 (1996); acc-phys/9602001;\
R. Raja, A. Tollestrup, [*Phys. Rev.*]{} [**D58**]{}, 013005 (1998); C. Ankenbrandt [*et al.,*]{} [*Phys. Rev. ST Accel. Beams*]{} [**2**]{}, 081001 (1999); M. Alsharo’a [*et al.*]{}, [*ibid.*]{} [**6**]{}, 081001 (2003). R. Davis [*et al.*]{} (Homestake), [*Phys. Rev. Lett.*]{} [**20**]{}, 1205 (1968); [*ApJ*]{} [**496**]{}, 505 (1998);\
Y. Fukuda [*et al.*]{} (SuperK), [*Phys. Rev. Lett.*]{} [**81**]{}, 1562 (1998); [**93**]{}, 101801 (2004);\
Q. R. Ahmad [*et al.*]{} (SNO), [*Phys. Rev. Lett.*]{} [**89**]{}, 011301 (2002); [**92**]{}, 181301 (2004);\
K. Eguchi [*et al.*]{} (KamLAND), [*Phys. Rev. Lett.*]{} [**90**]{}, 021802 (2003); hep-ex/0406035;\
M. Ahn [*et al.*]{} (K2K), [*Phy. Rev. Lett.*]{} [**90**]{}, 041801 (2003); E. Aliu, hep-ex/0411038. S. Geer, [*Phys. Rev.*]{} [**D57**]{}, 6989 (1998); C. Albright [*et al.,*]{} hep-ex/0008064; V. Barger [*et al.,*]{} [*Phys. Rev. Lett.*]{} [**45**]{}, 2084 (1980); A. Cervera [*et al.,*]{} [*Nucl. Phys.*]{} [**B579**]{}, 17 (2000). C. Albright [*et al.,*]{} physics/0411123; M. Maltoni [*et al.,*]{} hep-ph/0405172. V. Barger [*et al.,*]{} [*Phys. Rev. Lett.*]{} [**75**]{}, 1462 (1995); [*Phys. Rept.*]{} [**286**]{}, 1 (1997);\
D. Atwood and A. Soni, [*Phys. Rev.*]{} [**D52**]{}, 6271 (1995); J. F. Gunion, hep-ph/9802258. J. Ellis, LCWS 04, hep-ph/0409140; P. W. Higgs, [Phys. Rev.]{} [**145**]{}, 1156 (1966). R. Godang [*et al.,*]{} hep-ph/0411248; M. Cavaglia and S. Das, hep-th/0404050. R. C. Fernow, [*AIP Conf. Proc.*]{} [**721**]{}, 90 (2004); PAC99, eConf C990329, THP31.
A. A. Garren [*et al.,*]{} [*AIP Conf. Proc.*]{} [**297**]{}, 403 (1994). M. Craddock, [*CERN Cour.*]{} [**44N6**]{}, 23 (2004); D. J. Summers, physics/0411218;\
K. Symon, D. Kerst, L. Jones, L. Laslett, K. Terwilliger, [*Phys. Rev.*]{} [**103**]{}, 1837 (1956);\
J. S. Berg, [*AIP Conf. Proc.*]{} [**642**]{}, 213 (2003); D. Trbojevic [*et al.,*]{} [*ibid.*]{} [**530**]{}, 333 (2000);\
S. Koscielniak, C. Johnstone, [*AIP Conf. Proc.*]{} [**721**]{}, 467 (2004); [*NIM*]{} [**A523**]{}, 25 (2004);\
E. Keil, A. Sessler, [*NIM*]{} (2005); Y. Mori, [*ICFA Beam Dyn. Newslett.*]{} [**29**]{}, 20 (2002). J. Eades and L. M. Simons, [*Nucl. Instrum. Meth.*]{} [**A278**]{}, 368 (1989). J. Eades [*et al.,*]{} [*Nucl.Phys.Proc.Suppl.*]{} [**8**]{}, 457 (1989); E. Aschenauer [*et al.,*]{} [*Sov. J. Nucl. Phys.*]{} [**55**]{}, 856 (1992); L. M. Simons [*et al.,*]{} [*Springer Proc. Phys.*]{} [**59**]{}, 33 (1992); [*Phys. Scripta*]{} [**T22**]{}, 90 (1988); [*Hyperfine Interact.*]{} [**81**]{}, 253 (1993); [*Phys. Bl.*]{} [**48**]{}, 261 (1992); [*Nucl. Instrum. Meth.*]{} [**B87**]{}, 293 (1994); D. Horv[á]{}th [*et al.,*]{} [**B85**]{}, 736 (1994). P. DeCecco [*et al.,*]{} [*Nucl. Instrum. Meth.*]{} [**A394**]{}, 287 (1997); [email protected]. A. Girard [*et al.,*]{} [*Rev. Sci. Instrum.*]{} [**75**]{}, 1381 (2004); A. Zelenski [*et al.,*]{} 1535. H. Busch, [*Annalen Phys.*]{} [**81**]{}, 974 (1926); A. W. Chao and M. Tigner, [*Handbook of Accelerator Physics and Engineering,*]{} page 101 (1999).
[^1]: [email protected] Supported by DE–FG02–91ER40622.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper presents a method for the automatic extraction of subgrammars to control and speeding-up natural language generation NLG. The method is based on explanation-based learning EBL. The main advantage for the proposed new method for NLG is that the complexity of the grammatical decision making process during NLG can be vastly reduced, because the EBL method supports the adaption of a NLG system to a particular use of a language.'
author:
- |
Günter Neumann\
DFKI GmbH\
Stuhlsatzenhausweg 3\
66123 Saarbrücken, Germany\
[[email protected]]{}
title: 'Applying Explanation-based Learning to Control and Speeding-up Natural Language Generation'
---
\#1
Introduction
============
In recent years, a Machine Learning technique known as Explanation-based Learning EBL [@Mitchelletal:86; @HarmelenBundy:88; @Mintonetal:89] has successfully been applied to control and speeding-up natural language parsing [@Rayner:88; @SamuelssonRayner:91; @Neumann:94a; @Samuelsson:94; @SrinivasJoshi:95; @RaynerCarter:96]. The core idea of EBL is to transform the derivations (or [*explanations*]{}) computed by a problem solver (e.g., a parser) to some generalized and compact forms, which can be used very efficiently for solving similar problems in the future. EBL has primarily been used for parsing to automatically specialize a given source grammar to a specific domain. In that case, EBL is used as a method for adapting a general grammar and/or parser to the sub-language defined by a suitable training corpus [@RaynerCarter:96].
A specialized grammar can be seen as describing a domain-specific set of prototypical constructions. Therefore, the EBL approach is also very interesting for natural language generation (NLG). Informally, NLG is the production of a natural language text from computer-internal representation of information, where NLG can be seen as a complex—potentially cascaded—decision making process. Commonly, a NLG system is decomposed into two major components, viz. the strategic component which decides ‘what to say’ and the tactical component which decides ‘how to say’ the result of the strategic component. The input of the tactical component is basically a semantic representation computed by the strategic component. Using a lexicon and a grammar, its main task is the computation of potentially all possible strings associated with a semantic input. Now, in the same sense as EBL is used in parsing as a means to control the range of possible strings as well as their degree of ambiguity, it can also be used for the tactical component to control the range of possible semantic input and their degree of [*paraphrases*]{}.
In this paper, we present a novel method for the automatic extraction of subgrammars for the control and speeding-up of natural language generation. Its main advantage for NLG is that the complexity of the (linguistically oriented) decision making process during natural language generation can be vastly reduced, because the EBL method supports adaption of a NLG system to a particular language use. The core properties of this new method are:
- prototypical occuring grammatical constructions can automatically be extracted;
- generation of these constructions is vastly sped up using simple but efficient mechanisms;
- the new method supports [*partial*]{} matching, in the sense that new semantic input need not be completely covered by previously trained examples;
- it can easily be integrated with recently developed chart-based generators as described in, e.g., [@Neumann:94c; @Kay:96; @Shemtov:96].
The method has been completely implemented and tested with a broad-coverage HPSG-based grammar for English (see sec. \[properties\] for more details).
Foundations
===========
The main focus of this paper is tactical generation, i.e., the mapping of structures (usually representing semantic information eventually decorated with some functional features) to strings using a lexicon and a grammar. Thus stated, we view tactical generation as the inverse process of parsing. Informally, EBL can be considered as an intelligent storage unit of example-based generalized parts of the grammatical search space determined via training by the tactical generator.[^1] Processing of similar new input is then reduced to simple lookup and matching operations, which [*circumvent*]{} re-computation of this already known search space.
We concentrate on constraint-based grammar formalism following a sign-based approach considering linguistic objects (i.e., words and phrases) as utterance-meaning associations [@hpsg2]. Thus viewed, a grammar is a formal statement of the relation between utterances in a natural language and representations of their meanings in some logical or other artificial language, where such representations are usually called [*logical forms*]{} [@Shieber:93]. The result of the tactical generator is a feature structure (or a set of such structures in the case of multiple paraphrases) containing among others the input logical form, the computed string, and a representation of the derivation.
In our current implementation we are using TDL, a typed feature-based language and inference system for constraint-based grammars . TDL allows the user to define hierarchically-ordered types consisting of type and feature constraints. As shown later, a systematic use of type information leads to a very compact representation of the extracted data and supports an elegant but efficient generalization step.
[$${\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h1\/}} \\ }
{\mbox{{{\sc index}}} & {{\it e2\/}} \\ }
{\mbox{{{\sc liszt}}} & {{\it {\mbox{$\left\langle {\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it SandyRel }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h4\/}} \\ }
{\mbox{{{\sc inst}}} & {{\it x5\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it GiveRel }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h1\/}} \\ }
{\mbox{{{\sc event}}} & {{\it e2\/}} \\ }
{\mbox{{{\sc act}}} & {{\it x5\/}} \\ }
{\mbox{{{\sc preparg}}} & {{\it x6\/}} \\ }
{\mbox{{{\sc und}}} & {{\it x7\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it TempOver }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h1\/}} \\ }
{\mbox{{{\sc event}}} & {{\it e2\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}} \right.$}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it Some }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h9\/}} \\ }
{\mbox{{{\sc bv}}} & {{\it x7\/}} \\ }
{\mbox{{{\sc restr}}} & {{\it h10\/}} \\ }
{\mbox{{{\sc scope}}} & {{\it h11\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
\/}} \\ }
{\mbox{{{\sc }}} & {{\it {\mbox{$\left. {\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it ChairRel }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h10\/}} \\ }
{\mbox{{{\sc inst}}} & {{\it x7\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it To }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h12\/}} \\ }
{\mbox{{{\sc arg}}} & {{\it v13\/}} \\ }
{\mbox{{{\sc prep}}} & {{\it x6\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it KimRel }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h14\/}} \\ }
{\mbox{{{\sc inst}}} & {{\it x6\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}}
\right\rangle$}}\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}$$ ]{}
[$${\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc liszt}}} & {{\it {\mbox{$\left\langle {\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it SandyRel }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h4\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it GiveRel }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h1\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it TempOver }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h1\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it Some }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h9\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
\right.$}}\/}} \\ }
{\mbox{{{\sc }}} & {{\it {\mbox{$\left. {\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it ChairRel }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h10\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it To }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h12\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}},
{\mbox{{\begin{tabular}{@{}l@{}}
$\mbox{\it KimRel }^{{\mbox{\begin{math}
\setlength{\arraycolsep}{1mm}
\renewcommand{\arraystretch}{1} \hspace*{-0.35em}
\left[
\begin{array}{@{}l@{~}l@{}} \\[-0.16in] {\mbox{{{\sc handel}}} & {{\it h14\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}}$
\end{tabular}}}}
\right\rangle$}}\/}} \\ } \\[-0.16in] \end{array}
\right] \hspace*{-0.05em}
\end{math}}}$$ ]{}
We are adapting a “flat” representation of logical forms as described in [@Kay:96; @Copestakeetal:96]. This is a minimally structured, but descriptively adequate means to represent semantic information, which allows for various types of under-/overspecification, facilitates generation and the specification of semantic transfer equivalences used for machine translation [@Copestakeetal:96; @Shemtov:96].[^2]
Informally, a flat representation is obtained by the use of extra variables which explicitly represent the relationship between the entities of a logical form and scope information. In our current system we are using the framework called [*minimal recursion semantics*]{} (MRS) described in [@Copestakeetal:96]. Using their typed feature structure notation figure \[exam-mrs\] displays a possible MRS of the string “Sandy gives a chair to Kim” (abbreviated where convenient).
The value of the feature [liszt]{} is actually treated like a set, i.e., the relative order of the elements is immaterial. The feature [handel]{} is used to represent scope information, and [index]{} plays much the same role as a lambda variable in conventional representations (for more details see [@Copestakeetal:96]).
Overview of the method {#overview}
======================
The above figure displays the overall architecture of the EBL learning method. The right-hand part of the diagram shows the linguistic competence base (LCB) and the left the EBL-based subgrammar processing component (SGP).
LCB corresponds to the tactical component of a general natural language generation system NLG. In this paper we assume that the strategic component of the NLG has already computed the MRS representation of the information of an underlying computer program. SGP consists of a training module TM, an application module AM, and the subgrammar, automatically determined by TM and applied by AM.
Briefly, the flow of control is as follows: During the training phase of the system, a new logical form $mrs$ is given as input to the LCB. After grammatical processing, the resulting feature structure $fs(mrs)$ (i.e., a feature structure that contains among others the input MRS, the computed string and a representation of the derivation tree) is passed to TM. TM extracts and generalizes the derivation tree of $fs(mrs)$, which we call the template $templ(mrs)$ of $fs(mrs)$. $templ(mrs)$ is then stored in a [*decision tree*]{}, where indices are computed from the MRS found under the root of $templ(mrs)$. During the application phase, a new semantic input $mrs'$ is used for the retrieval of the decision tree. If a candidate template can be found and successfully instantiated, the resulting feature structure $fs(mrs')$ constitutes the generation result of $mrs'$.
Thus described, the approach seems to facilitate only exact retrieval and matching of a new semantic input. However, before we describe how [*partial matching*]{} is realized, we will demonstrate in more detail the exact matching strategy using the example MRS shown in figure \[exam-mrs\].
#### Training phase
The training module TM starts right after the resulting feature structure $fs$ for the input MRS $mrs$ has been computed. In the first phase, TM extracts and generalizes the derivation tree of $fs$, called the template of $fs$. Each node of the template contains the rule name used in the corresponding derivation step and a generalization of the local MRS. A generalized MRS is the abstraction of the [liszt]{} value of a MRS where each element only contains the (lexical semantic) type and [handel]{} information (the [handel]{} information is used for directing lexical choice (see below)).
In our example $mrs$, figure \[exam-mrs-g\] displays the generalized MRS $mrs_g$. For convenience, we will use the more compact notation:
{(SandyRel h4), (GiveRel h1),\
(TempOver h1), (Some h9),\
(ChairRel h10), (To h12), (KimRel h14)}
Using this notation, figure \[dtree\] (see next page) displays the template $templ(mrs)$ obtained from $fs$. Note that it memorizes not only the rule application structure of a successful process but also the way the grammar mutually relates the compositional parts of the input MRS.
In the next step of the training module TM, the generalized MRS $mrs_g$ information of the root node of $templ(mrs)$ is used for building up an index in a decision tree. Remember that the relative order of the elements of a MRS is immaterial. For that reason, the elements of $mrs_g$ are alphabetically ordered, so that we can treat it as a sequence when used as a new index in the decision tree.
The alphabetic ordering has two advantages. Firstly, we can store different templates under a common prefix, which allows for efficient storage and retrieval. Secondly, it allows for a simple efficient treatment of MRS as sets during the retrieval phase of the application phase.
#### Application phase
The application module AM basically performs the following steps:
1. Retrieval: For a new MRS $mrs'$ we first construct the [*alphabetically sorted*]{} generalized MRS $mrs^{'}_{g}$. $mrs^{'}_{g}$ is then used as a path description for traversing the decision tree. For reasons we will explain soon, traversal is directed by [*type subsumption*]{}. Traversal is successful if $mrs^{'}_{g}$ has been completely processed and if the end node in the decision tree contains a template. Note that because of the alphabetic ordering, the relative order of the elements of new input $mrs'$ is immaterial.
2. Expansion: A successfully retrieved template $templ$ is expanded by [*deterministically*]{} applying the rules denoted by the non-terminal elements from the top downwards in the order specified by $templ$. In some sense, expansion just re-plays the derivation obtained in the past. This will result in a grammatically fully expanded feature structure, where only lexical specific information is still missing. But note that through structure sharing the terminal elements will already be constrained by syntactic information.[^3]
3. Lexical lookup: From each terminal element of the [*unexpanded*]{} template $templ$ the type and [handel]{} information is used to select the corresponding element from the input MRS $mrs'$ (note that in general the MRS elements of the $mrs'$ are much more constrained than their corresponding elements in the generalized MRS $mrs^{'}_{g}$). The chosen input MRS element is then used for performing lexical lookup, where lexical elements are indexed by their relation name. In general this will lead to a set of lexical candidates.
4. Lexical instantiation: In the last step of the application phase, the set of selected lexical elements is unified with the constraints of the terminal elements in the order specified by the terminal yield. We also call this step [*terminal-matching*]{}. In our current system terminal-matching is performed from left to right. Since the ordering of the terminal yield is given by the template, it is also possible to follow other selection strategies, e.g., a semantic head-driven strategy, which could lead to more efficient terminal-matching, because the head element is supposed to provide selectional restriction information for its dependents.
A template together with its corresponding index describes all sentences of the language that share the same derivation and whose MRS are consistent with that of the index. Furthermore, the index and the MRS of a template together define a normalization for the permutation of the elements of a new input MRS. The proposed EBL method guarantees [*soundness*]{} because retaining and applying the original derivation in a template enforces the full constraints of the original grammar.
#### Achieving more generality
So far, the application phase will only be able to re-use templates for a semantic input which has the same semantic type information. However, it is possible to achieve more generality, if we apply a further abstraction step on a generalized MRS. This is simply achieved by selecting a [*supertype*]{} of a MRS element instead of the given specialized type.
The type abstraction step is based on the standard assumption that the word-specific lexical semantic types can be grouped into classes representing morpho-syntactic paradigms. These classes define the upper bounds for the abstraction process. In our current system, these upper bounds are directly used as the supertypes to be considered during the type abstraction step. More precisely, for each element $x$ of a generalized MRS $mrs_g$ it is checked whether its type $T_x$ is subsumed by an upper bound $T_s$ (we assume disjoint sets). Only if this is the case, $T_s$ replaces $T_x$ in $mrs_g$.[^4] Applying this type abstraction strategy on the MRS of figure \[exam-mrs\], we obtain:
{(Named h4), (ActUndPrep h1),\
(TempOver h1), (Some h9),\
(RegNom h10), (To h12), (Named h14)}
where e.g., [Named]{} is the common supertype of [SandyRel]{} and [KimRel]{}, and [ActUndPrep]{} is the supertype of [GiveRel]{}. Figure \[dtree2\] shows the template $templ_g$ obtained from $fs$ using the more general MRS information. Note, that the MRS of the root node is used for building up an index in the decision tree.
Now, if retrieval of the decision tree is directed by type subsumption, the same template can be retrieved and potentially instantiated for a wider range of new MRS input, namely for those which are [*type compatible*]{} wrt. subsumption relation. Thus, the template $templ_g$ can now be used to generate, e.g., the string “Kim gives a table to Peter”, as well as the string “Noam donates a book to Peter”.
However, it will not be able to generate a sentence like “A man gives a book to Kim”, since the retrieval phase will already fail. In the next section, we will show how to overcome even this kind of restriction.
Partial Matching {#details}
================
The core idea behind partial matching is that in case an exact match of an input MRS fails we want at least as many subparts as possible to be instantiated. Since the instantiated template of a MRS subpart corresponds to a phrasal sign, we also call it a [*phrasal template*]{}. For example, assuming that the training phase has only to be performed for the example in figure \[exam-mrs\], then for the MRS of “A man gives a book to Kim”, a partial match would generate the strings “a man” and “gives a book to Kim”.[^5] The instantiated phrasal templates are then combined by the tactical component to produce larger units (if possible, see below).
#### Extended training phase
The training module is adapted as follows: Starting from a template $templ$ obtained for the training example in the manner described above, we extract recursively all possible subtrees $templ_s$ also called [*phrasal templates*]{}. Next, each phrasal template is inserted in the decision tree in the way described above.
It is possible to direct the subtree extraction process with the application of [*filters*]{}, which are applied to the whole remaining subtree in each recursive step. By using these filters it is possible to restrict the range of structural properties of candidate phrasal templates (e.g., extract only saturated NPs, or subtrees having at least two daughters, or subtrees which have no immediate recursive structures). These filters serve the same means as the “chunking criteria” described in [@RaynerCarter:96].
During the training phase it is recognized for each phrasal template $templ_s$ whether the decision tree already contains a path pointing to a previously extracted and already stored phrasal template $templ^{'}_{s}$, such that $templ_s = templ^{'}_{s}$. In that case, $templ_s$ is not inserted and the recursion stops at that branch.
#### Extended application phase
For the application module, only the retrieval operation of the decision tree need be adapted.
Remember that the input of the retrieval operation is the sorted generalized MRS $mrs_g$ of the input MRS $mrs$. Therefore, $mrs_g$ can be handled like a sequence. The task of the retrieval operation in the case of a partial match is now to potentially find all subsequences of $mrs_g$ which lead to a template.
In case of exact matching strategy, the decision tree must be visited only once for a new input. In the case of partial matching, however, the decision tree describes only possible [*prefixes*]{} for a new input. Hence, we have to recursively repeat retrieval of the decision tree as long as the remaining suffix is not empty. In other words, the decision tree is now a finite representation of an infinite structure, because implicitly, each endpoint of an index bears a pointer to the root of the decision tree.
Assuming that the following template/index pairs have been inserted into the decision tree: , , . Then retrieval using the path $abcd$ will return all three templates, retrieval using $aabbcd$ will return template $t_1$ and $t_3$, and $abc$ will only return $t_1$.[^6]
#### Interleaving with normal processing
Our EBL method can easily be integrated with normal processing, because each instantiated template can be used directly as an already found sub-solution. In case of an agenda-driven chart generator of the kind described in [@Neumann:94a; @Kay:96], an instantiated template can be directly added as a [*passive edge*]{} to the generator’s agenda. If passive edges with a wider span are given higher priority than those with a smaller span, the tactical generator would try to combine the largest derivations before smaller ones, i.e., it would prefer those structures determined by EBL.
Implementation {#properties}
==============
The EBL method just described has been fully implemented and tested with a broad coverage HPSG-based English grammar including more than 2000 fully specified lexical entries.[^7] The TDL grammar formalism is very powerful, supporting distributed disjunction, full negation, as well as full boolean type logic.
In our current system, an efficient chart-based bidirectional parser is used for performing the training phase. During training, the user can interactively select which of the parser’s readings should be considered by the EBL module. In this way the user can control which sort of structural ambiguities should be avoided because they are known to cause misunderstandings. For interleaving the EBL application phase with normal processing a first prototype of a chart generator has been implemented using the same grammar as used for parsing.
First tests has been carried out using a small test set of 179 sentences. Currently, a parser is used for processing the test set during training. Generation of the extracted templates is performed solely by the EBL application phase (i.e., we did not considered integration of EBL and chart generation). The application phase is very efficient. The average processing time for indexing and instantiation of a sentence level template (determined through parsing) of an input MRS is approximately one second.[^8] Compared to parsing the corresponding string the factor of speed up is between 10 to 20. A closer look to the four basic EBL-generation steps: indexing, instantiation, lexical lookup, and terminal matching showed that the latter is the most expensive one (up to 70% of computing time). The main reasons are that 1.) lexical lookup often returns several lexical readings for an MRS element (which introduces lexical non-determinism) and 2.) the lexical elements introduce most of the disjunctive constraints which makes unification very complex. Currently, terminal matching is performed left to right. However, we hope to increase the efficiency of this step by using head-oriented strategies, since this might help to re-solve disjunctive constraints as early as possible.
Discussion
==========
The only other approach I am aware of which also considers EBL for NLG is [@Samuelsson:95a; @Samuelsson:95b]. However, he focuses on the compilation of a logic grammar using LR-compiling techniques, where EBL-related methods are used to optimize the compiled LR tables, in order to avoid spurious non-determinisms during normal generation. He considers neither the extraction of a specialized grammar for supporting controlled language generation, nor strong integration with the normal generator.
However, these properties are very important for achieving high applicability. Automatic grammar extraction is worthwhile because it can be used to support the definition of a [*controlled*]{} domain-specific language use on the basis of training with a general source grammar. Furthermore, in case exact matching is requested only the application module is needed for processing the subgrammar. In case of normal processing, our EBL method serves as a speed-up mechanism for those structures which have “actually been used or uttered”. However, completeness is preserved.
We view generation systems which are based on “canned text” and linguistically-based systems simply as two endpoints of a contiguous scale of possible system architectures (see also [@Daleetal:94]). Thus viewed, our approach is directed towards the automatic creation of application-specific generation systems.
Conclusion and Future Directions {#future-work}
================================
We have presented a method of automatic extraction of subgrammars for controlling and speeding up natural language generation (NLG). The method is based on explanation-based learning (EBL), which has already been successfully applied for parsing. We showed how the method can be used to train a system to a specific use of grammatical and lexical usage.
We already have implemented a similar EBL method for parsing, which supports on-line learning as well as statistical-based management of extracted data. In the future we plan to combine EBL-based generation and parsing to one [*uniform*]{} EBL approach usable for high-level performance strategies which are based on a strict interleaving of parsing and generation (cf. [@NeumannVanNoord:94c; @Neumann:94a]).
Acknowledgement
===============
The research underlying this paper was supported by a research grant from the German Bundesministerium für Bildung, Wissenschaft, Forschung und Technologie (BMB+F) to the DFKI project [paradime]{} FKZ ITW 9704.
I would like to thank the HPSG people from CSLI, Stanford for their kind support and for providing the HPSG-based English grammar. In particular I want to thank Dan Flickinger and Ivan Sag. Many thanks also to Walter Kasper for fruitful discussions.
Copestake, A., D. Flickinger, R. Malouf, S. Riehemann, and I. Sag. 1996. Translation using minimal recursion semantics. In [*Proceedings, 6th International Conference on Theoretical and Methodological Issues in Machine Translation*]{}.
Dale, R., W. Finkler, R. Kittredge, N. Lenke, G. Neumann, C. Peters, and M. Stede. 1994. Report from working group 2: Lexicalization and architecture. In W. Hoeppner, H. Horacek, and J. Moore, editors, [*Principles of Natural Language Generation*]{}, Dagstuhl-Seminar-Report; 93. [Schlo[ß]{}]{} Dagstuhl, Saarland, Germany, Europe, pages 30–39.
Kay, M. 1996. Chart generation. In [*34th Annual Meeting of the Association for Computational Linguistics*]{}, Santa Cruz, Ca.
Krieger, Hans-Ulrich and Ulrich Schäfer. 1994. —a type description language for constraint-based grammars. In [*Proceedings of the 15th International Conference on Computational Linguistics, COLING-94*]{}, pages 893–899.
Minton, S., J. G. Carbonell, C. A. Knoblock, D. R.Kuokka, O. Etzioni, and Y.Gi. 1989. Explanation-based learning: A problem solving perspective. , 40:63–115.
Mitchell, T., R. Keller, and S. Kedar-Cabelli. 1986. Explanation-based generalization: a unifying view. , 1:47–80.
Neumann, G. 1994a. Application of explanation-based learning for efficient processing of constraint based grammars. In [*Proceedings of the Tenth IEEE Conference on Artificial Intelligence for Applications*]{}, pages 208–215, San Antonio, Texas, March.
Neumann, G. 1994b. . thesis, Universit[ät]{} des Saarlandes, Germany, Europe, November.
Neumann, G. and G. van Noord. 1994. Reversibility and self-monitoring in natural language generation. In Tomek Strzalkowski, editor, [*Reversible Grammar in Natural Language Processing*]{}. Kluwer, pages 59–96.
Pollard, C. and I. M. Sag. 1994. . Center for the Study of Language and Information Stanford.
Rayner, M. 1988. Applying explanation-based generalization to natural language processing. In [*Proceedings of the International Conference on Fifth Generation Computer Systems*]{}, Tokyo.
Rayner, M. and D. Carter. 1996. Fast parsing using pruning and grammar specialization. In [*34th Annual Meeting of the Association for Computational Linguistics*]{}, Morristown, New Jersey.
Samuelsson, C. 1994. . thesis, Swedish Institute of Computer Science, Kista, Sweden, Europe.
Samuelsson, C. 1995a. An efficient algorithm for surface generation. In [*Proceedings of the 14th International Joint Conference on Artificial Intelligence*]{}, pages 1414–1419, Montreal, Canada.
Samuelsson, C. 1995b. Example-based optimization of surface-generation tables. In [*Proceedings of Recent Advances in Natural Language Processing*]{}, Velingrad, Bulgaria, Europe.
Samuelsson, C. and M. Rayner. 1991. Quantitative evaluation of explanation-based learning as an optimization tool for a large-scale natural language system. In [*IJCAI-91*]{}, pages 609–615, Sydney, Australia.
Shemtov, H. 1996. . In [*Proceedings of the 16th International Conference on Computational Linguistics (COLING)*]{}, pages 919–924, Kopenhagen, Denmark, Europe.
Shieber, S. M. 1993. The problem of logical-form equivalence. , 19:179–190.
Srinivas, B. and A. Joshi. 1995. Some novel applications of explanation-based learning to parsing lexicalized tree-adjoining grammars. In [*33th Annual Meeting of the Association for Computational Linguistics*]{}, Cambridge, MA.
van Harmelen, F. and A. Bundy. 1988. Explanation-based generalization=partial evaluation. , 36:401–412.
[^1]: In case a reversible grammar is used the parser can even be used for processing the training corpus.
[^2]: But note, our approach does not depend on a flat representation of logical forms. However, in the case of conventional representation form, the mechanisms for indexing the trained structures would require more complex abstract data types (see sec. \[details\] for more details).
[^3]: It is possible to perform the expansion step off-line as early as the training phase, in which case the application phase can be sped up, however at the price of more memory being taken up.
[^4]: Of course, if a very fine-grained lexical semantic type hierarchy is defined then a more careful selection would be possible to obtained different degrees of type abstraction and to achieve a more domain-sensitive determination of the subgrammars. However, more complex type abstraction strategies are then needed which would be able to find appropriate supertypes automatically.
[^5]: If we would allow for an exhaustive partial match (see below) then the strings “a book” and “Kim” would additionally be generated.
[^6]: It is possible to parameterize our system to perform an [*exhaustive*]{} or a [*non-exhaustive*]{} strategy. In the non-exhaustive mode, the longest matching prefixes are preferred.
[^7]: This grammar has been developed at CSLI, Stanford, and kindly be provided to the author.
[^8]: EBL-based generation of all possible templates of an input MRS is less than 2 seconds. The tests have been performed using a Sun UltraSparc.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We discuss the current status of our automatic perturbation theory program as applied to Fermilab Fermions. We give an overview of our methods, a discussion of tree level matching, and one loop results for the coefficients of the higher dimension kinetic operators.'
address: 'Department of Physics, Simon Fraser University, Burnaby, British Columbia, Canada'
author:
- 'Matthew A. Nobes and Howard Trottier'
title: One Loop Renormalization of Fermilab Fermions
---
A dominant source of error in current lattice calculations is the errors due to finite lattice spacing $a$. One way to reduce these errors is to use improved actions. The general structure for an improved action is $$\mathcal{L} = \mathcal{L}_{0} + \sum_{n=0}^{\infty} c_{n}(g_{0},m_{0}) \mathcal{L}_{n}$$ The series here is an expansion in the dimensionality of the various operators. Each of these terms comes with a new coupling constant $c_{n}$.
In order to use this type of action we must do two things, we must truncate the series at some specified order in $n$ and we must calculate the new couplings somehow. Fixing the new couplings can be done in a number of ways, for example, the first step could be to make an expansion in powers of $m_{0}$ for light quarks, or powers of $1/m_{0}$ for heavy quarks. The remaining dependence on $g_{0}$ can be determined perturbativly.
It is desirable to have a method of improvement which does not rely on light or heavy quark mass expansions. One such approach is the Fermilab approach [@mainfermi] which orders the expansion operators by dimension only (this amounts to a small momentum expansion). The unimproved fermilab action consists of dimension two and four operators, $$\begin{aligned}
S_{0} & = & \int dx \bar{\psi} \left\lbrace m_{0} + \frac{1+\gamma_{0}}{2} D^{-}_{0} -
\frac{1-\gamma_{0}}{2}D^{+}_{0} \right. \nonumber \\
& + & \left. \zeta \mathbf{\gamma}\cdot\mathbf{D} -\frac{r_{s}\zeta}{2} \mathbf{\triangle} \right\rbrace \psi \end{aligned}$$ For the definitions of the various derivatives we refer the reader to [@mainfermi]. This action includes an additional redundant dimension five operator, whose coefficient $r_{s}\zeta$ can be tuned to remove the fermion doubling problem.
To improve to $\mathcal{O}(a^{2})$ a large number of of dimension five and six operators must be added. The identification and tree level matching of these operators is discussed in [@bugra], we will not repeat it here. This report is concerned primarily with the one loop determination of the coefficients of the dimension six kinetic energy operators, $ \lbrace \mathbf{\gamma} \cdot \mathbf{D} , D^{2} \rbrace$ and $ \gamma_{i} D_{i}^{3}$. For this matching we will set all of the other improvement coefficients (again see [@bugra] for a full list) to zero. This leaves the action, $$\label{action}
S = S_{0} + \int dx \bar{\psi} \left[ c_{1} \lbrace \mathbf{\gamma} \cdot \mathbf{D} , D^{2} \rbrace
+ c_{2} \gamma_{i} D_{i}^{3} \right] \psi.$$ For these kinetic operators the determination of the coefficients is very straightforward. For any lattice action, the quark energy can always be expanded in powers of the three momentum $$\label{fullenergy}
E(\mathbf{p}) = M_{1} + \frac{\mathbf{p}^{2}}{2M_{2}} - \frac{w_{4}}{6} \sum_{i}p_{i}^{4}
- \frac{(\mathbf{p}^{2})^{2}}{8M_{4}^{3}} + \mathcal{O}(p^{6})$$ By tuning the three couplings $\zeta$, $c_{1}$ and $c_{2}$, we can imposed various conditions on this action. The coupling $\zeta$ can be tuned to ensure that $M_{1}=M_{2}$, $c_{1}$ and $c_{2}$ can be tuned to ensure that $M_{2}=M_{4}$ and $w_{4}=0$. The latter condition restores rotational symmetry to $\mathcal{O}(a^{2})$. Because it only produces an overall shift in the zero of energy it is not necessary to tune $M_{1}=M_{2}$ using $\zeta$. Rather, we’ll just set $\zeta=1$ for convenience.
With $\zeta=1$ we have two remaining improvement conditions, $M_{2}=M_{4}$ and $w_{4}=0$. We impose these in perturbation theory. The tree level calculation has been preformed in [@bugra]. Setting $r_{s}=\zeta=1$, the tree level values are $$\begin{aligned}
\label{treec1}
c^{(0)}_{1}(m_{0}) & = & \frac{m_{0}(2+m_{0})}{32} \nonumber \\
& \times & \left[\left(\frac{1}{1+m_{0}}+\frac{2}{m_{0}(2+m_{0})}\right)^{3}\right. \nonumber \\
& - & \frac{1}{(1+m_{0})^{2}} - \frac{8(1+m_{0})}{m_{0}^{2}(2+m_{0})^{2}} \nonumber \\
& - & \left. \frac{8}{m_{0}^{3}(2+m_{0})^{3}} - \frac{4}{m_{0}^{2}(2+m_{0})^{2}} \right],\end{aligned}$$ and, $$\begin{aligned}
\label{treec2}
c^{(0)}_{2}(m_{0}) & = & \frac{-m_{0}(2+m_{0})}{12} \nonumber \\
& \times & \left[\frac{1}{4(1+m_{0})}+\frac{2}{m_{0}(2+m_{0})}\right].\end{aligned}$$ In keeping with the Fermilab approach, these coefficients have their full dependence on the bare mass. Notationally, bracketed superscripts on any quantity denote the order of the expansion in the bare coupling. For example, the rest mass $M_{1}$ has the expansion $$M_{1} = M_{1}^{(0)} + g_{0}^{2} M_{1}^{(2)} + g_{0}^{4} M_{1}^{(4)} + \ldots$$
The one loop calculations detailed below have been preformed using our automatic perturbation theory methods. The core of this method is the use of the Lüscher – Weisz vertex generation algorithm [@luscher], with some straightforward modifications to include arbitrary quark actions. The vertex rules generated by this method can be used to construct Feynman diagrams, and VEGAS can be used to preform the loop sums. Additionally we use triple twisted periodic boundary conditions, with $L=200$, as an infrared regulator. Further details on how these types of calculations are preformed can be found in [@nobes].
\[fig:treenergy\]
The automatically generated vertex rules can be used to check the tree level calculations in [@bugra]. Figure \[fig:treenergy\] shows the ratio of masses extracted from the automatically generated quark propagator over a range of momenta. Correct tree level matching of $c_{1}^{(0)}$ and $c_{2}^{(0)}$ should give equation (\[fullenergy\]) with $M_{2}^{(0)}=M_{4}^{(0)}$ over the full range of momenta. Clearly the improved action satisfies this, whereas the unimproved action does not. We have also preformed this tree level checking for the temporal one gluon vertex, verifying the expressions for the coefficients $c_{e}$, $c_{6}$ and $c_{8}$ in [@bugra]. Work is currently underway to fix and confirm the remaining tree level coefficients.
We turn now to the one loop calculation of $c_{1}$ and $c_{2}$. In order to fix these coefficients we must compute the quark self energy to one loop. Details of how to preform this calculation can be found in [@mertens]. Our calculation mirrors [@mertens], apart from the different action ((\[action\]) with (\[treec1\]) and (\[treec2\])), and the method of obtaining the Feynman rules. Expanding (\[fullenergy\]) in powers of the bare coupling gives the one loop energy (recall $M_2^{(0)}=M_{4}^{(0)}$) $$\begin{aligned}
\label{e2}
E^{(2)} & = & M_{1}^{(2)} - \mathbf{p}^{2} \frac{M_{2}^{(2)}}{2\left(M_{2}^{(0)}\right)^{2}} \nonumber \\
& - & w_{4}^{(2)} \sum_{i} p_{i}^{4} + (\mathbf{p}^{2})^{2}
\frac{3M_{4}^{(2)}}{8\left(M_{2}^{(0)}\right)^{2}}.\end{aligned}$$ In terms of the quantities that appear here, the improvement conditions are, $$\label{impcon}
M_{2}^{(2)} = M_{4}^{(2)}, \quad w_{4}^{(2)}=0.$$
The one loop rest mass $M_{1}^{(2)}$ is easy to evaluate numerically. Figure \[fig:m12\] shows its value over a wide range of bare masses. The limit of large bare masses is approaching the value of $M_{1}^{(2)} = 0.1681(4)$ reported in [@mertens].
To compute $c_{1}$ and $c_{2}$ we start by computing $E^{(2)}$ using the tree level coefficients. Quantities computed with only the tree level values for $c_{1}$ and $c_{2}$ will be denoted with bars. Setting $p_{y}=p_{z}=0$, and computing $\bar{E}^{(2)}$ over a range of (small) $p_{x}$ allows $\bar{M}_{2}^{(2)}$ and the combination $\bar{\beta} = \frac{3\bar{M}_{4}^{(2)}}{8\left(M_{2}^{(0)}\right)} - \bar{w}_{4}^{(2)}$ to be extracted from fits to the data. To separate $\bar{M}_{4}^{(2)}$ from $\bar{w}_{4}^{(2)}$ we compute $\bar{E}^{(2)}$ at fixed $p_{x}$ over a range of $p_{y}$. The cross term $p_{x}^{2}p_{y}^{2}$ has coefficient $\frac{3\bar{M}_{4}^{(2)}}{4\left(M_{2}^{(0)}\right)}$ which can be extracted from a fit to the numerical values of the one loop energy, as illustrated in Figure \[fig:ebarx\]. This procedure gives the following values ($m_{0}=1$), $$\begin{aligned}
M_{1}^{(2)} = 0.08561(60) & \quad & \bar{M}_{2}^{(2)} = 0.234(10) \nonumber \\
\bar{M}_{4}^{(2)} = 0.263(47) & \quad & \bar{w}_{4}^{(2)} = -0.241(52) \nonumber\end{aligned}$$
\[fig:ebarx\]
The coefficients $c_1^{(2)}$ and $c_2^{(2)}$ provide the counterterms that are necessary to correct for these one-loop contributions to the energy, in order to satisfy the improvement conditions (\[impcon\]) Hence we obtain the preliminary values (recall $m_{0}=1$): $$c_1^{(2)} = 0.22(20) \quad c_2^{(2)} = -2.20(56).$$ These values ensure the matching conditions are satisfied and give the value $M_{2}^{(2)}=M_{4}^{(2)}=0.232(10)$.
To conclude, this report illustrates the application of automatic perturbation theory to the Fermilab fermion action. These techniques can be used, either to do or to check, tree level matching, as well as one loop matching. We have presented results for the one loop rest mass $M_{1}^{(2)}$ along with a first matching calculation for $c_{1}$ and $c_{2}$. Work is now underway to extend the one loop matching presented here to more values of the bare mass at higher precision, and more coefficients.
We thank A. Kronfeld, A. El-Khadra and P. Mackenzie for fruitful discussion. HDT is supported by NSERC.
[99]{} A.X. El-Khadra, A.S. Kronfeld, and P.B. Mackenzie, Phys. Rev. **D55**, 3933 (1997). M. Oktay, talk presented at this conference. M.Lüscher and P.Weisz, Nucl. Phys. **B266**, 309 (1986). M. Nobes, H. Trottier, G.P. Lepage, and Q. Mason, Nucl.Phys. Proc. Suppl. **106**, 838 (2002) B.P.G. Mertens, A.S.Kronfeld and A.X. El-Khadra, Phys. Rev. **D58**, 034505 (1998).
|
{
"pile_set_name": "ArXiv"
}
|
---
address:
- |
Laboratoire de Mathématiques de Versailles, Université de Versailles-Saint Quentin en Yvelines, 45 avenue des États-Unis\
78035 Versailles Cedex\
France
- 'School of Mathematics, University of East Anglia, Norwich NR4 7TJ, United Kingdom'
author:
- 'V. Sécherre'
- 'S. Stevens'
title: |
Smooth representations of $\GL_m(\D)$\
VI : semisimple types
---
Introduction {#introduction .unnumbered}
============
In [@BK1], Bushnell and Kutzko described a general approach to understanding the category of smooth (complex) representations of a reductive $p$-adic group $\G$: the theory of *types*. This is based on the Bernstein decomposition of the category [@Be] into indecomposable full subcategories, indexed by pairs $(\L,\pi)$, with $\L$ a Levi subgroup of $\G$ and $\pi$ an irreducible cuspidal representation of $\L$, up to a certain equivalence relation. A *type* for such a subcategory is a pair $(\K,\rho)$, with $\K$ a compact open subgroup of $\G$ and $\rho$ an irreducible representation of $\K$, such that the irreducible representations in the subcategory are precisely those irreducible representations of $\G$ which contain $\rho$. In this case, there is an equivalence of categories between the subcategory and the category of left modules over the spherical Hecke algebra $\Hh(\G,\rho)$. Thus one can classify the representations of $\G$ by first constructing a type for each subcategory, and then computing the spherical Hecke algebras.
This programme has been completed for general linear groups over a $p$-adic field (Bushnell–Kutzko [@BK; @BKsemi]), for special linear groups over a $p$-adic field (Bushnell–Kutzko [@BKSL1; @BKSL2], Goldberg–Roche [@GR1; @GR2]) up to the computation of a two-cocycle in the description of the Hecke algebra, and for three-dimensional unitary groups in odd residual characteristic (Blasco [@Bla]). In this paper, following previous work in [@VS1; @VS2; @VS3; @SS4; @BSS], we complete the programme for inner forms of general linear groups. The Hecke algebras that arise are all naturally isomorphic to products of affine Hecke algebras of type $A$.
Let $\D$ be a division algebra over a locally compact nonarchimedean local field $\F$, and let $\G=\GL_m(\D)$, with $m$ a positive integer; we will also think of $\G$ as the group of automorphisms of a right $\D$-vector space $\V$. Denote by $\Rr(\G)$ the category of smooth complex representations of $\G$. In order to describe our results more precisely, we begin by recalling the Bernstein decomposition [@Be], in the language of [@BK1]. For $\L$ a Levi subgroup of $\G$, denote by $\X(\L)$ the complex torus of *unramified characters* of $\L$: that is, smooth homomorphisms $\L\to\CC^\times$ which are trivial on all compact subgroups of $\L$. For $\pi$ an irreducible cuspidal representation of $\L$, we write $[\L,\pi]_\G$ for the *$\G$-inertial equivalence class* of $(\L,\pi)$: that is, the set of pairs $(\L',\pi')$, consisting of a Levi subgroup $\L'$ and an irreducible cuspidal representation $\pi'$ of $\L'$, such that $(\L,\pi)$ and $(\L',\pi'\otimes\chi')$ are $\G$-conjugate, for some unramified character $\chi'\in\X(\L')$. We denote by $\BB(\G)$ the set of $\G$-inertial equivalence classes of pairs $(\L,\pi)$ as above.
To $\ss=[\L,\pi]_\G$ an inertial equivalence class we associate a full subcategory $\Rr^\ss(\G)$ of $\Rr(\G)$, whose objects are those representations all of whose subquotients have cuspidal support in $\ss$. Then the Bernstein decomposition says that $$\Rr(\G) = \prod_{\ss\in\BB(\G)}\Rr^\ss(\G).$$
Bushnell–Kutzko’s theory of types [@BK1] is a strategy to understand the subcategories in this decomposition. For $\ss\in\BB(\G)$, an *$\ss$-type* is a pair $(\K,\rho)$, with $\K$ a compact open subgroup of $\G$ and $(\rho,\Ww)$ an irreducible smooth representation of $\K$, such that the irreducible representations in $\Rr^\ss(\G)$ are precisely those irreducible representations of $\G$ which contain $\rho$. In that case, there is an equivalence of categories $$\begin{aligned}
\Rr^\ss(\G) &\to& \Hh(\G,\rho)\hbox{-Mod} \\
\Vv &\mapsto& \Hom_\K(\Ww,\Vv),\end{aligned}$$ where $\Hh(\G,\rho)$ is the convolution algebra of compactly supported $\End_\CC(\check{\Ww})$-valued function $f$ of $\G$ which satisfy $f(hgk)=\check\rho(h)f(g)\check\rho(k)$, for $h,k\in\K$, $g\in\G$ (the spherical Hecke algebra).
Our main result is the following:
\[thm:main\] Let $\ss\in\BB(\G)$. There exists an $\ss$-type $(\K,\rho)$ in $\G$, such that $$\Hh(\G,\rho)\cong\bigotimes_{i=1}^l \Hh(r_i,q_{\F}^{f_i}),$$ for some positive integers $l$ and $r_i,f_i$, for $1\le i\le l$.
Here $q_\F$ denotes the cardinality of the residue field of $\F$, and $\Hh(n,q)$ is the affine Hecke algebra of type $\tilde A_{n-1}$ with parameter $q$. In particular, the category of modules of such algebras is completely understood, through the work of Kazhdan–Lusztig [@KL]. The values of $l,r_i,f_i$ can be described as follows.
To $\pi$ an irreducible cuspidal representation of $\G=\GL_m(\D)$, we may associate two invariants. First there is the *torsion number $n(\pi)$*, the (finite) number of unramified characters $\chi$ of $\G$ such that $\pi\simeq\pi\chi$. Second, we have the *reducibility number $s(\pi)$*: writing $\tilde\G=\GL_{2m}(\D)$ and $\tilde\P$ for the standard parabolic subgroup of $\tilde\G$ with Levi subgroup $\G\times\G$, it is the unique positive real number such that the induced representation $\Ind_{\tilde\P}^{\tilde\G}\pi\otimes\pi\nu^{s(\pi)}$ (with respect to normalized parabolic induction) is reducible, where $\nu(g)=|\Nrd(g)|_\F$, $\Nrd$ denotes the reduced norm of $\A$ over $\F$, and $|\cdot|_\F$ is the normalized absolute value on $\F$ (see [@VSU0 Theorem 4.6]).
Now suppose $\ss=[\L,\pi]_\G$ is an inertial equivalence class. The Levi subgroup $\L$ is the stabilizer of some decomposition $\V=\bigoplus_{j=1}^r\V^j$ into subspaces, which gives an identification of $\L$ with $\prod_{j=1}^r\GL_{m_j}(\D)$, where $m_j=\dim_\D\V^j$. We can then write $\pi=\bigotimes_{j=1}^r\pi_j$, for $\pi_j$ an irreducible cuspidal representation of $\G_j=\GL_{m_j}(\D)$. We define an equivalence relation on $\{1,\ldots,r\}$ by $$\label{eqn:equivrel}
j\sim k\ \Longleftrightarrow\ m_j=m_k\hbox{ and }[\G_j,\pi_j]_{\G_j}=[\G_k,\pi_k]_{\G_k},$$ where we have identified $\G_j$ with $\G_k$ whenever $m_j=m_k$. We may, and do, assume that $\pi_j=\pi_k$ whenever $j\sim k$, since this does not change the inertial class $\ss$. Denote by $S_1,\ldots,S_l$ the equivalence classes. For $i=1,\ldots,l$, we set $$r_i=\#S_i,\qquad f_i= n(\pi_j)s(\pi_j),\hbox{ for any }j\in S_i.$$ These are then the parameters appearing in the Hecke algebras of the Main Theorem.\[thm:main\]
We now describe in more detail the construction of the types. The starting point is the construction of the irreducible cuspidal representations of $\G$, which was achieved in [@VS3; @SS4]. In [@VS3], generalizing the work of Bushnell–Kutzko for $\GL_n(\F)$ [@BK], the first author constructed a set of pairs $(\J,\l)$ called *simple types*. Amongst these are the *maximal simple types*, which give rise to cuspidal representations: if $(\J,\l)$ is a maximal simple type then $\l$ extends to a representation $\tilde\l$ of its normalizer $\tilde\J$ and the compactly-induced representation $\cInd_{\tilde\J}^{\G}\tilde\l$ is irreducible and cuspidal. The main result of [@SS4] is that all irreducible cuspidal representations of $\G$ arise in this way. Here we prove more:
\[thm:cuspidal\] Let $\pi$ be an irreducible cuspidal representation of $\G$ and $\ss=[\G,\pi]_\G$. There is a maximal simple type $(\J,\l)$ contained in $\pi$, and any such is an $\ss$-type. Moreover, it is unique up to $\G$-conjugacy: that is, if $(\J_i,\l_i)$, for $i=1,2$, are maximal simple types contained in $\pi$ then there exists $g\in\G$ such that ${}^g\J_1=\J_2$ and ${}^g\l_1=\l_2$.
The uniqueness (up to conjugacy) is proved in Corollary \[cor:maxsimpleconj\]. We remark that, in the case of $\GL_n(\F)$, the uniqueness here follows from an “intertwining implies conjugacy” result for simple types. In the case of $\G$ there is no such result for two reasons: Firstly there is an extra invariant of a simple type, the *embedding type* (see paragraph \[SS.embed\]) and simple types with different embedding types may intertwine but be non-conjugate. Secondly there is an action of a galois group on simple types, and any two types in the same orbit will intertwine but may be non-conjugate; this phenomenon arises already for level zero representations in [@GSZ]. Nonetheless, the embedding type of a *sound simple type* (see §\[S.IIC\]) is determined by any irreducible representation containing it. Moreover, in the case of maximal simple types (which are always sound), the galois action can be realized by conjugation by an element of the normalizer of $\J$, and the uniqueness follows from this.
We turn to the case of a general $\G$-inertial equivalence class $\ss=[\L,\pi]_\G$, for which we have the corresponding cuspidal $\L$-inertial equivalence class $\ss_\L=[\L,\pi]_\L$. In [@BK1], Bushnell and Kutzko present a general framework for constructing an $\ss$-type from an $\ss_\L$-type $(\J_\L,\l_\L)$: the theory of *covers*. We do not recall precisely the definition of a cover here, only that it should be a pair $(\J_\G,\l_\G)$ which has an Iwahori decomposition with respect to any parabolic subgroup with Levi component $\L$, such that the Hecke algebra $\Hh(\G,\l_\G)$ contains a suitable invertible element. If one has such a cover $(\J_\G,\l_\G)$ then it is an $\ss$-type.
The normalizer $\N_\G(\L)$ acts on $\BB(\L)$ by conjugation and there is a Levi subgroup $\M$ of $\G$ which is minimal for the property of containing the $\N_\G(\L)$-stabilizer of $\ss_\L$. Then we also have an $\M$-inertial equivalence class $\ss_\M=[\L,\pi]_\M$. The strategy now is first to construct a cover $(\J_\M,\l_\M)$ of $(\J_\L,\l_\L)$, and then a cover $(\J_\G,\l_\G)$ of $(\J_\M,\l_\M)$ – by transitivity of covers, this will give the required $\ss$-type.
In our situation, we do indeed have an $\ss_\L$-type: Writing $\L=\prod_{j=1}^r\GL_{m_j}(\D)$ and $\pi=\bigotimes_{j=1}^r\pi_j$ as above, for $\pi_j$ an irreducible cuspidal representation of $\G_j=\GL_{m_j}(\D)$, there is a maximal simple type $(\J_j,\l_j)$ which is a $[\G_j,\pi_j]_{\G_j}$-type, by Theorem \[thm:cuspidal\]. Then, putting $$\J_\L=\prod_{j=1}^r \J_j,\qquad \l_\L=\bigotimes_{j=1}^r \l_j,$$ it is clear that $(\J_\L,\l_\L)$ is an $\ss_\L$-type. The Levi subgroup $\M$ is then that defined by the equivalence relation : it is the stabilizer of the decomposition $\V=\bigoplus_{i=1}^l \Y^i$, where $\Y^i=\bigoplus_{j\in S_i}\V^j$.
The first case to consider is when $\M=\G$, that is, the case $l=1$ of the Main Theorem. In this situation we have the following result, which summarizes [@VS3 Proposition 5.5, Théorème 4.6] and [@SS4 Théorème 5.23]:
\[thm:simplecover\] Let $\pi_0$ be an irreducible cuspidal representation of $\G_0=\GL_{m_0}(\D)$, with $m=rm_0$, and let $(\J_0,\l_0)$ be a maximal simple type which is a $[\G_0,\pi_0]_{\G_0}$-type. Let $\L={\G_0}^r$ be a Levi subgroup of $\G$ with irreducible cuspidal representation $\pi={\pi_0}^{\otimes r}$, and put $\ss=[\L,\pi]_\G$. Put $\J_\L={\J_0}^r$ and $\l_\L={\l_0}^{\otimes r}$. Then there is an $\ss$-type $(\J_\G,\l_\G)$ which is a cover of $(\J_\L,\l_\L)$, and $$\Hh(\G,\l_\G)\simeq \Hh(r,q_\F^f),$$ where $f=n(\pi_0)s(\pi_0)$. Moreover, there is a simple type $(\J,\l)$ in $\G$ such that $\l=\Ind_{\J_\G}^{\J}\l_\G$.
Now we turn to the general case of arbitrary $\M$. In order to describe the covering process, we need to recall some detail of the structure of simple types.
Let $(\J,\l)$ be a simple type contained in an irreducible representation $\pi$ of $\G$. There is a particular filtration of pro-$p$ subgroups $\{\H^{t+1}:t\ge 0\}$ of $\J$ such that $\l$ restricts to a multiple of a character $\theta^{(t)}$ on $\H^{t+1}$, and $\theta^{(0)}\ |\ \H^{t+1}=\theta^{(t)}$. These characters are called *simple characters of level $t$*. Simple characters were the main object of study of [@VS1; @BSS] and they exhibit remarkable functorial properties, as in the case $\D=\F$ [@BK; @BH]. In particular, it is possible to transfer them to the multiplicative group of other central simple $\F$-algebras. A convenient and powerful way to express this is in terms of *endo-classes* (see [@BH; @BSS] and §\[S.common\]): the simple character $\theta^{(t)}$ determines and endo-class $\Theta^{(t)}$, which depends only on the representation $\pi$.
Now let $\pi=\bigotimes_{j=1}^r \pi_j$ be a cuspidal representation of $\L$ as above and denote by $\Theta_j^{(t)}$ the endo-class of level $t$ determined by $\pi_j$. (We are assuming a normalization of the index in the filtrations.) For each integer $t\ge 0$, we define an equivalence relation on $\{1,\ldots,r\}$ by $$j \sim_t k \Longleftrightarrow \Theta_j^{(t)}\hbox{ is endo-equivalent to }\Theta_k^{(t)}.$$ As for the equivalence relation , this determines a Levi subgroup $\M_t$. Note that $$\M \subseteq \M_0;\qquad \M_t\subseteq\M_{t'}\hbox{, for }t\ge t';\qquad\hbox{and }\M_t=\G\hbox{, for sufficiently large }t.$$ It is useful to extend the notation and put $\M_t=\M$ for $t<0$. Of course, although the Levi subgroups are indexed by an integer $t$, there are only finitely many in $\{\M_t:t\in\ZZ\}$.
Theorem \[thm:simplecover\] provides a cover $(\J_\M,\l_\M)$ of $(\J_\L,\l_\L)$ in $\M$ and the Main Theorem now follows from the transitivity of covers and:
\[thm:coverhecke\] For $t\ge t'$, there are a cover $(\J_{\M_{t'}},\l_{\M_{t'}})$ of $(\J_{\M_t},\l_{\M_t})$ in $\G$ and a support-preserving isomorphism of Hecke algebras $\Hh(\M_{t},\l_{\M_{t}})\simeq\Hh(\M_{t'},\l_{\M_{t'}})$.
As always in the theory of covers, the difficulty is in defining the groups $\J_{\M_t}$. In fact, many covers were already constructed in [@SS4 §4] and we must show that we can put ourselves in the situation of *loc. cit.*. For this, we need to use the notion of a *common approximation* of simple characters from [@BKsemi], which is essentially a reinterpretation of the notion of endo-class.
We end the introduction with a summary of the contents of each section. Section \[S.notation\] consists of basic definitions, as well as recalling a very useful technique from [@BSS] for reducing proofs to easier situations. Section \[S.pairs\] concerns simple strata and pairs, while section \[S.characters\] concerns simple characters; these are the technical heart of the paper, in particular the translation principle Theorem \[thm:translation\] which is needed to cope with the fact that a simple character may be defined relative to several inequivalent simple strata. Along the way, we prove a generalization of a conjecture in [@BSS] on the embedding type of a simple character. Section \[S.common\] concerns the relationship between endo-classes and common approximations. Section \[S.simple\] recalls basic results about simple types but in the more general situation of lattice sequences which is needed later, and we prove the uniqueness results in section \[S.IIC\]. Finally, the general construction of a cover is given in section \[S.Semisimple\].
Much of the material here is necessarily technical. A reader who is already familiar with the situation (and common notations) for the case $\D=\F$ and is interested only in seeing the main results could probably get by reading only sections \[S.IIC\] and \[S.Semisimple\].
Notation and preliminaries {#S.notation}
==========================
Let $\F$ be a nonarchimedean locally compact field. For $\K$ a finite extension of $\F$, or more generally a division algebra over a finite extension of $\F$, we denote by $\Oo_{\K}$ its ring of integers, by $\p_{\K}$ the maximal ideal of $\Oo_{\K}$ and by $k_\K$ its residue field.
For $u$ a real number, we denote by $\lceil{u}\rceil$ the smallest integer which is greater than or equal to $u$, and by $\lfloor{u}\rfloor$ the greatest integer which is smaller than or equal to $u$, that is, its integer part.
All representations considered are smooth and complex.
Let $\A$ be a simple central $\F$-algebra, and let $\V$ be a simple left $\A$-module. The algebra $\End_{\A}(\V)$ is an $\F$-division algebra, the opposite of which we denote by $\D$. Considering $\V$ as a right $\D$-vector space, we have a canonical isomorphism of $\F$-algebras between $\A$ and $\End_{\D}(\V)$.
An *$\Oo_{\D}$-lattice sequence* on $\V$ is a map $$\La:\ZZ \to \{\hbox{$\Oo_\D$-lattices of $\V$}\}$$ which is decreasing (that is, $\La(k)\supseteq\La(k+1)$ for all $k\in\ZZ$) and such that there exists a positive integer $e=e(\La|\Oo_\D)$ satisfying $\La(k+e)=\La(k)\p_\D$, for all $k\in\ZZ$. This integer is called the *$\Oo_\D$-period* of $\La$ over $\Oo_\D$.
If $\La(k)\supsetneq\La(k+1)$ for all $k\in\ZZ$, then the lattice sequence $\La$ is said to be *strict*.
Associated with an $\Oo_{\D}$-lattice sequence $\La$ on $\V$, we have an $\Oo_{\F}$-lattice sequence on $\A$ defined by: $$\PP_{k}(\La)=\{a\in\A\ |\ a\La_{i}\subseteq\La_{i+k},\ i\in\ZZ\},
\quad k\in\ZZ.$$ The lattice $\AA(\La)=\PP_0(\La)$ is a hereditary $\Oo_\F$-order in $\A$, and $\mathfrak{P}(\La)=\PP_1(\La)$ is its Jacobson radical. They depend only on the set $\{\La(k)\mid k\in\ZZ\}$.
We denote by $\KK(\La)$ the *$\mult\A$-normalizer* of $\La$, that is the subgroup of $\mult\A$ made of all elements $g\in\mult\A$ for which there is an integer $n\in\ZZ$ such that $g(\La(k))=\La(k+n)$ for all $k\in\ZZ$. Given $g\in\KK(\La)$, such an integer is unique: it is denoted $\v_{\La}(g)$ and called the *$\La$-valuation* of $g$. This defines a group homomorphism $\v_{\La}$ from $\KK(\La)$ to $\ZZ$. Its kernel, denoted $\U(\La)$, is the group of invertible elements of $\AA(\La)$. We set $\U_0(\La)=\U(\La)$ and, for $k\>1$, we set $\U_k(\La)=1+\PP_k(\La)$.
Let $\E$ be a finite extension of $\F$ contained in $\A$. We denote by $e(\E/\F)$ and $f(\E/\F)$ the ramification index and residue class degree respectively.
An $\Oo_\D$-lattice sequence $\La$ on $\V$ is said to be *$\E$-pure* if it is normalized by $\E^{\times}$. The centralizer of $\E$ in $\A$, denoted $\B$, is a simple central $\E$-algebra. We fix a simple left $\B$-module $\V_\E$ and write $\D_\E$ for the algebra opposite to $\End_{\B}(\V_\E)$. By [@SS4 Théorème 1.4] (see also [@Br1 Theorem 1.3]), given an $\E$-pure $\Oo_{\D}$-lattice sequence on $\V$, there is an $\Oo_{\D_\E}$-lattice sequence $\Ga$ on $\V_\E$ such that: $$\PP_{k}(\La)\cap\B=\PP_{k}(\Ga), \quad k\in\ZZ.$$ It is unique up to translation of indices, and its $\B^\times$-normalizer is $\KK(\La)\cap\B^{\times}$.
A *stratum* in $\A$ is a quadruple $[\La,n,m,\b]$ made of an $\Oo_\D$-lattice sequence $\La$ on $\V$, two integers $m,n$ such that $0\<m\<n-1$ and an element $\b\in\PP_{-n}(\La)$.
For $i=1,2$, let $[\La,n,m,\b_i]$ be a stratum in $\A$. We say these two strata are *equivalent* if $\b_2-\b_1\in\PP_{-m}(\La)$.
Given a stratum $[\La,n,m,\b]$ in $\A$, we denote by $\E$ the $\F$-algebra generated by $\b$. This stratum is said to be *pure* if $\E$ is a field, if $\La$ is $\E$-pure and if $\v_{\La}(\b)=-n$. Given a pure stratum $[\La,n,m,\b]$, we denote by $\B$ the centralizer of $\E$ in $\A$. For $k\in\ZZ$, we set: $$\mathfrak{n}_k(\b,\La)=\{x\in\AA(\La)\ |\ \b x-x\b\in\PP_k(\La)\}.$$ The smallest integer $k\>\v_{\La}(\b)$ such that $\mathfrak{n}_{k+1}(\b,\La)$ is contained in $\AA(\La)\cap\B+\PP(\La)$ is called the *critical exponent* of the stratum $[\La,n,m,\b]$, denoted $k_0(\b,\La)$.
The stratum $[\La,n,m,\b]$ is said to be *simple* if it is pure and if we have $m\<-k_0(\b,\La)-1$.
Given $n\> 0$ and $\La$ an $\Oo_\D$-lattice sequence, there is another stratum which plays a very similar role to simple strata, namely the *null stratum* $[\La,n,n-1,0]$. In particular, a simple stratum $[\La,n,n-1,\b]$ is equivalent to a null stratum if and only if $\b\in\PP_{1-n}(\La)$.
Let $[\La,n,m,\b]$ be a simple stratum in $\A$. The *affine class* of $\La$ is the set of all $\Oo_{\D}$-lattice sequences on $\V$ of the form: $$a\La+b:k\mapsto\La(\lceil(k-b)/a\rceil),$$ with $a,b\in\ZZ$ and $a\>1$. The period of $a\La+b$ is $a$ times the period $e(\La|\Oo_\D)$ of $\La$. The *affine class* of the stratum $[\La,n,m,\b]$ is the set of all (simple) strata of the form $$[\La',n',m',\b],$$ where $\La'=a\La+b$ is in the affine class of $\La$, $n'=an$ and $m'$ is any integer such that $\lfloor m'/a\rfloor=m$.
In the course of the paper, there will be several objects associated to a simple stratum $[\La,n,m,\b]$, in particular simple characters (see §\[S.characters\]). By a straightforward induction (see [@BSS Lemma 2.2]), these objects depend only on the affine class of the stratum.
This article makes use of a number of results of Grabitz [@G] which are based on the following definition.
A simple stratum $[\La,n,m,\b]$ in $\A$ is *sound* if $\La$ is strict, $\AA\cap\B$ is principal and $\KK(\AA)\cap\mult\B=\KK(\AA\cap\B)$, where $\AA=\PP_0(\La)$ is the hereditary $\Oo_\F$-order defined by $\La$.
The condition on $\AA\cap\B$ forces $\AA$ to be a principal $\Oo_\F$-order. In the split case, a simple stratum $[\La,n,m,\b]$ is sound if and only if $\La$ is strict and $\AA$ is principal.
{#SS:passtoddag}
In [@BSS §2.7], we developed a process to reduce many proofs to the case of sound strata, which we recall briefly here: Let $[\La,n,m,\b]$ be a simple stratum in $\A$ and let $e$ denote the period of $\La$ over $\Oo_{\D}$. Write $\B$ for the centralizer of the field $\E=\F(\b)$ in $\A$, fix a simple left $\B$-module $\V_{\E}$ and write $\D_{\E}$ for the $\E$-algebra opposite to the algebra of $\B$-endomorphisms of $\V_{\E}$. Let $\Ga$ denote an $\Oo_{\D_{\E}}$-lattice sequence on $\V_{\E}$ such that $\PP_{k}(\La)\cap\B=\PP_{k}(\Ga)$ for $k\in\ZZ$, and let $e'$ denote its period over $\Oo_{\D_{\E}}$. We fix an integer $l$ which is a multiple of $e$ and $e'$ and set: $$\La^{\ddag}:k\mapsto\La(k)\oplus\La(k+1)\oplus\cdots\oplus\La(k+l-1),$$ which is a strict $\Oo_{\D}$-lattice sequence on $\V^{\ddag}=\V\oplus\dots\oplus\V$ ($l$ times). Thus we can form the simple stratum $[\La^{\ddag},n,m,\b]$ in $\A^{\ddag}=\End_{\D}(\V^{\ddag})$, where $\b$ is the block diagonal element $\diag(\b,\ldots,\b)\in
\A^l\subseteq\A^\ddag$. By [@BSS Lemma 2.17], the stratum $[\La^{\ddag},n,m,\b]$ is sound.
As we have mentioned, there will be several objects associated to a simple stratum $[\La,n,m,\b]$ through the course of the paper. If one identifies $\A$ with the “$(1,1)$-block” of $\A^\ddag$ and intersect (or restrict) these objects for $[\La^{\ddag},n,m,\b]$ to $\A$ one recovers the corresponding objects for $[\La,n,m,\b]$ (see, for example, [@SS4 Théorème 2.17]). Using this, in several proofs we write something like: “by passing to $\La^\ddag$ we may assume we are in the sound case” (Lemma \[lem:cruxstrata\], Proposition \[prop:derived\], Lemma \[lem:crux\])). By this we mean that we may prove the result for the sound stratum $[\La^{\ddag},n,m,\b]$ and then deduce the result for $[\La,n,m,\b]$ by intersecting with $\A$. In general, it is safe to do this provided there is no conjugation involved in the statement. An example of this is given already in [@BSS Theorem 4.16].
Simple strata and simple pairs {#S.pairs}
==============================
{#SS.embed}
We begin by recalling the definition of a *type* of embedding, from [@BG; @BSS].
We fix a simple central $\F$-algebra $\A$ and a simple left $\A$-module $\V$, and denote by $\D$ the opposite algebra of $\End_\A(\V)$. An *embedding* in $\A$ is a pair $(\E,\La)$ consisting of a finite field extension $\E$ of $\F$ contained in $\A$ and an $\E$-pure $\Oo_\D$-lattice sequence $\La$ on $\V$. Given such a pair, we denote by $\E^\flit$ the maximal finite unramified extension of $\F$ which is contained in $\E$ and whose degree divides the reduced degree of $\D$ over $\F$.
Two embeddings $(\E_i,\La_i)$ are *equivalent* if there is an element $g\in\G$ such that $\La_1$ is in the translation class of $g\La_2$ and $\E_1^\flit=g\E_2^\flit g^{-1}$. An equivalence class for this relation is called an *embedding type* in $\A$.
\[lem:embedtype\] Let $(\E,\La)$ be an embedding and put $e=e(\E/F)$, $f=f(\E/\F)$. Let $\E'$ be a finite field extension of $\F$ such that $e(\E'/F)=e$ and $f(\E'/F)=f$. Then there is an embedding $\ii:\E'\hookrightarrow\A$ such that $(\ii(\E'),\La)$ is an embedding with the same embedding type as $(\E,\La)$.
When $\La$ is strict, this is [@BG Corollary 3.16(ii)]. For the general case, we fix a simple right $\E\otimes_\F\D$-module $\rS$ and put $\A(\rS)=\End_\D(\rS)$. Let $\B$ be the commutant of $\E$ in $\A$, and let $\D_\E$ be the commutant of $\E$ in $\A(\rS)$. We also fix a decomposition $\V=\V^1\oplus\cdots\oplus\V^l$ into simple right $\E\otimes_\F\D$-modules (which are all copies of $\rS$) such that the lattice sequence $\La$ decomposes into the direct sum of the $\La^j=\La\cap\V^j$, for $j\in\{1,\ldots,l\}$. From [@VS3 §1.3], after choosing identifications $\V^j\simeq\rS$ for each $j$, we have an $\F$-algebra embedding $\ii:\A(\rS)\to\A$. Denote by $\SS$ the unique (up to translation) $\E$-pure strict $\Oo_\D$-lattice sequence on $\rS$. By the strict case, there is an embedding $\rho:\E'\hookrightarrow\A(\rS)$ such that $(\rho(\E'),\SS)$ is an embedding with the same embedding type as $(\E,\SS)$. (Indeed, by [@BSS Remark 2.12], any embedding $(\rho(\E'),\SS)$ has the same embedding type as $(\E,\SS)$.) By conjugating the embedding, we may assume $\rho(\E')^\flit=\E^\flit$. Then the embedding $\ii\circ\rho$ has the required property.
{#section-5}
We recall the definitions of *simple pair* and *endo-equivalence* from [@BSS Definitions 1.4, 1.7] (see also [@BH Definition 1.5]):
A *simple pair* over $\F$ is a pair $(k,\b)$ consisting of a non-zero element $\b$ of some finite extension of $\F$ and an integer $0\<k\<-k_{\F}(\b)-1$.
Let $\A$ be a simple central $\F$-algebra and $\V$ be a simple left $\A$-module. A *realization* of a simple pair $(k,\b)$ in $\A$ is a stratum in $\A$ of the form $[\La,n,m,\h(\b)]$ made of:
1. a homomorphism $\h$ of $\F$-algebra from $\F(\b)$ to $\A$;
2. an $\Oo_{\D}$-lattice sequence $\La$ on $\V$ normalized by the image of $\F(\b)^{\times}$ under $\h$;
3. an integer $m$ such that $\left\lfloor m/e_{\h(\b)}(\La)\right\rfloor=k$.
The integer $-n$ is then the $\La$-valuation of $\h(\b)$. By [@VS1 Proposition 2.25] we have: $$k_0(\h(\b),\La)=e_{\h(\b)}(\La)k_{\F}(\b),$$ which implies that any realization of a simple pair is a simple stratum.
1. For $i=1,2$, let $(k_{i},\b_{i})$ be a simple pair over $\F$. We say that these pairs are *endo-equivalent*, denoted: $$(k_{1},\b_1)\thickapprox(k_{2},\b_2),$$ if $k_1=k_2$ and $[\F(\b_1):\F]=[\F(\b_2):\F]$, and if there exists a simple central $\F$-algebra $\A$ together with realizations $[\La,n_i,m_i,\h_i(\b_i)]$ of $(k_i,\b_i)$ in $\A$, with $i=1,2$, which intertwine in $\A$.
This defines an equivalence relation on simple pairs, from the following Proposition:
\[prop:IICsimplepairs\] For $i=1,2$, let $(k,\b_{i})$ be a simple pair over $\F$, and suppose these pairs are endo-equivalent. Let $\A$ be a simple central $\F$-algebra and let $[\La,n_i,m_i,\h_i(\b_i)]$ be a realization of $(k,\b_i)$ in $\A$, for $i=1,2$. These strata then intertwine in $\A$.
Moreover, if $n_1=n_2$, $m_1=m_2$, and $(F[\h_i(\b_i)],\La)$ have the same embedding type then these strata are conjugate in $\KK(\La)$.
{#section-6}
Let $[\La,n,m,b]$ be a stratum in a simple central $\F$-algebra $\A=\End_\D(\V)$ and let $[\widetilde\La,n,m,b]$ be the induced stratum in the split central simple $\F$-algebra $\widetilde\A=\End_\F(\V)$, where $\widetilde\La$ denotes the $\Oo_\F$-lattice sequence defined by $\La$. By [@VS1 Théorème 2.23], this latter stratum is simple if and only the first one is, and in this case they are realizations of the same simple pair over $\F$.
We fix a uniformizer $\w_\F$ of $\F$ and set $$\y=\y(b,\La)=\varpi_\F^{n/g}b^{e/g},$$ where $e=e(\La|\Oo_\F)$ and $g=\gcd(n,e)$. Set $$\overline \y = \overline \y(b,\La) = \y+\PP_1(\widetilde\La),$$ considered as an element of $\PP_0(\widetilde\La)/\PP_1(\widetilde\La)$.
With notation as above:
1. the characteristic polynomial of $\overline \y$ (in $k_\F[X]$) is called the *characteristic polynomial of the stratum $[\La,n,m,b]$*;
2. the minimum polynomial of $\overline \y$ (in $k_\F[X]$) is called the *minimum polynomial of the stratum $[\La,n,m,b]$*.
\[rem:charminpol\]
1. Since $\overline \y$ depends only on the equivalence class of the stratum $[\La,n,n-1,b]$ (and the choice of the uniformizer $\w_\F$), the same is true of the minimum and characteristic polynomials.
2. If $b$ normalizes $\La$ then, by [@Br Lemma 2.1.9], the element $\y$ depends only on the strict lattice sequence whose image is the image of $\La$; hence the same is true of the minimum and characteristic polynomials.
3. The characteristic polynomial of a stratum may also be computed as the reduction modulo $\p_\F$ of the characteristic polynomial of $\y$ in $\widetilde\A$; of course, the same is *not* true for the minimum polynomial.
\[prop:simpleminpol\] Let $[\La,n,n-1,b]$ be a stratum in $\A$. Then $[\La,n,n-1,b]$ is equivalent to a simple stratum if and only if its minimum polynomial is irreducible and not $X$. Moreover, if $[\La,n,n-1,\b]$ is a simple stratum equivalent to $[\La,n,n-1,b]$ then, writing $\E=\F[\b]$, we have $$\Oo_\E+\PP_1(\La) = \Oo_\F[\y(b,\La)]+\PP_1(\La).$$
Note first that both conditions imply that the $b$ normalizes $\La$: if the minimum polynomial of $[\La,n,n-1,b]$ is irreducible and not $X$ then $\overline\y(b,\La)$ is invertible so $\y(b,\La)$ normalizes $\La$, whence so does $b$. Hence, using Remark \[rem:charminpol\](ii), we may (and do) assume in the proof that $\La$ is strict. Also, the final assertion is clear since, by the minimality of $\b$, the element $\y(\b,\La)+\p_\E$ generates the extension $k_\E/k_\F$, and $\y(\b,\La)+\PP_1(\La)=\y(b,\La)+\PP_1(\La)$.
Suppose $[\La,n,n-1,b]$ is equivalent to a simple stratum $[\La,n,n-1,\b]$ and put $\E=\F[\b]$. Then $[\widetilde\La,n,n-1,\b]$ is also simple; in particular, $\overline
\y(b,\La)=\overline \y(\b,\La)$ is a non-zero element of $k_\E$ in $\PP_0(\widetilde\La)/\PP_1(\widetilde\La)$ so has irreducible minimum polynomial not $X$.
For the converse, suppose $[\La,n,n-1,b]$ has irreducible minimum polynomial $\overline f(X)\in k_\F[X]$ and put $\d=\deg(\overline
f(X))$. Since $[\La,n,n-1,b]$ has characteristic polynomial which is a power of $\overline f(X)$, it is non-split fundamental. Now the proof follows that of [@Br Theorem 3.2.1] and we only sketch the difference so this proof should be read alongside *loc. cit.* – in particular, we will refer to notations used in the proofs there.
Following the ideas of [@Br §3], we treat first the simpler case when $\overline f(X)$ is also irreducible as an element of $k_\D[X]$ – that is, when $d$ is coprime to $\d$. We fix $\L/\F$ a maximal unramified subfield of $\D$, so that $k_\L=k_\D$.
Let $f(X)\in\Oo_\F[X]$ be any monic polynomial which reduces modulo $\p_\F$ to give $\overline f(X)$. We choose a matrix $\overline\gamma\in\M_\d(k_\F)$ with minimum polynomial $\overline f(X)$. In [@Br Definition 3.2.4], Broussous defines the notion of *$\overline\g$-standard form*, which we will use here. Since $[\La,n,n-1,b]$ has characteristic polynomial which is a power of $\overline f(X)$, it is non-split fundamental and, by [@Br Proposition 3.2.5] there is a $u\in\U(\La)$ such that $[\La,n,n-1,ubu^{-1}]$ is equivalent to a stratum in $\overline\g$-standard form. Since the property of being equivalent to a simple stratum is unchanged by conjugation in $\U(\La)$, we may as well assume $[\La,n,n-1,b]$ is itself in $\overline\g$-standard form.
Now we follow the proof of [@Br Theorem 3.2.1] in this case. In *op. cit.* p.221, an element $\b$ is defined and the proof of *op. cit.* Proposition 3.2.8 shows that there is $u\in\U(\La)$ such that $[\La,n,n-1,\b]$ is equivalent to $[\La,n,n-1,ubu^{-1}]$. (More precisely, [@Br Proposition 3.2.9] is applied to the $\Oo_\L$-order $\M_{n_i}(\Oo_\L)$, in the notation there.) Moreover, $\b$ is minimal over $\F$ by [@Br Lemma 3.2.10] so $[\La,n,n-1,b]$ is equivalent to the simple stratum $[\La,n,n-1,u^{-1}\b u]$.
Finally suppose we are in the general case where $\overline f(X)$ is not irreducible in $k_\L[X]$ and we decompose $\overline f(X)=\overline p_0(X)\cdots\overline p_{s-1}(X)$ into irreducibles. Now we follow [@Br §3.3] to reduce to previous case. The proof is essentially identical (but easier) so we will not repeat it – the only point is that, in [@Br Proposition 3.3.2], the orders $\mathfrak A_0$ and $\mathfrak B_0$ can be taken to be equal, by the case where $\overline f(X)$ is irreducible, and then the lattice sequences denoted $\Mm^1$ and $\Ll$ are equal, which implies that $\Mm$ is the lattice sequence here denoted $\La$ and, in the notation of [@Br Lemma 3.3.10], we have $\mathfrak A=\mathfrak A'$.
For $j=1,\ldots,r$, let $[\La^j,n,n-1,\b_j]$ be a simple stratum in $\A^j=\End_\D(\V^j)$ with $e(\La^j|\Oo_\F)=e$. Put $\V=\V^1\oplus\cdots\oplus\V^r$, and set $\La=\La^1\oplus\cdots\oplus\La^r$, a lattice sequence in $\V$ of $\Oo_\F$-period $e$. Write $\A=\End_\D(\V)$ and denote by $\e^j$ the idempotents in $\PP_0(\La)$ corresponding to the decomposition of $\V$. We put $\b=\sum_{j=1}^r \b_j$. Then $[\La,n,n-1,\b]$ is a stratum in $\A$.
\[cor:minpol\] With notation as above, suppose the strata $[\La^j,n,n-1,\b_j]$ are all equivalent to simple (or null) strata. Then they have the same minimum polynomial if and only if $[\La,n,n-1,\b]$ is equivalent to a simple (or null) stratum.
Writing $\A=\bigoplus_{i,j}\End_\D(\V^j,\V^i)$, it is clear that $\y(\b,\La)$ is block diagonal of the form $\diag(\y(\b_1,\La^1),\ldots,\y(\b_r,\La^r))$. In particular, the minimum polynomial of $\overline \y(\b,\La)$ is the $\gcd$ of the minimum polynomials of $\overline \y(\b_j,\La^j)$. The result is now immediate from Proposition \[prop:simpleminpol\], with the case of null strata coming from the case where the minimum polynomial is $X$.
\[cor:minpolpair\] Let $(k,\b)$ be a simple pair over $\F$ and let $[\La^1,n_1,m_1,\varphi_1(\b)]$ be a realization in some simple central $\F$-algebra $\A^1$. The minimum polynomial of $[\La^1,n_1,m_1,\varphi_1(\b)]$ depends only on the endo-equivalence class of the pair $(-k_\F(\b)-1,\b)$.
First note that Corollary \[cor:minpol\] essentially says that the minimum polynomial is independent of the realization. For suppose $[\La^2,n_2,m_2,\varphi_2(\b)]$ is another realization in a simple central $\F$-algebra $\A^2$. Since the minimum polynomial depends only on the induced strata in $\widetilde\A^1$ and $\widetilde\A^2$, we may as well suppose that both algebras are split – that is, $\A^j=\End_\F(\V^j)$, for some $\F$-vector space $\V^j$, $j=1,2$. Moreover, by scaling we may assume that $\La^1$ and $\La^2$ have the same period so that $n_1=n_2=n$. Put $m=\max\{m_1,m_2\}$.
Now set $\V=\V^1\oplus\V^2$ and use the notation of Corollary \[cor:minpol\]; also let $\varphi=\varphi_1+\varphi_2$ denote the diagonal embedding of $\F[\b]$ in $\A$. The stratum $[\La,n,m,\varphi(\b)]$ is then a realization of the simple pair $(k,\b)$ so it is simple and $[\La,n,n-1,\varphi(\b)]$, being pure, is equivalent to a simple stratum. Hence, by Corollary \[cor:minpol\], the strata $[\La^j,n,n-1,\varphi_j(\b)]$ have the same minimum polynomial.
Finally, suppose $(k,\g)$ is a simple pair endo-equivalent to $(k,\b)$. Let $[\La,n,m,\varphi(\b)]$ and $[\La,n,m,\rho(\g)]$ be realizations in some split simple central $\F$-algebra $\A=\End_\F(\V)$. By Proposition \[prop:IICsimplepairs\], these strata are conjugate by some $u\in\KK(\La)$ so, by conjugating the embedding $\rho$, we may assume that $[\La,n,n-1,\varphi(\b)]$ is equivalent to $[\La,n,n-1,\rho(\g)]$. In particular, they have the same minimum polynomial.
{#section-7}
The following is a generalization of [@BK Lemma 2.4.11], [@G Lemma 1.9]:
\[lem:scalar\] Let $[\La,n,n-1,b]$ be a stratum in $\A$. It is intertwined by every element of $\G$ if and only if $(b+\PP_{1-n}(\La))\cap\F\ne\emptyset$.
The proof follows the same scheme as that of [@BK Lemma 2.4.11] (see *op. cit.* pp.77–78) so we only sketch the argument. Suppose $[\La,n,n-1,b]$ is intertwined by every element of $\G$. If $b\in\PP_{1-n}(\La)$ then $0\in(b+\PP_{1-n}(\La))\cap\F$ so we assume $b\not\in\PP_{1-n}(\La)$. Then $b$ defines a non-zero map $\overline\b$ in $$\PP_{-n}(\La)/\PP_{1-n}(\La) = \bigoplus_{i=0}^{e-1}\Hom_{k_\D}(\La(i)/\La(i+1),\La(i-n)/\La(i-n+1)),$$ where $e=e(\La|\Oo_\D)$ is the $\Oo_\D$-period of $\La$. (Note that there is, in general, redundancy in this sum: the spaces $\Hom_{k_\D}(\La(i)/\La(i+1),\La(i-n)/\La(i-n+1))$ may be $0$.)
Since $\overline\b$ is non-zero, by moving $\La$ in its translation class we can suppose that it defines a non-zero map in $\Hom_{k_\D}(\La(0)/\La(1),\La(-n)/\La(1-n))$. If $e\nmid n$ then we can find $g\in\G\cap\PP_0(\La)$ such that $g$ induces the identity map on $\La(0)/\La(1)$ but the zero map on $\La(i)/\La(i+1)$, for $1\le
i\le e-1$. But then $\overline{gb}$ is zero on $\La(0)/\La(1)$, while $\overline{bg}$ coincides with $\overline b$ on $\La(0)/\La(1)$, so is non-zero, which contradicts the assumption that $g$ intertwines $[\La,n,n-1,b]$.
Thus $e$ divides $n$ and we put $t=-n/e$. Fix $\L/\F$ a maximal unramified subfield of $\D$, and $\w_\D$ a uniformizer of $\D$ which normalizes $\L$ (so acts via conjugation as a generator of $\Gal(\L/\F)$) and such that $\w_\D^d=\W_\F$. Then the coset $b\w_\D^{-t}+\PP_1(\La)$ is intertwined by every $g\in\G$ which commutes with $\w_\D$. In particular, since elementary matrices (with respect to a suitable basis) commute with $\w_\D$, we find that $(b\w_\D^{-t}+\PP_1)\cap\Oo_\D\ne\emptyset$ and, since conjugation by $\w_\D$ acts by Frobenius on $k_\D$, the fact that $\w_\D$ intertwines implies that $(b\w_\D^{-t}+\PP_1)\cap\Oo_\F\ne\emptyset$. Thus $b\equiv
\w_\D^t\l\pmod{\PP_{1-n}(\La)}$, for some $\l\in\Oo_\F^\times$. Finally, the fact that every element of $\Oo_\L^\times$ intertwines the stratum implies that conjugation by $\w_\D^t$ acts trivially on $k_\D$, so $d$ divides $t$ and $\w_\D^t\l\in\F$, as required.
The converse is trivial.
{#section-8}
Let $\A$ be a simple central $\F$-algebra and $\V$ be a simple left $\A$-module. Let $[\La,n,m,\b]$ be a simple stratum, set $\E=\E_\b=\F[\b]$, and let $\B=\B_\b$ denote the $\A$-centralizer of $\E$. We identify $\widetilde\A=\End_\F(\V)$ with $\A\otimes_\F\End_\A(\V)$. From [@SS4 Définition 2.25] (see also [@Br §4.2]) a *tame corestriction relative to $\E/\F$* on $\A$ is a $(\B,\B)$-bimodule homomorphism $s=s_\b:\A\to\B$ such that $\widetilde s=s\otimes\id_{\End_\A(\V)}$ is a tame corestriction relative to $\E/\F$ on $\widetilde\A$, in the sense of [@BK Definition 1.3.3].
\[lem:cruxstrata\] Let $[\La,n,m,\b_i]$ be equivalent simple strata. Then, putting $\E_i=\F[\b_i]$, we have:
1. $\Oo_{\E_1}+\PP_1(\La)=\Oo_{\E_2}+\PP_1(\La)$;
2. there are tame corestrictions $s_i$ on $\A$ relative to $\E_i/\F$ such that, for all $k\in\ZZ$ and $x\in\PP_k(\La)$, $$s_1(x) \equiv s_2(x) \pmod{\PP_{k+1}(\La)};$$
3. there are prime elements $\varpi_i$ of $\E_i$ such that $\varpi_1\U^1(\La)=\varpi_2\U^1(\La)$;
4. the pairs $(\E_i,\La)$ have the same embedding type.
We begin by proving (i)-(iii), for which we may assume that $\La$ is strict by passing first to $\La^\ddag$ (cf. paragraph \[SS:passtoddag\]). To return to $\Lambda$, notice that the condition in (iii) is equivalent to $\w_1\equiv\w_2\pmod{\PP_{e+1}(\La)}$, with $e=e(\La|\Oo_{\E_i})$, so for (i) and (iii) we can simply intersect with $\A$. The same is true for (ii) since tame corestrictions also decompose by blocks, by [@SS4 Proposition 2.26].
The simple strata $[\widetilde\La,n,m,\b_i]$ in $\widetilde\A$ are equivalent so, by [@BK Lemma 2.4.12], we have the results corresponding to (i) and (ii) in $\widetilde\A$, while (iii) is a by-product of the proof (see also [@BKcc Lemma 5.2] and its proof). Intersecting with $\A$ gives the result here.
In the case of strict sequences, (iv) is given by [@BG Lemma 5.2]. Moreover, writing $\Ll$ for the strict lattice sequence with the same image as $\La$, since $\PP_0(\La)=\PP_0(\Ll)$ and $\PP_1(\La)=\PP_1(\Ll)$, the same proof (using *op. cit.* Lemma 2.3.6) works in the general case to show that the maximal unramified subextensions of $\E_i$ are conjugate in $\U^1(\Ll)=\U^1(\La)$.
{#section-9}
Now let $\V_\E$ be a simple left $\B$-module, let $\D_\E$ be the algebra opposite to $\End_\B(\V_\E)$, and let $\Ga=\Ga_\b$ be the unique (up to translation) $\Oo_{\D_\E}$-lattice sequence $\V_\E$ such that $\PP_k(\La)\cap\B = \PB_k(\Ga)$, for all $k\in\ZZ$.
A *derived stratum* of $[\La,n,m,\b]$ is a stratum of the form $[\Ga,m,m-1,s(b)]$, for some $b\in\PP_{-m}(\La)$ and some tame corestriction $s$ on $\A$ relative to $\E/\F$.
The following result is a slight strengthening of [@SS4 Proposition 3.30]:
\[prop:refinement\] Let $[\La,n,m,\b]$ be a simple stratum and let $b\in\PP_{-m}(\La)$ be such that the derived stratum $[\Ga,m,m-1,s(b)]$ is equivalent to a simple (or null) stratum $[\Ga,m,m-1,c]$. Then there is a simple stratum $[\La,n,m-1,\b']$ equivalent to $[\La,n,m-1,\b+b]$ and, moreover, for any such stratum, writing $\E'=\F[\b']$ and $\E_1=\F[\b,c]$, we have:
1. $e(\E'/\F)=e(\E_1/\F)$, $f(\E'/\F)=f(\E_1/\F)$ and $k_0(\b',\La)=\begin{cases} k_0(\b,\La) &\hbox{ if }c\in\E, \\
-m &\hbox{ otherwise;} \end{cases}$
2. $
\Oo_{\E'}+\PP_1(\La) = \Oo_{\E_1}+\PP_1(\La) = \Oo_{\E}[\y(s(b))] + \PP_1(\La).
$
The first assertion is proved in [@SS4 Proposition 3.30], under the hypothesis that the derived stratum $[\Ga,m,m-1,s(b)]$ is itself simple. If it is only equivalent to the simple stratum $[\Ga,m,m-1,c]$ then $c-s(b)\in\PB_{1-m}(\Ga)$ so, by [@SS4 Proposition 2.29], there is $d\in\PP_{1-m}(\La)$ such that $s(d)=c-s(b)$. Replacing $b$ by $b+d$ we reduce to the case that the derived stratum is simple and the result here follows.
For (i), we may pass first to $\La^\ddag$, where the result follows from [@G Proposition 9.5].
For (ii), the second equality follows from Proposition \[prop:simpleminpol\], while the independence of $\Oo_{\E'}+\PP_1(\La)$ on the choice of $\beta'$ comes from Lemma \[lem:cruxstrata\]. In particular, we need only find a single $\b'$ for which the first equality holds.
We fix a simple right $\E_1\otimes_\F\D$-module $\rS$ and put $\A(\rS)=\End_\D(\rS)$. Let $\C$ be the commutant of $\E_1$ in $\A$, and let $\D_1$ be the commutant of $\E_1$ in $\A(\rS)$. We also fix a decomposition $\V=\V^1\oplus\cdots\oplus\V^l$ into simple right $\E_1\otimes_\F\D$-modules (which are all copies of $\rS$) such that the lattice sequence $\La$ decomposes into the direct sum of the $\La^j=\La\cap\V^j$, for $j\in\{1,\ldots,l\}$. From [@VS3 §1.3], after choosing identifications $\V^i\simeq \rS$, we have an $\F$-algebra embedding $\ii:\A(\rS)\to\A$ and an isomorphism of $(\A(\rS),\C)$-bimodules $$\A(\rS)\otimes_{\D_1}\C \to \A.$$ We denote by $\B(\rS)$ the commutant of $\E$ in $\A(\rS)$ and let $\rS_\E$ be a simple left $\B(\rS)$-module. By [@SS4 Lemme 3.31], the tame corestriction $s$ on $\A$ relative to $\E/\F$ takes the form $s_1\otimes\id_\C$, for $s_1$ a tame corestriction on $\A(\rS)$ relative to $\E/\F$.
Denote by $\SS$ the unique (up to translation) $\E_1$-pure strict $\Oo_\D$-lattice sequence on $\rS$, and by $\SS_\b$ the unique (up to translation) $\Oo_{\D_\E}$-lattice sequence on $\rS_\E$ compatible with the filtration from $\SS$. Set $n_0=-v_{\SS}(\b)$ and $m_0=-v_{\SS}(c)$ and pick $b_0\in\PP_{-m_0}(\SS)$ such that $s_1(b_0)=c$. By [@SS4 Lemme 3.32], the stratum $[\SS,n_0,m_0-1,\b+b_0]$ is pure, so equivalent to a simple stratum $[\SS,n_0,m_0-1,\b+b']$, with $s_1(b')\in c+\PB_{1-m_0}(\SS_\b)$. We put $\E'=\F[\b+b']$.
Suppose first that $c\in\E$. Then, by (i), we have $k_0(\b+b',\SS)=k_0(\b,\SS)$ so that $[\SS,n_0,m_0,\b]$ and $[\SS,n_0,m_0,\b+b']$ are equivalent simple strata. Now Lemma \[lem:cruxstrata\](i) implies $$\label{eqn:oEinAScase1}
\Oo_{\E'}+\PP_1(\SS) = \Oo_{\E_1}+\PP_1(\SS).$$
Now suppose $c\not\in\E$. Let $x\in\Oo_{\E'}$ and put $r=-k_0(\b,\SS)$, which is strictly greater than $m_0$. Then certainly $a_\b(x)\in\PP_{-m_0}(\SS)$ so, by [@SS4 Proposition 2.29], we can write $x=\g+y$, with $\g\in\PB_0(\SS)$ and $y\in\PP_{r-m_0}(\SS)\subseteq\PP_1(\SS)$. Thus $$0\ =\ (\b+b')(\g+y)-(\g+y)(\b+b')\ \equiv\ a_\b(y)+b'\g-\g b' \pmod{\PP_{1-m_0}(\SS)}.$$ Applying $s_1$ and using $s_1(b')\in c+\PB_{1-m_0}(\SS)$, we see that $a_{c}(\g)\in\PB_{1-m_0}(\SS_\b)$. Now $k_0(c,\SS_\b)=-m_0$ so we deduce that $\g\in\Oo_{\D_1}+\PB_1(\SS_\b)$, since $\D_1$ is the commutant of $\E_1=\E[c]$ in $\B(\rS)$. In particular, we see that $\Oo_{\E'}\subseteq\Oo_{\D_1}+\PP_1(\SS)$ so that the image of the residue field $k_{\E'}$ in $\PP_0(\SS)/\PP_1(\SS)$ is contained in the image of $k_{\D_1}$. Since (the images of) $k_{\E'}$ and $k_{\E_1}$ are then subfields of $k_{\D_1}$ of the same cardinality (by (i)), they must coincide and we deduce again that $$\label{eqn:oEinAS}
\Oo_{\E'}+\PP_1(\SS) = \Oo_{\E_1}+\PP_1(\SS).$$
Finally, we must translate this back to $\A$, using the embedding $\ii:\A(\rS)\to\A$; we will identify $\b$ and $c$ with their images under $\ii$. The image of the simple stratum $[\SS,n_0,m_0-1,\b+b']$ under $\ii$ is a simple stratum $[\La,n,m-1,\b+\ii(b')]$, and we have $$s(\ii(b')) = \ii(s_1(b'))\ \equiv c \equiv s(b) \pmod{\PB_{1-m}(\Ga)},$$ since $\ii(\PP_{1-m_0}(\SS))\subseteq\PP_{1-m}(\La)$. Thus, as in the end of the proof of [@SS4 Proposition 3.30], there exists $h\in\U_1(\La)$ such that, putting $\b'=h^{-1}(\b+\ii(b'))h$, we get a simple stratum $[\La,n,m-1,\b']$ equivalent to $[\La,n,m-1,\b+b]$. Then $\F[\b']=h^{-1}\ii(\E')h$ and $$\Oo_{\F[\b']} = h^{-1}\ii(\Oo_{\E'})h \equiv \ii(\Oo_{\E'}) \pmod{\PP_1(\La)}.$$ Finally, by , we have $\ii(\Oo_{\E'})\equiv \ii(\Oo_{\E_1}) \pmod{\PP_1(\La)}$ and, since we have identified $\E_1$ with $\ii(\E_1)$, we deduce $$\Oo_{\F[\b']} + \PP_1(\La) = \Oo_{\E_1} + \PP_1(\La).$$ This completes the proof of Proposition \[prop:refinement\].
{#section-10}
We also have a converse to Proposition \[prop:refinement\]:
\[prop:derived\] Let $[\La,n,m,\b]$ be a pure stratum equivalent to the simple stratum $[\La,n,m,\g_1]$. Let $s_1$ be a tame corestriction on $\A$ relative to $\F[\g_1]/\F$. Then the derived stratum $[\Ga_1,m,m-1,s_1(\b-\g_1)]$ is equivalent to a simple (or null) stratum.
We will need the following Lemma, which is in fact a special case of the Proposition.
\[lem:simplederived\] Let $[\La,n,m,\b_i]$ be equivalent simple strata and let $s_1$ be a tame corestriction on $\A$ relative to $\F[\b_1]/\F$. Then there exists $\d\in\F[\b_1]$ such that $s_1(\b_1-\b_2)\equiv
\d\pmod{\PP_{1-m}(\La)}$.
By passing to $\La^\ddag$, we may assume we are in the strict sound case. The proof is then identical to that of [@BK (2.4.10)], replacing [@BK Proposition 1.4.6] by [@Br Proposition 4.3.3] and [@BK Theorem 1.5.8] by [@Br Proposition 4.1.1].
By passing to an equivalent stratum, we may assume that the stratum $[\La,n,m-1,\b]$ is simple. If $[\La,n,m,\b]$ is also simple then the result follows from Lemma \[lem:simplederived\]; thus we may assume $k_0(\b,\La)=-m$. By passing to $\La^\ddag$, we may assume we are in the strict sound case. We write $\E_1=\F[\g_1]$ and $\B_1$ for the $\A$-centralizer of $\E_1$.
By [@G Propositions 3.8, 9.3], there exists a simple stratum $[\La,n,m,\g_2]$ equivalent to $[\La,n,m,\b]$ such that the derived stratum $[\Ga_2,m,m-1,s_2(\b-\g_2)]$ is equivalent to a simple (or null) stratum, where $s_2$ is a tame corestriction on $\A$ relative to $\F[\g_2]/\F$. Moreover, this derived stratum is non-scalar, by Proposition \[prop:refinement\], since $k_0(\b,\La)
> k_0(\g_2,\La)$, and thus, by Proposition \[prop:simpleminpol\], it has irreducible minimum polynomial.
We write $\E_2=\F[\g_2]$ and $\B_2$ for the $\A$-centralizer of $\E_2$. By Lemma \[lem:cruxstrata\], we may assume that the tame corestriction $s_2$ is chosen such that, for all $k\in\ZZ$ and $x\in\PP_k(\La)$, $$s_1(x) \equiv s_2(x) \pmod{\PP_{k+1}(\La)}.$$ We also use Lemma \[lem:cruxstrata\] to choose uniformizers $\w_i$ for $\E_i$ such that $\w_1\U^1(\La)=\w_2\U^1(\La)$. Again by Lemma \[lem:cruxstrata\], the residue fields $k_{\E_i}$ have a common image in $\PP_0(\La)/\PP_1(\La)$ so that we may identify them. Moreover, $\PB_0(\Ga_i)/\PB_1(\Ga_i)$ have a common image in $\PP_0(\La)/\PP_1(\La)$: by [@SS4 Proposition 2.29], the maps $s_i:\PP_k(\La)\to\PP_k(\Ga_i)$ are surjective, for all $k\in\ZZ$, so $\PB_0(\Ga_i)/\PB_1(\Ga_i)$ are the common image of $s_i$ in $\PP_0(\La)/\PP_1(\La)$.
Put $b_i=\b-\g_i$. By the choices of $\w_i$ and $s_i$, we have $$\y(s_1(b_2),\Ga_1) \equiv \y(s_2(b_2),\Ga_2) \pmod{\PP_1(\La)}.$$ In particular (given the identifications we have made), we see that the two strata $[\Ga_i,m,m-1,s_i(b_2)]$ have the same minimum and characteristic polynomials. In particular, $[\Ga_1,m,m-1,s_1(b_2)]$ has irreducible minimum polynomial so, by Proposition \[prop:simpleminpol\], it is equivalent to a simple stratum $[\Ga_1,m,m-1,c]$.
Finally, by Lemma \[lem:simplederived\], the stratum $[\Ga_1,m,m-1,s_1(\g_2-\g_1)]$ is equivalent to a scalar stratum $[\Ga_1,m,m-1,\d]$, whence $[\Ga_1,m,m-1,s_1(b_1)]$ is equivalent to the simple stratum $[\Ga_1,m,m-1,c+\d]$, as required.
Simple characters and refinement {#S.characters}
================================
{#section-11}
Let $\A$ be a central simple $\F$-algebra and let $[\La,n,0,\b]$ be a simple stratum in $\A$. To this simple stratum, in [@SS4 §2.4], one attaches compact open subgroups $\H(\b,\La)\subseteq\J(\b,\La)$ of $\mult\A$, together with filtrations $$\H^{m+1}(\b,\La)=\H(\b,\La)\cap\U^{m+1}(\La),\quad
\J^{m+1}(\b,\La)=\J(\b,\La)\cap\U^{m+1}(\La),\qquad m\ge 0,$$ and a finite set $\Cc(\La,0,\b)$ of characters of $\H^{1}(\b,\La)$, called *simple characters of level $0$*, depending on the choice of an additive character $$\psi_\F:\F\to\mathbb{C}^{\times}$$ which is trivial on $\p_{\F}$ but not on $\Oo_{\F}$, and which will now be fixed once and for all.
By restriction to $\H^{m+1}(\b,\La)$, we get also a set $\Cc(\La,m,\b)$ of *simple characters of level $m$*. If $\lfloor n/2\rfloor\<m$, then $\H^{m+1}(\b,\La)=\U^{m+1}(\La)$, and the set $\Cc(\La,m,\b)$ reduces to the single character $\psi_\b$ of $\U^{m+1}(\La)$ defined by $$\psi_\b:x\mapsto\psi_\F\circ\tr_{\A/\F}(\b(x-1)),$$ where $\tr_{\A/\F}$ denotes the reduced trace of $\A$ over $\F$, which depends only on the equivalence class of $[\La,n,m,\b]$. More generally, for any $m$, the subgroup $\H^{m+1}(\b,\La)$ and the set $\Cc(\La,m,\b)$ depend only on the equivalence class of $[\La,n,m,\b]$.
Note that we will use the following common convention: the trivial character of the group $\U^{t+1}(\La)$ will be considered as a simple character for the trivial stratum $[\La,t,t,0]$.
{#section-12}
Various properties of simple characters can be found in [@SS4; @BSS]. For now we recall two of them, the first of which is a special case of the intertwining formula [@SS4 Théorème 2.24]:
Let $\A$ be a central simple $\F$-algebra, let $[\La,n,0,\b]$ be a simple stratum in $\A$ and let $\t\in\Cc(\La,0,\b)$. Then, writing $\B$ for the $\A$-centralizer of $\b$ as usual, we have $$\I_\G(\t)=\J^1(\b,\La)\mult\B\J^1(\b,\La).$$
The following fundamental result is one of the main results of [@BSS]:
\[prop:IICwithK\] Let $\A$ be a simple central $\F$-algebra. For $i=1,2$, let $[\La,n,m,\b_i]$ be a simple stratum in $\A$, and let $\t_{i}\in\Cc(\La,m,\b_i)$ be a simple character. Write $\K_i$ for the maximal unramified extension of $\F$ contained in $\F(\b_i)$. Assume that $\t_1$ and $\t_2$ intertwine in $\mult\A$ and that the $(\F[\b_i],\La)$ have the same embedding type. Then there is an element $u\in\KK(\La)$ such that:
1. $\K_1=u\K_2u^{-1}$;
2. $\Cc(\La,m,\b_1)=\Cc(\La,m,u\b_2u^{-1})$;
3. $\t_2(x)=\t_1(uxu^{-1})$, for all $x\in\H^{m+1}(\b_2,\La)$.
{#section-13}
One of the technical difficulties with simple characters is that they do not determine the simple stratum used to define them: that is, we may have $\Cc(\La,m,\b_1)\cap\Cc(\La,m,\b_2)\ne\emptyset$, for inequivalent strata $[\La,n,m,\b_i]$ (though certain invariants of the strata are equal – see later). In order to cope with this, we need the following translation principle, which is the main result of this section.
\[thm:translation\] Let $[\La,n,m,\g_i]$ be simple strata with $\Cc(\La,m,\g_1)\cap\Cc(\La,m,\g_2)\ne\emptyset$. Let $[\La,n,m-1,\b_1]$ be a simple stratum such that $[\La,n,m,\b_1]$ is equivalent to $[\La,n,m,\g_1]$. Then there is a simple stratum $[\La,n,m-1,\b_2]$ such that $[\La,n,m,\b_2]$ is equivalent to $[\La,n,m,\g_2]$ and $\Cc(\La,m-1,\b_1)=\Cc(\La,m-1,\b_2)$.
The proof of this translation principle, which will take up most of the remainder of this section, begins with the following special case, in which $\b_1=\g_1$:
\[lem:translation\] Let $[\La,n,m,\g_i]$ be simple strata with $\Cc(\La,m,\g_1)\cap\Cc(\La,m,\g_2)\ne\emptyset$. Then $\H^m(\g_1,\La)=\H^m(\g_2,\La)$ and there is a simple stratum $[\La,n,m,\b_2]$ equivalent to $[\La,n,m,\g_2]$ such that $\Cc(\La,m-1,\b_2)=\Cc(\La,m-1,\g_1)$.
From [@BSS Lemma 4.12], we have already that $\H^m(\g_1,\La)=\H^m(\g_2,\La)$. The remainder of the proof is *mutatis mutandis* that of [@BK Theorem 3.5.9]: we replace [@BK Theorem 3.3.2] by [@SS4 Théorème 2.23], [@BK Lemma 2.4.11] by Lemma \[lem:scalar\], [@BK Theorem 2.2.8] by Proposition \[prop:refinement\], [@BK 3.3.20] by [@SS4 Proposition 2.15], and [@BK 3.5.8] by [@BSS Theorem 4.16]; for the proof of [@BK Lemma 3.5.10] we replace [@BK Corollary 3.3.17] by [@SS4 Proposition 2.24] and [@BK Proposition 3.3.9] by [@SS4 Lemme 2.30(2)].
{#section-14}
The technical crux of the translation principle is contained in the following lemma:
\[lem:crux\] Let $[\La,n,m-1,\b_i]$ be simple strata with $\Cc(\La,m-1,\b_1)\cap\Cc(\La,m-1,\b_2)\ne\emptyset$. Then, putting $\E_i=\F[\b_i]$, we have:
1. $\Oo_{\E_1}+\PP_1(\La)=\Oo_{\E_2}+\PP_1(\La)$;
2. the pairs $(\E_i,\La)$ have the same embedding type;
3. there are prime elements $\varpi_i$ of $\E_i$ such that $\varpi_1\U^1(\La)=\varpi_2\U^1(\La)$;
4. there are tame corestrictions $s_i$ on $\A$ relative to $\E_i/\F$ such that, for all $k\in\ZZ$ and $x\in\PP_k(\La)$, $$s_1(x) \equiv s_2(x) \pmod{\PP_{k+1}(\La)}.$$
Note that (ii) in this lemma answers Conjecture 4.17 of [@BSS] – indeed, it is a generalization of that conjecture since here we do not assume that the sequence $\La$ is strict. Also, the hypothesis $\Cc(\La,m-1,\b_1)\cap\Cc(\La,m-1,\b_2)\ne\emptyset$ is equivalent to $\Cc(\La,m-1,\b_1)=\Cc(\La,m-1,\b_2)$, by [@BSS Theorem 4.16].
In the split case when $\La$ is strict, this is [@BKcc Lemma 5.2], while the case of arbitrary $\La$ follows by passing to $\La^\ddag$.
We proceed by induction on $m$. When $m=n$ the result is immediate from Lemma \[lem:cruxstrata\]. Note that $k_0(\b_i,\La)$ is independent of $i$, by [@BSS Lemma 4.7]. If $k_0(\b_i,\La)<-m$ then again the result is clear from the induction hypothesis, since the conclusions (i)–(iv) are independent of $m$, so we assume $k_0(\b_i,\La)=-m>-n$. By replacing $[\La,n,m-1,\b_1]$ by an equivalent stratum, Lemma \[lem:translation\] says we may (and do) assume $\Cc(\La,0,\b_1)=\Cc(\La,0,\b_2)$ without affecting the conclusion of the lemma, by Lemma \[lem:cruxstrata\].
For $i=1,2$, let $[\La,n,m,\g_i]$ be a simple stratum equivalent to $[\La,n,m,\b_i]$. Then $$\Cc(\La,m,\g_1)=\Cc(\La,m,\b_1)=\Cc(\La,m,\b_2)=\Cc(\La,m,\g_2)$$ so, by induction applied to the simple strata $[\La,n,(m+1)-1,\g_i]$, (i)–(iv) are satisfied with $\E_{\g_i}=\F[\g_i]$ in place of $\E_i$, and we pick uniformizers $\w_{\g_i}$ and tame corestrictions $s_{\g_i}$ satisfying (iii), (iv). Moreover, by replacing $[\La,n,m,\g_1]$ by an equivalent stratum, Lemma \[lem:translation\] says we may assume $\Cc(\La,0,\g_1)=\Cc(\La,0,\g_2)$.
Put $c_i=\b_i-\g_i$ and let $\t\in\Cc(\La,m-1,\b_1)=\Cc(\La,m-1,\b_2)$. By [@SS4 Proposition 2.15], we have $\t=\vartheta_i\psi_{c_i}$, for some $\vartheta_i\in\Cc(\La,m-1,\g_i)$. Hence $$\vartheta_1 = \vartheta_2\psi_{c_2-c_1}$$ and $\vartheta_1,\vartheta_2\in\Cc(\La,m-1,\g_1)$ both restrict to the same character $\vartheta\in\Cc(\La,m,\g_1)$. Since $\vartheta_1,\vartheta_2$ are both intertwined by $\B_{\g_1}^\times$, the same is true of $\psi_{c_2-c_1}$. In particular, restricting to $\H^m(\g_1,\La)\cap\B_{\g_1}^\times=\U^m(\Ga_{\g_1})$, we see that $\psi_{s_{\g_1}(c_2-c_1)}|\U^m(\Ga_{\g_1})$ is intertwined by all of $\B_{\g_1}^\times$ and, by Lemma \[lem:scalar\], $(s_{\g_1}(c_2-c_1)+\PB_{1-m}(\Ga_{\g_1}))\cap
\E_{\g_1}\ne\emptyset$. In particular, the stratum $[\Ga_{\g_1},m,m-1,s_{\g_1}(c_2-c_1)]$ is equivalent to a simple (or null) scalar stratum.
By Proposition \[prop:refinement\], there is a simple stratum $[\La,n,m-1,\g_1']$ equivalent to the stratum $[\La,n,m-1,\g_1+(c_2-c_1)]$. Since $\vartheta_2\in\Cc(\La,m-1,\g_1)$, by [@SS4 Proposition 2.15] we have $\vartheta_1=\vartheta_2\psi_{c_2-c_1}=\vartheta_2\psi_{\g'_1-\g_1}\in\Cc(\La,m-1,\g'_1)$. Hence $\Cc(\La,m-1,\g'_1)=\Cc(\La,m-1,\g_1)$. Moreover, putting $c_1'=\b_1-\g_1'$ we see that $c_2-c'_1\in\PP_{1-m}(\La)$; in particular, for any tame corestriction $s_{\g'_1}$ on $\A$ relative to $\E_{\g_1'}/\F$, we have $s_{\g_1'}(c_2-c_1')\in\PP_{1-m}(\La)$.
Thus, replacing $\g_1$ by $\g_1'$ we may assume that $s_{\g_1}(c_2-c_1)\in\PP_{1-m}(\La)$. By (iv), we also have $s_{\g_2}(c_2-c_1)\in\PP_{1-m}(\La)$. In particular, looking at the derived strata $[\Ga_{\g_i},m,m-1,s_{\g_i}(c_j)]$, with $i,j\in\{1,2\}$, and using the inductive hypotheses (iii), (iv), we get $$\y(s_{\g_1}(c_1)) \equiv
\y(s_{\g_1}(c_2)) \equiv
\y(s_{\g_2}(c_2)) \pmod{\PP_1(\La)}.$$ (The elements $\y$ here are computed with respect to the uniformizers $\w_{\g_i}$ satisfying (iii).) By Proposition \[prop:derived\] the derived strata $[\Ga_{\g_i},m,m-1,s_{\g_i}(c_i)]$ are equivalent to simple or null strata so, by Proposition \[prop:refinement\] (applied to the strata $[\La,n,m,\g_i]$ and $\b'=\b_i$) and the inductive hypothesis (i), we have $$\Oo_{\E_1}+\PP_1(\La) =
\Oo_{\E_{\g_1}}[\y(s_{\g_1}(c_1))]+\PP_1(\La) =
\Oo_{\E_{\g_2}}[\y(s_{\g_2}(c_2))]+\PP_1(\La) =
\Oo_{\E_2}+\PP_1(\La)$$ and we have proved (i). Now (ii) follows exactly as in the proof of Lemma \[lem:cruxstrata\] (see also [@BG Lemma 5.2]); indeed the proof gives the existence of $u\in\U^1(\La)$ such that $u^{-1}\K_1 u=\K_2$, where $\K_i$ is the maximal unramified subextension of $\E_i/\F$.
For the remainder, we may pass to $\La^\ddag$ and assume we have soundly embedded strict lattice sequences with $\PP_0(\La)$ principal, as in the proof of Lemma \[lem:cruxstrata\]. (By [@BSS Proposition 4.11], we have $\Cc(\La^\ddag,m-1,\b_1)=\Cc(\La^\ddag,m-1,\b_2)$; cf. the proof of *op. cit.* Theorem 4.16.) Recall that we have $\t\in\Cc(\La,m-1,\b_1)=\Cc(\La,m-1,\b_2)$, which we extend to a common simple character in $\Cc(\La,0,\b_1)=\Cc(\La,0,\b_2)$. Then $\t^u\in\Cc(\La,0,u^{-1}\b_1
u)$ surely intertwines $\t\in\Cc(\La,0,\b_2)$ so, by Proposition \[prop:IICwithK\], there is $g\in\KK(\La)$ such that $\t^{ug}=\t$ and $(ug)^{-1}\K_1 (ug)=\K_2$. Since $ug$ then normalizes $\t$, we have $ug\in\KK(\Ga_{\b_2})\J^1(\b_b,\La)$.
In particular, there is $x\in\KK(\Ga_{\b_2})$ such that $h=ugx\in\J^1(\b_2,\La)$, $h^{-1}\K_1 h=\K_2$ and $\t^h=\t$. Thus, replacing $\b_1$ by $h^{-1}\b_1 h$, we may assume that $\K_1=\K_2=\K$, without affecting the conclusion of the Lemma (since $h\in
\U^1(\La)$).
Now we will utilise the *interior lifting* and *base change* processes of [@BSS] to reduce to the split case.
We suppose first that we are in the special case $\K=\F$, that is $\E_i/\F$ is totally ramified. Fix an unramified extension $\L/\F$ which splits $\A$, so that $\L_i=\E_i\otimes_\F\L$ is a field, for $i=1,2$. The algebra $\overline\A=\A\otimes_\F\L$ is then a split simple $\L$-central algebra and we choose a simple left $\overline\A$-module $\overline\V$. There is a unique (up to translation) strict $\Oo_\L$-lattice sequence $\overline\La$ on $\overline\A$ such that $\PP_k(\overline\La)=\PP_k(\La)\otimes_{\Oo_\F}\Oo_\L$, for all $k\in\ZZ$ (see [@VS1 §2.2]). Identifying $\A$ with $\A\otimes_\F
1 \subseteq \overline\A$, we get strata $[\overline\La,n,m-1,\b_i]$, which are simple.
Denote by $\Cc(\overline\La,m-1,\b_i)$ the set of simple characters with respect to the character $\psi_\F\circ\tr_{\L/\F}$. The base change process from [@BSS §7.2] gives rise to injective $\KK(\La)$-equivariant maps $$\bb^i = \bb^i_{\L/\F}:\Cc(\La,m-1,\b_i) \to \Cc(\overline\La,m-1,\b_i).$$ Moreover, by [@BSS Proposition 7.6], we have $\bb^1(\t)=\bb^2(\t)$ so $\Cc(\overline\La,m-1,\b_1)\cap\Cc(\overline\La,m-1,\b_2)\ne\emptyset$. In particular, by the split case we get uniformizers $\w^\L_i$ of $\L_i$ such that $$\label{eqn:wL}
\w^\L_1 + \PP_{e+1}(\overline\La) = \w^\L_2 + \PP_{e+1}(\overline\La),$$ with $e=e(\overline\La|\Oo_\L)=e(\La|\Oo_\F)$, and tame corestrictions $s_i^\L$ on $\overline\A$ relative to $\L_i/\L$ for the character $\psi_\F\circ\tr_{\L/\F}$ such that, for all $k\in\ZZ$ and $x\in\PP_k(\overline\La)$, $$s_1^\L(x) \equiv s_2^\L(x) \pmod{\PP_{k+1}(\overline\La)}.$$ Multiplying through by a unit, we see that we may assume $\w^\L_1=\w_1$ is a uniformizer of $\E_1$ and $\w^\L_2=\w_2\z$, for some uniformizer $\w_2$ of $\E_2$ and $\z\in\Oo_\L^\times$ a root of unity of order coprime to $p$. Thus we have $$\w_1\w_2^{-1}\ \equiv\ \z \pmod{\PP_1(\overline\La)}.$$ Now the Galois group $\Gal(\L/\F)$ acts on $\overline A$, fixing $\w_1\w_2^{-1}$, so the image of $\z$ in $k_\L$ is fixed by $\Gal(k_\L/k_\F)$. In particular, $\z\in\Oo_\F^\times$ so, replacing $\w_2$ by $\w_2\z$, we get $$\w_1 + \PP_{e+1}(\overline\La) = \w_2 + \PP_{e+1}(\overline\La),$$ and intersecting with $\A$ completes the proof of (iii).
The argument for the tame corestrictions is similar: We check that, if $s_i$ is an arbitrary tame corestriction on $\A$ relative to $\E_i/\F$, then $s_i\otimes 1$ is a tame corestriction on $\overline\A$ relative to $\L_i/\L$. By [@SS4 Proposition 2.26], $s_i^\L$ and $s_i\otimes 1$ differ by a unit $u_i$ in $\Oo_{\L_i}$ and, changing by a root of unity, we can assume $u_1\in 1+\p_{\L_1}$. We have $u_2\equiv\z\pmod{\p_{\L_i}}$, for some root of unity $\z\in\Oo_\L^\times$. Then, for all $k\in\ZZ$ and $a\in\PP_k(\La)$ $$s_1(a)\otimes 1\equiv s_1^\L(a\otimes 1) \equiv s_2^\L(a\otimes 1) \equiv s_2(a)\otimes \z \pmod{\PP_{k+1}(\overline\La)}.$$ Again, the Galois group $\Gal(\L/\F)$ acts on $\overline A$, fixing $s_1(a)\otimes 1$, so $$s_2(a)\otimes \z \equiv s_2(a)\otimes \z^\s
\pmod{\PP_{1}(\overline\La)}, \hbox{\qquad for all
}a\in\PP_0(\La),\ \s\in\Gal(\L/\F).$$ By [@SS4 Proposition 2.29], the map $s_2:\PP_0(\overline\La)\to\PP_0(\overline\Ga_2)$ is surjective so, as above, we deduce that $\z\in\Oo_\F^\times$ and, after replacing $s_2$ by $\z s_2$, we may assume $\z=1$. Finally, intersecting with $\A$ completes with proof of (iv).
Finally we consider the case where $\K\ne\F$. Denote by $\C$ the $\A$-centralizer of $\K$, fix a simple left $\C$-module $\W$, and let $\D_\K$ be the algebra opposite to $\End_\C(\W)$. Let $\Ga_\K$ be the unique (up to translation) $\Oo_{\D_\K}$-lattice sequence on $\W$ such that $$\PP_k(\La)\cap\C = \PP_k(\Ga_\K), \qquad k\in\ZZ.$$ Then $[\Ga_\K,n,m-1,\b_i]$ is a simple stratum in $\C$, for $i=1,2$, by [@BSS Proposition 5.2]. From [@BSS Theorem 5.8, Proposition 6.12], we get *interior lifting* maps $$\ll^i = \ll^i_{\K/\F}:\Cc(\La,m-1,\b_i) \to \Cc(\Ga_\K,m-1,\b_i),$$ which are injective and $\KK(\Ga_\K)$-equivariant. Moreover, by [@BSS Proposition 6.13], we have $\ll^1(\t)=\ll^2(\t)$ so that $\Cc(\Ga_\K,m-1,\b_1)\cap\Cc(\Ga_\K,m-1,\b_2)\ne\emptyset$. From the case $\K=\F$ above, we find uniformizers $\w_i$ of $\E_i$ with $\w_1\U^1(\Ga_\K)=\w_2\U^1(\Ga_\K)$; in particular, $\w_1\U^1(\La)=\w_2\U^1(\La)$ which proves (iii). We also get tame corestrictions $s^\K_i$ on $\C$ relative to $\E_i/\K$ satisfying (iv): for all $k\in\ZZ$ and $x\in\PP_k(\Ga_\K)$, $$s_1^\K(x) \equiv s_2^\K(x) \pmod{\PP_{k+1}(\Ga_\K)}.$$ Finally, if $s_\K$ is any tame corestriction on $\A$ relative to $\K/\F$ then $s_i=s_i^\K\circ s_\K$ are tame corestrictions on $\A$ relative to $\E_i/\F$, which satisfy (iv) since $\PP_{k+1}(\Ga_\K)\subseteq\PP_{k+1}(\La)$.
{#section-15}
Now we are ready to complete the proof of the translation principle.
For $i=1,2$, we have simple strata $[\La,n,m,\g_i]$ such that $\Cc(\La,m,\g_1)\cap\Cc(\La,m,\g_2)\ne\emptyset$; these sets of simple characters are then equal, by [@BSS Theorem 4.16]. Moreover, by Lemma \[lem:translation\], we may replace $[\La,n,m,\g_1]$ by an equivalent stratum so that $\Cc(\La,m-1,\g_1)=\Cc(\La,m-1,\g_2)$.
Let $\B_i$ denote the $\A$-centralizer of $\F[\g_i]$, let $\V_i$ be a simple left $\B_i$-module, and let $\D_i$ be the algebra opposite to $\End_{\B_i}(\V_{\F[\g_i]})$. Denote by $\Ga_i$ the unique (up to translation) $\Oo_{\D_i}$-lattice sequence in $\V_i$ such that $$\PP_k(\La)\cap\B_i = \PB_k(\Ga_i),\quad k\in\ZZ.$$
We use Lemma \[lem:crux\] to choose uniformizers $\w_i$ for $\F[\g_i]$ such that $\w_1\U^1(\La)=\w_2\U^1(\La)$, and tame corestrictions $s_i$ on $\A$ relative to $\F[\g_i]/\F$ such that, for all $k\in\ZZ$ and $x\in\PP_k(\La)$, $$s_1(x) \equiv s_2(x) \pmod{\PP_{k+1}(\La)}.$$ Again by Lemma \[lem:crux\], the residue fields $k_{\F[\g_i]}$ have a common image in $\PP_0(\La)/\PP_1(\La)$ so that we may identify them. Moreover, $\PB_0(\Ga_i)/\PB_1(\Ga_i)$ have a common image in $\PP_0(\La)/\PP_1(\La)$, as in the proof of Proposition \[prop:derived\].
Now let $[\La,n,m-1,\b_1]$ be a simple stratum such that $[\La,n,m,\b_1]$ is equivalent to $[\La,n,m,\g_1]$. If the stratum $[\La,n,m,\b_1]$ is itself simple then the result follows from Lemma \[lem:translation\] (applied with $\b_1$ in place of $\g_1$), so we assume this is not the case. We write $\b_1=\g_1+b$, with $b\in\PP_{-m}(\La)$, and pick a simple character $\vartheta\in\Cc(\La,m-1,\g_1)$, so that $\vartheta\psi_b$ is a simple character in $\Cc(\La,m-1,\b_1)$.
By Proposition \[prop:derived\], the derived stratum $[\Ga_1,m,m-1,s_1(b)]$ is equivalent to a simple stratum, which is non-scalar by Proposition \[prop:refinement\], since $k_0(\b_1,\La)
> k_0(\g_1,\La)$. Thus, by Proposition \[prop:simpleminpol\], $[\Ga_1,m,m-1,s_1(b)]$ has irreducible minimum polynomial. However, by the choices of $\w_i$ and $s_i$, we have $$\y(s_1(b),\Ga_1) \equiv \y(s_2(b),\Ga_2) \pmod{\PP_1(\La)}.$$ In particular (given the identifications we have made), we see that the strata $[\Ga_i,m,m-1,s_i(b)]$ have the same minimum and characteristic polynomials. In particular, $[\Ga_2,m,m-1,s_2(b)]$ has irreducible minimum polynomial so, by Proposition \[prop:simpleminpol\], it is equivalent to a simple stratum.
Finally, by Proposition \[prop:refinement\], there is a simple stratum $[\La,n,m-1,\b_2]$ equivalent to $[\La,n,m-1,\g_2+b]$ and, by [@SS4 Proposition 2.15], we have $\vartheta\psi_b\in\Cc(\La,m-1,\b_2)$. In particular, $\Cc(\La,m-1,\b_1)\cap\Cc(\La,m-1,\b_2)\ne\emptyset$ so, by [@BSS Theorem 4.16], we have $\Cc(\La,m-1,\b_1)=\Cc(\La,m-1,\b_2)$ as required.
{#section-16}
We will need one corollary of the translation principle, which is in fact a generalization of Proposition \[prop:derived\]:
\[cor:derived\] Let $[\La,n,m,\g]$ be a simple stratum and let $[\La,n,0,\b]$ be a simple stratum such that $\Cc(\La,m,\b)\cap\Cc(\La,m,\g)\ne\emptyset$. Suppose $\vartheta\in\Cc(\La,m-1,\g)$ and $\theta\in\Cc(\La,m-1,\b)$ coincide on $\H^{m+1}(\g,\La)=\H^{m+1}(\b,\La)$. Then there is $c\in\PP_{-m}(\La)$ such that $\theta=\vartheta\psi_c$ and, for any such $c$, the derived stratum $[\Ga_\g,m,m-1,s_\g(c)]$ is equivalent to a simple (or null) stratum.
The entire statement depends only on the equivalence class of the stratum $[\La,n,m-1,\b]$ so, by replacing with an equivalent stratum, we may assume this stratum is simple. Moreover, that such $c\in\PP_{-m}(\La)$ exists is clear so we need only prove that the derived stratum is simple.
By the translation principle (Theorem \[thm:translation\]), there is a simple stratum $[\La,n,m-1,\b']$ such that $[\La,n,m,\b']$ is equivalent to $[\La,n,m,\g]$ and $\Cc(\La,m-1,\b')=\Cc(\La,m-1,\b)$. Then, by [@SS4 Proposition 2.15], we have $\t|\H^m(\b',\La)=\vartheta'\psi_{c'}$, for some $\vartheta'\in\Cc(\La,m-1,\g)$ and $c'=\b'-\g$. Moreover, the derived stratum $[\Ga_\g,m,m-1,s_\g(c')]$ is simple (or null), by Proposition \[prop:derived\].
Now we have two simple characters, $\vartheta',\vartheta|\H^m(\g,\La)$ in $\Cc(\La,m-1,\g)$, which differ by the character $\psi_{c-c'}$. Since the simple characters are intertwined by $\B_\g^\times$, so is $\psi_{c-c'}$ and, in particular, its restriction to $\U^m(\Ga_\g)$. Then, by Lemma \[lem:scalar\], there is a $\d\in \F[\g]$ such that $s_\g(c-c')-\d\in\PB_{1-m}(\Ga)$. In particular, $[\Ga_\g,m,m-1,s_\g(c)]$ is equivalent to $[\Ga_\g,m,m-1,s_\g(c')+\d]$, which is simple (or null).
Endo-classes and common approximations {#S.common}
======================================
In this section, we collect together some results concerning endo-classes of ps-characters and their relationship with common approximations (see [@BKsemi §8]). Much of this is implicit in [@BKsemi] in the split case.
{#section-17}
Let $(k,\b)$ be a simple pair and, for $i=1,2$, let $[\La_i,n_i,m_i,\h_i(\b)]$ be a realization in a simple central $\F$-algebra $\A_i$. According to [@VS1 §3.3], there is a canonical *transfer* map $$\tau:\Cc(\La_1,m_1,\h_1(\b)) \to \Cc(\La_2,m_2,\h_2(\b)).$$ Denote by $\boldsymbol{\EuScript{C}}_{(k,\b)}$ the set of pairs $([\La,n,m,\h(\b)],\t)$ made of a realization $[\La,n,m,\h(\b)]$ of $(k,\b)$ in a simple central $\F$-algebra and a simple character $\t\in\Cc(\La,m,\h(\b))$. Then the transfer maps $\tau$ induce an equivalence relation on $\boldsymbol{\EuScript{C}}_{(k,\b)}$.
A *potential simple character* over $\F$ (or *ps-character* for short) is a triple $(\Theta,k,\b)$ made of a simple pair $(k,\b)$ over $\F$ and an equivalence class $\Theta$ in $\boldsymbol{\EuScript{C}}_{(k,\b)}$.
When the context is clear, we will often denote by $\Theta$ the ps-character $(\Theta,k,\b)$. Given a realization $[\La,n,m,\h(\b)]$ of $(k,\b)$, we will denote by $\Theta(\La,m,\h)$ the simple character $\t$ such that the pair $([\La,n,m,\h(\b)],\t)$ belongs to $\Theta$.
For $i=1,2$, let $(\Theta_{i},k_{i},\b_{i})$ be a ps-character over $\F$. We say that these ps-characters are *endo-equivalent*, denoted: $$\Theta_1\thickapprox\Theta_2,$$ if $k_1=k_2$ and $[\F(\b_1):\F]=[\F(\b_2):\F]$, and if there exist a simple central $\F$-algebra $\A$ and realizations $[\La,n_i,m_i,\h_i(\b_i)]$ of $(k_i,\b_i)$ in $\A$, for $i=1,2$, such that the simple characters $\Theta_1(\La,m_1,\h_1)$ and $\Theta_2(\La,m_2,\h_2)$ intertwine in $\mult\A$.
That this defines an equivalence relation on ps-characters follows from a major result in [@BSS]:
\[prop:IICpschars\] For $i=1,2$, let $(\Theta_{i},k_{i},\b_{i})$ be a ps-character over $\F$, and suppose that $\Theta_{1}\thickapprox\Theta_{2}$. Let $\A$ be a simple central $\F$-algebra and let $[\La,n_i,m_i,\h_i(\b_i)]$ be realizations of $(k_{i},\b_i)$ in $\A$, for $i=1,2$. Then $\t_1=\Theta_1(\La,m_1,\h_1)$ and $\t_2=\Theta_2(\La,m_2,\h_2)$ intertwine in $\A^{\times}$.
In the situation of Proposition \[prop:IICpschars\], if $(F[\h_i(\b_i)],\La)$ have the same embedding type then we can apply Proposition \[prop:IICwithK\] to conclude that the realizations $\t_1$, $\t_2$ are actually conjugate.
We will use the common convention that, for each $t\ge 0$, there is the *trivial ps-character $\Theta^{(t)}_0$*, whose realization on any lattice sequence $\La$ is the trivial character of the group $\U^{t+1}(\La)$; then $\{\Theta^{(t)}_0\}$ forms a singleton equivalence class under endo-equivalence.
{#section-18}
Let $[\La,n,0,\b]$ be a simple stratum in $\A=\End_\D(\V)$. Let $\t\in\Cc(\La,0,\b)$ be a simple stratum and denote by $(\T,0,\b)$ the ps-character that it determines – that is, $\t$ is a realization of $\T$.
For $t\ge 0$, let $[\La,n,t,\b^{(t)}]$ be a simple stratum equivalent to the pure stratum $[\La,n,t,\b]$ and write $\E^{(t)}=\F[\b^{(t)}]$. Then the restriction $\t|{\H^{t+1}(\b,\La)}$ is a simple character in $\Cc(\La,t,\b^{(t)})$ and we denote by $(\T^{(t)},k^{(t)},\b^{(t)})$ the ps-character determined by this restriction, with $k^{(t)}=\lfloor t/e(\La|\Oo_{\E^{(t)}})\rfloor$.
[**Remark.**]{} We allow the case $t\ge n$, in which case we interpret the stratum $[\La,n,t,\b]$ to be equivalent to a null stratum $[\La,t,t,0]$ and $\T^{(t)}$ is the trivial ps-character.
\[lem:PSembed\] With notation as above let $(\T_\g,k,\g)$ be a ps-character which is endo-equivalent to $\T^{(t)}$. Then there is an embedding $\iota_\g:\F[\g]\hookrightarrow \A$ such that $$\t|{\H^{t+1}(\b,\La)} = \T_\g(\La,t,\iota_\g).$$
Put $\E_\g=\F[\g]$. Since the ps-characters $\T_\g$, $\T^{(t)}$ are endo-equivalent, the fields $\E_\g$, $\E^{(t)}$ have the same invariants by [@BSS Lemma 4.8]: $$e(\E_\g/\F)=e(\E^{(t)}/\F), \ \
f(\E_\g/\F)=f(\E^{(t)}/\F),
\hbox{ \ and \ }
k_\F(\g)=k_\F(\b^{(t)}).$$ Then, by Lemma \[lem:embedtype\], there is an embedding $\iota_\g:\K_\g\hookrightarrow\A$ such that $[\La,n,t,\iota_\g(\g)]$ is a pure stratum with the same embedding type as $[\La,n,t,\b^{(t)}]$, which is simple since $k_\F(\g)=k_\F(\b^{(t)})$ and $[\La,n,t,\b^{(t)}]$ is simple.
Finally, since the ps-characters are endo-equivalent, by Propositions \[prop:IICpschars\] and \[prop:IICwithK\], the realization $\T_\g(\La,t,\iota_\g)$ is conjugate to $\t|{\H^{t+1}(\b,\La)}$ by some $u\in\KK(\La)$. Conjugating our embedding $\iota_\g$ by $u$ gives the desired embedding.
For $j=1,\ldots,r$, let $\A^j=\End_\D(\V^j)$, let $[\La^j,n_j,0,\b_j]$ be a simple stratum in $\A^j$, write $\E_j=\F[\b_j]$, and let $\t_j\in\Cc(\La^j,0,\b_j)$. We normalize so that the lattice sequences $\La^j$ have the same $\Oo_\F$-period $e$.
As above, for $t \ge 0 $, let $[\La^j,n_j,t,\b_j^{(t)}]$ be a simple stratum equivalent to the pure stratum $[\La^j,n_j,t,\b_j]$, and write $\E_j^{(t)}=\F[\b_j^{(t)}]$. The restriction $\t_j|{\H^{t+1}(\b_j,\La^j)}$ is a simple character in $\Cc(\La^j,t,\b_j^{(t)})$ and we denote by $(\T_j^{(t)},k_j^{(t)},\b_j^{(t)})$ the ps-character determined by this restriction, with $k_j^{(t)}=\lfloor t/e(\La|\Oo_{\E_j^{(t)}})\rfloor$.
We put $\V=\V^1\oplus\cdots\oplus\V^r$, and set $\La=\La^1\oplus\cdots\oplus\La^r$, a lattice sequence in $\V$ of $\Oo_\F$-period $e$. Write $\A=\End_\D(\V)$ and denote by $\e^j$ the idempotents in $\PP_0(\La)$ corresponding to the decomposition of $\V$. We put $\b=\sum_{j=1}^r \e^j\b_j\e^j$. Then $[\La,n,0,\b]$ is a stratum in $\A$, with $n=\max_j n_j$. We write $\LL$ for the stabilizer in $\G=\Aut_\D(\V)$ of the decomposition $\V=\V^1\oplus\cdots\oplus\V^r$, and $\A_\LL$ for its stabilizer in $\A$.
A *common approximation of $(\t_j)$ of level $t$ on $\La$* is a pair $([\La,n,t,\g],\vartheta)$ consisting of: a simple stratum $[\La,n,t,\g]$ with $\g\in\A_\LL$ and $0\le t\le n$, such that $$\H^{t+1}(\g,\La)\cap\LL =\prod_{j=1}^r\H^{t+1}(\b_j,\La^j);$$ and a simple character $\vartheta\in\Cc(\La,0,\g)$ such that $$\vartheta|\H^{t+1}(\g,\La)\cap\LL = \t_1\otimes\cdots\otimes\t_r.$$
When we have such a common approximation, we will identify $\g$ with its images $\e^j\g\e^j$ in $\A^j$.
\[lem:commonPS\] Let $0\le t\le n$. Then the following are equivalent:
- There is a common approximation of $(\t_j)$ of level $t$ on $\La$.
- The ps-characters $\T_j^{(t)}$ are endo-equivalent.
(ii)$\Rightarrow$(i) Let $(\T_\g,k,\g)$ be a ps-character which is endo-equivalent to all $\T_j^{(t)}$. Then, by Lemma \[lem:PSembed\], for each $j$ there is an embedding $\ii_j:\F[\g]\hookrightarrow \A^j$ such that $\T_\g(\La^j,t,\ii_j)=\t_j|\H^{t+1}(\b_j,\La^j)$. Denote by $\ii$ the diagonal embedding $\ii:\F[\g]\hookrightarrow \bigoplus_{j=1}^r \A^j
\subseteq \A$ and let $\vartheta$ be any simple character in $\Cc(\La,0,\ii(\g))$ which restricts to $\T_\g(\La,t,\ii)$ on $\H^{t+1}(\g,\La)$. Then $([\La,n,t,\ii(\g)],\vartheta)$ is a common approximation as required.
(i)$\Rightarrow$(ii) Suppose $([\La,n,t,\g],\vartheta)$ is a common approximation of $(\t_j)$ of level $t$ on $\La$. Then the characters $\t_j|\H^{t+1}(\b_j,\La^j)$ are simple characters in $\Cc(\La^j,t,\g)$ and, by [@SS4 Théorème 2.17], these characters are all transfers of each other relative to $\g$; hence the corresponding ps-characters (which are supported by the simple pair $(k,\g)$, with $k=\lfloor
t/e(\La|\Oo_{\F[\g]})\rfloor$) are endo-equivalent.
We suppose now that there is a common approximation $([\La,n,t,\g],\vartheta)$ of $(\t_j)$. Denote by $\B_\g$ the $\A$-centralizer of $\E_\g=\F[\g]$, by $\V_\g$ a simple left $\B_\g$-module, by $\D_\g$ the opposite algebra to $\End_{\B_\g}(\V_\g)$, and by $s_\g$ a tame corestriction on $\A$. Note that, since $\g\in\A_\LL$, the restriction of $s_\g$ to $\A^j$ is also a tame corestriction (see [@SS4 Proposition 2.26]). Also, the idempotents $\e^j$ lie in $\B_\g$ so correspond to a decomposition $\V_\g=\V^1_\g\oplus\cdots\oplus\V^r_\g$. Let $\Ga_\g$ be an $\Oo_{\D_\g}$-lattice sequence in $\V_\g$ such that $\PP_n(\La)\cap\B_\g=\PB_n(\Ga_\g)$, for all $n\in\Z$, and put $\Ga_\g^j=\Ga_\g\cap\V_\g^j$, for $1\le j\le r$.
Since $\t_j$ and $\vartheta$ coincide on $\H^{t+1}(\g,\La)$, Corollary \[cor:derived\] says that there is $c_j\in\PP_{-t}(\La^j)$ such that $\t_j|\H^{t}(\beta_j,\La^j) = \vartheta\psi_{c_j}$, and that the derived stratum $[\Ga^j_\g,t,t-1,s_\g(c_j)]$ is equivalent to a simple (or null) stratum. The following result is a generalization of Corollary \[cor:minpolpair\].
\[cor:endocommon\] In the situation above, the derived strata $[\Ga^j_\g,t,t-1,s_\g(c_j)]$ have the same minimum polynomial if and only if the ps-characters $\T_j^{(t-1)}$ are endo-equivalent.
Suppose first that the minimum polynomials of the derived strata $[\Ga^j_\g,t,t-1,s_\g(c_j)]$ are all the same. Note that they are irreducible since these strata are equivalent to simple strata. Then the derived stratum $[\Ga_\g,t,t-1,s_\g(c)]$ is equivalent to a simple stratum by Corollary \[cor:minpol\] and, by Proposition \[prop:refinement\], there is a simple stratum $[\La,n,t-1,\g']$ equivalent to $[\La,n,t-1,\g+c]$, so that $\vartheta\psi_c\in\Cc(\La,t-1,\g')$. Then, for any $\vartheta'\in\Cc(\La,0,\g')$ extending $\vartheta\psi_c$, the pair $([\La,n,t-1,\g'],\vartheta')$ is a common approximation of level $t-1$. Hence, by Lemma \[lem:commonPS\], the ps-characters $\T_j^{(t-1)}$ are endo-equivalent.
Conversely, suppose the ps-characters $\T_j^{(t-1)}$ are endo-equivalent so, by Lemma \[lem:commonPS\], there is a common approximation $([\La,n,t-1,\g'],\vartheta')$ of level $t-1$. We then have $\vartheta'=\vartheta\psi_c$ and, by Corollary \[cor:derived\], the derived stratum $[\Ga_\g,t,t-1,s_\g(c)]$ is equivalent to a simple (or null) stratum. Hence, by Corollary \[cor:minpol\], the derived strata $[\Ga^j_\g,t,t-1,s_\g(c_j)]$ have the same minimum polynomial.
Simple types {#S.simple}
============
In this section we recall some results from [@VS3] concerning simple types. In later sections we will need these in slightly more generality than in *op. cit.* – in particular, in the case where we have a non-strict lattice sequence. Already in the case of $\GL(n,\F)$, simple types on non-strict lattice sequence are required in [@BKsemi], although this is not immediately apparent. The proofs are mostly identical to those in [@VS2; @VS3].
{#section-19}
Let $[\La,n,0,\b]$ be a simple stratum in $\A=\End_\D(\V)$, and use all the usual notation from the previous sections. Since $\b$ is fixed, we will omit it from the notations; when $\La$ is fixed, we will omit that also.
\[eta\] Let $\t\in\Cc(\La,0,\b)$ be a simple character. Then there is a unique irreducible representation $\eta$ of $\J^1$ which contains $\t$; moreover, $\eta|\H^1$ is a multiple of $\t$, the dimension $\dim(\eta)=(\J^1:\H^1)^{1/2}$ and $$\dim \I_g(\eta)=\begin{cases} 1
&\hbox{ if }g\in\J^1\B^\times\J^1, \\
0 &\hbox{ otherwise.} \end{cases}$$
The proof of all but the final assertion is identical to that of [@BK Proposition 5.1.1)], replacing [@BK Theorem 3.4.1)] by [@SS4 Proposition 2.31]. The proof of the final assertion is identical to that of [@BK Proposition 5.1.8)], replacing the exact sequences there by those of [@SS4 Proposition 2.27].
For $i=1,2$, let $[\La^i,n_i,0,\b]$ be a simple stratum in $\A$, let $\t_i\in\Cc(\La^i,0,\b)$, and let $\eta_i$ be the unique irreducible representation of $\J^1(\b,\La^i)$ which contains $\t_i$. Then $$\frac{\dim(\eta_1)}{\dim(\eta_2)} =
\frac{(\J^1(\b,\La^1):\J^1(\b,\La^2))}{(\U^1(\La^1)\cap\B:\U^1(\La^2)\cap\B)}.$$
Again, the proof is identical to that of [@BK Proposition 5.1.2], replacing the exact sequence there with [@SS4 Proposition 2.27].
{#section-20}
Recall that a *$\b$-extension of $\t$* is a representation $\k$ of $\J$ which extends the representation $\eta$ given by Lemma \[eta\] and such that $I_G(\k)\supset \B^\times$. In the case that $\La$ is strict, the existence of $\b$-extensions is given by [@VS2 Théorème 2.28]. Using this, we proceed here via a simplified version of the compatibility argument used there.
\[def:compat\] Let $[\La,n,0,\b]$ be a simple stratum in $\A$ and let $\La'$ be an $E$-pure lattice sequence in $\V$ such that $\PP_0(\La)=\PP_0(\La')$. Let $\t,\t'$ be simple characters which are realizations of the same ps-character on $\La,\La'$ respectively, and let $\k,\k'$ be extensions of the representations $\eta,\eta'$ given by Lemma \[eta\] respectively. We say that $\k,\k'$ are *compatible* (or *mutually coherent*) if $$\Ind_{\J}^{(\U(\La)\cap\B)\U^1(\La)} \k \simeq
\Ind_{\J(\b,\La')}^{(\U(\La)\cap\B)\U^1(\La)} \k'.$$
With the notations of Definition \[def:compat\], the notion of compatibility induces a bijection $$\left\{\hbox{$\b$-extensions of $\t$}\right\} \longleftrightarrow
\left\{\hbox{$\b$-extensions of $\t'$}\right\}.$$ In particular, there is a $\b$-extension $\k$ of $\t$, and then the set of $\b$-extension of $\t$ is given by $$\left\{\k\otimes\left(\chi\circ\N_{\B/\E}\right) : \chi\in
\widehat{\U_\E/\U^1_\E}\right\}.$$
The first assertion follows as in [@VS2 Lemmes 2.23,2.24] (cf. also [@BK Proposition 5.25]). Now taking $\La'$ to be the strict lattice sequence in $\V$ with the same image as $\La$, the final assertion follows from [@VS2 Théorème 2.28].
{#SS.etaP}
We continue with a simple stratum $[\La,n,0,\b]$ and a simple character $\t\in\Cc(\La,0,\b)$, together with the unique irreducible representation $\eta$ of $\J^1$ containing $\t$. Let $\V_\E$ be a simple left $\B$-module, let $\D_\E$ be the opposite algebra to $\End_\B(\V_\E)$, and set $m_\E=\dim_{\D_\E}\V_\E$. We write $\Ga$ for the unique (up to translation) $\Oo_{\D_\E}$-lattice sequence on $\V_\E$ such that $\PP_k(\La)\cap\B=\PP_k(\Ga)$, for all $k\in\ZZ$.
We suppose given a decomposition $\V=\V^1\oplus\cdots\oplus\V^r$ which is *subordinate to $\PP_0(\Ga)$* in the sense of [@SS4 Définition 5.1]: that is, it is a decomposition of $\E\otimes\D$-bimodules and, writing $\e^j$ for the idempotents of $\PP_0(\Ga)$ defined by the decomposition and $m_j=\dim_{\D_\E}\e^j\V_\E$, there is an isomorphism of $E$-algebras $\Psi:\B\to\M_{m_\E}(\D_\E)$ such that:
1. for $1\le j\le r$, the idempotent $\Psi(\e^j)$ is $\I^j=\diag(0,\ldots,\Id_{m_j},\ldots,0)$:
2. The hereditary order $\Psi(\PB_0(\Ga))$ is the $\Oo_\E$-subalgebra of $\M_{m_\E}(\Oo_{\D_\E})$ consisting of matrices whose reduction modulo $\p_{\D_\E}$ is upper triangular by blocks of size $(m_1,\ldots,m_r)$.
Note then that $\La^j=\La\cap\V^j$ is in the affine class of a strict lattice sequence of $\Oo_{\D_\E}$-period $1$ in $\V^j$.
Let $\P$ be the parabolic subgroup of $\G$ stabilizing the flag $$\{0\}\subset \V^1\subset \V^1\oplus \V^2 \subset \cdots \subset \V,$$ and write $\P=\LL\N$, where $\LL$ is the stabilizer of the decomposition $\V=\bigoplus_{j=1}^r \V^j$ and $\N$ is the unipotent radical. Write $\P_-=\LL\N_-$ for the opposite parabolic relative to $\LL$.
We define the groups $$\J_\P=\H^1\left(\J\cap\P\right),\quad
\J^1_\P=\H^1\left(\J^1\cap\P\right),\quad
\H^1_\P=\H^1\left(\J^1\cap\N\right),$$ and define the character $\t_\P$ of $\H^1_\P$ by $\t_\P(hu)=\t(h)$, for $h\in\H^1$ and $u\in\J^1\cap\N$. We also put $\J_\LL=\J\cap\LL$ and $\J^1_\LL=\J^1\cap\LL$ and notice that, since the decomposition is subordinate to $\PP_0(\La)\cap\B$, we have $$\J_\P/\J^1_\P \simeq \J_\LL/\J^1_\LL \simeq
\U(\Ga)/\U^1(\Ga) \simeq \J/\J^1.$$ In particular, given a representation of $\U(\Ga)$ trivial on $\U^1(\Ga)$, we can also regard it as a representation of $\J_\P$ (respectively $\J_\LL$, $\J$) trivial on $\J^1_\P$ respectively $\J^1_\LL$, $\J^1$).
The following Proposition summarizes the results of [@SS4 Propositions 5.3–5] (see also *op. cit.* §5.8); the results there are in the case that $\La$ is strict but, given our preliminary results above, identical proofs apply in the general case.
Let $\eta_\P$ denote the natural representation of $\J^1_\P$ on on the $\J\cap\N$-invariants of $\eta$. Then $\eta_\P$ is the unique irreducible representation of $\J^1_\P$ which contains $\t_\P$. Moreover, $\Ind_{\J_\P}^{\J^1}\eta_\P$ is isomorphic to $\eta$ and $$\dim \I_g(\eta_\P)=\begin{cases} 1
&\hbox{ if }g\in\J_\P^1\B^\times\J_\P^1, \\
0 &\hbox{ otherwise.} \end{cases}$$
{#section-21}
Now let $\k$ be a $\b$-extension of $\t$ and let $\k_\P$ denote the natural representation of $\J_\P$ on on the $\J\cap\N$-invariants of $\k$. The proof of [@SS4 Proposition 5.8] (see also *op. cit.* §5.8) again generalizes to the non-strict case and gives:
\[prop:kP\] $\k_\P$ is an irreducible representation of $\J_\P$ with the following properties:
1. $\k_\P|\H^1(\b,\La)$ is a multiple of $\t$;
2. $\k_\P$ is trivial on $\J_\P\cap\N$ and $\J_\P\cap\N_-$;
3. $\k_\P|\J_\LL \simeq \k_1\otimes\cdots\otimes\k_r$, for some $\b$-extensions $\k_j$ containing $\t_j$;
4. $\I_\G(\k_\P)=\J_\P\B^\times\J_\P$;
5. if $m_j=m_k$ then $\k_j\simeq\k_k$.
6. if $\xi$ is an irreducible representation of $\U(\Ga)$ trivial on $\U^1(\Ga)$, then $$\Ind_{\J_\P}^\J (\k_\P\otimes\xi) \simeq \k\otimes\xi,
\qquad\hbox{and}\qquad
\I_\G(\k_\P\otimes\xi)=\J_\P\I_{\B^\times}(\xi)\J_\P.$$
The only property not given by [@SS4 Proposition 5.8 or §5.8] is (v), where imitate the proof of [@BK Corollary 7.2.6]. There is a permutation matrix in $w\in\M_{m_\E}(\D_\E)$ (which we have identified with $\B$ via $\Psi$ above) which swaps $\V^j$ with $\V^k$, and leaves all other $\V^i$ fixed. In particular, it lies in $\B$ and normalizes $\J_\P\cap\LL$. Then $w$ intertwines $\k_\P$ so normalizes $\k_\P|\J_\LL$ and hence induces an isomorphism between $\k_j$ and $\k_k$.
{#section-22}
Finally, we consider the case where all $m_j$ are equal to some integer $s$, so that $\U(\Ga)/\U^1(\Ga)\simeq
\GL_s(k_{\D_\E})^r$ and let $\xi$ be the inflation to $\U(\Ga)$ of the representation $\sigma^{\otimes r}$, for $\sigma$ an irreducible cuspidal representation of $\GL_s(k_{\D_\E})$.
We put $\l=\k\otimes\xi$, $\l_\P=\k_\P\otimes\xi$ and $\l_\LL=\l_P|\J_\LL$.
\[simplecase\] The pair $(\J_\P,\l_\P)$ is a cover of $(\J_\LL,\l_\LL)$ and $$\Hh(\G,\l_\P)\cong \Hh(r,q_{\D_\E}^{s}).$$
We remark that the parameter $q_{\D_\E}^{s}$ here is the same as that in Theorem \[thm:simplecover\] of the introduction, by [@VSU0 Theorem 4.6].
In the case that $\La$ is strict, this is given by [@VS3 Proposition 5.5, Théorème 4.6]. The idea of the proof is to reduce to this case, as in the proof of [@VS3 Théorème 5.6]. Let $\Ll$ be the strict lattice sequence with the same image as $\La$ and we make the same constructions for $\Ll$, which we denote with the superscript $^\Ll$. In particular, we choose a $\b$-extension $\k^\Ll$ compatible with $\k$. Hence (as in [@VS3 Proposition 4.5]) we have a support-preserving isomorphism $$\Hh(\G,\l) \cong \Hh(\G,\l^\Ll).$$ Moreover, we have $\l=\Ind_{\J_\P}^\J \l_\P$ and $\l^\Ll=\Ind_{\J^\Ll_\P}^{\J^\Ll} \l^\Ll_\P$, so we get a support-preserving isomorphism $$\Hh(\G,\l_\P) \cong \Hh(\G,\l^\Ll_\P).$$ Then the assertions follow from (the proof of) [@VS3 Proposition 5.5].
A pair $(\J,\l)$ as in this paragraph is called a *simple type*; if $r=1$ then it is called a *maximal simple type*.
Recall that the main result of [@SS4] (Théorèmes 5.21, 5.23) is that every irreducible cuspidal representation of $\G$ contains a maximal simple type.
Intertwining and conjugacy {#S.IIC}
==========================
In this section we consider the unicity of the simple type contained in an irreducible representation $\pi$ of $\G=\GL_m(\D)$. That is, we suppose the inertial class $\ss(\pi)$ of $\pi$ is homogeneous: there are a positive integer $r$ dividing $m$, an irreducible cuspidal representation $\rho$ of the group $\G_0=\GL_{m/r}(\D)$ and unramified characters $\chi_i$ of $\G_0$, with $i\in\{1,\dots,r\}$, such that $\pi$ is isomorphic to a quotient of the normalized parabolically induced representation $\rho\chi_1\times\dots\times\rho\chi_r$. Unlike the situation for $\D=\F$, the simple type is *not* uniquely determined up to conjugacy in general, as there is a galois action we must take into account.
We consider the set $\SS$ of *sound simple types* $$\SS=\left\{\,\hbox{\begin{minipage}{0.6\textwidth}simple types $(\J(\b,\La),\l)$ such that $\La$ is strict, $\PP_0(\La)$ is principal and $(\F[\b],\La)$ is soundly embedded\end{minipage}}\,\right\}.$$ For $(\J,\l)\in\SS$, we use all the associated notation of §\[S.simple\]: that is, there are a (sound) simple stratum $[\La,n,0,\b]$, a simple character $\t\in\Cc(\La,0,\b)$, a $\b$-extension $\k$ and a representation $\xi$ which is the inflation to $\U(\Ga)$ of the representation $\overline\xi=\s^{\otimes r}$ of $$\U(\Ga)/\U^1(\Ga) \simeq \GL_s(k_{\D_\E})^r,$$ for $\s$ an irreducible cuspidal representation of $\GL_s(k_{\D_\E})$. Note that, implicit in the isomorphism above are the choice of a decomposition $\V=\V^1\oplus\cdots\oplus\V^r$ subordinate to $\PP_0(\Ga)$ and the choice of an $\E$-algebra isomorphism $\Psi:\B\to\M_{m_\E}(\D_\E)$ as in paragraph \[SS.etaP\].
We fix a uniformizer $\w$ of $\D_\E$. The Galois group $\Gg=\Gal(k_{\D_\E}/k_\E)$ identifies, via reduction, with the group generated by ${\rm Ad}(\w)$, the inner automorphism given by conjugation by $\w$. The Galois group $\Gg$ acts on the representations of $\GL_s(k_{\D_\E})$. Moreover, a different choice of $\E$-algebra isomorphism $\Psi$ could result in a change in the identification $\U(\Ga)/\U^1(\Ga) \simeq \GL_s(k_{\D_\E})^r$ by conjugating each factor by an element of $\Gg$, rather than just by an inner automorphism. Thus we define $[\s]$ to be the orbit of $\s$ under the action of $\Gg$ and set $$[\l] = \left\{\,\hbox{\begin{minipage}{0.58\textwidth}equivalence classes of representations $\k\otimes\xi'$, for $\xi'$ the inflation to $\U(\Ga)$ of $\s_1\otimes\cdots\otimes\s_r$, with $\s_i\in[\s]$\end{minipage}}\,\right\}.$$
We also define an equivalence relation on $\SS$ by: $(\J,\l) \sim (\J,\l')$ if and only if there is an irreducible representation $\pi$ of $\G$ such that $\pi$ contains both $\l$ and $\l'$.
\[thm:simpleconj\] Let $(\J,\l)$ and $(\J',\l')$ be sound simple types. Then $(\J,\l) \sim (\J',\l')$ if and only if there exists $g\in\G$ such that $^g\J'=\J$ and $\left[{}^g\l'\right]=[\l]$.
The proof follows that of [@BK Theorem 5.7.1]. Suppose first that $(\J,\l) \sim (\J',\l')$ and use all notation as above, with a prime $'$ to indicate the corresponding objects for $\J',\l')$, in particular writing $\E'=\F[\b']$. We also write $\Theta$ for the ps-character defined by $\t$. Then [@BSS Theorems 9.2, 9.3] imply that:
- $\Theta'$ is endo-equivalent to $\Theta$;
- $(\E,\La)$ and $(\E',\La')$ have the same embedding type.
In particular, by Propositions \[prop:IICpschars\] and \[prop:IICwithK\], and the definition of embedding type, there is $g\in\G$ such that $g\La'$ is in the translation class of $\La$, $^g\H^1(\b',\La')=\H^1(\b,\La)$ and $^g\t'=\t$. Replacing $\l'$ by $^g\l'$ we may assume that $g=1$, so that $\t'=\t$; moreover, since changing $\La$ in its translation class affects nothing, we may assume $\La'=\La$.
Now the $\U(\La)$ intertwining of $\t$ is $\J(\b,\La)$ so we get $\J'=\J$ and $\J^1(\b',\La')=\J^1(\b,\La)$. By unicity in Lemma \[eta\], we get $\eta'=\eta$. Moreover, since the intertwining of $\t$ is $\J\B^\times\J=\J(\B')^\times\J$, the $\b$-extension $\k$ is also a $\b'$-extension and we may assume $\k'=\k$.
As in the proof of [@BK Theorem 5.7.1], the cuspidality of $\overline\xi$ can be interpreted in purely group-theoretic terms. In particular, if we identify $\J/\J^1$ with $\GL_s(k_{\D_\E})^r$, then $\overline{\xi'}$ decomposes as $\s'_1\otimes\cdots\otimes\s'_r$ with all $\s'_i$ cuspidal.
Now $\l$, $\l'$ are contained in some irreducible representation $\pi$ of $\G$ and are therefore intertwined by some $x\in\G$. Since $\l$, $\l'$ both restrict to a multiple of $\t$ on $\H^1$, we have also that $x$ intertwines $\t$ and thus $x\in\J^1\B^\times\J^1$. In particular, we may assume $x\in\B^\times$. Then, arguing as in [@BK Proposition 5.3.2], we see that $x$ intertwines $\xi$ with $\xi'$, when we interpret them as representations of $\U(\Ga)$.
To finish, we argue again as in the proof of [@BK Theorem 5.7.1], using results from [@GSZ]. In particular, we will use some notation from [@GSZ §0.8], writing $\tilde\W_\B$ for the generalized affine Weyl group in $\B^\times$, which we have identified with $\GL_{m_\E}(\D_\E)$ via $\Psi$. By the affine Bruhat decomposition, we may assume $x\in\tilde\W_\B$.
Since $\overline\xi$ and $\overline{\xi'}$ are cuspidal, the same proof as that of [@GSZ Proposition 1.2] shows that $x$ normalizes $\LL\cap\U(\Ga)$, where $\LL$ is the Levi subgroup of $\G$ which is the stabilizer of the decomposition $\V=\V^1\oplus\cdots\oplus\V^r$. Likewise, [@GSZ Lemma 1.5] implies that $$\Hom_{\U(\Ga)\cap x^{-1}\U(\Ga)x}(\xi',\xi^x) = \Hom_{\U(\Ga)/\U^1(\Ga)}(\overline{\xi'},\overline\xi^x).$$ Now $\overline\xi^x=\s_1\otimes\cdots\otimes\s_r$, with each $\s_i\in[\s]$ and, since $\overline{\xi'}$, $\overline\xi^x$ are irreducible, we deduce $\s'_i\simeq\s_i\in[\s]$. Hence the equivalence class of $\l'$ is in $[\l]$, so $[\l']=[\l]$, as required.
The converse is given by [@SS4 Proposition 5.19].
Although, in general, equivalent sound simple types are not conjugate, they are in the special case of cuspidal representations:
\[cor:maxsimpleconj\] Suppose $(\J,\l)$, $(\J',\l')$ are maximal simple types. Then $(\J,\l) \sim (\J',\l')$ if and only if there exists $g\in\G$ such that $^g\J'=\J$ and $^g\l'\simeq\l$.
Suppose $(\J,\l)$ and $(\J',\l')$ are equivalent maximal simple types. By Theorem \[thm:simpleconj\], there there exists $g\in\G$ such that $^g\J'=\J$ and $\left[{}^g\l'\right]=[\l]$. That is, as in the proof of Theorem \[thm:simpleconj\], we can write $^g\l'\simeq \k\otimes\xi'$, with $\overline{\xi'}\simeq\overline\xi^\g$, for some $\g\in\Gg$. But the action of $\g$ can be realized as conjugation by a power of $\w$, which normalizes $\Ga$, so there is $y\in\KK(\Ga)$ such that $^y\xi'\simeq\xi$. Since $y$ also normalizes $\k$, we deduce that $^{yg}\l'\simeq\l$.
This also completes the proof of Theorem \[thm:cuspidal\] of the introduction.
Semisimple types {#S.Semisimple}
================
Suppose we have cuspidal representations $\pi_j$ of $\G^j=\Aut_\D(\V^j)$, for $1\le j\le r$, and we think of $\LL=\prod_{j=1}^r\G_j$ as a Levi subgroup in $\G=\Aut_\D(\V)$, where $\V=\bigoplus_{j=1}^r \V^j$. The aim of this section is to prove Theorem \[thm:coverhecke\]: a maximal simple type for $(\LL,\pi_1\otimes\cdots\otimes\pi_r)$ admits a cover, with an explicitly computable Hecke algebra.
For $j=1,\ldots,r$, the cuspidal representation $\pi_j$ contains a (maximal) simple type $(\J(\b_j,\La^j),\l_j)$, where $[\La^j,n_j,0,\b_j]$ is a simple stratum in $\A^j=\End_\D(\V^j)$, $\t_j$ is a simple character of $\H^1(\b_j,\La^j)$, $\k_j$ is a $\b_j$-extension containing $\t_j$, $\sigma_j$ is a cuspidal representation of $\J(\b_j,\La^j)/\J^1(\b_j,\La^j)$ and $\l_j=\k_j\otimes\sigma_j$. We write $(\T_j,0,\b_j)$ for the ps-character defined by $\t_j$. Then our type in $\LL$ is $(\J_\LL,\l_\LL)$, given by $$\J_\LL = \prod_{j=1}^r \J(\b_j,\La^j),\quad
\l_\LL = \l_1\otimes\cdots\otimes\l_r.$$
For $j=1,\ldots,r$, we write $\B^j$ for the $\A_j$-centralizer of $\b_j$. Then $\B_j$ has the form $\End_{\D_{\E_j}}(\W^j)$, for some right $\D_{\E_j}$-vector space $\W^j$. We write $\Ga^j$ for the unique strict $\Oo_{\D_{\E_j}}$-lattice sequence in $\W^j$ such that $\PB_0(\Ga^j)=\PP_0(\La^j)\cap\B^j$. Since $\pi_j$ is cuspidal, $\Ga^j$ is a sequence of $\Oo_{\D_{\E_j}}$-period $1$ and then the normalizer in $\G^j\cap\B^j$ of $\Ga^j$ is just $\KK(\La^j)\cap\B^j$; moreover, $\La^j$ is a strict lattice sequence and $\PP_0(\La^j)$ is a principal order in $\A^j$.
The homogeneous case
--------------------
We suppose first that the $(\T_j,0,\b_j)$ are all endo-equivalent to some fixed ps-character $(\T,0,\b)$. By Lemma \[lem:PSembed\], for each $j$ there is a realization $\T(\La^j,0,\ii_j)$ equal to $\t_j$. Hence we may (and do) assume that all $\t_j$ are defined relative to the same simple pair $(0,\b)$ and are realizations of the same ps-character $\T$.
We have $\V=\bigoplus_{j=1}^r \V^j$ and put $\W=\bigoplus_{j=1}^r
\W^j$, so that $\B=\End_{\D_\E}(\W)$. Write $\e^j$ for the idempotent in $\B$ with image $\W^j$ and kernel $\bigoplus_{i\ne j}\W^i$. As an element of $\A$ it has image $\V^j$ and kernel $\bigoplus_{i\ne
j}\V^i$.
Let $\Ga$ be an $\Oo_{\D_{\E}}$-lattice sequence in $\W$ such that:
1. $\Ga\cap\W^j$ is in the affine class of $\Ga^j$, and
2. $\Ga$ is subordinate to the decomposition $\W=\bigoplus_{j=1}^r \W^j$.
Note that condition (ii) is generically satisfied: that is, amongst all lattice sequences satisfying (i), those also satisfying (ii) are dense (in the building of $\B^\times$). A particular example of such a sequence is given by $$\Ga(k)\ =\ \bigoplus_{j=1}^{r}
\Ga^j\left(\left\lfloor\frac{k+j}{r}\right\rfloor\right),\qquad k\in\ZZ,$$ which is strict of $\Oo_{\D_{\E}}$-period $r$ but not principal in general.
We fix $\Ga$ satisfying (i),(ii) and let $\La$ be the corresponding $\Oo_\D$-lattice sequence in $\V$ given by [@SS4 Théorème 1.7]. Then $\e^j\in\PB_0(\Ga)\subseteq\PP_0(\La)$ so $\La$ is decomposed by (indeed subordinate to) the decomposition $\V=\bigoplus_{j=1}^r \V^j$ and the lattice sequence $r\mapsto
\La(r)\cap \V^j$ is in the affine class of $\La^j$. In fact, replacing $\La^j$ by this sequence changes nothing in the construction of the type $(\J(\b_j,\La^j),\l_j)$ so we may (and do) assume this done, in which case $\La=\bigoplus_{j=1}^r\La^j$.
Let $\P$ be the parabolic subgroup of $\G$ stabilizing the flag $$\{0\}\subset \V^1\subset \V^1\oplus \V^2 \subset \cdots \subset \V,$$ and write $\P=\LL\N$, where $\LL$ is the stabilizer of the decomposition $\V=\bigoplus_{j=1}^r \V^j$ and $\N$ is the unipotent radical. Write $\P_-=\LL\N_-$ for the opposite parabolic relative to $\LL$.
Now $[\La,n,0,\b]$ is a simple stratum in $\A=\End_\D(\V)$, for a suitable integer $n$, and we have $\H^1(\b,\La)\cap\LL\cong \prod_{j=1}^r \H^1(\b,\La^j)$, with similar decompositions for $\J^1$ and $\J$. Moreover, $\t=\T(V,0,\La)$ is a simple character such that $$\t|\H^1(\b,\La)\cap\LL = \t_1\otimes\cdots\otimes\t_r.$$ As in §\[S.simple\], we define the groups $$\J_\P=\H^1(\b,\La)\left(\J(\b,\La)\cap\P\right),\qquad
\J^1_\P=\H^1(\b,\La)\left(\J^1(\b,\La)\cap\P\right),$$ noting that $\J_\LL=\J_\P\cap\LL$.
Let $\k_\P$ be the irreducible representation of $\J_\P$ given by Proposition \[prop:kP\], so that $\k_\P|\J_\LL \simeq
\bigotimes_{j=1}^r \k'_j$, for some $\b$-extensions $\k'_j$ containing $\t_j$. Then we can choose the decompositions $\l_j=\k_j\otimes\sigma_j$ of the maximal simple types above so that $\k_j=\k'_j$, which we assume done.
We define an equivalence relation $\sim$ on $\{1,\ldots,r\}$ by $$j\sim k\ \iff\ \sigma_j\simeq\sigma_k^\gamma,\hbox{ for some
}\gamma\in\Gal(k_{\D_\e}/k_\E),$$ and denote by $I_1,\ldots,I_l$ the equivalence classes. Put $r_i=\#
I_i$ and define $s_i$ by $\J(\b,\La^j)/\J^1(\b,\La^j)\simeq\GL_{s_i}(k_{\D_\E})$, for any $j\in
I_i$. Note also that, by conjugating the types $\l_j$ by a suitable element of $\D_\E^\times$, we may (and do) assume that $\sigma_j\simeq\sigma_k$ whenever $j\sim k$. Put $\Y^i=\bigoplus_{j\in I_i}\V^j$ and denote by $\M$ the Levi subgroup which stabilizes the decomposition $\V=\bigoplus_{i=1}^l \Y^i$. Note that this is the Levi subgroup $\M$ of the introduction.
Now $\J_\P/\J^1_\P\cong\prod_{j=1}^r \J(\b,\La^j)/\J^1(\b,\La^j)$ so we can define a representation $\sigma$ of $\J_\P$ inflated from $\bigotimes_{j=1}^r \sigma_j$. Then we put $\l_\P=\k_\P\otimes\sigma$. We put $\J_\M=\J_\P\cap \M$ and $\l_\M=\l_\P|\J_\M$.
\[prop:hom\] The pair $(\J_\P,\l_\P)$ is a cover of $(\J_\M,\l_\M)$, which is a cover of $(\J_\LL,\l_\LL)$. Moreover, we have a support-preserving Hecke algebra isomorphism $$\Hh(\M,\l_\M)\cong \Hh(\G,\l_\P)$$ and $$\Hh(\M,\l_\M)\cong\bigotimes_{i=1}^l \Hh(r_i,q_{\D_\E}^{s_i}).$$
The proof of the first assertion is identical to that of [@SS4 Proposition 5.17]; indeed the proof there shows that $\I_\G(\l_\P)\subseteq
\J_\P(\B^\times\cap\M)\J_\P$ and then the first Hecke algebra isomorphism also follows from [@BK1 7.2].
That $(\J_\M,\l_\M)$ is a cover of $(\J_\LL,\l_\LL)$ follows from Proposition \[simplecase\], as does the second Hecke algebra isomorphism.
The general case
----------------
We now treat the general case, where the ps-characters $\T_j$ are not all endo-equivalent. We define an equivalence relation $\sim$ on $\{1,\ldots,r\}$ by $$j\sim k\ \iff\ \T_j,\T_k\hbox{ are endo-equivalent},$$ and denote by $I_1,\ldots,I_l$ the equivalence classes. We put $\Y^{(i)}=\bigoplus_{j\in I_i}\V^j$. As in the homogeneous case, we assume that, for fixed $i$, every $j\in I_i$ has the same ps-character, defined relative to the same simple pair. We write $\La^{(i)}$ for an $\Oo_\D$-lattice sequence in $\Y^{(i)}$, as in the homogeneous case. By changing in their affine class, we may (and do) assume that all the $\La^{(i)}$ have the same $\Oo_\F$-period; as in the homogeneous case, we suppose also that we have replaced the lattice sequences $\La^j$ with sequences in their affine class, so that $\La^{(i)}=\bigoplus_{j\in I_i}\La^j$.
We write $\M$ for the Levi subgroup of $\G$ which stabilizes the decomposition $\V=\bigoplus_{i=1}^l \Y^{(i)}$, and $(\J_\M,\l_\M)$ for the cover of $(\J_\LL,\l_\LL)$ in $\M$, given by the homogeneous case in Proposition \[prop:hom\]. Note that this $\M$ is now *not* the Levi subgroup of the introduction, but one rather larger.
We put $\La=\bigoplus_{i=1}^l \La^{(i)}$, a (not necessarily strict) $\Oo_\D$-lattice sequence in $\V$, and $\b=\sum_{j=1}^r \b_j\in\A=\End_\D(\V)$. Then $[\La,n,0,\b]$ is a (non-simple) stratum in $\A$, for a suitable integer $n$.
For $0\le t\le n$, we write $\T_j^{(t)}$ for the ps-character defined by the character $\t_j|\H^{t+1}(\beta_j,\La^j)$.
\[thm:general\] There is a cover $(\K,\tau)$ of the type $(\J_\M,\l_\M)$ with the following properties:
1. $\U^{n+1}(\La)\subseteq \K\subseteq \U(\La)$;
2. if the ps-characters $\T_j^{(t)}$, for $1\le j\le r$, are endo-equivalent, and $([\La,n,0,\gamma],\vartheta,t)$ is a common approximation of $(\t_1,\ldots,\t_r)$, then
1. $\K$ contains and normalizes $\H^{t+1}(\gamma,\La)\cdot
\left(\H^t(\gamma,\La)\cap\M\right)$;
2. $\tau|\H^{t+1}(\gamma,\La)$ is a multiple of $\vartheta$;
3. $\tau|\H^t(\gamma,\La)\cap\LL$ is a multiple of $\t_1\otimes\cdots\otimes\t_r$;
3. there is a support-preserving isomorphism of Hecke algebras $\Hh(\M,\l_\M)\simeq\Hh(\G,\tau)$.
The proof is by induction on $r$, the case $r=1$ being empty. So let $r>1$ and suppose that $t\ge 0$ is minimal such that the ps-characters $\T_j^{(t)}$ are endo-equivalent. Let $([\La,n,0,\gamma],\vartheta,t)$ be a common approximation of $(\t_1,\ldots,\t_r)$. If $t=0$ then the Theorem is given by Proposition \[prop:hom\], so we assume $t>0$. We allow $t=n$ (that is, the ps-characters $\T_j^{(n-1)}$ are not all endo-equivalent) in which case $\vartheta$ is the trivial character of $\U^{n+1}(\La)$. We use the notation of §\[S.common\].
For $1\le j\le r$, let $c_j\in\PP_{-t}(\La^j)$ be such that $\t_j|\H^{t+1}(\beta_j,\La^j) =
\vartheta\psi_{c_j}$. By Corollary \[cor:derived\], the derived stratum $[\Ga^j_\g,t,t-1,s_\g(c_j)]$ is equivalent to a simple (or null) stratum and we write $\phi_j(X)$ for the minimum polynomial of this stratum (so that the characteristic polynomial is a power of $\phi_j(X)$). We define an equivalence relation on $\{1,\ldots,r\}$ by $$j\sim_t k\ \iff\
\T^{(t-1)}_j,\T^{(t-1)}_k\hbox{ are endo-equivalent}.$$ Note that, by Corollary \[cor:endocommon\], we have $j\sim_t k$ if and only if $\phi_j(X)=\phi_k(X)$. Let $J$ denote an equivalence class for $\sim_t$ for which the minimum polynomial is not $X$; then $J$ is a union of certain equivalence classes $I_i$ but is not the whole of $\{1,\ldots,r\}$, or else we would contradict the minimality of $t$.
Set $\Z=\bigoplus_{j\in J} \V^j$ and $\Z'=\bigoplus_{j\not\in J}\V^j$; let $\bar\M$ be the Levi subgroup which stabilizes the decomposition $\V=\Z\oplus\Z'$ and let $\bar\P=\bar\M\bar\N$ be a parabolic subgroup with Levi component $\bar\M$ and opposite $\bar\P^-=\bar\M\bar\N^-$. By the inductive hypothesis, we have a cover $(\K_{\bar\M},\tau_{\bar\M})$ of $(\K_{\M},\tau_{\M})$ satisfying the conditions of the theorem (with $\bar\M$ in place of $\G$). We define the group $\K$ by $$\K = \K_{\bar\M} \H^t(\g,\La)
\cdot (\U^1(\La)\cap\B)\Om_{q-t+1}(\g,\La)\cap\bar\N,$$ where $\Om_{q-t+1}(\g,\La)$ is the group defined in [@SS4 §2.8]. Then [@SS4 Corollaire 4.6] says that there is a unique irreducible representation $\tau$ of $\K$ such that $(\K,\tau)$ is a cover of $(\K_{\bar\M},\tau_{\bar\M})$, which has all the required properties by transitivity of covers. \[Note that, although it is assumed in [@SS4 §4] that the lattice sequence $\La$ is strict, this extra condition is never used.\]
Now Theorem \[thm:coverhecke\] of the introduction follows from Theorem \[thm:general\] and Proposition \[prop:hom\], whence the Main Theorem.
[10]{}
– [[“]{}Le “centre” de [B]{}ernstein[”]{}]{}, in *Representations of reductive groups over a local field*, Travaux en Cours, Hermann, Paris, 1984, Written by P. Deligne, p. 1–32.
– [[“]{}Description du dual admissible de [${\rm
U}(2,1)(F)$]{} par la théorie des types de [C]{}. [B]{}ushnell et [P]{}. [K]{}utzko[”]{}]{}, *Manuscripta Math.* **107** (2002), no. 2, p. 151–186.
– [[“]{}Hereditary orders and embeddings of local fields in simple algebras[”]{}]{}, *J. Algebra* **204** (1998), no. 1, p. 324–336.
[— ]{}, [[“]{}Minimal strata for [${\rm GL}(m,D)$]{}[”]{}]{}, *J. Reine Angew. Math.* **514** (1999), p. 199–236.
– [[“]{}Pure elements and intertwining classes of simple strata in local central simple algebras[”]{}]{}, *Comm. Algebra* **28** (2000), no. 11, p. 5405–5442.
– [[“]{}Smooth representations of [${\rm GL}_m(D)$]{}. [V]{}. [E]{}ndo-classes[”]{}]{}, arXiv:1004.5032v1, April 2010.
– [[“]{}Local tame lifting for [${\rm GL}(N)$]{}. [I]{}. [S]{}imple characters[”]{}]{}, *Inst. Hautes Études Sci. Publ. Math.* (1996), no. 83, p. 105–233.
– *The admissible dual of [${\rm GL}({\rm N})$]{} via compact open subgroups*, Princeton University Press, Princeton, NJ, 1993.
[— ]{}, [[“]{}The admissible dual of [${\rm SL}(N)$]{}. [I]{}[”]{}]{}, *Ann. Sci. École Norm. Sup. (4)* **26** (1993), no. 2, p. 261–280.
[— ]{}, [[“]{}The admissible dual of [${\rm SL}(N)$]{}. [II]{}[”]{}]{}, *Proc. London Math. Soc. (3)* **68** (1994), no. 2, p. 317–379.
[— ]{}, [[“]{}Simple types in [${\rm GL}(N)$]{}: computing conjugacy classes[”]{}]{}, in *Representation theory and analysis on homogeneous spaces ([N]{}ew [B]{}runswick, [NJ]{}, 1993)*, Contemp. Math., vol. 177, Amer. Math. Soc., Providence, RI, 1994, p. 107–135.
[— ]{}, [[“]{}Smooth representations of reductive $p$-adic groups: structure theory via types[”]{}]{}, *Proc. London Math. Soc. (3)* **77** (1998), no. 3, p. 582–634.
[— ]{}, [[“]{}Semisimple types in [${\rm GL}\sb n$]{}[”]{}]{}, *Compositio Math.* **119** (1999), no. 1, p. 53–97.
– [[“]{}Types in [${\rm SL}_n$]{}[”]{}]{}, *Proc. London Math. Soc. (3)* **85** (2002), no. 1, p. 119–138.
[— ]{}, [[“]{}Hecke algebras and [${\rm SL}_n$]{}-types[”]{}]{}, *Proc. London Math. Soc. (3)* **90** (2005), no. 1, p. 87–131.
– [[“]{}Simple characters for principal orders in [${\rm
M}_m(D)$]{}[”]{}]{}, *J. Number Theory* **126** (2007), no. 1, p. 1–51.
– [[“]{}Level zero types and [H]{}ecke algebras for local central simple algebras[”]{}]{}, *J. Number Theory* **91** (2001), no. 1, p. 92–125.
– [[“]{}Proof of the [D]{}eligne-[L]{}anglands conjecture for [H]{}ecke algebras[”]{}]{}, *Invent. Math.* **87** (1987), no. 1, p. 153–215.
– [[“]{}Représentations lisses de [${\rm
GL}(m,D)$]{}. [I]{}. [C]{}aractères simples[”]{}]{}, *Bull. Soc. Math. France* **132** (2004), no. 3, p. 327–396.
[— ]{}, [[“]{}Représentations lisses de [${\rm GL}(m,D)$]{}. [II]{}. [$\beta$]{}-extensions[”]{}]{}, *Compos. Math.* **141** (2005), no. 6, p. 1531–1550.
[— ]{}, [[“]{}Représentations lisses de [${\rm GL}_m(D)$]{}. [III]{}. [T]{}ypes simples[”]{}]{}, *Ann. Sci. École Norm. Sup. (4)* **38** (2005), no. 6, p. 951–977.
[— ]{}, [[“]{}Proof of the [T]{}adić conjecture ([U]{}0) on the unitary dual of [${\rm GL}_m(D)$]{}[”]{}]{}, *J. Reine Angew. Math.* **626** (2009), p. 187–203.
– [[“]{} Représentations lisses de [${\rm GL}_m(D)$]{}. [IV]{}. [R]{}eprésentations supercuspidales[”]{}]{}, *J. Inst. Math. Jussieu* **7** (2008), no. 3, p. 527–574.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'For systems in an externally controllable time-dependent potential, the optimal protocol minimizes the mean work spent in a finite-time transition between two given equilibrium states. For overdamped dynamics which ignores inertia effects, the optimal protocol has been found to involve jumps of the control parameter at the beginning and end of the process. Including the inertia term, we show that this feature not only persists but that even delta peak-like changes of the control parameter at both boundaries make the process optimal. These results are obtained by analyzing two simple paradigmatic cases: First, a Brownian particle dragged by a harmonic optical trap through a viscous fluid and, second, a Brownian particle subject to an optical trap with time-dependent stiffness. These insights could be used to improve free energy calculations via either thermodynamic integration or “fast growth” methods using Jarzynski’s equality.'
author:
- 'Alex Gomez-Marin$^{1,2}$, Tim Schmiedl$^{2}$, and Udo Seifert$^{2}$'
title: Optimal protocols for minimal work processes in underdamped stochastic thermodynamics
---
Introduction
============
The free energy difference $\Delta F$ between two equilibrium states is an important quantity in isothermal statistical mechanics. Strategies to extract $\Delta F$ from experiments or computer simulations are traditionally based on either thermodynamic integration or thermodynamic perturbation [@zwan54] which use one infinitesimally slow transition or many infinitesimally fast transitions, respectively, between the two equilibrium states. A decade ago, Jarzynski proposed the remarkable relation $$e^{-\Delta F / T} = {\left \langle e^{- W / T} \right \rangle}$$ which interpolates between these extreme cases using nonequilibrium work values $W$ obtained from trajectories of finite time transitions between the equilibrium states at temperature $T$ (with Boltzmann’s constant $k_B = 1$ throughout the paper) [@jarz97; @jarz97a]. This exact relation which holds for any time-dependent driving described by an external control parameter $\lambda(t)$ has been extended to various fluctuation theorems [@croo99; @croo00; @humm01; @hata01; @seif05a]. Although these (necessarily irreversible) finite time transitions occur in nonequilibrium, the equilibrium quantity $\Delta F$ can be inferred from a sufficient number of trajectories either from computer simulations [@hend01; @humm01a; @shir03; @park04; @mara06] or real experiments [@liph02; @coll05]. However, the convergence of the involved exponential average causes problems for far out of equilibirium transitions where the work $W$ is substantially larger than the free energy difference $\Delta F$ [@gore03]. In this regime, the exponential average is dominated by low work values which are very rarely sampled [@jarz06]. As a remedy, several path sampling techniques biasing the dynamics for low work have been proposed [@ytre04; @atil04; @lech06]. It is, however, still under debate [@ober05; @ytre06; @lech07] for which systems fast growth techniques are superior to refined “conventional” approaches such as umbrella sampling [@torr77] or flat histogram methods [@wang01]. Though valuable for computer simulations, it is hard to imagine how to bias dynamics in real experiments, where, however, apparatus drift may prevent long measurements necessary for thermodynamic integration [@coll05; @mara07] and thus render fast growth methods competitive.
Both for thermodynamic integration and “fast growth” methods employing Jarzynski’s equality (or some variant), efficiency gains can be achieved by optimizing the driving scheme $\lambda(t)$. For thermodynamic integration, where the work $W \geq \Delta F$ is taken as an estimator for $\Delta F$, it is obvious that a minimal work gives the best result [@koni05]. In the case of fast growth methods, the statistics for free energy estimates quite generally also improves with smaller mean work [@atil04; @jarz06]. Fast growth simulations even allow to combine data from different driving schemes in a straightforward way [@minh06].
The minimization of the work spent in a finite time process can, however, also be seen in the context of minimal energy dissipation. On a macroscopic scale, the optimization of the work (or power) exerted in a macroscopic cyclic process has been discussed for quite a while [@curz75; @band82; @andr84; @hoff03]. On a microscopic scale, fluctuations will also affect optimal cyclic processes [@schm08] which may become relevant for constructing optimal nano machines.
For overdamped Langevin dynamics, the optimal protocol leading to a minimal mean work in a finite time $t$ has been calculated analytically for harmonic potentials [@schm07]. Surprisingly, the optimal protocol shows jumps at the beginning and end of the finite time transition. Since most molecular dynamics (MD) simulations of the dynamics of biomolecules are on time-scales where inertia plays an important role (see [@soto07] for a review on steered MD), it is an interesting question how these results transfer to underdamped dynamics. In particular, it is important to know whether the jumps are a result of having neglected inertia.
In this paper, we investigate the role of inertia for two previously introduced paradigmatic processes. In Sect. \[case1\], we calculate optimal driving schemes for an underdamped Brownian particle dragged through a viscous fluid by harmonic optical tweezers. In Sect. \[case2\], we study an underdamped Brownian particle subject to an optical trap with time-dependent stiffness. In both cases, we compare our findings with the corresponding results in the overdamped limit [@schm07]. We find that the optimal protocol still involves jumps. Even more surprisingly, the optimal protocol includes delta peaks at the beginning and end of the process.
Case Study I: The moving trap {#case1}
=============================
Optimal protocol
----------------
We consider a Brownian particle of mass $m$ dragged through a viscous fluid with friction coefficient $\gamma$ by a harmonic potential $$V(x,t)=\frac{k}{2}(x-\lambda(t))^2,$$ where $k$ is the (constant) trap stiffness. The focus of the optical trap $\lambda(t)$ is changed time-dependently from an initial position $\lambda_i = 0$ to a final position $\lambda_f$. Including inertial effects, the Langevin equation reads $$\label{langm}
m\ddot{x}=-\gamma \dot{x} -k(x-\lambda(t)) + \eta(t),$$ where thermal fluctuations are modeled by Gaussian white noise $$\langle \eta(t) \eta(t') \rangle = {2 T} {\gamma} \delta(t-t').$$ The mean work spent in the process of total duration $t_f$ is given by $$W \equiv \int_{0}^{t_f}dt \left\langle \frac{\partial V(x,t)}{\partial t} \right\rangle,$$ where the average $\langle \dots \rangle$ is over the intitial thermal distribution and over the noise history. In the present case the total (mean) work reduces to $$\label{work0}
W = \int_{0}^{t_f}dt k\dot{\lambda}(\lambda - u ) = k \int_0^{tf} \lambda \dot u + \frac k 2 \lambda_f^2 - k\left [ \lambda u \right]_0^{t_f} ,$$ where, for simplicity in the notation, we have defined the mean position of the particle as $u\equiv \langle x \rangle$. This quantity $u(t)$ depends on the whole history of $\lambda(t)$ and thus, the work $W$ is a non-local functional of the protocol $\lambda(t)$. However, in analogy to the overdamped limit [@schm07], we can express the work as a local functional of the mean particle position $u$. By averaging the evolution equation (\[langm\]) we have $$\label{l1}
\lambda=u +\gamma \dot{u}/k+m\ddot{u}/k,
\label{lam_u}$$ which inserted in Eq. (\[work0\]) leads to $$\begin{aligned}
W =\left[ \frac{m}{2}\dot{u}^2 +\frac{m^2}{2k}\ddot{u}^2 +\frac{m\gamma}{k}\dot{u}\ddot{u} +\frac{\gamma^2}{2k}\dot{u}^2 \right]_{0}^{t_f}
+\gamma \int_{0}^{t_f}dt \dot{u}^2 .\end{aligned}$$
The only term remaining in the integral, $\dot{u}^2$, is identical to the one in the overdamped limit, while the boundary terms are different. In complete analogy to the overdamped case, we now proceed in two steps. First, we calculate the optimal shape $u(t)$ minimizing only the integral given initial values $u(0^+) = 0$ and $\dot u(0^+) = A$. Note that despite the initial equilibrium value $\dot u(0) = 0$, we are free to choose $\dot u(0^+) = A$ since the necessary “kink” in $u(t)$ at $t = 0$ does not contribute to the integral. Similarly, at the end of the protocol (at $t=t_f$) there can be another jump in the velocity. In a second step, we adjust the constant $A$ to yield the minimal total work. First, from the Euler-Lagrange equation corresponding to the Lagrangian $\dot{u}^2 $ (and subject to the initial conditions just mentioned), we find $$u(t)=At
\label{u}$$ for $0< t < t_f$. In contrast to the overdamped case, we cannot determine all the boundary terms at $t_f$ from the evolution equation. Thus, $C \equiv \dot{u}(t_f) $ is another free parameter. With $\ddot{u}(t_f)=[k(\lambda_f-At_f)-\gamma C ]/m$, we get the total work as a function of the yet unknown constants $A$ and $C$ $$\begin{aligned}
W(A,C) =\frac{m}{2}C^2+ \frac{k}{2} (\lambda_f-At_f)^2 +\gamma \int_{0}^{t_f}dt A^2.\end{aligned}$$ The work is clearly minimal for $C^*=0$, where the asterisk will denote optimal from now on. The remaining terms then read $$\begin{aligned}
W(A) = \frac{k}{2} (\lambda_f-At_f)^2 +\gamma t_f A^2 ,\end{aligned}$$ which, surprisingly, is exactly the same expression that was found in the overdamped limit. Minimizing this expression with respect to $A$ leads to $$\begin{aligned}
A^*=\frac{\lambda_f}{2\gamma/k + t_f}\end{aligned}$$ which yields the work $$\begin{aligned}
\label{W*}
W^*=k\lambda_f^2 \frac{1}{2+ {k t_f} / {\gamma}}.\end{aligned}$$ Inserting Eq. (\[u\]) into Eq. (\[lam\_u\]), we find the optimal protocol $$\lambda^*(t)=\lambda_f \frac{kt/\gamma+1}{kt_f/\gamma+2},
\label{mov_lam*}$$ for $0 < t < t_f$ implying symmetrical jumps $$\Delta \lambda \equiv \lambda(0^+)-\lambda_i
=\lambda_f-\lambda(t_f^-)=
\lambda_f \frac{1}{kt_f/\gamma+2}$$ at the beginning and at the end of the process.
Superficially, this optimal protocol looks like the expression in the overdamped case [@schm07]. There is, however, a subtle difference arising from the presence of inertia terms. The optimal protocol forces the mean velocity to instantly jump at the beginning of the process from its initial equilibrium value $\dot{u}(0)=0$ to $\dot{u}(0^+)=A^*$. At the end of the protocol, the optimal strategy consists in setting back the mean velocity to zero $\dot{u}(t_f)\equiv C^*=0$. Due to the second time derivative in the equation of motion such jumps in the velocity, which require delta functions in the acceleration, imply a delta-type singularity in the protocol. Specifically, in Eq. (\[l1\]), the jumps in $\dot u$ imply a $\delta$-function for $\ddot{u}$ and hence a $\delta$ function in $\lambda(t)$. The optimal protocol \[Eq. (\[mov\_lam\*\])\] thus becomes $$\lambda^*(t) = \lambda_f \frac{kt/\gamma+1}{kt_f/\gamma+2} + {\frac{m \lambda_f}{2\gamma + k t_f}} \left [ \delta(t) - \delta(t-t_f) \right ],$$ as shown in Fig. \[fig1\]. In the overdamped limit, $m \to 0$, the delta peaks vanish.
Physical origin of singularities in the optimal protocol
--------------------------------------------------------
The benefit of having jumps in the optimal protocol can be understood intuitively as follows. From the perspective of minimal dissipation, it is obvious that the particle should be dragged at a constant (mean) velocity from the beginning rather than being accelerated during a finite time. This initial jump in the velocity of the particle can only be achieved by a finite initial difference $\lambda(0) - u(0)$, corresponding to a jump in $\lambda$ at $t = 0$. In the present underdamped regime, a velocity jump corresponds to a jump in the (mean) particle momentum which can only be achieved by a delta peak in the force, corresponding to a delta peak in the protocol.
The final jump is harder to grasp intuitively. In fact, it stems from focussing on the minimal work rather than on the minimal (mean) dissipation (or entropy production). If we had searched for the minimal entropy production (as defined in [@seif05a]), we would have found an optimal protocol without a final jump. In the present minimization, at the final time $t_f$, the particle is still in non-equilibrium with respect to the final potential $V(x, \lambda(t_f))$. Relaxation to equilibrium leads to further dissipation [*after*]{} time $t_f$ which has, however, already been paid for by the total work since at constant $\lambda$ no work is exerted anymore. A smaller final particle position $u(t_f)$ leads to a longer relaxation time which can decrease the total dissipation of the combined process (nonequilibrium transition and relaxation).
The final delta peak corresponds to setting the final velocity to zero. This decreases the kinetic energy of the particle and thus is beneficial for a small work. It also explains the surprising fact that, according to Eq. (\[W\*\]), we do not have to pay any extra cost for having inertia. During the initial singularity, the exerted work is stored in the (mean) kinetic energy of the particle. This contribution is fully recovered during the final singularity where the kinetic energy of the particle is set back to the equilibrium value.
Comparison to a linear protocol
-------------------------------
Without prior knowledge, one might have expected a continuous linear protocol $$\lambda^{\rm lin}(t) \equiv \lambda_f t / t_f$$ to yield the lowest work. In the overdamped limit, the work for a linear protocol was at most $14 \%$ larger than for the optimal protocol. We now check how much smaller the value of the optimal work $W^*$ is compared to a linear protocol if we include inertia. First, we rescale the system in order to compactly write the relevant combination of parameters. With the rescaled mass $\tilde{m}\equiv mk/\gamma^2$, the energy scale $e\equiv k\lambda_f^2$ and a rescaled time $\tilde{t}\equiv t_fk/\gamma$, the work can be written as $W=e\tilde{W}(\tilde{t},\tilde{m})$, with the optimal work $W^*=e / \left( 2+\tilde{t}\right)$.
Solving the second order differential equation of motion (\[lam\_u\]) using the linear protocol $\lambda^{\rm lin}(t)$, we find the ratio: $$\frac{W^{\rm{lin}}}{W^*}
= \left \lbrace \begin{array}{l} \frac {2+\tilde{t}} {\tilde{t}^2} \left ( \theta_0+\tilde{t}
-e^{- \frac {\tilde{t}} {2\tilde{m}}} \left [ \theta_0 \cosh \left ( \nu \tilde{t} \right ) + \theta_1 \sinh(\nu \tilde{t}) \right ] \right ) ~~~\tilde m < \frac 1 4 \\ \frac {2+\tilde{t}} {\tilde{t}^2} \left (
\theta_0 + \tilde{t}
- e^{- \frac {\tilde{t}} {2\tilde{m}}} \left [ \theta_0 ~\cos \left ( \nu \tilde{t} \right ) ~+~ \theta_1~ \sin(\nu \tilde{t}) \right ] \right ) ~~~\tilde m > \frac 1 4 \end{array} \right.$$ with $$\begin{aligned}
\nu = \frac {\sqrt{\left | 4\tilde{m} -1 \right | }} {2\tilde{m}} $$ and $$\begin{aligned}
\theta_0=\tilde{m} - 1
,\;\; \theta_1=\frac{3\tilde{m}-1}{2\tilde{m}\nu} .$$ In Fig. \[fig2\], we plot the ratio ${W^{\rm{lin}}} / {W^*}$ as a function of rescaled time $\tilde{t}$ and mass $\tilde{m}$. This result shows that the optimal protocol significantly reduces the work spent in the process compared to a linear protocol.
Case Study II: The Stiffening trap {#case2}
==================================
In the first case study, only the averaged quantity $u = {\left \langle x \right \rangle}$ appeared in the work and thus the same result could have been obtained from a deterministic damped dynamics. We next examine a second case study where fluctuations are important. We consider a Brownian particle of mass $m$ in an optical trap with time dependent stiffness $\lambda(t)$ which is driven from an initial value $\lambda(0) = \lambda_i$ to a final value $\lambda(t_f) = \lambda_f$ in a finite time $t_f$. The time dependent potential $$V(x,t)=\frac{\lambda(t)}{2}x^2,$$ leads to the underdamped Langevin equations $$\begin{aligned}
\dot{x} & = & p/m \\ \nonumber
\dot{p} & = & - \gamma p/m -\lambda(t)x + \eta(t),\end{aligned}$$ where $p$ is the momentum of the particle and the noise $\eta(t)$ has the same properties introduced in the first case study. Again our main goal is to find the protocol for which the corresponding total (mean) work $$\label{W1}
W = \int_{0}^{t_f}dt \dot{\lambda} \frac{ \langle x^2 \rangle }{2}$$ is minimal. Note that the mean squared position $$w \equiv \langle x^2 \rangle$$ of the particle is non-trivially coupled to the mean squared momentum $$z \equiv \langle p^2 \rangle$$ and to the position-momentum correlation $$y\equiv\langle xp \rangle .$$ Their time evolution is governed by the set of coupled differential equations $$\begin{aligned}
\dot{w} & = &2y/m \label{wdot} , \\
\dot{z} &= &-2\lambda y -2 \gamma z/m +2\gamma T \label{zdot} ,
\\
\dot{y} & = & z/m-\lambda w - \gamma y/m. \label{ydot}\end{aligned}$$
Unlike both the moving trap (with and without inertia) and the stiffening trap in the overdamped limit, the present case is much more involved since one cannot eliminate the protocol and write the work as a function of one variable only. We thus express the work as a time-local functional of $w(t)$ and $z(t)$. Solving Eqs. (\[wdot\]) and (\[ydot\]) for $\lambda$ yields $$\begin{aligned}
\label{Lstiff}
\lambda=\frac{1}{w}[z/m-\gamma\dot{w}/2-m\ddot{w}/2]\end{aligned}$$ which, inserted in Eq. (\[W1\]) and after partial integration, leads to $$\label{W20}
W =
\left[ \frac{\lambda w}{2}+\frac{m\dot{w}^2}{8w} \right]_{0}^{t_f} + \frac{1}{2}\int_{0}^{t_f}dt \mathcal{L}
\label{W_stiff}$$ with the “Lagrangian” $$\label{Lagrangian}
\mathcal{L}= \frac{ \gamma \dot{w}^2}{2w}-\frac{z\dot{w}}{mw}+
\frac{m\dot{w}^3}{4w^2}.$$
We proceed in two steps analogously to the moving trap. We first minimize the integral in Eq. (\[W\_stiff\]) for given initial conditions and then optimize with respect to remaining free parameters. The integrand $\mathcal{L}$ depends on $w$ (and $\dot{w}$) but also on $z$. The variables $w$ and $z$ are not independent. Eliminating $\lambda$ and $y$ from the equations of motion (\[wdot\]), (\[zdot\]), and (\[ydot\]), we find the physical constraint $$\label{constr}
\mathcal{G}\equiv z\dot{w}-m\gamma\dot{w}^2 /2-m^2\dot{w}\ddot{w}/2-2\gamma Tw +2\gamma wz/m+w\dot{z}=0.$$ A detailed analysis of the solution of this optimization problem using Euler-Lagrange equations is given in Appendix \[app1\].
The results for both the rescaled optimal protocol $\lambda^* ( t / t_f) / \lambda_i$ and the optimal work $W^* / T$ depend on the dimensionless quantities $$\tilde{t}\equiv t_f \lambda_i /\gamma ,~~ \tilde{\lambda}\equiv \lambda_f /\lambda_i, ~~\tilde{m}\equiv \lambda_i m /\gamma^2.$$ An extensive analysis of the optimal protocol as a function of all three parameters is out of scope. Since the overdamped limit ($\tilde m \to 0$) has been discussed previously, we focus on the behaviour as a function of $\tilde m$ for given $\tilde \lambda = 2$, $\tilde t = 1$. Given these parameters, the optimization problem can be solved numerically and the corresponding total work $W$ can be calculated. In Fig. \[lam\*\]a, we plot the value of the minimal work $W^*$ (obtained from the optimal protocol) as a function of the rescaled dimensionless mass $\tilde{m}$ and compare it to other benchmark protocols. All work values are bounded from below by the free energy difference $\Delta F = \left ( \ln 2 / 2 \right ) T \simeq 0.35 T$. Quite generally, work values are also bounded from above by the work for an immediate jump $W^{\rm{jp}} \equiv \lim_{t\to 0} W^* = T / 2$. We study (i) a linear protocol, (ii) a protocol leading to a parabolic mean-squared position $$w(t) = \frac T {\lambda_i} (1 + c t)^2
\label{w_harm}$$ with optimized parameter $c$ and optimized final delta peak, and (iii) a protocol leading to $$w(t) = 1 + a t^3 \left ( 1 - e^{-1 / \left [ 0.01+(5 t/t_f)^2 \right ] } \right ) + b t^5 + c t^7 + d t^9
\label{w_poly}$$ without any discontinuities (except for a final jump) but with free parameters $a, b, c, d$. The work arising from protocol (i) lies significantly above the optimal protocol. Protocol (ii) implies (optimized) jumps and delta peaks at the beginning and end. The work for protocol (ii) and the optimal work almost coincide. The inset shows that the optimal work is in fact slightly smaller than the work obtained for the protocol (ii). The difference to the numerically obtained exact solution $W^*$ decreases for decreasing $\tilde m$ which is consistent with the analytical finding that protocol (ii) is optimal in the overdamped limit. The fact that protocol (ii) which involves optimized singularities but not the optimized shape is so close to the optimal work highlights the importance of jumps for the optimal protocol. Protocol (iii) has no delta peaks and no initial jump but mimics these features approximately since the parameters $a, b, c, d$ have been optimized, see Fig. \[lam\*\]b. These trial protocols show that jumps and delta peak-like singularities can decrease the total work and confirms that our numerical solution of the Euler-Lagrange equations is the solution of the optimization problem.
Finally, the explicit shape of the optimal protocol $\lambda^*(t)$ can be reconstructed numerically from Eq. (\[Lstiff\]), see Fig. \[lam\*\]c. It displays initially a delta peak upwards accompanied with a jump $\Delta \lambda$ and, finally, a delta peak downwards together with another jump $\Delta \lambda '$. Such discontinuities in the protocol are a consequence of the discontinuities in $z$, $\dot{w}$ and $\ddot{w}$. The first singularity is “needed” to suddenly increase $\langle p^2\rangle$ from its equilibrium value and also to change the derivative of $\langle x^2\rangle$, which is proportional to the correlation $\langle xp \rangle$. Note that both size and direction of the jumps strongly depend on the rescaled mass $\tilde m$ as shown in Fig. \[lam\*\]d. For small $\tilde m$, the protocol jumps upwards (as also observed in the overdamped regime [@schm07]) while for large $\tilde m$, the protocol jumps downwards.
Concluding Discussion
=====================
In summary, we have calculated optimal protocols yielding the minimal mean work for underdamped Langevin dynamics in two different model potentials. Surprisingly, these optimal protocols involve jumps and delta peaks at the initial and final times $t_i=0$ and $t_f$. While we have shown that the singularities in the optimal protocol appear for harmonic potential, there is no reason to believe that this feature generically vanishes for anharmonic potentials. In fact, in the overdamped limit, a recent study has shown that initial and final jumps are also present in a simple anharmonic potential [@then08]. At first sight, such singularities seem to be unphysical since neither jumps nor delta peaks can be implemented in real experiments. Still our theoretical result is an important insight because it implies that there exists no optimal continuous protocol. Every such protocol could be improved by even steeper gradients mimicking the jumps and delta peaks at the beginning and end. If there was an experimental constraint on the allowed maximum rate of change in $\lambda$, $|\dot \lambda| < r$, the minimal work would still be achieved by a protocol which looks roughly like the optimal one, with the jumps and delta peaks replaced by their best approximation consistent with $|\dot \lambda| < r$ (*e. g.* steep straight lines instead of jumps). Thus, it should be possible to exploit our results for real experiments.
Our results may also be used in steered MD simulations. Even though we here have calculated optimal protocols for underdamped Langevin dynamics, there is no reason to believe that other thermostats frequently used in MD simulations would yield qualitatively different results for the optimal protocol. We have neglected memory effects by assuming white noise. While there are systems for which the underdamped Langevin equation is an appropriate physical description [@doua06], it might still be interesting to see how our results are altered when considering memory effects.
The optimal protocol for a minimal mean work is not strictly equivalent to a protocol leading to an optimized free energy estimate. However, it has been found that the latter shares the same features (jumps at the boundaries) for overdamped Langevin dynamics [@dell08]. Moreover, the optimal protocol leads to improved estimates of the free energy difference in both of our (underdamped) case studies. For case study I, the work distribution is Gaussian [@mazo99; @tani08] for which it has been shown that the error in the estimate of the free energy difference decreases with decreasing mean work [@gore03]. In our second case study, for which the work distribution is no longer Gaussian, the error in the estimate of the free energy difference is indeed lower for the optimal protocol compared to a linear protocol, see Fig. \[fig\_dF\] and its caption for technical details. For thermodynamic integration, it is obvious that a minimum mean work leads to the best estimate of the free energy difference. We thus conjecture that appropriate singularities at the boundaries generically improve free energy calculations from either fast growth methods or thermodynamic integration.
For determining the optimal protocol for an unknown potential, we envisage an adaptive procedure in which trial protocols (including estimated singularities) are successively improved in an iterative fashion guided by the monitored work values. It might also prove beneficial to use the optimal moving trap protocol (case study I) rather than a linear protocol in simulations of (protein) pulling experiments.
Solution of the optimization problem in case study II {#app1}
=====================================================
In this appendix, we give a detailed analysis of the numerical solution of the optimization problem. In order to minimize the integral in Eq. (\[W20\]), the constraint \[Eq. (\[constr\])\] is included in an effective Lagrangian $\mathcal{L}_{\rm{eff}}\equiv \mathcal{L} - \alpha(t)\mathcal{G}$ through a Lagrange multiplier $\alpha(t)$. Then the Euler-Lagrange equations whose solutions minimize the integral in Eq. (\[W20\]) are obtained from $$\label{EL_gen}
\frac{\partial \mathcal{L}_{\rm{eff}}}{\partial w}+
\frac{d^2}{dt^2}\frac{\partial \mathcal{L}_{\rm{eff}}}{\partial \ddot{w}}=
\frac{d}{dt}\frac{\partial \mathcal{L}_{\rm{eff}}}{\partial \dot{w}}
, \;\;\;\;
\frac{\partial \mathcal{L}_{\rm{eff}}}{\partial z}=\frac{d}{dt}\frac{\partial \mathcal{L}_{\rm{eff}}}{\partial \dot{z}},$$ which, together with the constraint $\mathcal{G} = 0$, define a system of three differential equations for $w$, $z$ and $\alpha$. By defining the useful new variable $$\label{def_mu}
\mu \equiv zw-\frac{m^2}{4}\dot{w}^2$$ we can write the initially cumbersome differential equations (\[EL\_gen\]) after a tedious manipulation in the following reduced form $$\begin{aligned}
\ddot{w} & = & \frac{\dot{w}^2}{2w} - \frac{2}{m^2}\frac{\mu}{w} +2Tw\alpha +\frac{2T}{m} \label{ELv} , \\
\dot{\mu} & = & - \frac{2\gamma}{m}\mu+2\gamma T w , \\
\dot{\alpha} & = & \frac{2\gamma}{m}\alpha+ \frac{1}{m}\frac{\dot{w}}{w^2}.\end{aligned}$$ These equations have no analytical solution but they can easily be solved numerically for given initial conditions $w(0^+)$, $\dot{w}(0^+)$, $\mu(0^+)$ and $\alpha(0^+)$. It is important to note that some of these initial conditions are not fixed by the initial equilibrium conditions $w(0) = T/\lambda_i$, $\dot w (0) = 0$, $\ddot w(0) = 0$, $z(0) = m T$, but can be realized by additional discontinuities in the respective quantities at the boundaries. If such discontinuities do not change the value of the integral in Eq. (\[W\_stiff\]), they do not affect the optimization of the integral via the Euler-Lagrange equations and hence the respective initial conditions should (in a first step) be treated as free parameters. Since the Langrangian does not depend on $\ddot w(t)$, discontinuities in $\dot w(t)$ and $\ddot w(t)$ can occur at the boundaries. However, a jump in the mean squared position $w(t)$ would affect the integral in Eq. (\[W\_stiff\]) and thus $w(t)$ must be chosen to be continuous at the boundaries, enforcing $w(0^+)=w(0)\equiv w_0=T/\lambda_i$. Likewise, discontinuities in $z(t)$ can occur at the boundaries. However, the initial values $z(0^+)$ and $\dot w(0^+)$ are related by the constraint $\mathcal{G}=0$. Integrating this constraint $$\lim_{\epsilon \to 0} \int_{t-\epsilon}^{t +\epsilon} dt' \mathcal{G}=0$$ leads to $$\label{magicrel}
\left[ wz \right]^{t^+}_{t^-}=\frac{m^2}{4}\left[\dot{w}^2\right]^{t^+}_{t^-}.$$ When applied at $t=0$ it yields $$\label{magicrel0}
\frac{T}{\lambda_i}[z(0^+)-mT]=\frac{m^2}{4}\dot{w}^2(0^+).$$ We consider a (possible) discontinuity through the parameter $s_1$ in $$\label{z0+}
z(0^+)\equiv mT s_1.$$ With Eq. (\[magicrel0\]), the jump in the derivative of $w$ at the initial time as a function of $s_1$ becomes $$\dot{w}(0^+)=\pm 2T \sqrt{\frac{s_1-1}{m\lambda_i}}.$$ In the case in which $\lambda_i<\lambda_f$, the correct sign is the negative one. Note that the last equation implies $s_1>1$, so that at the initial time and given the equilibrium initial distribution, it is not possible to have a decrease in the mean squared momentum. From Eqs. (\[magicrel0\]) and (\[def\_mu\]) we also find $$\mu(0^+)= mT^2 /\lambda_i.$$
Secondly, we define a new free parameter $s_2$ in $$\dot{y}(0^+)\equiv T s_2,$$ which, from the evolution equation (\[ydot\]), directly yields $\ddot{w}(0^+)= \frac{2T}{m} s_2$. Then, writing Eq. (\[ELv\]) at $t=0^+$ and inserting the above values, the initial value of the Lagrange multiplier needed to solve the Euler-Lagrange equations is $$\alpha(0^+)=\frac{\lambda_i}{mT}(s_2-s_1+1).$$ Last, from the evolution equations (\[wdot\])-(\[ydot\]) we find the relative value of the initial jump in the protocol as a function of $s_1$ and $s_2$: $$\frac{\Delta \lambda_i}{\lambda_i}\equiv
\frac{\lambda(0^+)-\lambda_i}{\lambda_i}=s_1-s_2-1+ \gamma \sqrt{\frac{s_1-1}{m\lambda_i}}.$$
At the end of the process, the value of $z(t_f)$ is allowed to jump again. Recalling Eq. (\[magicrel\]) applied now at the final time $t=t_f$ and isolating $z(t_f)$, we obtain $$z(t_f)=z(t_f^-)+\frac{m^2}{4}\frac{\dot{w}(t_f)^2-\dot{w}(t_f^-)^2}{w(t_f)}.$$ Every quantity on the right hand side of the last equation except for $\dot{w}(t_f)$ is fixed by the solution of the Euler-Lagrange equation. The minimum value for $z(t_f)$, which leads to the minimal contribution to the work in Eq. (\[W2\]), is reached for $\dot{w}(t_f)\equiv s_3=0$.
For a comparison of the present case with its overdamped analogue [@schm07], one can formally integrate the differential equations for $\mu$ and $\alpha$ and plug them into Eq. (\[ELv\]) to obtain the following integro-differential equation for $w$, $$\label{intdif}
\left( \ddot{w}-\frac{\dot{w}^2}{2w}\right) = \frac{2T} {m} \left [
f(t)\mathcal{A}-\frac{\mathcal{B}}{f(t)}+f(t)(1+s_2-s_1) \right ]$$ where $f(t)\equiv \frac{w(t)}{w_0}e^{2\gamma t/m}$ and $$\mathcal{A}=1-\frac{2\gamma}{m} \int_0^{t} \frac{1}{f(t')} dt',
\;\;\;\;\;
\mathcal{B}=1+\frac{2\gamma}{m} \int_0^{t} f(t') dt'.$$ In the overdamped limit, the Euler-Lagrange equation is given by $\ddot{w}-\dot{w}^2 / 2w=0$. Including inertia leads to nonvanishing terms on the right hand side of Eq. (\[intdif\]). However, taking the corresponding overdamped limit $\tilde m \to 0$ in Eq. (\[intdif\]) yields the overdamped Euler-Lagrange equation only after optimizing the parameters $s_1$ and $s_2$.
Combining Eqs. (\[wdot\])-(\[ydot\]), the work $W$ \[Eq. (\[W1\])\] can be written as $$\begin{aligned}
\label{W2}
W =
\left[ \frac{\lambda w}{2} +\frac{z}{2m} \right]_{0}^{t_f} -\frac{\gamma T}{m}t_f + \frac{\gamma}{m^2}\int_0^{t_f}dt z.\end{aligned}$$ To calculate the integral, we insert the solution of the Euler-Lagrange equations for $z$, which depends on $s_1$ and $s_2$. Then, we need to insert the boundary values for $w$ and $z$ at $t=0$ and $t=t_f$. In a last step, the work is optimized with respect to the free parameters $s_1$ and $s_2$.
[40]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , , , ****, ().
, , , , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, .
, , ().
, ****, ().
Figure captions {#figure-captions .unnumbered}
---------------
Figure \[fig1\] : Scheme of the optimal mean position $u^*(t)$ and protocol $\lambda^*(t)$. Figure \[fig2\] : Ratio between mean work $W^{\rm{lin}}$ spent using the continuous linear protocol $\lambda^{\rm{lin}}(t)$ and optimal work $W^*$ as a function of the dimensionless parameters $\tilde{m}\equiv m k/\gamma^2$ and $\tilde{t} \equiv t_f k/\gamma$. Figure \[lam\*\] : Optimization results for case study II for $\tilde{t} = 1$ and $\tilde {\lambda} = 2$. (a) Mean work $W^*$ in units of $T$ as a function of the rescaled mass $\tilde{m}$ compared to (i) a linear protocol, (ii) a protocol leading to $w(t)$ given by Eq. (\[w\_harm\]), and (iii) a continuous (except for a final jump) protocol leading to $w(t)$ given by Eq. (\[w\_poly\]) with adjusted parameters to yield a minimal work (see main text for details). (b) Protocol (iii) with optimized parameters for $\tilde m = 1$. (c) Optimal protocol $\lambda^*(t)$ for $\tilde{m} = 0.5$. (d) Jump heights $\Delta \lambda$ and amplitudes $\mathcal{D}$ of delta peaks (in rescaled time $t / t_f$) for the optimal protocol as a function of the rescaled mass $\tilde m$. Figure \[fig\_dF\] : Comparison of free energy estimates for case study II for a linear protocol and a continous approximation to the optimal protocol for $\tilde t = 1, \tilde m = 1, \tilde \lambda = 2$. The data were obtained from Langevin simulation of $10^6$ trajectories for each protocol with $\gamma = 1, m = 1, T=1, \lambda_i = 1$. (a) Distribution $P(W)$ of work values $W$ for the two protocols shown in the inset: (i) the linear protocol $\lambda(t) = 1+t$ and (ii) the linear protocol with additional continuously approximated delta singularities. The columns show the free energy difference $\Delta F \simeq 0.3466$ and the mean work values ${\left \langle W^{(i)} \right \rangle} \simeq 0.4700 (\pm 0.0006)$, ${\left \langle W^{(ii)} \right \rangle} \simeq 0.4270 (\pm 0.0005)$ which both are consistent with a direct evaluation based on Eq. (\[W1\]). (b) Histogram of $10^5$ Jarzynski estimates for the free energy difference $\Delta F^{\rm{est}} \equiv - (1 / \beta) \ln \left [\sum_{i=1}^N \exp (-\beta W_i) / N \right ]$ obtained from $N = 10$ single trajectory work values $W_i$ each. The mean squared error (MSE) of these estimates consists of two parts [@gore03]: the systematic error (bias) $B = {\left \langle \Delta F^{\rm{est}} \right \rangle} - \Delta F$ and the statistical error $\sigma = \sqrt{\rm{Var} (\Delta F^{\rm{est}} )}$. The columns show the free energy difference and the mean value of the estimates obtained from the two protocols. Since the bias ($B^{(i)} = 0.0066 (\pm 0.0004)$, $B^{(ii)} = 0.0052 (\pm 0.0003)$) can be neglected for both protocols, the MSE is dominated by the statistical error ($\sigma^{(i)} = 0.119$, $\sigma^{(ii)} = 0.104$) which is smaller for protocol (ii).
$\;\;\;\;\;\;\;\;\;\;$
1
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We describe a fabrication process for devices with few quantum bits (qubits), which are suitable for proof-of-principle demonstrations of silicon-based quantum computation. The devices follow the Kane proposal to use the nuclear spins of $^{31}$P donors in $^{28}$Si as qubits, controlled by metal surface gates and measured using single electron transistors (SETs). The accurate registration of $^{31}$P donors to control gates and read-out SETs is achieved through the use of a self-aligned process which incorporates electron beam patterning, ion implantation and triple-angle shadow-mask metal evaporation.'
address:
- Centre for Quantum Computer Technology
- '1) School of Physics, University of New South Wales, Sydney 2052, AUSTRALIA'
- '2) School of Electrical Engineering, University of New South Wales, Sydney 2052, AUSTRALIA'
author:
- |
T.M. Buehler$^{1}$, R.P. McKinnon$^{1}$, N.E. Lumpkin$^{1}$, R. Brenner$^{1}$,\
D.J. Reilly$^{1}$, L.D. Macks$^{1}$, A.R. Hamilton$^{1}$, A.S. Dzurak$^{2}$ and R.G. Clark$^{1}$
title: 'Self-aligned fabrication process for silicon quantum computer devices'
---
{width="7.5cm"}
Large-scale quantum computers [@a1] promise a massive increase in the speed of important computational tasks such as prime factorisation [@a2] and database searching [@a3]. One of the most promising schemes, which is capable of being scaled up to processors with many quantum bits (qubits), is the silicon-based solid-state quantum computer (SSQC), first proposed by Kane [@a4]. Such devices require the positioning of single dopant atoms to better than 10 nm: a highly challenging task. However the paradigm has a major advantage over competing proposals in that the fabrication strategy is compatible with industry-standard silicon metal-oxide-semiconductor (MOS) processing technology.
A silicon-based SSQC is depicted schematically in Figure 1. The nuclear spin (I) of single $^{31}$P donor atoms (I=1/2) embedded in $^{28}$Si (I=0) constitutes the qubit. ’A’-gates above the donors enable single-qubit operations by controlling the hyperfine interaction between electron and nucleus. ’J’-gates between donors enable two-qubit operations by controlling the exchange interaction between adjacent donors [@a4]. During the qubit read-out process, nuclear spin information is transferred to electron spin states as described by Kane [@a4]. Under the influence of an asymmetric gate bias, the Pauli exclusion principle dictates that electron motion between two P donors (to form a two-electron bound state - see Fig. 1) will only take place if the electrons are in a spin-singlet state. Qubit read-out is achieved using a charge-sensitive single electron transistor (SET) to detect whether or not this charge transfer occurs, thereby allowing determination of the two-electron spin state [@a4]. The device must be operated at temperatures of 1 K, or below, in order to minimise qubit decoherence and to ensure that all electron spins are in their ground state.
Realization of the structure shown in Fig. 1 requires not only the ability to form ordered arrays of $^{31}$P atoms in $^{28}$Si, but also a capability to accurately align these atoms to their control gates and read-out SETs. These requirements remain relevant to variations on the original Kane design, including those utilizing either nuclear-spin qubits [@a5] or electron-spin qubits [@a6]. Recently, it has been demonstrated [@a7] that P qubit arrays on Si can be fabricated using an atom-structuring approach in which a scanned probe is used to pattern a monohydride resist on a silicon surface. While this approach can be extended to produce high-precision P donor arrays on a large scale, a number of challenging steps, including subsequent epitaxial overgrowth and gate registration, have yet to be demonstrated.
This paper describes an approach to SSQC fabrication which has two key advantages over other schemes: (i) all of its steps are based upon existing semiconductor processes; and (ii) it employs a self-alignment technique for registering single donors to control gates and read-out SETs. The scheme described here is designed for few-qubit devices which will be critical for determining the feasibility of a SSQC. It is also possible to envisage modifications which would allow scale-up of the process for many-qubit array fabrication. Here we demonstrate the fabrication of an integrated array of control gates and SETs, and present a detailed strategy for self-aligning these structures to individual P atoms.
{width="7.5cm"}
For the purposes of this paper, we consider a six-donor, four-qubit device as shown in Figure 2(a). We first outline the conceptual approach for the process, then focus on a particular implementation employing poly-methylmethacrylate (PMMA), poly-methacrylic acid (PMAA) and polymethylglutarimide (PMGI) resists.
Surface metallisations for the device are defined using multi-angle shadow evaporation, which is already a well-established technique for fabricating Al/Al$_{2}$O$_{3}$ SETs [@a8]. A single high-resolution electron beam lithography (EBL) step is used to create the black pattern (Pattern A) shown in Fig. 2(b). Following resist development, triple-angle shadow evaporation [@a9] and oxidation are performed to produce Al and Al/Al$_{2}$O$_{3}$ structures which constitute all of the gates and SETs shown in Fig. 2(a). Although some extraneous features are produced by the shadow evaporation process, these will not significantly affect the performance of the device. The resist structure used for the shadow evaporation step (Resist A) must be able to produce a wide cavity beneath fine features in a surface layer [@a9] and would generally be comprised of two or more resist layers of differing sensitivity, as discussed later.
A linear array of self-aligned donors can be realized by incorporating an additional EBL step using a different resist (Resist B). The locations of the six donors in the device may be defined using the grey pattern (Pattern B) shown in Fig 2(b). The points at which Patterns A and B cross define the donor sites. Access channels for ion implantation, which extend to the substrate at these donor sites, may be opened by either wet development or dry etching. A critical requirement is that the two resist structures utilize different developer solutions or different dry etch processes.
{width="7.5cm"}
Partial etching or development of Resist A, along with complete etching or development of Resist B produces a profile such as that shown in Figure 3(a) for ion implantation. Ions must be implanted at low energy (of order keV) to ensure donors are located 5-20 nm below the Si/SiO$_{2}$ interface and with an areal dose such that on average one ion is implanted at each donor site. By Poissonian statistics, the probability that a single ion lands in each of N sites is given [@a6] by 0.367$^{N}$. As N increases, the device yield would fall to low levels without additional strategies aimed at controlling ion placement. One such strategy involves the use of on-chip ion impact detectors which electrically register the arrival of each ion in the device, during the implant process [@a10]. This approach ensures that exactly N donors will be implanted in an intended N-donor device. Recent results using this technique demonstrate the detection of single $^{31}$P+ ions of energy 15 keV entering an i-Si substrate [@a10]. This ion energy leads to an average implant depth of 20 nm, suitable for a SSQC device. In order to scale-up this process for the construction of many-qubit processors it will be necessary to localise the ion beam to a spot size smaller than the spacing between donor sites and move the beam between each site until a single ion implant event is registered. Such localisation could be achieved using a focused ion beam (FIB) or a moveable mask with a nanomachined aperture [@a11].
Although ion implantation is a convenient technique for introducing donors into Si, there are two aspects of the process which could impair the reliable operation of a SSQC, namely straggle of P donors from their desired locations and damage to the Si lattice. The effect of straggle has been estimated using SRIM software (Stopping and Range of Ions in Matter) and for typical donor depths of 10 nm we find a lateral straggle of order 5 nm [@a12]. These values suggest that it will be necessary to individually tune the gate voltages that control each donor qubit, to take into account its specific location. Lateral straggle can be improved somewhat by channelling the implant along a crystal axis, however, the amorphous SiO$_{2}$ barrier layer will serve to randomise the incident angle and reduce channelling.
There is limited information on damage induced by $single$ implanted donors in Si, although molecular dynamics calculations [@a13] indicate that considerable damage may be automatically removed by self-annealing in the first picosecond after impact. Residual damage can be removed by annealing after the metallisation of the gates and removal of the resist layers. In the worst case, it may be necessary to anneal at temperatures as high as 900°C in order to remove impact damage and activate the P donor electrons. Given that SET tunnel barriers are likely to be degraded above 400°C during long anneal steps, it may be necessary to use a rapid thermal anneal (RTA) process for these devices. Alternatively, SETs could be fabricated in a separate EBL step after annealing [@a10]. Although the alignment of SETs to donors would be compromised in this case, read-out is still expected to be successful due to the demonstrated high sensitivity of SETs. An advantage of post-fabrication of the SETs is that refractory metals could be used for the gates. Some refractory metals, such as tungsten, have high melting points and also exhibit low diffusivity through SiO$_{2}$ barrier layers.
Following ion implantation, complete development of Resist A produces a cavity suitable for multi-angle shadow evaporation (see Fig. 3b). Any remaining Resist B on the floor of the cavity is also removed at this point. The triple-angle shadow evaporation process depicted in Fig. 3(b) produces self-aligned gate and SET structures with A-gates located directly above implanted donors, thus realizing the complete device shown in Fig. 2(a).
We now describe a particular example of the fabrication strategy outlined above which employs only wet chemical processing. Firstly, PMGI (Resist B) is prepared on a Si/SiO$_{2}$ substrate. Pattern B is exposed using EBL and developed with an aqueous developer. Next, a bilayer (Resist A) of PMMA and PMMA/PMAA copolymer is prepared on top of the PMGI and Pattern A is exposed. Partial development of the bilayer with an organic solvent (MIBK/IPA) is then carried out to create narrow trenches down to the PMGI. This developer solution does not affect the PMGI. Fig. 3(c) shows the profile of a partially developed bilayer cavity. Where the two EBL patterns overlap, channels extending to the substrate surface are created, through which ion implantation may be performed.
As a trial of this process, we have exposed and developed a series of lines in bilayer resist, on top of and perpendicular to a series of lines exposed and developed in PMGI (Fig. 4a). The resulting structure was etched using an HF solution and then the resist layers removed. Atomic force microscope (AFM) imaging (Figs. 4b and 4c) confirmed that etch pits in the Si/SiO$_{2}$ substrate were formed only where the two sets of lines overlapped.
{width="7.5cm"}
Following ion implantation, the bilayer is fully developed and any PMGI remaining in the cavity is removed with a solvent which does not affect the bilayer. Fig. 3(d) shows the profile of a typical bilayer cavity. Triple-angle Al shadow evaporation is then performed with in-situ oxidation between the first and second angle evaporations to form the Al/Al$_{2}$O$_{3}$ tunnel junctions required for SET operation. Removal of any remaining resist then leaves the structure shown in Fig. 2(a) on the substrate surface.
{width="7.5cm"}
The triple-angle evaporation process described above has been demonstrated using Pattern A on a bilayer resist, producing the device shown in Figure 5(a). Fig. 5(b) shows a side-view schematic of the same device indicating the various metal layers deposited during the three evaporation steps. The intended donor sites are indicated although ion implantation was not carried out on this trial device. The 50 nm-wide lines in Fig. 5(a) are the A-gates, which lie directly above the intended donor sites. The J-gates on this device are slightly wider than the A-gates due to imperfect overlaying of metal lines. The gate spacing of this device is sufficiently small to fabricate an electron-spin SSQC [@a6], however, it will be necessary to reduce metal linewidths to 5-10 nm to realize a nuclear-spin-based device [@a4; @a5]. Such linewidths have been realized using single layer EBL resists [@a14] and work is underway to achieve this using our multilayer process.
{width="7.5cm"}
We are currently integrating ion implantation into the process to produce a fully configured few-qubit SSQC device. In the first instance, a two-donor device (Fig. 6a) will be implanted. Fig. 6(b) shows the EBL patterns required to produce such a device. Fig. 6(c) is an SEM image of a demonstration metallisation of the gate and SET array formed by triple-angle shadow evaporation using Pattern A. The same device is shown on a larger scale in Fig. 6(d) revealing connections leading out to macroscopic bond pads.
{width="7.5cm"}
Figure 7(a) shows a detailed SEM image of a typical SET fabricated by double-angle shadow evaporation on a bilayer resist. Fig. 7(b) shows the clear Coulomb blockade oscillations measured in such devices. The high transconductance $(dG/dV_{g})$ demonstrates the sensitivity of these devices to single charge transfer events, and thus their suitability for use as qubit read-out devices in the Kane scheme. Recently, we have used two such devices in a $twin$-SET configuration to measure controlled electron transfer between two metallic dots, joined by a tunnel barrier [@a15].
Although here we have focused on a particular process implementation using PMGI, PMMA and PMMA/PMAA, there are a number of possible variations. To obtain the highest possible aspect ratios for the ion implantation channels, dry etching could be used in place of the wet development procedures discussed above. For example, instead of a standard bilayer structure, Resist A could be composed of a Ge layer sandwiched between two organic resists [@a16]. Pattern A could be exposed in the top resist layer and developed, then reactive ion etching with CF$_{4}$ gas used to remove the exposed Ge [@a16]. Next, another gas could be used to anisotropically etch the underlying organic resist, thereby defining vertical-walled ion implantation channels. In another variation, the order of Resists A and B could be reversed. After ion implantation, Resist B would be removed, allowing for complete development of Resist A, followed by triple-angle shadow evaporation.
In conclusion, we have designed a process involving two EBL exposures, ion implantation and triple-angle shadow evaporation capable of producing self-aligned gate, SET and donor structures for a few-qubit SSQC. The process can utilize a variety of different resists, development and etch procedures. We have made preliminary test structures comprising gates and read-out SETs and have demonstrated that cavities for ion implantation may be formed in an organic trilayer resist structure. Work on incorporating ion implantation into our process is currently underway.
Acknowledgement {#acknowledgement .unnumbered}
===============
This work was supported by the Australian Research Council, the Australian Government and by the U.S. National Security Agency (NSA), Advanced Research and Development Activity (ARDA) and the Army Research Office (ARO) under contract number DAAD19-01-1-0653.
[99]{} Steane A [*Rep. Prog. Phys.*]{} [**61**]{} 117 (1998)
Shor P W [*Proc. 35th Ann. Symp. on Foundations of Computer Science, ed. S. Goldwasser (IEEE Computer Society, Los Alamitos CA)*]{} 124 (1994)
Grover L K [*Phys. Rev. Lett.*]{} [**79**]{} 325 (1997)
Kane B E [*Nature*]{} [**393**]{} 133 (1998)
Kane B E [*Fortschr. Phys.*]{} [**48**]{} 1023 (2000); [*see also*]{} quant-ph/0003031
Vrijen R, Yablonovitch E, Wang K, Jiang H W, Balandin A, Roychowdhury V, Mor T and DiVincenzo D [*Phys. Rev. A*]{} [**62**]{} 012306 (2000)
O’Brien J L, Schofield S R, Simmons M Y, Clark R G, Dzurak A S, Curson N J, Kane B E, McAlpine N S, Hawley M E and Brown G W [*Phys. Rev. B*]{} [**64**]{} 161401 (2001)
Fulton T A and Dolan G J [*Phys. Rev. Lett.*]{} [**59**]{} 109 (1987)
Dolan G J [*Appl. Phys. Lett.*]{} [**31**]{} 337 (1977)
McKinnon R P, Stanley F E, Macks L D, Mitic M, Gauja E, Peceros K, Buehler T M, Dzurak A S, Clark R G, Yang C, Jamieson D N, Prawer S, Pakes C I and McCallum J C, [*to be submitted to Appl. Phys. Lett.*]{}; Yang C, Jamieson D N, Hearne S M, Pakes C I, Prawer S D, Stanley F E, Macks L D M, Gauja E, Dzurak A S and Clark R G, [*in preparation*]{}
Jamieson D N, Prawer S, Andrienko I, Brett D A, Millar V, Yang C and Cimmino A, [*Proc. Int. Conf. on Experimental Implementation of Quantum Computation, ed. R.G. Clark and A.R. Hamilton*]{} (2001)
Jamieson D N [*(private communication)*]{}
Marks N A, Hee Lee K and McKenzie D R [*Proc. Int. Conf. on Experimental Implementation of Quantum Computation, ed. R.G. Clark and A.R. Hamilton*]{} (2001)
Chen W and Ahmed H [*Appl. Phys. Lett.*]{} [**62**]{} 1499 (1993)
Buehler T B, Brenner R, Reilly D J, Hamilton A R, Dzurak A S and Clark R G, accepted for [*Smart Materials and Structures*]{} (2002)
Nakamura Y, Klein D L and Tsai J S [*Appl. Phys. Lett.*]{} [**68**]{} 275 (1996)
|
{
"pile_set_name": "ArXiv"
}
|
---
bibliography:
- 'main.bib'
title:
- |
Learning a Behavior Model of Hybrid Systems Through Combining Model-Based Testing and Machine Learning\
(Full Version)[^1]
- ' Learning a Behavior Model of Hybrid Systems Through Combining Model-Based Testing and Machine Learning '
---
[ ]{}
#### Acknowledgment. {#acknowledgment. .unnumbered}
This work is supported by the TU Graz LEAD project “Dependable Internet of Things in Adverse Environments”. It is also partially supported by ECSEL Joint Undertaking under Grant No.: 692455.
[^1]: This work is an extended preprint of the conference paper “Learning a Behavior Model of Hybrid Systems Through Combining Model-Based Testing and Machine Learning” accepted for presentation at IFIP-ICTSS 2019, the 31^st^ International Conference on Testing Software and Systems in Paris, France.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Two-loop $\beta$-function and anomalous dimension are calculated for $N=1$ supersymmetric quantum electrodynamics, regularized by higher derivatives in the minimal subtraction scheme. The result for two-loop contribution to the $\beta$-function appears to be equal to 0, does not depend on the form of regularizing term and does not lead to anomaly puzzle. Two-loop anomalous dimension can be also made independent on parameters of higher derivative regularization by a special choice of subtraction scheme.'
author:
- 'A.Soloshenko[^1] and K.Stepanyantz [^2]'
title: 'Two-loop renormalization of $N=1$ supersymmetric electrodynamics, regularized by higher derivatives.'
---
6.1in
8.70in .03 in .03in
.03in =-12pt
[*Moscow State University, physical faculty,\
department of theoretical physics.\
$117234$, Moscow, Russia*]{}
Introduction.
=============
Investigation of quantum corrections in supersymmetric theories is a very important and complicated problem. In principle, supersymmetric theories have better ultraviolet behavior, than nonsupersymmetric models. For example, in $N=2$ Yang-Mills theory perturbative divergences are present only in one-loop diagrams. In principle it follows from the fact, that in supersymmetric theories axial anomaly and anomaly of energy-momentum tensor trace belong to the same supermultiplet [@Ferrara; @Clark; @Piquet1; @Piquet2]. The axial anomaly is known to be completely defined by the one-loop approximation [@Bardeen; @Slavnov_Book], while the trace anomaly is proportional to $\beta$-function [@Adler_Collins]. Therefore, due to the supersymmetry the $\beta$-function should be also defined by the one-loop approximation. The same arguments can be applied to $N=1$ supersymmetric theories. However, explicit calculations of radiative corrections show, that the $\beta$-function in $N=1$ supersymmetric models has contributions from higher loops [@Tarasov; @Grisaru; @Caswell; @Avdeev]. This contradiction is usually called “anomaly puzzle” and was investigated in a large number of papers, for example [@12; @13; @14; @15; @16; @18; @19; @20; @21; @22; @NSVZ; @Arkani].
Usually different proposals to solve anomaly puzzle require to fix the form of the $\beta$-function in all orders of perturbation theory. For example, in theories with matter the $\beta$-function should be related with the anomalous dimension. For the first time such $\beta$-function was obtained by Novikov, Shifman, Vainshtein and Zakharov (NSVZ) from investigation of structure of instanton corrections [@NSVZ_Instanton]. Later this result was checked by explicit calculations, which were usually made by the dimensional reduction technique [@Siegel]. Two-loop $\beta$-function, obtained in this regularization, is shown to coincide with a prediction, following from NSVZ exact expression. However, three-loop $\beta$-function [@ThreeLoop1; @ThreeLoop2; @ThreeLoop3] does not agree with it. Nevertheless, the deviations can be removed by a redefinition of the coupling constant [@Jones], the possibility of such redefinition being highly nontrivial [@JackJones]. In principle it is possible to relate $\overline{\mbox{DR}}$ scheme and NSVZ scheme order by order [@North] in the perturbation theory.
It is necessary to especially mention paper [@Arkani], in which $\beta$-function is shown to depend on the normalization of matter and gauge superfields. In particular, NSVZ $\beta$-function can be obtained after a special rescaling of these superfields, which reduces kinetic terms to the canonical normalization. Otherwise, (without rescalings) $\beta$-function is argued to be completely defined by the one-loop approximation. So, it is really quite possible to obtain zero contributions of higher loops. The problem is how to calculate the corrections. In principle, it is possible to look for the regularization or renormalization scheme, in which $\beta$-function is equal to the one-loop result or coincides with NSVZ exact $\beta$-function. For example, the calculation of super Yang-Mills two-loop $\beta$-function in the differential renormalization [@DiffR] was made in [@Mas]. Another interesting possibility is using of higher covariant derivative regularization [@Slavnov; @Bakeyev]. For the supersymmetric Yang-Mills theory the Lagrangian of the regularized theory was constructed in [@West_Paper]. For electrodynamics construction of the regularized Lagrangian is simpler, because instead of covariant derivatives it is necessary to use usual derivatives. However, calculation of diagrams, regularized by higher covariant derivatives is rather complicated. In particular, explicit calculation of the one-loop quantum correction [^3] for the (nonsupersymmetric) Yang-Mills theory was made rather recently [@Martin; @Asorey] and gives the same result as the dimensional regularization. In principle it is possible to prove, that one-loop calculations using higher covariant derivative regularization (certainly, complemented by the additional regularization for one-loop diagrams) always give the same result as the dimensional regularization [@PhysLett]. Investigation of two-loop corrections in theories, regularized by higher derivatives has not yet been done.
In this paper we try to understand features of higher derivative regularization in supersymmetric theories and calculate two-loop renormgroup functions for massless $N=1$ supersymmetric quantum electrodynamics in this regularization using minimal subtraction scheme.
The paper is organized as follows:
In Section \[Section\_SUSY\_QED\] we introduce notations and remind some information about $N=1$ supersymmetric electrodynamics. In the next Section \[Section\_Covariant\_Derivatives\] the considered model is regularized by adding of higher derivative term. After it we describe the quantization procedure for the constructed theory. Two-loop $\beta$-function and anomalous dimension are calculated in Section \[Section\_Two\_Loop\]. Agreement of the results with renormgroup equations is checked in Section \[Section\_Renormgroup\]. A brief summary and discussion are presented in Conclusion. Technical details of calculations, including expressions for all Feinman diagrams, can be found in the Appendix.
Supersymmetric quantum electrodynamics.
=======================================
\[Section\_SUSY\_QED\]
$N=1$ supersymmetric massless electrodynamics in the superspace is described by the following action: [^4]
$$\label{SQED_Action}
S_0 = \frac{1}{4 e^2} \mbox{Re}\int d^4x\,d^2\theta\,W_a C^{ab} W_b
+ \frac{1}{4}\int d^4x\, d^4\theta\,
\Big(\phi^* e^{2V}\phi +\tilde\phi^*
e^{-2V}\tilde\phi\Big).$$
Here $\phi$ and $\tilde\phi$ are chiral superfields, which in components can be written as
$$\begin{aligned}
\label{Phi_Superfield}
&& \phi(y,\theta) = \varphi(y) + \bar\theta (1+\gamma_5) \psi(y)
+ \frac{1}{2}\bar\theta (1+\gamma_5)\theta f(y);\nonumber\\
&&\tilde \phi(y,\theta) = \tilde \varphi(y)
+ \bar\theta (1+\gamma_5) \tilde \psi(y)
+ \frac{1}{2}\bar\theta (1+\gamma_5)\theta \tilde f(y),\end{aligned}$$
where $y^\mu = x^\mu + i\bar\theta\gamma^\mu\gamma_5\theta/2$ are chiral coordinates, $\varphi$ and $\tilde\varphi$ are complex scalar fields, $\psi$ and $\tilde\psi$ are Maiorana spinors, which can be unified in a Dirac spinor
$$\Psi = \frac{1}{\sqrt{2}}\Big((1+\gamma_5)\psi+(1-\gamma_5)\tilde\psi\Big),$$
and $f$ and $\tilde f$ are auxiliary complex scalar fields.
$V$ is a real abelian superfield, which is a supersymmetric generalization of the gauge field. In components this superfield can be written as
$$\begin{aligned}
\label{V_Superfield}
&& V(x,\theta) = C(x)+i\sqrt{2}\bar\theta\gamma_5\xi(x)
+\frac{1}{2}(\bar\theta\theta)K(x)
+\frac{i}{2}(\bar\theta\gamma_5\theta)H(x)
+\frac{1}{2}(\bar\theta \gamma^\mu \gamma_5\theta) A_\mu(x)
+\nonumber\\
&& + \sqrt{2} (\bar\theta\theta) \bar\theta
\Big(i\gamma_5\chi(x)
+\frac{1}{2}\gamma^\mu\gamma_5\partial_\mu\xi(x)\Big)
+ \frac{1}{4} (\bar\theta\theta)^2 \Big(D(x)
-\frac{1}{2}\partial^2 C(x)\Big).\end{aligned}$$
The chiral superfield $W_a$ is a supersymmetric generalization of the field strength tensor and in the abelian case is defined as
$$\label{W_Superfield}
W_a = \frac{1}{16} \bar D (1-\gamma_5) D\Big[(1+\gamma_5)D_a V\Big],$$
where the supersymmetric covariant derivative $D$ is written as
$$D = \frac{\partial}{\partial\bar\theta} - i\gamma^\mu\theta\,\partial_\mu.$$
Model (\[SQED\_Action\]) is invariant under supersymmetric gauge transformations
$$\label{Gauge_Transformations}
V \to V - \frac{1}{2}(A+A^+);
\qquad \phi\to e^{A}\phi;\qquad \tilde\phi\to e^{-A} \tilde\phi,$$
where $A$ is an arbitrary chiral superfield. In principle, it is possible to choose Wess-Zumino gauge, in which the superfield $V$ is written as
$$\begin{aligned}
V(x,\theta) =
\frac{1}{2}(\bar\theta \gamma^\mu \gamma_5\theta) A_\mu(x)
+ i\sqrt{2} (\bar\theta\theta) \bar\theta\gamma_5\chi(x)
+ \frac{1}{4} (\bar\theta\theta)^2 D(x).\end{aligned}$$
However, this gauge is not supersymmetric. That is why we do not use it for calculation of quantum corrections.
Higher derivative regularization of $N=1$ supersymmetric electrodynamics.
=========================================================================
\[Section\_Covariant\_Derivatives\]
To regularize model (\[SQED\_Action\]) by higher derivatives let us first modify its action by the following way:
$$\begin{aligned}
\label{Regularized_SQED_Action}
&& S_0 \to S = S_0 + S_{\Lambda}
=\vphantom{\frac{1}{2}}\nonumber\\
&&\qquad
= \frac{1}{4 e^2} \mbox{Re}\int d^4x\,d^2\theta\,W_a C^{ab}
\Big(1 + \frac{\partial^{2n}}{\Lambda^{2n}}\Big) W_b
+\nonumber\\
&&\qquad\qquad\qquad\qquad\qquad\qquad
+ \frac{1}{4}\int d^4x\, d^4\theta\,
\Big(\phi^* e^{2V}\phi +\tilde\phi^* e^{-2V}\tilde\phi\Big).\qquad\end{aligned}$$
Note, that the considered model is abelian and the superfield $W^a$ is gauge invariant. Therefore, a regularizing term should contain usual derivatives instead of the covariant ones.
It is convenient to introduce operators
$$\begin{aligned}
&& \bar D^2 \equiv \frac{1}{2} \bar D (1-\gamma_5) D;\qquad
D^2 \equiv \frac{1}{2} \bar D (1+\gamma_5) D;\nonumber\\
&&\qquad
\Pi_{1/2} \equiv - \frac{1}{16\partial^2} D_a \Big(C(1+\gamma_5)\Big)^{ab}
\bar D^2 D_b,\end{aligned}$$
which satisfy identities
$$\begin{aligned}
\label{Identity1}
&& D^2 \bar D^2 + \bar D^2 D^2
= -16\Pi_{1/2}\partial^2 - 16\partial^2;\\
\label{Identity2}
&& \frac{1}{2} \bar D \gamma^{\mu} \gamma_5 D\,
\bar D \gamma^{\nu} \gamma_5 D
+ \frac{1}{2} \bar D \gamma^{\nu} \gamma_5 D\,
\bar D \gamma^{\mu} \gamma_5 D
= - 16\eta^{\mu\nu}\Pi_{1/2} \partial^2 - 16 \partial^\mu \partial^\nu.
\qquad\end{aligned}$$
Then the first term in action (\[Regularized\_SQED\_Action\]) can be presented in the following form:
$$\begin{aligned}
&& S_{\mbox{\footnotesize gauge}}\equiv
\frac{1}{4e^2}\mbox{Re}\int d^4x\, d^2\theta\,
W_a C^{ab} \Big(1 + \frac{\partial^{2n}}{\Lambda^{2n}}\Big)W_b
=\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
= - \frac{1}{4e^2} \int d^4x\, d^4\theta\, V \Pi_{1/2}\partial^2
\Big(1 + \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V.\qquad\quad\end{aligned}$$
The gauge invariance (\[Gauge\_Transformations\]) can be fixed by addition of the terms
$$S_{\mbox{\footnotesize gf}} = - \frac{1}{64 e^2}\int d^4x\,d^4\theta\,
\Bigg(V D^2 \bar D^2
\Big(1 + \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
+ V \bar D^2 D^2
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V\Bigg),$$
which are invariant under supersymmetry transformations. Then due to identity (\[Identity1\]) the kinetic term for the gauge field is written in the most simple form:
$$S_{\mbox{\footnotesize gauge}}
+ S_{\mbox{\footnotesize gf}} = \frac{1}{4 e^2}\int d^4x\,d^4\theta
V\partial^2 \Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V.$$
Due to the gauge invariance (\[Gauge\_Transformations\]) the renormalized action of the considered model (without gauge fixing term) can be presented as
$$\begin{aligned}
\label{Two_Loop_S_Ren}
&& S_{\mbox{\footnotesize ren}} =
\frac{1}{4 e^2} Z_3(\Lambda/\mu)\,
\mbox{Re}\int d^4x\,d^2\theta\,W_a C^{ab}
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) W_b
+\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad
+ \frac{1}{4} Z(\Lambda/\mu) \int d^4x\, d^4\theta\,
\Big(\phi^* e^{2V}\phi +\tilde\phi^* e^{-2V}\tilde\phi\Big),\qquad\end{aligned}$$
where $e=e(\Lambda/\mu)$ is a renormalized coupling constant. A bare coupling constant $e_0$ is defined by the equation
$$\frac{1}{e_0^2} = \frac{1}{e^2} Z_3(\Lambda/\mu)$$
and does not depend on $\Lambda/\mu$. The $\beta$-function and anomalous dimension in our notations are defined as
$$\beta = \frac{d}{d\ln\mu}\Bigg(\frac{e^2}{4\pi}\Bigg);
\qquad\quad
\gamma = \frac{d\ln Z}{d\ln\mu}.$$
At the first sight, the generating functional can be written in the following form:
$$\begin{aligned}
\label{Z}
&& Z = \int DV\,D\phi\,D\tilde \phi\,
\exp\Bigg\{i\Bigg[\frac{1}{4 e^2} \int d^4x\,d^4\theta\, V\partial^2
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
-\nonumber\\
&& \qquad\qquad\qquad\qquad
- \frac{1}{4 e^2} \Big(Z_3(\Lambda/\mu)-1\Big)
\int d^4x\,d^4\theta\, V \Pi_{1/2}\partial^2
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
+\nonumber\\
&& + \frac{1}{4} Z(\Lambda/\mu) \int d^4x\,d^4\theta\,
\Big(\phi^* e^{2V}\phi
+ \tilde\phi^* e^{-2V}\tilde\phi \Big)
+ \int d^4x\,d^4\theta\,J V
+\nonumber\\
&& \qquad\qquad\qquad\qquad
+ \int d^4x\,d^2\theta\, \Big(j\,\phi + \tilde j\,\tilde\phi \Big)
+ \int d^4x\,d^2\bar\theta\,
\Big(j^*\phi^* + \tilde j^* \tilde\phi^* \Big)\Bigg]\Bigg\}.\qquad\quad\end{aligned}$$
(Below we slightly modify this expression.) Note, that the considered case corresponds to the gauge group $U(1)$ and, therefore, diagrams with ghost loops are absent.
Taking into account, that
$$\label{Chiral_Identity}
\bar D^2 D^2 \phi = - 16\partial^2\phi$$
for any chiral superfield $\phi$ and that
$$\int d^4x\,d^2\theta = -\frac{1}{2} \int d^4x\,D^2,$$
generating functional (\[Z\]) can be presented as
$$\begin{aligned}
&& Z = \int DV\,D\phi\,D\tilde \phi\,
\exp\Bigg\{i \int d^4x\,d^4\theta\,
\Bigg(\frac{1}{4 e^2} V \partial^2
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
-\nonumber\\
&& - \frac{1}{4 e^2} \Big(Z_3(\Lambda/\mu)-1\Big)\,
V \Pi_{1/2}\partial^2
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
+ \frac{1}{4} Z(\Lambda/\mu)\Big(\phi^* e^{2V}\phi
+ \tilde\phi^* e^{-2V}\tilde\phi\Big)
+\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
+ J V + \phi \frac{D^2}{8\partial^2} j
+ \tilde\phi \frac{D^2}{8\partial^2}\tilde j
+ \phi^* \frac{\bar D^2}{8\partial^2} j^*
+ \tilde\phi^* \frac{\bar D^2}{8\partial^2} \tilde j^*\Bigg) \Bigg\}.
\qquad\nonumber\\\end{aligned}$$
In order to calculate this functional we should present the argument of the exponent as a sum of a part $S_Q$, quadratic in fields, and interaction $S_I$, where
$$\begin{aligned}
\label{Quadratic}
&& S_Q = \frac{1}{4 e^2} \int d^4x\,d^4\theta\,\Bigg(V\partial^2
\Big(1 + \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
+ \frac{1}{4} \Big(\phi^* \phi +\tilde\phi^* \tilde\phi\Big)
+\nonumber\\
&& \qquad\qquad\qquad\qquad
+ J V + \phi \frac{D^2}{8\partial^2} j
+ \tilde\phi \frac{D^2}{8\partial^2}\tilde j
+ \phi^* \frac{\bar D^2}{8\partial^2} j^*
+ \tilde\phi^* \frac{\bar D^2}{8\partial^2} \tilde j^* \Bigg);
\nonumber\\
\label{Interaction}
&& S_I = \sum\limits_{n=1}^\infty \frac{1}{4 n!}\int d^4x\,d^4\theta\,
\Big(\phi^* (2V)^{n} \phi + \tilde\phi^* (-2V)^n\tilde\phi\Big)
-\nonumber\\
&& \qquad\qquad
- \Big(Z_3(\Lambda/\mu)-1\Big) \frac{1}{4 e^2}\int d^4x\,d^4\theta\,
V\Pi_{1/2} \partial^2 \Big(1+\frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
+\nonumber\\
&& \qquad\qquad\qquad\qquad
+ \Big(Z(\Lambda/\mu) - 1\Big) \frac{1}{4}\int d^4x\,d^4\theta\,
\Big(\phi^* e^{2V} \phi + \tilde\phi^* e^{-2V}\tilde\phi\Big).\qquad\quad
\vphantom{\Bigg(}\end{aligned}$$
(Last two terms of $S_I$ will generate vertexes with insersions of counterterms, obtained in previous orders of perturbation theory.)
Then the generating functional can be written as
$$\begin{aligned}
\label{Z_With_SI}
&& Z = \exp\Bigg\{i S_I\Bigg(\frac{1}{i}\frac{\delta}{\delta J},\,
\frac{1}{i}\frac{\delta}{\delta j},\ldots\Bigg)\Bigg\}
\int DV\,D\phi\,D\tilde \phi\,\exp\Big(i S_Q\Big).\end{aligned}$$
Note, that the differentiation over chiral superfields in our notations is defined as follows:
$$\frac{\delta j(\theta_x,x)}{\delta j(\theta_y,y)} = - \frac{\bar D^2}{2}
\delta^4(\theta_x-\theta_y)\,\delta^4(x-y),$$
so that
$$\frac{\delta}{\delta j(\theta_y,y)}
\int d^4x\,d^2\theta_x\,j(x,\theta_x)\,\phi(x,\theta_x)
= \phi(y,\theta_y).$$
The integral, remaining in equation (\[Z\_With\_SI\]), is Gaussian and can be easily calculated:
$$\begin{aligned}
\label{Perturbative_Z }
&& Z = \exp\Bigg\{i S_I\Bigg(\frac{1}{i}\frac{\delta}{\delta J},\,
\frac{1}{i}\frac{\delta}{\delta j},\ldots\Bigg)\Bigg\}
\times\nonumber\\
&& \qquad\qquad\quad
\times
\exp\Bigg\{ i \int d^4x\,d^4\theta\,\Bigg(
j^* \frac{1}{\partial^2} j
+ \tilde j^* \frac{1}{\partial^2} \tilde j
- J \frac{e^2}{\partial^2
\Big(1+\partial^{2n}/\Lambda^{2n}\Big)} J\Bigg)\Bigg\}\qquad\end{aligned}$$
Expansion of this expression in powers of $J$, $j$ and $\tilde j$ gives a series of perturbation theory.
However, introducing of higher derivative term does not eliminate all divergences. Really, in Appendix \[Appendix\_Degree\_Of\_Divergence\] we check, that the superficial degree of divergence for model (\[Regularized\_SQED\_Action\]) is equal to
$$\label{Degree_Of_Divergence}
\omega_\Lambda = 2 - 2n (L-1) - E_\phi (n+1),$$
where $L$ is a number of loops and $E_\phi$ is a number of external $\phi$-lines. Note, that $\omega_\Lambda$ does not depend on a number of external $V$-lines $E_V$. Therefore even after introducing of the higher derivative term with $n\ge 2$, divergences are present in one-loop diagrams. In order to regularize them [@Slavnov_Book] it is necessary to insert Pauli-Villars determinants in generating functional (\[Z\]), so that
$$\begin{aligned}
\label{Modified_Z}
&& Z = \int DV\,D\phi\,D\tilde \phi\,
\prod\limits_i \Big(\det PV(V,M_i)\Big)^{c_i}
\exp\Bigg\{i\Bigg[\frac{1}{4 e^2} \int d^4x\,d^4\theta\, V\partial^2
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
-\nonumber\\
&& - \frac{1}{4 e^2} \Big(Z_3(\Lambda/\mu)-1\Big) \int d^4x\,d^4\theta\,
V \Pi_{1/2}\partial^2
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) V
+\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
+ \frac{1}{4} Z(\Lambda/\mu) \int d^4x\,d^4\theta\,
\Big(\phi^* e^{2V}\phi
+ \tilde\phi^* e^{-2V}\tilde\phi \Big)
+\nonumber\\
&&
+ \int d^4x\,d^4\theta\,J V
+ \int d^4x\,d^2\theta\, \Big(j\,\phi + \tilde j\,\tilde\phi \Big)
+ \int d^4x\,d^2\bar\theta\,
\Big(j^*\phi^* + \tilde j^* \tilde\phi^* \Big)\Bigg]\Bigg\},\end{aligned}$$
where
$$\begin{aligned}
&& \Big(\det PV(V,M)\Big)^{-1} = \int D\Phi\,D\tilde \Phi\,
\exp\Bigg\{i\Bigg[ Z(\Lambda/\mu) \frac{1}{4} \int d^4x\,d^4\theta\,
\Big(\Phi^* e^{2V}\Phi
+\qquad\nonumber\\
&& + \tilde\Phi^* e^{-2V}\tilde\Phi \Big)
+ \frac{1}{2}\int d^4x\,d^2\theta\, M \tilde\Phi \Phi
+ \frac{1}{2}\int d^4x\,d^2\bar\theta\, M \tilde\Phi^* \Phi^*
\Bigg]\Bigg\}\end{aligned}$$
and the coefficients $c_i$ satisfy equations
$$\sum\limits_i c_i = 1;\qquad \sum\limits_i c_i M_i^2 = 0.$$
Below we will also assume, that $M_i = a_i\Lambda$, where $a_i$ are constants. Insersion of such Pauli-Villars determinants allows to cancel remaining divergences in all one-loop diagrams, including diagrams with insersions of counterterms.
Repeating the above arguments it is possible to write the Pauli-Villars determinants in the following form
$$\begin{aligned}
&& \Big(\det PV(V,M)\Big)^{-1}
= \exp\Bigg\{i (S_{PV})_I\Bigg(V,\,
\frac{1}{i}\frac{\delta}{\delta {\bf j}},\ldots\Bigg)\Bigg\}
\exp\Bigg\{ i \int d^4x\,d^4\theta\,
\times
\nonumber\\
&& \times
\Bigg(
{\bf j}^* \frac{1}{\partial^2+M^2} {\bf j}
+ \tilde {\bf j}^* \frac{1}{\partial^2+M^2} \tilde {\bf j}
+ {\bf j} \frac{M}{\partial^2+M^2}\frac{D^2}{4\partial^2} \tilde {\bf j}
+ {\bf j}^* \frac{M}{\partial^2+M^2} \frac{\bar D^2}{4\partial^2}
\tilde {\bf j}^*
\Bigg)\Bigg\}\Bigg|_{{\bf j},\tilde {\bf j}=0},\qquad\end{aligned}$$
where
$$\begin{aligned}
&& (S_{PV})_I
= \sum\limits_{n=1}^\infty \frac{1}{4 n!}\int d^4x\,d^4\theta\,
\Big(\Phi^* (2V)^{n} \Phi + \tilde\Phi^* (-2V)^n\tilde\Phi\Big)
+\nonumber\\
&& \qquad\qquad\qquad\qquad
+ \Big(Z(\Lambda/\mu) - 1\Big) \frac{1}{4}\int d^4x\,d^4\theta\,
\Big(\Phi^* e^{2V} \Phi + \tilde\Phi^* e^{-2V}\tilde\Phi\Big),\qquad\quad\end{aligned}$$
that allows to find their perturbative expansions. Then it is possible to calculate the generating functional $Z$ according to the prescription
$$\begin{aligned}
\label{Generating_Functional_Z}
&& Z = \prod\limits_i
\Bigg(\det PV\Big(\frac{1}{i}\frac{\delta}{\delta J},M_i\Big)\Bigg)^{c_i}
\exp\Bigg\{i S_I\Bigg(\frac{1}{i}\frac{\delta}{\delta J},\,
\frac{1}{i}\frac{\delta}{\delta j},\ldots\Bigg)\Bigg\}
\times\nonumber\\
&& \qquad\qquad\qquad
\times
\exp\Bigg\{ i \int d^4x\,d^4\theta\,\Bigg(
j^* \frac{1}{\partial^2} j
+ \tilde j^* \frac{1}{\partial^2} \tilde j
- J \frac{e^2}{\partial^2
\Big(1+\partial^{2n}/\Lambda^{2n}\Big)} J\Bigg)\Bigg\}.\qquad\end{aligned}$$
The generating functional for connected Green functions in our notations is written as
$$\label{W}
W = - i\ln Z,$$
and the effective action is defined by making a Legender transformation
$$\label{Gamma}
\Gamma = W - \int d^4x\,d^4\theta\,J V
- \int d^4x\,d^2\theta\, \Big(j\,\phi + \tilde j\,\tilde\phi \Big)
- \int d^4x\,d^2\bar\theta\,
\Big(j^*\phi^* + \tilde j^* \tilde\phi^* \Big),$$
where $J$, $j$ and $\tilde j$ should be expressed in terms of $V$, $\phi$ and $\tilde\phi$ through solving of the equations
$$V = \frac{\delta W}{\delta J};\qquad
\phi = \frac{\delta W}{\delta j};\qquad
\tilde\phi = \frac{\delta W}{\delta\tilde j}.$$
Expressions for Feinman diagrams in the coordinate representation can be found expanding generating functional (\[Generating\_Functional\_Z\]) and substituting the result into equations (\[W\]) and (\[Gamma\]). Certainly, after this procedure $\Gamma$ will contain only 1PI-diagrams. Expressions for Feinman diagrams in the momentum space can be then obtained by Fourier transformations. Performing the calculations we used this algorithm and tried to avoid direct application of Feinman rules in order to be complitely sure in the correctness of numerical factors for all diagrams. However, for the sake of completeness we formulate Feinman rules for the considered theory, which allow to verify structure of expressions for the diagrams.
1\. External lines correspond to a factor
$$\prod\limits_{E}
\int \frac{d^4p_{{}_{E_V}}}{(2\pi)^4} V(p_{{}_{E_V}})
\int \frac{d^4p_{{}_{E_\phi}}}{(2\pi)^4} \phi(p_{{}_{E_\phi}})
\cdot \ldots\cdot
(2\pi)^4 \delta\Big(\sum\limits_{E} p_{{}_E}\Big),$$
where the index $E$ numerates external momentums.
2\. Each internal line of $V$-superfield corresponds to
$$\frac{2e^2}{(k^2+i0) \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)} \,
\delta^4(\theta_1-\theta_2).$$
3\. Each internal line $\phi-\phi^*$ or $\tilde\phi-\tilde\phi^*$ corresponds to
$$-\frac{1}{4 (k^2+i0)}\bar D^2 D^2 \delta^4(\theta_1-\theta_2).$$
(Note, that in the considered theory the action is quadratic in matter superfields, that allows to formulate Feinman rules in a bit different manner, than for, say, Wess-Zumino model.)
4\. Pauli-Villars fields are present only in the closed loops. Each internal line $\Phi-\Phi^*$ or $\tilde\Phi-\tilde\Phi^*$ corresponds to
$$- \frac{1}{4(k^2-M_i^2+i0)}\, \bar D^2 D^2 \delta^4(\theta_1-\theta_2).$$
Internal lines $\Phi-\tilde\Phi$ and $\Phi^*-\tilde\Phi^*$ corresponds to
$$\frac{M_i}{k^2-M_i^2+i0}\,\bar D^2 \delta^4(\theta_1-\theta_2)
\quad\mbox{and}\quad
\frac{M_i}{k^2-M_i^2+i0}\, D^2 \delta^4(\theta_1-\theta_2)$$
respectively. Also it is necessary to add ${\displaystyle - \sum\limits_i c_i}$ for each closed loop of Pauli-Villars fields.
5\. Each loop gives integration over a loop momentum ${\displaystyle \int \frac{d^4k}{(2\pi)^4}}$.
6\. Each vertex gives integration over the corresponding $\theta$: ${\displaystyle \int d^4\theta}$.
7\. There are numerical factors, which can be calculated expanding generating functional (\[Generating\_Functional\_Z\]).
Calculation of two-loop renormgroup functions.
==============================================
\[Section\_Two\_Loop\]
Let us calculate two-loop $\beta$-function and anomalous dimension for a model, described by action (\[Regularized\_SQED\_Action\]). In the two-loop approximation $\beta$-function can be found after calculation of diagrams with $E_V=2$, $E_\phi=0$, presented at Figure \[Figure\_Beta\_Diagrams\]. Note, that each graph at this figure corresponds to a diagram with internal $\phi$-line, a diagram with internal $\tilde\phi$-line and a set of diagrams with internal lines of Pauli-Villars fields. As an example in Appendix \[Appendix\_One\_Loop\] we present detailed calculation of one-loop diagrams. Expressions obtained for the other diagrams are collected in Appendix \[Appendix\_Diagrams\]. Each of these diagrams has the following structure:
$$\int d^4\theta \frac{d^4p}{(2\pi)^4}
\Bigg(V(-p,\theta)\, \partial^2 \Pi_{1/2} V(p,\theta)\,f_1(p,\Lambda)
+ V(-p,\theta) V(p,\theta)\,f_2(p,\Lambda) \Bigg).$$
Terms proportional to ${\displaystyle \int d^4\theta\,V(-p,\theta) V(p,\theta)}$ are not gauge invariant and should disappear after summing of all Feinman diagrams, that is very convenient for checking correctness of the calculations. The other terms can be written as
$$- \mbox{Re} \int d^2\theta \frac{d^4p}{(2\pi)^4}
W_a(-p,\theta) C^{ab} W_b(p,\theta)\,
\sum\limits_{\mbox{\footnotesize diagrams}} f_1(p,\Lambda).$$
Having performed the calculations, in the Minkowsky space we obtained, that the result for two-loop contribution to the effective action, corresponding to the two-point Green function of the gauge field, can be written as
$$\label{Two_Loop_Effective_Action_V}
\quad\Delta\Gamma^{(2)}_{V} = \mbox{Re} \int d^2\theta\,
\frac{d^4p}{(2\pi)^4} W_a(p) C^{ab} W_b(-p)
\Big(f_{\mbox{\footnotesize 1-loop}} + f_{\mbox{\footnotesize 2-loop}}
+ f_{\mbox{\footnotesize PV}} + f_{\mbox{\footnotesize Konishi}} \Big),
\quad$$
where (for simplicity we omit $+i0$ in propagators)
$$\label{1-Loop}
f_{\mbox{\footnotesize 1-loop}} =
- \frac{i}{2}\Bigg(\int\frac{d^4k}{(2\pi)^4} \frac{1}{k^2 (k+p)^2}
- \sum\limits_i c_i \int\frac{d^4k}{(2\pi)^4}
\frac{1}{(k^2-M_i^2)\Big((k+p)^2-M_i^2\Big)}\Bigg)$$
is a total one-loop contribution, including contributions of diagrams with internal lines of Pauli-Villars fields;
$$\label{2-Loop}
f_{\mbox{\footnotesize 2-loop}}
= - e^2 \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}
\frac{(k+p+q)^2+q^2-k^2-p^2}{k^2\Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big)
(k+q)^2 (k+p+q)^2 q^2 (q+p)^2}$$
is a sum of diagrams (\[Two\_Loop\_Diagram1\]) – (\[Two\_Loop\_Diagram6\]) with internal lines of $\phi$ and $\tilde\phi$ fields;
$$\begin{aligned}
\label{PV}
&& f_{\mbox{\footnotesize PV}} = e^2 \sum\limits_i c_i\,
\int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\,
\frac{1}{k^2 \Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big)}
\times\nonumber\\
&& \times
\Bigg[\frac{(k+p+q)^2+q^2-k^2-p^2}{
\Big((k+q)^2-M_i^2\Big) \Big((k+p+q)^2-M_i^2\Big) \Big(q^2-M_i^2\Big)
\Big((q+p)^2-M_i^2\Big)}
+\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
+ \frac{4 M_i^2}{\Big((k+q)^2-M_i^2\Big) \Big(q^2-M_i^2\Big)^2
\Big((q+p)^2-M_i^2\Big)} \Bigg] \qquad\quad\end{aligned}$$
is a contribution of diagrams (\[Two\_Loop\_Diagram1\]) – (\[Two\_Loop\_Diagram6\]) with internal lines of Pauli-Villars fields;
$$\label{Konishi}
f_{\mbox{\footnotesize Konishi}} =
- \frac{i e^2}{2\pi^2}\ln\frac{\Lambda}{\mu}\,
\sum\limits_i c_i
\int \frac{d^4k}{(2\pi)^4}\,\frac{M_i^2}{(k^2-M_i^2)^2
\Big((k+p)^2-M_i^2\Big)}\qquad\quad$$
is a total contribution of diagrams (\[Counterterms\_Diagram1\]) – (\[Counterterms\_Diagram4\]) with insersions of one-loop counterterms. ($M_i = a_i\Lambda$ are masses of Pauli-Villars fields.)
Similarly, anomalous dimension can be found after calculation of diagrams with $E_V=0$, $E_\phi=2$, presented at Figure \[Figure\_Anomalous\_Dimension\_Diagrams\]. Calculation of the one-loop diagram is described in Appendix \[Appendix\_One\_Loop\]. Results for the other diagrams are presented in Appendix \[Appendix\_Diagrams\]. The total two-loop contribution in the Minkowsky space can be written in the following form:
$$\begin{aligned}
\label{Two_Loop_Effective_Action_Phi}
&& \hspace*{-5mm}
\Delta\Gamma^{(2)}_{\phi} =
\int d^4\theta\,\frac{d^4p}{(2\pi)^4}\,
\Big(\phi^*(p,\theta)\,\phi(-p,\theta)
+ \tilde\phi^*(p,\theta)\,\tilde\phi(-p,\theta)\Big)
\times\nonumber\\
&& \hspace*{-5mm}
\times
\Bigg\{
i \int\frac{d^4k}{(2\pi)^4}
\frac{e^2}{2 k^2 (k+p)^2 \Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big)}
-\nonumber\\
&& \hspace*{-5mm}
- \int \frac{d^4k}{(2\pi)^4}\,\frac{d^4q}{(2\pi)^4}\,
\frac{e^4}{k^2 q^2 (k+p)^2 (q+p)^2
\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)
\Big(1+ (-1)^n q^{2n}/\Lambda^{2n}\Big)}
-\nonumber\\
&&\hspace*{-5mm}
- \int \frac{d^4k}{(2\pi)^4}\,\frac{d^4q}{(2\pi)^4}\,
\frac{e^4}{\displaystyle k^2 q^2 (q+p)^2 (k+q+p)^2
\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)
\Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big)}
+\nonumber\\
&&\hspace*{-5mm}
+ \frac{i e^4}{4\pi^2}\ln \frac{\Lambda}{\mu}
\int \frac{d^4k}{(2\pi)^4}\,\frac{1}{k^2 (k+p)^2
\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)}
+ \int \frac{d^4k}{(2\pi)^4}\,\frac{d^4q}{(2\pi)^4}\,
\times\nonumber\\
&&\hspace*{-5mm}
\times
\frac{e^4 (k+q+2p)^2}{k^2 (k+p)^2 q^2 (q+p)^2 (k+q+p)^2
\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)
\Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big)}
-\nonumber\\
&&\hspace*{-5mm}
- \int \frac{d^4q}{(2\pi)^4}
\,\frac{e^4}{q^2 (q+p)^2 \Big(1+ (-1)^n q^{2n}/\Lambda^{2n}\Big)^2}
\Bigg(\int \frac{d^4k}{(2\pi)^4}\, \frac{1}{k^2 (k+q)^2}
-\nonumber\\
&&\qquad\qquad\qquad\qquad\qquad\qquad
- \sum\limits_i c_i
\int \frac{d^4k}{(2\pi)^4}\,\frac{1}{(k^2-M_i^2)\Big((k+q)^2-M_i^2\Big)}
\Bigg).\qquad\end{aligned}$$
Results (\[Two\_Loop\_Effective\_Action\_V\]) and (\[Two\_Loop\_Effective\_Action\_Phi\]) are evidently invariant under supersymmetry transformations, because they can be written as integrals from products of superfields (\[Phi\_Superfield\]) and (\[W\_Superfield\]) over the superspace. The gauge invariance is absent in equation (\[Two\_Loop\_Effective\_Action\_Phi\]) because the calculations were made only for diagrams with $E_V=0$, $E_\phi=2$. Adding of terms, corresponding to diagrams with arbitrary $E_V$ and $E_\phi=2$, will certainly restore the gauge invariance.
Divergent parts of the integrals in equations (\[1-Loop\]) – (\[Konishi\]) are calculated in Appendix \[Appendix\_Integrals\]. Using results, obtained there, one can conclude, that the sum of contributions (\[1-Loop\]) and (\[2-Loop\]) gives NSVZ result for the $\beta$-function. Expression (\[PV\]) is shown in Appendix \[Appendix\_Integrals\] to be a finite constant and does not contribute to the $\beta$-function. However, sum of diagrams with insersions of one-loop counterterms (\[Konishi\]) is not zero and exactly cancels contribution (\[2-Loop\]). Actually, equation (\[Konishi\]) produces Konishi anomaly [@Konishi], calculated by using Pauli-Villars regularization according to a method, described in [@Bertlmann]. According to this method, an anomaly is equal to a contribution of diagrams with internal lines of Pauli-Villars fields in the limit $M_i\to 0$, while contributions of diagrams with internal lines of usual fields are equal to 0.
Divergent parts of integrals in equation (\[Two\_Loop\_Effective\_Action\_Phi\]) are also calculated in Appendix \[Appendix\_Integrals\].
Using results of Appendix \[Appendix\_Integrals\] it is easy to verify, that counterterms, needed to cancel two-loop divergences in the minimal subtraction scheme can be written as
$$\begin{aligned}
&& \Delta S = - \frac{1}{16\pi^2} \ln\frac{\Lambda}{\mu}
\,\mbox{Re}\int d^4x\,d^2\theta\,\,W_a C^{ab}
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) W_b
+\nonumber\\
&& + \frac{1}{4} \int d^4x\,d^4\theta\,
\Big(\phi^* e^{2V} \phi + \tilde\phi^* e^{-2V} \tilde\phi\Big)
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\quad
\times \smash{\Bigg\{
\frac{\alpha}{\pi} \ln\frac{\Lambda}{\mu}
+ \frac{\alpha^2}{\pi^2} \ln^2 \frac{\Lambda}{\mu}
- \frac{\alpha^2}{\pi^2}
\ln \frac{\Lambda}{\mu}\Bigg(\sum\limits_i c_i \ln \frac{M_i}{\Lambda}
+ \frac{3}{2} \Bigg)\Bigg\} },\qquad\vphantom{\int}\end{aligned}$$
that corresponds to
$$\begin{aligned}
\label{Two_Loop_E2}
&& \frac{4\pi^2}{e_0^2} =
\frac{\pi}{\alpha\Big(\Lambda/\mu\Big)}
- \ln\frac{\Lambda}{\mu}+O(\alpha^2);\\
\label{Two_Loop_Z}
&& Z\Big(\Lambda/\mu\Big)
= 1 + \frac{\alpha}{\pi}\,\ln\frac{\Lambda}{\mu}
+ \frac{\alpha^2}{\pi^2}\,\ln^2 \frac{\Lambda}{\mu}
- \frac{\alpha^2}{\pi^2}\,\ln\frac{\Lambda}{\mu}
\Bigg(\sum\limits_i c_i \ln \frac{M_i}{\Lambda}
+ \frac{3}{2} \Bigg)+ O(\alpha^3).\qquad\end{aligned}$$
Therefore the two-loop $\beta$-function and anomalous dimension of $N=1$ supersymmetric quantum electrodynamics, regularized by higher derivatives, are written as
$$\begin{aligned}
&& \beta = \frac{\alpha^2}{\pi} + O(\alpha^4);\nonumber\\
&& \gamma(\alpha) = - \frac{\alpha}{\pi} + \frac{\alpha^2}{\pi^2}
\Bigg(\sum\limits_i c_i \ln \frac{M_i}{\Lambda}
+ \frac{3}{2} \Bigg) + O(\alpha^3).\end{aligned}$$
In particular, two-loop contribution to the $\beta$-function appears to be 0, so that the beta function is completely defined by the one-loop approximation.
The anomalous dimension $\gamma(\alpha)$ in the two-loop approximation does not depend on $n$ or, by other words, on a form of regularizing term. However, it depends on the ratios of Pauli-Villars masses to the constant $\Lambda$. Nevertheless, the dependence on $M_i/\Lambda$ can be removed by addition of finite counterterms, proportional to $\ln M_i/\Lambda$:
$$\begin{aligned}
\label{Two_Loop_S_Ren_Modefied}
&& S_{\mbox{\footnotesize ren}} =
\frac{1}{4 e^2} Z_3(\Lambda/\mu)\,\mbox{Re}\int d^4x\,d^2\theta\,W_a C^{ab}
\Big(1+ \frac{\partial^{2n}}{\Lambda^{2n}}\Big) W_b
- \frac{1}{16\pi^2}\sum\limits_i c_i \ln \frac{M_i}{\Lambda}
\times\qquad\nonumber\\
&& \times
\,\mbox{Re}\int d^4x\,d^2\theta\,W_a C^{ab} W_b
+ Z(\Lambda/\mu)\frac{1}{4}\int d^4x\, d^4\theta\,
\Big(\phi^* e^{2V}\phi +\tilde\phi^* e^{-2V}\tilde\phi\Big),\qquad\end{aligned}$$
that corresponds to another subtraction scheme, in which anomalous dimension is equal to
$$\gamma(\alpha) = - \frac{\alpha}{\pi} + \frac{3\alpha^2}{2\pi^2}
+ O(\alpha^3)$$
and does not depend on both $n$ and $M_i/\Lambda$. In principle, in this scheme it is possible to consider, that $\Lambda\to\infty$, $M_i/\Lambda\to\infty$ instead of $M_i = a_i\Lambda$.
Comparing of the results with predictions of the renormalization group method. {#Section_Renormgroup}
==============================================================================
The obtained results can be checked by the renormalization group method. It is well known [@Collins], that in renormalizable theories terms proportional to $\ln^2\mu/\Lambda$ are completely defined by one-loop counterterms. Therefore, it is possible to calculate such terms by renormgroup equations and compare them with the result of calculation of Feinman graphs.
Using the notation
$$t = \ln \frac{\mu}{\Lambda}$$
for the considered model renormgroup equations can be written as
$$\label{Renormgroup}
Z(t) = \exp\Big\{ \int dt\,\gamma\Big(\alpha(t)\Big) \Big\};\qquad
t = \int \frac{d\alpha}{\beta(\alpha)}.$$
Because in the one-loop approximation the $\beta$-function is equal to
$$\beta(\alpha) = \alpha^2 \beta_1 + O(\alpha^3),$$
in the lowest order
$$\alpha(t) = \alpha_0 \Big(1+ \beta_1 \alpha_0\, t + O(\alpha_0^2)\Big),$$
where $\alpha_0 = \alpha(0)$. Expanding the anomalous dimension in powers of $\alpha$
$$\gamma(\alpha) = \alpha \gamma_1 + \alpha^2 \gamma_2 + O(\alpha^3)
= \gamma_1\Big(\alpha_0 + \beta_1 \alpha_0^2\, t\Big)
+ \alpha_0^2\,\gamma_2 + O(\alpha_0^3)$$
and substituting it to the first equation of (\[Renormgroup\]), we obtain, that
$$Z(t) = 1 + \gamma_1 \alpha_0\,t + \gamma_1 \beta_1\alpha_0^2\,t^2/2
+ \gamma_2 \alpha_0^2\,t + \gamma_1^2 \alpha_0^2\,t^2/2 + O(\alpha_0^3).$$
Taking into account, that according to the results of one-loop calculations $\gamma_1 = - 1/\pi$ and $\beta_1 = 1/\pi$, the function $Z$ should take the following form:
$$\begin{aligned}
&& Z(\Lambda/\mu) = 1 + \frac{\alpha_0}{\pi}\, \ln \frac{\Lambda}{\mu}
- \gamma_2 \alpha_0^2\, \ln \frac{\Lambda}{\mu} + O(\alpha_0^3)
=\nonumber\\
&& \qquad\qquad\qquad\qquad
= 1 + \frac{\alpha}{\pi}\,\ln \frac{\Lambda}{\mu}
+ \frac{\alpha^2}{\pi^2}\,\ln^2 \frac{\Lambda}{\mu}
- \gamma_2 \alpha^2\, \ln \frac{\Lambda}{\mu} + O(\alpha^3).\qquad\end{aligned}$$
Comparing this expression with equation (\[Two\_Loop\_Z\]) we see, that terms proportional to $\ln^2\mu/\Lambda$ coincide, that can be considered as a check of performed calculations.
Conclusion.
===========
In this paper we calculated two-loop $\beta$-function and anomalous dimension for $N=1$ supersymmetric massless quantum electrodynamics, regularized by higher derivatives. In particular, two-loop contribution to the $\beta$-function is found to be 0 and not to depend on the form of higher derivative term. As we mentioned above, this result follows from the fact, that the axial anomaly and the anomaly of energy-momentum tensor trace in the considered model belong to one supermultiplet. However, to obtain it we have to perform calculations using higher covariant derivative regularization. In principle, this regularization (complemented by the Pauli-Villars regularization for one-loop diagrams) allows to perform easy calculation of diagrams with insertion of counterterms, which are proportional to Konishi anomaly [@Konishi] and have nonzero contribution. Possibly the results of the paper allow to assume, that contributions of all higher loops to the $\beta$-function of $N=1$ supersymmetric electrodynamics, regularized by higher derivatives, are also equal to 0. However, to be completely sure in it, it is necessary to calculate scheme dependent three-loop $\beta$-function.
Two-loop anomalous dimension is found to be independent on the form of higher derivative term if one renormalizes the coupling constant in this term. However, $\gamma(\alpha)$ depends on the ratios of Pauli-Villars masses to the constant $\Lambda$. Nevertheless, it is possible to make anomalous dimension completely independent on parameters of higher derivative regularization ($n$ and $M/\Lambda$) if one introduces some finite counterterms in the renormalized action, that actually corresponds to a different choice of renormalization scheme.
Note, that the result for $\beta$-function does not contradict to NSVZ result
$$\beta(\alpha) = \frac{\alpha^2}{\pi^2}\Big(1-\gamma(\alpha)\Big),$$
because in considered model $\beta$-function depends on the normalization of the matter superfields [@Arkani]. In particular, after a scale transformation making matter superfields canonically normalized, it is possible to obtain NVSZ $\beta$-function. However, in the present paper this statement was not checked by explicit calculations and we hope to make it later.
Authors are very grateful to P.I.Pronin and A.A.Slavnov for valuable discussions.
[**Appendix.**]{}
Superficial degree of divergence for the supersymmetric quantum electrodynamics, regularized by higher derivatives.
===================================================================================================================
\[Appendix\_Degree\_Of\_Divergence\]
In order to calculate the superficial degree of divergence for an arbitrary diagram in the massless supersymmetric quantum electrodynamics let us introduce the following notations:
$L$ is a number of loops,
$I_V$ is a number of internal $V$-lines,
$I_\phi$ is a number of internal $\phi$-lines,
$E_V$ is a number of external $V$-lines,
$E_\phi$ is a number of external $\phi$-lines.
We will start with the result for the superficial degree of divergence of massless supersymmetric electrodynamics without higher derivatives [@Ferrara2]
$$\omega = 2 - E_\phi.$$
Adding of the higher derivative term changes only the propagator of $V$-superfield, which will be proportional to $k^{-2-2n}$ instead of $k^{-2}$ in the usual supersymmetric electrodynamics. Therefore, in the regularized theory the superficial degree of divergence is equal to
$$\label{Degree_Divergence}
\omega_\Lambda = \omega - 2 n I_V.$$
Taking into account, that $\phi$-lines are continuous, it is easy to prove the following identity:
$$\begin{aligned}
\label{Appendix_Degree_Identities}
%&& N_V = I_\phi + \frac{1}{2} E_\phi;\nonumber\\
&& L = I_V + 1 - \frac{1}{2} E_\phi.\end{aligned}$$
Therefore expression (\[Degree\_Divergence\]) can be finally rewritten as
$$\label{Appendix_Degree_Degree_Of_Divergence}
\omega_\Lambda = 2 - 2n (L-1) - E_\phi (n+1).\qquad$$
In principle this result can be also obtaind from the Feinman rules in the superspace, but the derivation is more complicated.
Calculation of one-loop $\beta$-function and anomalous dimension.
=================================================================
\[Appendix\_One\_Loop\]
Expanding generating functional (\[Generating\_Functional\_Z\]) and substituting it to effective action (\[Gamma\]) it is possible to find, that a one-loop diagram with $E_V=0$ and $E_\phi=2$ in the coordinate representation is written as
$$\begin{aligned}
\label{One_Loop_Phi_Diagram}
&& \smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga1.eps}}
\hspace*{-1.2cm}
=\vphantom{\int\limits_p}\nonumber\\
&& \vphantom{\int\limits^h}
= \frac{i}{8}\int d^8x_1 d^8x_2\,
\phi^*(x_1,\theta_1)\,\phi(x_2,\theta_2)
\frac{e^2}{\partial^2 \Big(1+\partial^{2n}/\Lambda^{2n}\Big)}
\delta^8_{12}
\frac{D^2 \bar D^2}{\partial^2} \delta^8_{12}.\qquad\end{aligned}$$
where
$$\int d^8x \equiv \int d^4x\, d^4\theta;\qquad
\delta^8_{12} \equiv \delta^4(x_1-x_2)\,\delta^4(\theta_1-\theta_2).$$
After Fourier transformation this expression in the Minkowsky space can be written as
$$\begin{aligned}
\label{One_Loop_Anomalous_Dimension_Calculation}
&& \frac{i}{8}\int d^4\theta_1 d^4\theta_2
\int \frac{d^4p}{(2\pi)^4}\, \phi^*(p,\theta_1)\,\phi(-p,\theta_2)
\int \frac{d^4k}{(2\pi)^4}
\times\nonumber\\
&&\qquad\qquad\quad
\times
\frac{e^2}{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)}
\delta^4(\theta_1-\theta_2)
\frac{1}{(k+p)^2} D^2 \bar D^2 \delta^4(\theta_1-\theta_2)
=\qquad\nonumber\\
&& = \frac{i}{2}\int d^4\theta \frac{d^4p}{(2\pi)^4}
\frac{d^4k}{(2\pi)^4}\,\phi^*(p,\theta)\,\phi(-p,\theta)
\frac{e^2}{k^2 \Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big) (k+p)^2},\end{aligned}$$
where we used the identity
$$\label{Delta_Theta_Identity}
\delta^4(\theta_1-\theta_2)
\,D^2 \bar D^2 \delta^4(\theta_1-\theta_2)
= 4\,\delta^4(\theta_1-\theta_2).$$
The integral over $d^4k$ in equation (\[One\_Loop\_Anomalous\_Dimension\_Calculation\]) can be calculated after the Wick rotation:
$$\frac{i}{2}\int \frac{d^4k}{(2\pi)^4}\,
\frac{e^2}{k^2 \Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big) (k+p)^2}
\to
- \frac{1}{2}\int \frac{d^4k}{(2\pi)^4}
\frac{e^2}{k^2 \Big(1 + k^{2n}/\Lambda^{2n}\Big) (k+p)^2}.$$
Then it is possible to perform calculation in the Eucliedian space and analitically continue the result for imaginary $p^0$.
The diagrams contributing to the one-loop $\beta$-function can be considered similarly.
From equations (\[Generating\_Functional\_Z\]), (\[W\]) and (\[Gamma\]) it is possible to find, that in the coordinate representation
$$\begin{aligned}
&& \smash{\epsfxsize6.0truecm\epsfbox[240 365 540 865]{dg6.eps}}
\hspace*{-3.4cm}
= -i\int d^8x_1 V(x_1)^2 \frac{D_1^2 \bar D_1^2}{4 \partial^2}\delta^8_{11}
= -i\int d^8x_1 V(x_1)^2 \frac{1}{\partial^2}\delta^4(x_1-x_1)
=\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad
= i\int d^4\theta \frac{d^4p}{(2\pi)^4}\,V(p,\theta)\,V(-p,\theta)
\int \frac{d^4k}{(2\pi)^4} \frac{1}{k^2},\quad\end{aligned}$$
where we take into account identity (\[Delta\_Theta\_Identity\]). The corresponding diagram with Pauli-Villars fields in the coordinate representation is written as
$$\begin{aligned}
&& \smash{\epsfxsize6.0truecm\epsfbox[240 365 540 865]{dpv3.eps}}
\hspace*{-3.4cm}
= i\sum\limits_i c_i \int d^8x_1 V(x_1)^2
\frac{D_1^2 \bar D_1^2}{4 (\partial^2+M_i^2)}\delta^8_{11}
=\\
&& \qquad\qquad\qquad\qquad\qquad
= - i\sum\limits_i c_i
\int d^4\theta \frac{d^4p}{(2\pi)^4}\,V(p,\theta)\,V(-p,\theta)
\int \frac{d^4k}{(2\pi)^4} \frac{1}{k^2-M_i^2}.\quad\nonumber\end{aligned}$$
(A bar at diagrams with Pauli-Villarse fields denotes a part of a vertex corresponding to $\Phi^*$ or $\tilde\Phi^*$.)
For the next diagram equations (\[Generating\_Functional\_Z\]), (\[W\]) and (\[Gamma\]) give the following expression:
$$\begin{aligned}
\label{Beta_Diagram1}
&&\quad\ \ \smash{\epsfxsize6.0truecm\epsfbox[270 390 570 890]{dg1.eps}}
\hspace*{-2.8cm}
= -\frac{i}{4} \int d^8x_1 d^8x_2\,V(x_1)\,V(x_2)\,
\frac{D_1^2 \bar D_1^2}{4 \partial^2} \delta^8_{12}\,
\frac{\bar D_1^2 D_1^2}{4 \partial^2} \delta^8_{12}
=\vphantom{\int\limits_p}\nonumber\\
&& \vphantom{\int\limits^h}\hspace*{-1.2cm}
= -\frac{i}{4} \int d^8x_1 d^8x_2\,V(x_2)\,
\frac{1}{4\partial^2}\delta^8_{12}\,
\bar D_1^2 D_1^2 \Big(V(x_1)
\frac{\bar D_1^2 D_1^2}{4 \partial^2} \delta^8_{12}\Big)
=\nonumber\\
&& \hspace*{-1.2cm}
= -\frac{i}{4} \int d^8x_1 d^8x_2\,V(x_2)\,
\frac{1}{4\partial^2}\delta^8_{12}\,
\Big(D_1^2 \bar D_1^2 + [\bar D_1^2, D_1^2]\Big) \Big(V(x_1)
\frac{\bar D_1^2 D_1^2}{4 \partial^2} \delta^8_{12}\Big)
=\\
&& \hspace*{-1.2cm}
= -\frac{i}{4} \int d^8x_1 d^8x_2\,V(x_2)\,
\frac{1}{4\partial^2}\delta^8_{12}\,
\Bigg( D_1^2 \bar D_1^2 V(x_1)
\frac{\bar D_1^2 D_1^2}{4 \partial^2} \delta^8_{12}
+\nonumber\\
&& \qquad\qquad\qquad\qquad
+ \partial_\mu \Big(
i \bar D_1\gamma^\mu\gamma_5 D_1 V(x_1)
\frac{\bar D_1^2 D_1^2}{\partial^2} \delta^8_{12}
- 4 V(x_1)\, \partial_\mu
\frac{\bar D_1^2 D_1^2}{\partial^2} \delta^8_{12}
\Big)\Bigg).\nonumber\end{aligned}$$
Here we took into accout, that
$$\delta^4(\theta_1-\theta_2)\, D_{a_1} \ldots D_{a_k}
\delta^4(\theta_1-\theta_2) = 0$$
for $k=0,1,2,3$ and used identities
$$[\bar D^2, D^2] = 4i \bar D\gamma^\mu \gamma_5 D\,\partial_\mu;
\qquad
\bar D\gamma_\mu\gamma_5 D\, \bar D^2
= 4i \partial_\mu \bar D^2.$$
Taking into accont equation (\[Delta\_Theta\_Identity\]) in the momentum representation expression (\[Beta\_Diagram1\]) can be written as
$$\begin{aligned}
&& - \frac{i}{16}\int d^4\theta \frac{d^4p}{(2\pi)^4}\frac{d^4k}{(2\pi)^4}
\frac{1}{k^2 (k+p)^2} V(p,\theta)
\times\nonumber\\
&& \qquad\qquad\qquad\qquad
\times
\Big(D^2 \bar D^2 + 4\bar D\gamma^\mu\gamma_5 D\,k_\mu
+ 16 (p+k)^\mu k_\mu\Big) V(-p,\theta).\qquad\vphantom{\frac{1}{2}}\end{aligned}$$
Let us note, that the integral
$$I_\mu \equiv \int \frac{d^4k}{(2\pi)^4}\frac{k_\mu}{k^2 (k+p)^2}$$
is proportional to $p_\mu$. Therefore, it can be presented as
$$\begin{aligned}
&& I_\mu = \frac{p_\mu p^\nu}{p^2} I_\nu
= \frac{p_\mu}{2 p^2}\int \frac{d^4k}{(2\pi)^4}\,
\frac{(k+p)^2-k^2-p^2}{k^2 (k+p)^2}
=\nonumber\\
&& \qquad\qquad\quad
= - \frac{1}{2} p_\mu \int \frac{d^4k}{(2\pi)^4}\frac{1}{k^2 (k+p)^2}
+ \frac{p_\mu}{2 p^2} \int \frac{d^4k}{(2\pi)^4}
\Bigg(\frac{1}{k^2}-\frac{1}{(k+p)^2}\Bigg).\qquad\end{aligned}$$
The last term can be omitted, because with the corresponding contribution of the diagram with Pauli-Villars fields it is proportional to
$$\quad\int \frac{d^4k}{(2\pi)^4}
\Bigg(\frac{1}{k^2} - \sum\limits_i c_i\,\frac{1}{k^2-M_i^2}\Bigg)
- \int \frac{d^4k}{(2\pi)^4} \Bigg(\frac{1}{(k+p)^2}
- \sum\limits_i c_i\,\frac{1}{(k+p)^2-M_i^2}\Bigg) = 0,\quad$$
where we take into account, that both integrals are convergent and it is possible to make in the second integral a substitution $k+p\to k$. Therefore, the considered diagram takes the following form:
$$\begin{aligned}
&& -\frac{i}{16}\int \frac{d^4k}{(2\pi)^4}\frac{1}{k^2 (k+p)^2}
V(p,\theta)\Big(D_1^2 \bar D_1^2 - 2 \bar D\gamma^\mu\gamma_5 D\,p_\mu
+ 16 (k+p)^\mu k_\mu\Big) V(-p,\theta)
=\nonumber\\
&& = -\frac{i}{16}\int \frac{d^4k}{(2\pi)^4}\frac{1}{k^2 (k+p)^2}
V(p,\theta)\Big(D_1^2 \bar D_1^2 + \frac{1}{2} [\bar D_1^2,D_1^2]
+ 16 (k+p)^\mu k_\mu\Big) V(-p,\theta)
=\nonumber\\
%&& = - \frac{i}{2} \int \frac{d^4k}{(2\pi)^4}\frac{1}{k^2 (k+p)^2}
%V(p,\theta)\Big(- \partial^2 \Pi_{1/2} + p^2
%+ 2 (k+p)^\mu k_\mu\Big) V(-p,\theta)
%=\nonumber\\
&& = \frac{i}{2} \int \frac{d^4k}{(2\pi)^4}\frac{1}{k^2 (k+p)^2}
V(-p,\theta)\Big(\partial^2 \Pi_{1/2} - (k+p)^2 - k^2 \Big) V(p,\theta).\end{aligned}$$
This diagram has two corresponding diagrams with the loop of Pauli-Villars fields, presented below. The first diagram is calculated similar to the massless case:
$$\begin{aligned}
&& \smash{\epsfxsize6.0truecm\epsfbox[220 390 520 890]{dpv1.eps}}
\hspace*{-1.8cm}
= \frac{i}{4} \sum\limits_i c_i \int d^4x_1 d^4x_2\,V(x_1)\,V(x_2)\,
\frac{D_1^2 \bar D_1^2}{4 (\partial^2+M_i^2)} \delta^8_{12}
\times\vphantom{\int\limits_p}\nonumber\\
&& \vphantom{\int\limits^h}\times
\frac{\bar D_1^2 D_1^2}{4 (\partial^2+M_i^2)} \delta^8_{12}
= \frac{i}{2} \sum\limits_i c_i \int d^4\theta \frac{d^4p}{(2\pi)^4}
\frac{d^4k}{(2\pi)^4}
\frac{1}{\Big(k^2-M_i^2\Big)\Big((k+p)^2-M_i^2\Big)}
\times\qquad\nonumber\\
&& \times
V(p,\theta) \Big(-\partial^2\Pi_{1/2} + (k+p)^2 + k^2 \Big) V(-p,\theta).
\vphantom{\Bigg)}\end{aligned}$$
Taking into account, that signs in $\phi^* V \phi$ and $\tilde\phi^* V \tilde\phi$ vertexes are different, the second diagram can be written in the following form:
$$\begin{aligned}
&& \smash{\epsfxsize6.0truecm\epsfbox[220 390 520 890]{dpv2.eps}}
\hspace*{-1.8cm}
= -\frac{i}{4}\sum\limits_i c_i
\int d^8x_1 d^8x_2\,V(x_1)\,V(x_2)\,
\frac{M_i D_1^2}{\partial^2+M_i^2}\delta^8_{12}
\times\vphantom{\int\limits_p}\nonumber\\
&& \vphantom{\int\limits^h}\times
\frac{M_i \bar D_1^2}{\partial^2+M_i^2} \delta^8_{12}
= -\frac{i}{4}\sum\limits_i c_i
\int d^4\theta_1 d^4\theta_2
\frac{d^4p}{(2\pi)^4} \frac{d^4k}{(2\pi)^4}\,
V(p,\theta_1)\,V(-p,\theta_2)\,
\frac{M_i^2}{k^2-M_i^2}
\times\nonumber\\
&& \times
\delta^4(\theta_1-\theta_2)\,\frac{1}{(k+p)^2-M_i^2}\,
\bar D_1^2 D_1^2
\delta^4(\theta_1-\theta_2)
= -i \sum\limits_i c_i
\int d^4\theta \frac{d^4p}{(2\pi)^4} \frac{d^4k}{(2\pi)^4}\,
\times\nonumber\\
&& \times
V(p,\theta)\, V(-p,\theta)\,
\frac{M_i^2}{\Big(k^2-M_i^2\Big)\Big((k+p)^2-M_i^2\Big)}.\end{aligned}$$
Collecting the results we see, that the sum of the one-loop diagrams is equal to
$$\begin{aligned}
&& \frac{i}{2}\int d^4\theta \frac{d^4p}{(2\pi)^2}\,
V(p,\theta)\,\partial^2\Pi_{1/2}V(-p,\theta)\,
\times\nonumber\\
&& \qquad\qquad\quad
\times
\int \frac{d^4k}{(2\pi)^2}\Bigg(\frac{1}{k^2 (k+p)^2}
-\sum\limits_i c_i
\frac{1}{\Big(k^2-M_i^2\Big)\Big((k+p)^2-M_i^2\Big)}\Bigg)
=\nonumber\\
&& = - \frac{i}{2}\mbox{Re} \int d^2\theta \frac{d^4p}{(2\pi)^2}\,
W_a(p,\theta) C^{ab} W_b(-p,\theta)\,
\times\nonumber\\
&& \qquad\qquad\quad
\times
\int \frac{d^4k}{(2\pi)^2}\Bigg(\frac{1}{k^2 (k+p)^2}
-\sum\limits_i c_i
\frac{1}{\Big(k^2-M_i^2\Big)\Big((k+p)^2-M_i^2\Big)}\Bigg).
\qquad\end{aligned}$$
Note, that all noninvariant terms, proportional to $V^2$, disappeared, that can be considered as a check of the correctness of the calculations.
After Wich rotation in the Eucliedian space the last integral over $d^4k$ can be written as
$$\frac{1}{2}\int \frac{d^4k}{(2\pi)^2}\Bigg(\frac{1}{k^2 (k+p)^2}
-\sum\limits_i c_i
\frac{1}{\Big(k^2+M_i^2\Big)\Big((k+p)^2+M_i^2\Big)}\Bigg).$$
This integral is calculated below in Appendix \[Appendix\_Integrals\]. In order to find the contribution to the effective action the result should be continued for imaginary $p_0$ and multiplied by
$$\mbox{Re}\int d^2\theta \frac{d^4p}{(2\pi)^4}\,
W_a(p,\theta)C^{ab} W_b(-p,\theta).$$
Two-loop Feinman diagrams are calculated in the similar way. However, the calculations are much more complicated and we do not describe them in details.
Diagrams, giving nontrivial contribution to the two-loop $\beta$-function and anomalous dimension.
==================================================================================================
\[Appendix\_Diagrams\]
Below we present expressions for all Feinman graphs, giving nontrivial contributions to the two-loop $\beta$-function. Each of these graphs corresponds to a set of diagrams, which consists of a diagram with internal $\phi$-line, a diagram with internal $\tilde\phi$-line and diagrams with internal lines of Pauli-Villars fields. In order to find contributions to the effective action it is necesary to add the factor
$$\int d^4\theta \frac{d^4p}{(2\pi)^4}.$$
Note, that for simplicity of notations we also omit $+i0$ in propagators.
$$\begin{aligned}
\label{One_Loop_Diagram1}
&& \smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dg1.eps}}
\hspace*{-1.2cm} = \frac{i}{2} \int \frac{d^4k}{(2\pi)^4}\Bigg[
\, V\partial^2\Pi_{1/2} V \frac{1}{k^2 (k+p)^2}
- V^2 \frac{1}{k^2}
-\nonumber\\
&& - V^2 \frac{1}{(k+p)^2}
- \sum\limits_i c_i \Bigg(
V\partial^2\Pi_{1/2} V \frac{1}{(k^2-M_i^2)\Big((k+p)^2-M_i^2\Big)}
- V^2 \frac{1}{k^2-M_i^2}
-\nonumber\\
&& - V^2 \frac{1}{(k+p)^2-M_i^2}
\Bigg)\Bigg];
\\
\nonumber\\
\nonumber\\
\nonumber\\
\label{One_Loop_Diagram2}
&& \smash{\epsfxsize6.0truecm\epsfbox[190 365 490 865]{dg6.eps}}
\hspace*{-2.0cm}
= i \int \frac{d^4k}{(2\pi)^4} V^2
\Bigg(\frac{1}{k^2}-\sum\limits_i c_i\,\frac{1}{k^2-M_i^2}\Bigg);
\\
\nonumber\\
\nonumber\\
\nonumber\\
\label{Two_Loop_Diagram1}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dg2.eps}}
\hspace*{-1.2cm} =
\int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^2}
\,\frac{e^2}{k^2 \Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big)}
\times\vphantom{\int\limits_p}\nonumber\\
&& \times
\Bigg[
\, V \partial^2\Pi_{1/2} V
\frac{4(k+p+q)^2-k^2-p^2}{
(k+q)^2 (k+p+q)^2 q^2 (q+p)^2}
- V^2 \frac{2}{(k+q)^2 q^2} \Bigg]
-\nonumber\\
&& -\sum\limits_i c_i
\int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^2}
\,\frac{e^2}{k^2 \Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big)}
\Bigg[
\,V\partial^2\Pi_{1/2}V
\times\nonumber\\
&& \times
\frac{4(k+p+q)^2-k^2-p^2-2 M_i^2}{
\Big((k+q)^2-M_i^2\Big) \Big((k+p+q)^2-M_i^2\Big) \Big(q^2-M_i^2\Big)
\Big((q+p)^2-M_i^2\Big)}
-\nonumber\\
&&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad\
- V^2 \frac{2}{\Big((k+q)^2-M_i^2\Big) \Big(q^2-M_i^2\Big)}
\Bigg];
\\
\nonumber\\
\nonumber\\
\label{Two_Loop_Diagram2}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dg3.eps}}
\hspace*{-1.2cm} = \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\, V^2
\frac{e^2}{k^2 \Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big)}
\times\\
&& \qquad\qquad\qquad\quad
\times
\Bigg[- \frac{2}{(k+q)^2 (q+p)^2}
+\sum\limits_i c_i
\frac{2}{\Big((k+q)^2-M_i^2\Big) \Big((q+p)^2-M_i^2\Big)} \Bigg];
\nonumber\\
\nonumber\\
\nonumber\\
\label{Two_Loop_Diagram3}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dg4.eps}}
\hspace*{-1.2cm}
= \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}
\frac{e^2}{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)}
\times\nonumber\\
&& \times
\Bigg[
\,V\partial^2\Pi_{1/2}V \frac{2}{(k+q)^2 q^2 (q+p)^2}
- V^2 \Bigg(\frac{2}{q^2 (k+q)^2}
+ \frac{2}{(k+q)^2 (q+p)^2}\Bigg)
\Bigg]
-\nonumber\\
&& -\sum\limits_i c_i \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}
\frac{e^2}{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)}
\times\nonumber\\
&& \times
\Bigg[
\,V\partial^2 \Pi_{1/2} V \frac{2 \Big(q^2+M_i^2\Big)}{
\Big((k+q)^2-M_i^2\Big) \Big(q^2-M_i^2\Big)^2 \Big((q+p)^2-M_i^2\Big)}
-\nonumber\\
&& \qquad\qquad\qquad\qquad\quad\ \
- V^2 \frac{2 \Big(q^2+M_i^2\Big) \Big(q^2 + (q+p)^2\Big)
- 8 M_i^2 q^2}{\Big((k+q)^2-M_i^2\Big)
\Big(q^2-M_i^2\Big)^2 \Big((q+p)^2-M_i^2\Big) } \Bigg];\qquad
\\
\nonumber\\
\nonumber\\
\nonumber\\
\label{Two_Loop_Diagram4}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dg5.eps}}
\hspace*{-1.2cm}
= \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}
\frac{e^2}{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)}
\times\vphantom{\int\limits_p}\nonumber\\
&& \times
\Bigg[
- V\partial^2\Pi_{1/2}V \frac{4}{(k+q)^2 q^2 (q+p)^2}
+ V^2 \Bigg(\frac{4}{q^2 (k+q)^2}
+ \frac{4}{(k+q)^2 (q+p)^2}\Bigg)
\Bigg]
-\vphantom{\int\limits^h}\nonumber\\
&& - \sum\limits_i c_i \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}
\frac{e^2}{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)}
\Bigg[
- V\partial^2\Pi_{1/2} V
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\quad
\times
\frac{4}{\Big((k+q)^2-M_i^2\Big)
\Big(q^2-M_i^2\Big) \Big((q+p)^2-M_i^2\Big)}
+\nonumber\\
&& + V^2 \Bigg(\frac{4}{\Big(q^2-M_i^2\Big) \Big((k+q)^2-M_i^2\Big)}
+ \frac{4}{\Big((k+q)^2-M_i^2\Big) \Big((q+p)^2-M_i^2\Big)}\Bigg)
\Bigg];
\\
\nonumber\\
\nonumber\\
\nonumber\\
\label{Two_Loop_Diagram5}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 365 490 865]{dg7.eps}}
\hspace*{-2.0cm}
= \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\,
\frac{e^2}{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)} V^2
\Bigg[- \frac{2}{q^2 (k+q)^2}
+\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad
+\sum\limits_i c_i\,
\frac{2}{\Big(q^2-M_i^2\Big) \Big((k+q)^2-M_i^2\Big)}
\Bigg];
\\
\nonumber\\
\nonumber\\
\label{Two_Loop_Diagram6}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 365 490 865]{dg8.eps}}
\hspace*{-2.0cm}
= \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\,
\frac{e^2}{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)} V^2
\Bigg[\,\frac{2}{q^2 (k+q)^2}
-\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\ \
-\sum\limits_i c_i\,
\frac{2 \Big(q^2+M_i^2\Big)}{\Big(q^2-M_i^2\Big)^2 \Big((k+q)^2-M_i^2\Big)}
\Bigg];
\\
\nonumber\\
\nonumber\\
\label{Counterterms_Diagram1}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dg9.eps}}
\hspace*{-1.2cm}
= -\frac{i e^2}{4\pi^2}\ln\frac{\Lambda}{\mu}
\int \frac{d^4k}{(2\pi)^4} \Bigg[
\, V\partial^2 \Pi_{1/2} V \frac{1}{k^2 (k+p)^2}
-\vphantom{\int\limits_p}\nonumber\\
&& - V^2 \Bigg(\frac{1}{k^2}+\frac{1}{(k+p)^2}\Bigg)
\Bigg]
+ \sum\limits_i c_i
\frac{i e^2}{4\pi^2}\ln\frac{\Lambda}{\mu}
\int \frac{d^4k}{(2\pi)^4} \Bigg[
\,V\partial^2\Pi_{1/2} V
\times\nonumber\\
&& \times
\frac{k^2+M_i^2}{(k^2-M_i^2)^2\Big((k+p)^2-M_i^2\Big)}
- V^2 \frac{(k^2+M_i^2)\Big(k^2 + (k+p)^2\Big)-4k^2 M_i^2}{
(k^2-M_i^2)^2\Big((k+p)^2-M_i^2\Big)}
\Bigg];
\\
\nonumber\\
\nonumber\\
\nonumber\\
\label{Counterterms_Diagram2}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dg10.eps}}
\hspace*{-1.2cm}
= \frac{i e^2}{4\pi^2}\ln\frac{\Lambda}{\mu}
\int \frac{d^4k}{(2\pi)^4} \Bigg[
\, V\partial^2\Pi_{1/2} V \frac{1}{k^2 (k+p)^2}
-\vphantom{\int\limits_p}\nonumber\\
&& - V^2 \Bigg(\frac{1}{k^2}+\frac{1}{(k+p)^2} \Bigg)
\Bigg]
- \sum\limits_i c_i
\frac{i e^2}{4\pi^2}\ln\frac{\Lambda}{\mu}
\int \frac{d^4k}{(2\pi)^4} \Bigg[
\,V\partial^2\Pi_{1/2}V
\times\nonumber\\
&& \times
\frac{1}{(k^2-M_i^2)\Big((k+p)^2-M_i^2\Big)}
- V^2 \Bigg(\frac{1}{k^2-M_i^2} + \frac{1}{(k+p)^2-M_i^2}\Bigg)
\Bigg];
\\
\nonumber\\
\nonumber\\
\nonumber\\
\label{Counterterms_Diagram3}
&& \smash{\epsfxsize6.0truecm\epsfbox[190 365 490 865]{dg11.eps}}
\hspace*{-2.0cm}
= \frac{i e^2}{4\pi^2}\ln\frac{\Lambda}{\mu}
\int \frac{d^4k}{(2\pi)^4} V^2 \Bigg(
-\frac{1}{k^2}+\sum\limits_i c_i\, \frac{k^2+M_i^2}{(k^2-M_i^2)^2}\Bigg);
\\
\nonumber\\
\nonumber\\
\nonumber\\
\label{Counterterms_Diagram4}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 365 490 865]{dg12.eps}}
\hspace*{-2.0cm}
= \frac{i e^2}{4\pi^2}\ln\frac{\Lambda}{\mu}
\int \frac{d^4k}{(2\pi)^4} V^2 \Bigg(\frac{1}{k^2}
- \sum\limits_i c_i\,\frac{1}{k^2-M_i^2}\Bigg).
\\
\nonumber\\
\nonumber\end{aligned}$$
Expressions for diagrams, giving nontrivial contributions to the two-loop anomalous dimension are presented below. Again, in order to obtain corresponding contributions to the effective action it is necessary to add the factor
$$\int d^4\theta \frac{d^4p}{(2\pi)^4} \Big(\phi^*(p,\theta)\,\phi(-p,\theta)
+ \tilde\phi^*(p,\theta)\,\tilde\phi(-p,\theta)\Big).$$
$$\begin{aligned}
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga1.eps}}
\hspace*{-1.2cm}
= \frac{i}{2}\int\frac{d^4k}{(2\pi)^4}
\frac{e^2}{k^2 \Big(1 + (-1)^n k^{2n}/\Lambda^{2n}\Big) (k+p)^2};
\\
\nonumber\\
\nonumber\\
\nonumber\\
&& \smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga2.eps}}
\hspace*{-1.2cm}
= - \int\frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\frac{e^4}
{q^2 \Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big)^2 (q+p)^2}
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\quad\
\times
\Bigg(\frac{1}{k^2 (k+q)^2} - \sum\limits_i c_i\,
\frac{1}{\Big(k^2-M_i^2\Big)\Big((k+q)^2-M_i^2\Big)}\Bigg)
+\nonumber\\
&&
+ 2\int\frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\frac{e^4}
{q^4\Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big)^2 (q+p)^2}
\Bigg(\frac{1}{k^2} - \sum\limits_i c_i\, \frac{1}{k^2 - M_i^2}\Bigg)
+\nonumber\\
&&
+\int\frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\frac{e^4}
{q^4 \Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big)^2}
\Bigg(\frac{1}{k^2 (k+q)^2}
-\\
&&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad
- \sum\limits_i c_i\,
\frac{1}{\Big(k^2-M_i^2\Big)\Big((k+q)^2-M_i^2\Big)}\Bigg);
\nonumber\\
\nonumber\\
\nonumber\\
&& \smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga3.eps}}
\hspace*{-1.2cm} =
- 2\int\frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}
\frac{e^4}
{q^4\Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big)^2 (q+p)^2}
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
\times
\Bigg(\frac{1}{k^2} - \sum\limits_i c_i\, \frac{1}{k^2 - M_i^2}\Bigg);
\\
\nonumber\\
\nonumber\\
\nonumber\\
\nonumber\\
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 365 490 865]{dga7.eps}}
\hspace*{-2.0cm}
= - \int\frac{d^4k}{(2\pi)^4} \frac{d^4q}{(2\pi)^4}
\frac{e^4} {q^4 \Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big)^2}
\Bigg(\frac{1}{k^2 (k+q)^2}
-\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
- \sum\limits_i c_i\,
\frac{1}{\Big(k^2 - M_i^2\Big)\Big((k+q)^2 - M_i^2\Big)} \Bigg);
\\
\nonumber\\
\nonumber\\
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga6.eps}}
\hspace*{-1.2cm}
= - 2\int\frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}
\times \nonumber\\
\\
&& \times \frac{e^4}
{k^2\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)
q^2 \Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big)
(q+p)^2(k+q+p)^2 };
\nonumber\\
\nonumber\\
\nonumber\\
&& \smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga0.eps}}
\hspace*{-1.2cm}
=\nonumber\\
\vphantom{\frac{1}{2}}\\
&&
= \int\frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\frac{e^4}
{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)
q^2 \Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big) (q+p)^2(k+q+p)^2 };
\nonumber\\
\nonumber\\
\nonumber\\
&&
\nonumber\\
&& \smash{\epsfxsize5.8truecm\epsfbox[190 390 490 890]{dga4.eps}}
\hspace*{-0.2cm}
=\nonumber\\
\vphantom{\frac{1}{2}}\\
&&
= - \int\frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\frac{e^4}
{k^2\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)
q^2 \Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big) (q+p)^2(k+p)^2};
\nonumber\\
\nonumber\\
\nonumber\\
\nonumber\\
&& \smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga5.eps}}
\hspace*{-1.2cm}
= \int\frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\times\nonumber\\
\\
&&
\times
\frac{e^4 (q+k+2p)^2}
{k^2 \Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big) q^2
\Big(1+(-1)^n q^{2n}/\Lambda^{2n}\Big) (k+q+p)^2 (q+p)^2 (k+p)^2};
\nonumber\\
\nonumber\\
\nonumber\\
\nonumber\\
&& \smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga8.eps}}
\hspace*{-1.2cm}
= -\frac{i}{8\pi^2}\ln\frac{\Lambda}{\mu}\int\frac{d^4k}{(2\pi)^4}
\frac{e^4}{k^2\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)(k+p)^2};
\nonumber\\
\\
\nonumber\\
\nonumber\\
&& \smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga9.eps}}
\hspace*{-1.2cm}
= - \frac{i}{8\pi^2}\ln\frac{\Lambda}{\mu}
\int\frac{d^4k}{(2\pi)^4}
\frac{e^4}{k^4\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)}
+\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
+ \frac{i}{8\pi^2}\ln\frac{\Lambda}{\mu}\int\frac{d^4k}{(2\pi)^4}
\frac{e^4}{k^2\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big)(k+p)^2};
\nonumber\\
\nonumber\\
\nonumber\\
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 390 490 890]{dga10.eps}}
\hspace*{-1.2cm}
= \frac{i}{4\pi^2}\ln\frac{\Lambda}{\mu}
\int\frac{d^4k}{(2\pi)^4}\frac{e^4}{k^2
\Big(1+(-1)^n k^{2n}/\Lambda^{2n}\Big) (k+p)^2};
\nonumber\\
\\
\nonumber\\
\nonumber\\
\nonumber\\
&&
\smash{\epsfxsize6.0truecm\epsfbox[190 365 490 865]{dga12.eps}}
\hspace*{-2.0cm}
= \frac{i}{8\pi^2} \ln\frac{\Lambda}{\mu}
\int\frac{d^4k}{(2\pi)^4}
\frac{e^4}{k^4\Big(1+ (-1)^n k^{2n}/\Lambda^{2n}\Big)}.
\\
\nonumber\end{aligned}$$
Calculation of integrals, regularized by higher derivatives.
============================================================
\[Appendix\_Integrals\]
Let us calculate divergent parts of integrals, which were encountered in equations (\[Two\_Loop\_Effective\_Action\_V\]) and (\[Two\_Loop\_Effective\_Action\_Phi\]). First it is necessary to perform the Wick rotation. After it the integrals, which are present in equations (\[Two\_Loop\_Effective\_Action\_V\]) and (\[Two\_Loop\_Effective\_Action\_Phi\]) will be proportional to
$$\begin{aligned}
\label{Integrals_Definition}
&& I_1 = \int d^4 k \frac{1}{k^2 (k+p)^2
\Big(1+ k^{2n}/\Lambda^{2n}\Big)};\nonumber\\
&& I_2 = \int d^4k\, \frac{1}{k^2 (k+p)^2}
- \sum\limits_i c_i
\int d^4k\,\frac{1}{(k^2+M_i^2)\Big((k+p)^2+M_i^2\Big)};
\nonumber\\
&& I_3 = \int d^4k\,d^4q
\frac{(k+p+q)^2+q^2-k^2-p^2}{k^2 \Big(1 + k^{2n}/\Lambda^{2n}\Big)
(k+q)^2 (k+p+q)^2 q^2 (q+p)^2};\nonumber\\
&& I_4 = \int d^4k\,\frac{M^2}{(k^2+M^2)^2 \Big((k+p)^2+M^2\Big)};
\nonumber\\
&& I_5 = \int d^4k\,d^4q\,
\frac{1}{k^2 (k+p)^2 q^2 (q+p)^2 \Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)}=I_1^2;\nonumber\\
&& I_6 = \int d^4k\,d^4q\,
\frac{(k+q+2p)^2}{k^2 (k+p)^2 q^2 (q+p)^2 (k+q+p)^2
\Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)};\qquad\nonumber\\
&& I_7 = \int d^4k\,d^4q\,
\frac{1}{\displaystyle k^2 q^2 (q+p)^2 (k+q+p)^2
\Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)};\nonumber\\
&& I_8 = \int d^4q
\,\frac{1}{q^2 (q+p)^2 \Big(1+ q^{2n}/\Lambda^{2n}\Big)^2}\, I_2(q/M);
\nonumber\\
&& I_9 = \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\,
\frac{1}{k^2 \Big(1 + k^{2n}/\Lambda^{2n}\Big)}
\times\nonumber\\
&& \qquad\qquad
\times
\frac{(k+p+q)^2+q^2-k^2-p^2}{
\Big((k+q)^2+M^2\Big) \Big((k+p+q)^2+M^2\Big) \Big(q^2+M^2\Big)
\Big((q+p)^2+M^2\Big)};\nonumber\\
&& I_{10} = \int \frac{d^4k}{(2\pi)^4}\frac{d^4q}{(2\pi)^4}\,
\frac{1}{k^2 \Big(1 + k^{2n}/\Lambda^{2n}\Big)}
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad
\times
\frac{M^2}{\Big((k+q)^2+M^2\Big) \Big(q^2+M^2\Big)^2
\Big((q+p)^2+M^2\Big)}.\end{aligned}$$
In order to calculate integral $I_1$ it is possible to use four-dimensional spherical coordinates
$$\begin{aligned}
&& k_1 = k\sin\theta_3\sin\theta_2\sin\theta_1;\nonumber\\
&& k_2 = k\sin\theta_3\sin\theta_2\cos\theta_1;\nonumber\\
&& k_3 = k\sin\theta_3\cos\theta_2;\nonumber\\
&& k_4 = k\cos\theta_3.\nonumber\\\end{aligned}$$
and direct fourth axis along $p^\mu$, so that the integrand will depend only on $\theta_3$ and
$$\begin{aligned}
&& \int d^4k = \int\limits_{0}^\infty k^3 dk\,
\int\limits_0^\pi d\theta_3\, \sin^2\theta_3
\int\limits_0^\pi d\theta_2\, \sin\theta_2
\int\limits_0^{2\pi} d\theta_1 =
4\pi \int\limits_{0}^\infty k^3 dk\,
\int\limits_0^\pi d\theta_3\, \sin^2\theta_3
=\qquad\nonumber\\
&&= \Big[x=\cos\theta_3\Big] =
4\pi \int\limits_{0}^\infty k^3 dk\,
\int\limits_{-1}^1 dx\, \sqrt{1-x^2}.\end{aligned}$$
Taking into account, that $k^\mu p_\mu = kp\cos\theta_3 = kpx$, the integral can be written as
$$\begin{aligned}
&& \int d^4k \frac{1}{k^2 (k+p)^2 \Big(1 + k^{2n}/\Lambda^{2n}\Big)}
=\nonumber\\
&&\qquad\qquad
= 4\pi \int\limits_0^\infty k\, dk\,
\int\limits_{-1}^{1} dx\,\frac{\sqrt{1-x^2}}{(k^2+2kpx+p^2)
\Big(1+k^{2n}/\Lambda^{2n}\Big)}
=\nonumber\\
&& \qquad\qquad\qquad\qquad
= 2\pi \int\limits_0^\infty k\, dk\,
\oint\limits_C dx\,\frac{\sqrt{1-x^2}}{(k^2+2kpx+p^2)
\Big(1+k^{2n}/\Lambda^{2n}\Big)},\qquad\end{aligned}$$
where the contour $C$ is presented at Figure \[Figure\_Contour\]. The integrand here has singularities at branch points $x = \pm 1$, a pole $x=\infty$ and a pole
$$x_0 = -\frac{k^2+p^2}{2kp}.$$
Then it is easy to see, that
$$\begin{aligned}
&& \oint\limits_C dx\,\frac{\sqrt{1-x^2}}{k^2+2kpx+p^2} =
2\pi i\, \mbox{Res}\Bigg(\frac{\sqrt{1-x^2}}{k^2+2kpx+p^2},
\,x=\infty\Bigg)
-\nonumber\\
&& - 2\pi i\,\mbox{Res}\Bigg(\frac{\sqrt{1-x^2}}{k^2+2kpx+p^2},
\,x=x_0\Bigg)
= 2\pi i\Bigg(-i\frac{k^2+p^2}{4 k^2p^2}
+ i \frac{|k^2-p^2|}{4k^2p^2}\Bigg).\qquad\end{aligned}$$
Therefore, the integral over angles is reduced to
$$\begin{aligned}
\label{Angle_Integral}
\oint dx\,\frac{\sqrt{1-x^2}}{k^2+2kpx+p^2} =
\left\{
\begin{array}{l}
{\displaystyle \frac{\pi}{k^2},\quad k\ge p;}\\
\\
{\displaystyle \frac{\pi}{p^2},\quad p\ge k}
\end{array}
\right.\end{aligned}$$
and finally
$$\begin{aligned}
&& I_1 = 2\pi^2 \int\limits_0^p dk\,\frac{k}{p^2}\,
\frac{1}{\Big(1+k^{2n}/\Lambda^{2n}\Big)}
+ 2\pi^2 \int\limits_p^\infty dk\,\frac{1}{k}\,
\frac{1}{\Big(1+k^{2n}/\Lambda^{2n}\Big)}
=\qquad\nonumber\\
&& = \pi^2 + o(1)
+ \frac{\pi^2}{n} \ln\frac{\Lambda^{2n}+p^{2n}}{p^{2n}}
= 2 \pi^2 \Big(\ln\frac{\Lambda}{p} + \frac{1}{2}\Big) + o(1).\end{aligned}$$
Integral $I_2$ can be calculated using standard methods [@Ramond]. First, using an identity
$$\frac{1}{ab} = \int\limits_0^1 dy\,\frac{1}{\Big(ay + b(1-y)\Big)^2},$$
it can be written as
$$\label{Rewritten_I2}
I_2 = \int\limits_0^1 dy \int d^4k\,\Bigg(
\frac{1}{\Big(k^2 + 2kp y + y p^2\Big)^2}
-\sum\limits_i c_i \frac{1}{\Big(k^2+2kp y + y p^2 + M_i^2\Big)^2}\Bigg).$$
Each of these integrals diverges, but their difference is finite. Therefore, to simplify calculations it is convenient to use an auxiliary regularization, for example, the dimensional regularization. Then the integrals in equation (\[Rewritten\_I2\]) can be easily taken:
$$\begin{aligned}
\label{I2_Result}
&& I_2 = \lim\limits_{D\to 4} \pi^2 \int\limits_0^1 dy\,
\frac{\Gamma(2-D/2)}{\Gamma(2)} \Bigg(\Big(y (1-y)\,p^2\Big)^{D/2-2}
-\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad
- \sum\limits_i c_i \Big(y (1-y)\,p^2 + M_i^2\Big)^{D/2-2}\Bigg)
=\nonumber\\
&&
= \pi^2 \sum\limits_i c_i
\int\limits_0^1 dy\,\ln\Bigg(1+\frac{M_i^2}{y(1-y)\,p^2}\Bigg)
=\\
&& \qquad\qquad\qquad\qquad
= 2\pi^2 \sum\limits_i c_i
\Bigg(\ln\frac{M_i}{p} + \sqrt{1+\frac{4M_i^2}{p^2}}
\mbox{arctanh}\sqrt{\frac{p^2}{4M_i^2 + p^2}}\Bigg).\qquad\nonumber\end{aligned}$$
where we take into account, that $\sum\limits_i c_i = 1$.
To calculate divergent part of the integral $I_3$ note, that $I_3 = I_3(p/\Lambda)$ and due to the logarithmical divergence
$$\label{I2_Expansion}
I_3 = a_1 \ln^2\frac{\Lambda}{p} + a_2 \ln\frac{\Lambda}{p}
+ \sum\limits_{i=0}^\infty b_i \Bigg(\frac{p^2}{\Lambda^2}\Bigg)^i.$$
If $a_1 = 0$, then it is possible to find
$$\label{A2_Expression}
a_2 = \lim\limits_{p\to 0}\,\frac{d I_3}{d\ln\Lambda}
= \int d^4k\,d^4q\,
\frac{4n\,k^{2n-2}}{\Lambda^{2n} \Big(1+k^{2n}/\Lambda^{2n}\Big)^2}
\,\frac{q^2 + k_\mu q_\mu}{(k+q)^4 q^4}.$$
If this limit does not exist, then $a_1\ne 0$. The integral in the right hand side of equation (\[A2\_Expression\]) can be taken, using four-dimensional spherical coordinates:
$$\begin{aligned}
\label{For_I3}
&& \int d^4q\,\frac{q^2 + k_\mu q_\mu}{(k+q)^4 q^4}
= 4\pi \int\limits_0^\infty dq \int\limits_{-1}^1 dx\,
\frac{(q + k x)\sqrt{1-x^2}}{(k^2 + 2kqx + q^2)^2}
=\nonumber\\
&& \qquad
= - 2\pi \int\limits_{-1}^1 dx \int\limits_0^\infty dq\,
\frac{d}{d q} \frac{\sqrt{1-x^2}}{(k^2 + 2kqx + q^2)}
= \frac{2\pi}{k^2} \int\limits_{-1}^1 dx\,\sqrt{1-x^2}
= \frac{\pi^2}{k^2},\qquad\end{aligned}$$
so that
$$a_2 = 4n\pi^2 \int d^4k\,
\frac{k^{2n-4}}{\Lambda^{2n} \Big(1+k^{2n}/\Lambda^{2n}\Big)^2} = 4\pi^4,$$
Therefore, from equation (\[I2\_Expansion\]) we conclude, that
$$I_3 = 4\pi^4 \ln\frac{\Lambda}{p} + O(1).$$
In order to calculate integral $I_4$ let us note, that
$$\int \frac{d^4k}{(2\pi)^4}\,\frac{M^2}{(k^2+M^2)^2 \Big((k+p)^2+M^2\Big)}
= f(p/M).$$
Therefore, instead of taking the limit $M\to \infty$ it is possible to take the limit $p\to 0$, so that
$$\begin{aligned}
&& I_4 = \int d^4k\,\frac{M^2}{(k^2+M^2)^3} + o(1) = \frac{\pi^2}{2}
+ o(1).\end{aligned}$$
Divergent part of integral $I_5$ can be also easily calculated, because
$$I_5 = I_1^2
= 4\pi^4\Bigg(\ln^2 \frac{\Lambda}{p} + \ln\frac{\Lambda}{p}\Bigg) + O(1).$$
To find a divergent part of $I_6$ let us consider
$$\begin{aligned}
&& \lim\limits_{p\to 0}\,\Lambda \frac{d}{d\Lambda} \Big(I_5 - I_6\Big)
=\nonumber\\
&& = \lim\limits_{p\to 0}\,\Lambda \frac{d}{d\Lambda}\int d^4k\,d^4q\,
\Bigg(1 - \frac{(k+q+2p)^2}{(k+q+p)^2} \Bigg)
\times\nonumber\\
&& \qquad\qquad\qquad\qquad
\times
\frac{1}{k^2 (k+p)^2 q^2 (q+p)^2 \Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)}
=\qquad\nonumber\\
&& = \lim\limits_{p\to 0} \int d^4k\,d^4q\,
\frac{- 2(k+q)p - 3p^2}{(k+q+p)^2}
\times\nonumber\\
&& \qquad\qquad\qquad\qquad
\times
\frac{4n q^{2n}/\Lambda^{2n}}{k^2 (k+p)^2 q^2 (q+p)^2
\Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)^2} = 0.\qquad\end{aligned}$$
(It is important to note, that all integrals here are convergent.) Therefore,
$$I_6 = I_1^2 + O(1)
= 4\pi^4\Bigg(\ln^2 \frac{\Lambda}{p} + \ln\frac{\Lambda}{p}\Bigg) + O(1).$$
A divergent part of $I_7$ can be calculated similarly:
$$\begin{aligned}
&& \lim\limits_{p\to 0}\,\Lambda \frac{d}{d\Lambda} \Big(I_5 - 2 I_7\Big)
=\nonumber\\
&& = \lim\limits_{p\to 0}\,\Lambda \frac{d}{d\Lambda}\int d^4k\,d^4q\,
\Bigg(\frac{1}{(k+p)^2} - \frac{2}{(k+q+p)^2} \Bigg)
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
\times
\frac{1}{k^2 q^2 (q+p)^2 \Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)}
=\qquad\nonumber\\
&& = \lim\limits_{p\to 0}\,\Lambda \frac{d}{d\Lambda}\int d^4k\,d^4q\,
\frac{q^2 + 2 (k+p)q - (k+p)^2}{(k+p)^2 (k+q+p)^2}
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
\times
\frac{1}{k^2 q^2 (q+p)^2 \Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)}
=\nonumber\\
&& = \lim\limits_{p\to 0}\,\Lambda \frac{d}{d\Lambda}\int d^4k\,d^4q\,
\frac{q^2 + 2 (k+p)q - (q+p)^2}{(k+p)^2 (k+q+p)^2}
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
\times
\frac{1}{k^2 q^2 (q+p)^2 \Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)}
=\nonumber\\
&& = \lim\limits_{p\to 0}\,\Lambda \frac{d}{d\Lambda}\int d^4k\,d^4q\,
\frac{2 kq - p^2}{(k+p)^2 (k+q+p)^2}
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\qquad
\times
\frac{1}{k^2 q^2 (q+p)^2 \Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)}
=\nonumber\\
&& = \int d^4k\,d^4q\,
\frac{2kq}{k^4 q^4 (k+q)^2}
\,\Lambda \frac{d}{d\Lambda}\Bigg[
\frac{1}{\Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)}\Bigg]
=\nonumber\\
&& \qquad\qquad\qquad\quad
= 8n\int d^4q\,d^4k\,
\frac{kq}{k^4 q^4 (k+q)^2}
\frac{q^{2n}/\Lambda^{2n}}{\Big(1+k^{2n}/\Lambda^{2n}\Big)
\Big(1+ q^{2n}/\Lambda^{2n}\Big)^2}.\qquad\end{aligned}$$
To calculate this integral we again use four-dimensional spherical coordinates and direct fourth axis along $q^\mu$. Then similar to the case, considered above, the integral over angles is reduced to
$$\begin{aligned}
&& 4\pi\int\limits_{-1}^{1} dx\,\frac{x \sqrt{1-x^2}}{k^2+2kqx+q^2}
= 2\pi \oint\limits_C dx\,\frac{x\sqrt{1-x^2}}{k^2+2kqx+q^2}
=\nonumber\\
&& = 4\pi^2 i\, \mbox{Res}\Bigg(\frac{x\sqrt{1-x^2}}{k^2+2kqx+q^2},
\,x=\infty\Bigg)
- 4\pi^2 i\,\mbox{Res}\Bigg(\frac{x\sqrt{1-x^2}}{k^2+2kqx+q^2},
\,x=x_0\Bigg)
=\qquad\nonumber\\
&& = 4\pi^2 i\Bigg(- \frac{i}{4 kq} + \frac{i (k^2+q^2)^2}{8 k^3 q^3}
- \frac{i|k^2-q^2| (k^2 + q^2)}{8 k^3 q^3}\Bigg),\end{aligned}$$
so that
$$\begin{aligned}
2\pi \oint dx\,\frac{x\sqrt{1-x^2}}{k^2+2kqx+q^2} =
\left\{
\begin{array}{l}
{\displaystyle - \frac{\pi^2 q}{k^3},\quad k\ge q;}\\
\\
{\displaystyle - \frac{\pi^2 k}{q^3},\quad q\ge k.}
\end{array}
\right.\end{aligned}$$
Therefore,
$$\begin{aligned}
&& \lim\limits_{p\to 0}\,\Lambda \frac{d}{d\Lambda} \Big(I_5 - 2 I_7\Big)
= - 16n\,\pi^4 \int\limits_0^\infty dq\,
\frac{q^{2n}/\Lambda^{2n}}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\times\nonumber\\
&&\qquad\qquad\qquad\quad\ \ \times
\Bigg(\int\limits_q^\infty dk\,
\frac{q}{k^3 \Big(1+ k^{2n}/\Lambda^{2n}\Big)}
+ \int\limits_0^q dk\,
\frac{k}{q^3 \Big(1+ k^{2n}/\Lambda^{2n}\Big)}
\Bigg)
=\nonumber\\
&&
= - 4n\,\pi^4 \int\limits_0^\infty dx\,
\frac{x^{n}}{(1+x^{n})^2}
\int\limits_0^{1/x}
\frac{dy}{1+ y^{-n}}
- 4n\,\pi^4 \int\limits_0^\infty dx\,
\frac{x^{n}}{(1+x^{n})^2}
\int\limits_0^{1/x}
\frac{dy}{1+ y^{n}}
=\nonumber\\
&&
= - 4n\,\pi^4 \int\limits_0^\infty dx\,
\frac{x^{n-1}}{(1+x^{n})^2}
= - 4\pi^4\end{aligned}$$
and finally
$$I_7 = \frac{1}{2} I_1^2 + 2\pi^4 \ln\frac{\Lambda}{p} + O(1)
= 2\pi^4\Bigg(\ln^2 \frac{\Lambda}{p} + 2\ln\frac{\Lambda}{p}\Bigg) + O(1).$$
Using equation (\[I2\_Result\]) integral $I_8$ can be written as
$$\begin{aligned}
\label{Integrals_DK_Integral}
&& I_8 = 2 \pi^2 \sum\limits_i c_i \int d^4q\, \frac{1}{q^2 (q+p)^2
\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\times\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\quad
\times
\Bigg(\ln \frac{M_i}{q}
+ \sqrt{1+\frac{4M_i^2}{q^2}}
\mbox{arctanh}\sqrt{\frac{q^2}{4M_i^2 + q^2}}\Bigg),\qquad\end{aligned}$$
where $M_i = a_i\Lambda$, $a_i$ being constants. To calculate the divergent part of this integral let us consider first an integral
$$I_f \equiv \int d^4q \frac{1}{q^2 (q+p)^2} f\Big(\Lambda/q\Big)
= I_f\Big(\Lambda/p\Big),$$
where $f$ is a function. Differentiating $I_f$ over $\ln \Lambda$ and setting then $p=0$, we obtain, that
$$\begin{aligned}
\label{Integrals_DIF}
&& \Lambda\frac{d I_f}{d\Lambda} \Bigg|_{p=0}
= \int d^4q \frac{1}{q^4}\,\Lambda\frac{d}{d\Lambda} f\Big(\Lambda/q\Big)
= - \int d^4q \frac{1}{q^3}\,\frac{d}{dq} f\Big(\Lambda/q\Big)
=\nonumber\\
&& \qquad\qquad\qquad\qquad\qquad\quad
= - 2\pi^2 \int\limits_0^\infty dq\, \frac{d}{dq} f\Big(\Lambda/q\Big)
= 2\pi^2 \Big(f(\infty) - f(0)\Big).\qquad\quad\end{aligned}$$
So, if the values $f(\infty)$ and $f(0)$ are finite, then
$$\label{Integrals_IF}
I_f = 2\pi^2 \Big(f(\infty) - f(0)\Big)\ln\frac{\Lambda}{p} + O(1).$$
If the function $f$ is taken to be
$$f\Big(\Lambda/q\Big) = \sum\limits_i c_i
\frac{2\pi^2}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\sqrt{1+\frac{4M_i^2}{q^2}}\,\mbox{arctanh}\sqrt{\frac{q^2}{4M_i^2 + q^2}},$$
then from equation (\[Integrals\_IF\]) we conclude, that
$$\begin{aligned}
\label{Integrals_Atan_Integral}
&& 2\pi^2 \sum\limits_i c_i \int d^4q \frac{1}{q^2 (q+p)^2
\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\sqrt{1+\frac{4M_i^2}{q^2}} \mbox{arctanh}\sqrt{\frac{q^2}{4M_i^2 + q^2}}
=\qquad\nonumber\\
&&
= 4\pi^4 \sum\limits_i c_i \ln\frac{\Lambda}{p} + O(1)
= 4\pi^4 \ln\frac{\Lambda}{p} + O(1).\qquad\end{aligned}$$
However, it is impossible to substitute in equation (\[Integrals\_IF\])
$$f\Big(\Lambda/q\Big)
= \frac{2\pi^2}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\sum\limits_i c_i \ln \frac{M_i}{q}$$
because $f(\infty)$ does not exist. Nevertheless, the function $f$ can chosen in following form:
$$\begin{aligned}
&& f\Big(\Lambda/q\Big)
= \Lambda\frac{d}{d\Lambda}
\Bigg(\frac{2\pi^2}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\sum\limits_i c_i \ln \frac{M_i}{q}\Bigg)
=\nonumber\\
&& \qquad\qquad\qquad\qquad\quad
= \frac{8\pi^2 n\,q^{2n}/\Lambda^{2n}}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^3}
\sum\limits_i c_i \ln \frac{M_i}{q}
+ \frac{2\pi^2}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^2},\qquad\end{aligned}$$
so that $f(0) = 0$ and $f(\infty) = 2\pi^2$. Then from equation (\[Integrals\_DIF\]) we obtain, that
$$\Lambda \frac{d}{d\Lambda}
2\pi^2\int d^4q \frac{1}{q^2 (q+p)^2
\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}\sum\limits_i c_i \ln\frac{M_i}{q}
= 4\pi^4 \ln \frac{\Lambda}{p} + O(1)$$
and, therefore,
$$\label{Integrals_LnQ2}
2\pi^2 \int d^4q \frac{1}{q^2 (q+p)^2
\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}\sum\limits_i c_i \ln\frac{M_i}{q}
= 2\pi^4 \ln^2\frac{\Lambda}{p}
+ O\Big(\ln \frac{\Lambda}{p}\Big).$$
Then it is necessary to calculate logarithmical divergences. For this purpose we subtract from integral (\[Integrals\_LnQ2\]) terms, proportional $\ln^2\Lambda/p$ and differentiate the result over $\ln\Lambda$:
$$\begin{aligned}
\label{Integrals_LnQ}
&& \lim\limits_{p\to 0}\,\Lambda\frac{d}{d\Lambda} \Bigg[
2\pi^2 \int d^4q \frac{1}{q^2 (q+p)^2
\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}\sum\limits_i c_i \ln\frac{M_i}{q}
- 2\pi^4 \ln^2\frac{\Lambda}{p} \Bigg]
=\nonumber\\
&& = \lim\limits_{p\to 0}\Bigg\{
- 2\pi^2 \int d^4q \frac{1}{q^2 (q+p)^2}\,q \frac{d}{dq}
\Bigg(\frac{1}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\sum\limits_i c_i \ln\frac{M_i}{q}\Bigg)
- 4\pi^4 \ln\frac{\Lambda}{p}\Bigg\}
=\nonumber\\
&& = \lim\limits_{p\to 0}\Bigg\{
- 4\pi^4 \int\limits_0^p dq\,\frac{q^2}{p^2}\, \frac{d}{dq}
\Bigg(\frac{1}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\sum\limits_i c_i \ln\frac{M_i}{q}\Bigg)
-\nonumber\\
&&\qquad\qquad\qquad\qquad\quad
- 4\pi^4 \int\limits_p^\infty dq\, \frac{d}{dq}
\Bigg(\frac{1}{\Big(1+q^{2n}/\Lambda^{2n}\Big)^2}
\sum\limits_i c_i \ln\frac{M}{q}\Bigg)
- 4\pi^4 \ln\frac{\Lambda}{p}\Bigg\}
=\nonumber\\
&&
= 2\pi^4 + 4\pi^4 \sum\limits_i c_i \ln \frac{M_i}{\Lambda}.\end{aligned}$$
From equations (\[Integrals\_DK\_Integral\]), (\[Integrals\_Atan\_Integral\]), (\[Integrals\_LnQ2\]) and (\[Integrals\_LnQ\]) we see, that the divergent part of $I_8$ is equal to
$$I_8 = 2\pi^4\Bigg(\ln^2\frac{\Lambda}{p}
+ 2\ln \frac{\Lambda}{p}\Big(\sum\limits_i c_i \ln \frac{M_i}{\Lambda}
+ \frac{3}{2}\Big)\Bigg) + O(1).$$
In order to prove, that integrals $I_9$ and $I_{10}$ are finite at $\Lambda\to\infty$, first note, that
$$I_9 = I_9\Big(p/\Lambda\Big);\qquad I_{10}=I_{10}\Big(p/\Lambda\Big).$$
Therefore, it is necessary to prove, that $I_9(p=0)$ and $I_{10}(p=0)$ are finite constants. Let us set $p=0$ and make a substitution
$$k^\mu = \Lambda K^\mu;\qquad q^\mu = \Lambda Q^\mu.$$
Taking into account, that $M = a\Lambda$, where $a$ is a finite constant, the considered integrals can be written as
$$\begin{aligned}
\label{KQ_I9}
&& I_9 = \int \frac{d^4K}{(2\pi)^4}\frac{d^4Q}{(2\pi)^4}\,
\frac{2 (K+Q)_\mu Q_\mu}{
K^2 \Big(1 + K^{2n}\Big) \Big((K+Q)^2+a^2\Big)^2 \Big(Q^2+a^2\Big)^2};
\qquad\\
\label{KQ_I10}
&& I_{10} = \int \frac{d^4K}{(2\pi)^4}\frac{d^4Q}{(2\pi)^4}\,
\frac{a^2}{K^2 \Big(1 + K^{2n}\Big)
\Big((K+Q)^2+a^2\Big) \Big(Q^2+a^2\Big)^3}.\end{aligned}$$
The integrals over $Q$ are evidently convergent. However, it is necessary to check, that, after taking these integrals, the remaining integration over $K$ will be also convergent. Possible divergences can arise at $K\to 0$ or at $K\to\infty$. In the limit $K\to 0$
$$\begin{aligned}
&& \int \frac{d^4Q}{(2\pi)^4}\,
\frac{2 (K+Q)_\mu Q_\mu}{
\Big((K+Q)^2+a^2\Big)^2 \Big(Q^2+a^2\Big)^2} \to
\int \frac{d^4Q}{(2\pi)^4}\,\frac{2 Q^2}{
(Q^2+a^2)^4} = \frac{1}{24\pi^2 a^2};
\qquad\quad\\
&& \int \frac{d^4Q}{(2\pi)^4}\,
\frac{a^2}{\Big((K+Q)^2+a^2\Big) \Big(Q^2+a^2\Big)^3}\to
\int \frac{d^4Q}{(2\pi)^4}\,
\frac{a^2}{(Q^2+a^2)^4} = \frac{1}{96\pi^2 a^2}.\end{aligned}$$
It means that, in equations (\[KQ\_I9\]) and (\[KQ\_I10\]) the integration over $K$ is convergent if $K\to 0$. Similarly, in the limit $K\to\infty$
$$\begin{aligned}
&& \int \frac{d^4Q}{(2\pi)^4}\,
\frac{2 (K+Q)_\mu Q_\mu}{
\Big((K+Q)^2+a^2\Big)^2 \Big(Q^2+a^2\Big)^2} \approx
\int \frac{d^4Q}{(2\pi)^4}\,
\frac{2 (K+Q)_\mu Q_\mu}{(K+Q)^4 Q^4} = \frac{1}{8\pi^2 K^2};
\qquad\quad\\
&& \int \frac{d^4Q}{(2\pi)^4}\,
\frac{a^2}{\Big((K+Q)^2+a^2\Big) \Big(Q^2+a^2\Big)^3} \approx
\int \frac{d^4Q}{(2\pi)^4}\,
\frac{a^2}{(K+Q)^2 (Q^2+a^2)^3}
=\\
&& = \frac{a^2}{8\pi^2}
\int\limits_0^K dQ \frac{Q^3}{K^2 (Q^2+a^2)^3}
+ \frac{a^2}{8\pi^2} \int\limits_K^\infty dQ \frac{Q}{(Q^2+a^2)^3}
= \frac{1}{32\pi^2 (K^2+a^2)} \approx \frac{1}{32\pi^2 K^2},\quad\nonumber\end{aligned}$$
where we used equations (\[For\_I3\]) and (\[Angle\_Integral\]). Therefore, due to the presence of higher derivative term the integration over $K$ in equations (\[KQ\_I9\]) and (\[KQ\_I10\]) is also convergent at $K\to\infty$. Thus the integrals $I_9$ and $I_{10}$ are proven to be finite in the limit $\Lambda\to\infty$.
Collecting the above results we can finally write integrals (\[Integrals\_Definition\]) in the following form:
$$\begin{aligned}
\label{Integrals_Results}
&& I_1 = 2 \pi^2 \Bigg(\ln\frac{\Lambda}{p} + \frac{1}{2}\Bigg)
+ o(1);\nonumber\\
&& I_2 = 2\pi^2\sum\limits_i c_i \Bigg(\ln\frac{M_i}{p}
+ \sqrt{1+\frac{4M_i^2}{p^2}}\,
\mbox{arctanh}\sqrt{\frac{p^2}{4M_i^2 + p^2}}\Bigg);
\nonumber\\
&& I_3 = 4\pi^4 \ln\frac{\Lambda}{p} + O(1);\vphantom{\Bigg(}\nonumber\\
&& I_4 = \frac{\pi^2}{2} + o(1);\nonumber\\
&& I_5 = 4\pi^4\Bigg(\ln^2 \frac{\Lambda}{p} + \ln\frac{\Lambda}{p}\Bigg)
+ O(1);\nonumber\\
&& I_6 = 4\pi^4\Bigg(\ln^2 \frac{\Lambda}{p} + \ln\frac{\Lambda}{p}\Bigg)
+ O(1);\nonumber\\
&& I_7 = 2\pi^4\Bigg(\ln^2 \frac{\Lambda}{p}
+ 2\ln\frac{\Lambda}{p}\Bigg) + O(1);\nonumber\\
&& I_8 = 2\pi^4\Bigg(\ln^2\frac{\Lambda}{p}
+ 2 \ln \frac{\Lambda}{p}\Big(\sum\limits_i c_i \ln \frac{M_i}{\Lambda}
+ \frac{3}{2}\Big)\Bigg) + O(1);\nonumber\\
&& I_9 = O(1);\vphantom{\frac{1}{2}}\nonumber\\
&& I_{10} = O(1).\vphantom{\frac{1}{2}}\end{aligned}$$
[100]{}
S.Ferrara, B.Zumino, [*Nucl.Phys.*]{} [**B 87**]{}, 207, (1975).
T.E.Clark, O.Piquet, K.Sibold, [*Nucl.Phys.*]{} [**B 143**]{}, 445, (1978).
O.Piquet, K.Sibold, [*Nucl.Phys.*]{} [**B 196**]{}, 428, (1982);
O.Piquet, K.Sibold, [*Nucl.Phys.*]{} [**B 196**]{}, 447, (1982).
S.L.Adler, W.A.Bardeen, [*Phys.Rev.*]{} [**182**]{}, 1517, (1969).
L.D.Faddeev, A.A.Slavnov, [*Gauge fields, introduction to quantum theory*]{}, second edition, Benjamin, Reading, 1990.
S.L.Adler, J.C.Collins, A.Duncan, [*Phys.Rev.*]{} [**D 15**]{}, 1712, (1977).
O.V.Tarasov, V.A.Vladimirov, [*Phys.Lett.*]{} [**96 B**]{}, 94, (1980).
M.T.Grisaru, M.Rocek, W.Siegel, [*Phys.Rev.Lett.*]{} [**45**]{}, 1063, (1980).
W.Caswell, D.Zanon, [*Phys.Lett.*]{} [**100 B**]{}, 152, (1980).
L.V.Avdeev, D.Zanon, [*Phys.Lett.*]{} [**112 B**]{}, 356, (1982).
D.Jones, [*Phys.Lett.*]{} [**123**]{}, (1983), 45.
M.Grisaru, P.West, [*Nucl.Phys.*]{} [**B254**]{}, (1985), 249.
A.Vainstein, V.Zakharov, V.Novikov, M.Shifman , [*ZhETF Pis’ma*]{}, [**40**]{}, (1984), 161.
P.Dreitenlohner, D.Maison, K.Stelle, [*Phys.Lett.*]{}, [**134B**]{}, (1984), 242.
D.Jones, L.Mezincescu, [*Phys.Lett.*]{}, [**136B**]{}, (1984), 283; [**138B**]{}, (1984), 242.
D.Kazakov, [*ZhETF Pis’ma*]{} [**41**]{}, (1985), 272.
D.Jones, L.Mezincescu, P.West, [*Phys.Lett.*]{}, [**151B**]{}, (1985), 219.
V.Novikov, M.Shifman, A.Vainstein, V.Zakharov, [*Phys.Lett.*]{} [**151B**]{}, (1985), 169.
M.Grisaru, B.Milewski, D.Zanon, [*Phys.Lett.*]{} [**157B**]{}, (1985), 174.
M.Grisaru, B.Milewski, D.Zanon, [*Nucl.Phys.*]{} [**B266**]{}, (1986), 589.
M.Shifman, A.Vainstein, [*Nucl.Phys.*]{} [**B277**]{}, (1986), 456.
N.Arkani-Hamed, H.Mirayama, [*Holomorphy, rescaling anomalies and exact $\beta$-functions in supersymmetric theories*]{}, hep-th/9707133.
V.Novikov, M.Shifman, A.Vainstein, V.Zakharov, [*Phys.Lett.*]{} [**166B**]{}, (1985), 329.
W.Siegel, [*Phys.Lett.*]{} [**84 B**]{}, (1979), 193.
L.Avdeev, O.Tarasov, [*Phys.Lett.*]{} [**112 B**]{}, (1982), 356.
I.Jack, D.Jones, C.North, [*Nucl.Phys.*]{} [**B 473**]{}, (1996), 308
I.Jack, D.Jones, C.North, [*Phys.Lett.*]{} [**386 B**]{}, (1996), 138
I.Jack, D.Jones and C.North, [*N=1 supersymmetry and the three loop anomalous dimension for the chiral superfield*]{}, hep-ph/9603386.
I.Jack, D.Jones, [*Regularization of supersymmetric theories*]{}, hep-ph/9707278.
I.Jack, D.R.T.Jones, C.G.North, [*Nucl.Phys.*]{} [**B 486**]{}, 479, (1997).
D.Z.Freedman, K.Johnson, J.I.Latorre, [*Nucl.Phys.*]{} [**B371**]{}, (1992), 353.
J.Mas, M.Perez-Victoria, C.Seijas, [*The beta function of N=1 SYM in differential renormalization*]{}, hep-th/0202082.
A.Slavnov, [*Nucl.Phys.*]{} [**B31**]{}, (1971), 301.
T.Bakeyev and A.Slavnov, [*Mod.Phys.Lett.*]{} [**A11**]{}, (1996), 1539.
P.West, [*Nucl.Phys.*]{} [**B 268**]{}, (1986), 113.
C.Martin, F.Ruiz Ruiz, [*Nucl.Phys.*]{} [**B 436**]{}, (1995), 645.
M.Asorey and F.Falceto, [*Phys.Rev.*]{} [**D 54**]{}, (1996), 5290.
P.Pronin and K.Stepanyantz, [*Phys.Lett.*]{} [**B414**]{}, (1997), 117.
K.Konishi, [*Phys.Lett.*]{} [**135B**]{}, (1984), 439.
R.Bertlmann, [*Anomalies in quantum field theory*]{}, Clarendon press, Oxford, 1996.
J.Collins, [*Renormalization*]{}, 1984.
S.Ferrara, O.Piquet, [*Nucl.Phys.*]{} [**B 93**]{}, (1975), 261
P.Ramond, [*Field Theory: A modern primer*]{}, The Benjamin Cummings Publishing Company, Inc. Massachusetts, 1981.
12.0truecm
12.0truecm
(0,0)(0,0) (7.5,-3.9)[$\mbox{Re}(x)$]{} (4.8,-0.9)[$\mbox{Im}(x)$]{} (3.2,-4.1)[$-1$]{} (5.6,-4.1)[$1$]{} (2,-3.9)[$x_0$]{}
9.0truecm
[^1]: E-mail:[email protected]$
[^2]: E-mail:[email protected]$
[^3]: Note, that introducing of a term with higher covariant derivatives does not lead to regularization of one-loop divergences. For the one-loop divergences it is necessary to use one more regularization, for example, introduce Pauli-Villars fields.
[^4]: In our notations the metric tensor in the Minkowski space-time has the diagonal elements (1, -1, -1, -1).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The observations of the photometric space telescopes CoRoT and Kepler show numerous Blazhko RR Lyrae stars which have non-repetitive modulation cycles. The phenomenon can be explained by multi-periodic, stochastic or chaotic processes. From a mathematical point of view, almost periodic functions describe all of them. We assumed band-limited almost periodic functions either for the light curves of the main pulsation or for the modulation functions. The resulting light curves can generally be described analytically and it can also be examined by numerical simulations. This presentation is a part of our systematic study on the modulation of pulsating stars [@benko09; @benko11; @benko12].'
author:
- 'József M. Benkő, and Margit Paparó'
bibliography:
- 'Benko\_proc.bib'
title: 'Connections between Quasi-periodicity and Modulation in Pulsating Stars'
---
Definitions
===========
When we speak about quasi-periodic signals in physics we generally think about signals which can be described mathematically by almost periodic functions and not quasi-periodic ones. Before defining them let we remind the well-known definition of periodic function. $x(t)$ real function is [*periodic*]{} with the period $P$ if $$x(t)=x(t+P), \ \ {\mathrm {or}} \ \ \vert x(t) - x(t+P) \vert = 0 \ \
\ \mathrm{for \ all \ } t.$$ Its Fourier representation is $$\label{F0}
x(t)=\frac{a_0}{2} +
\sum_{n=1}^{\infty}{a_n \sin \left( 2\pi n f_0 t + \varphi_n \right)},$$ where $f_0$, $a_n$, $\varphi_n$ are constants.
The definition of almost periodic function is quite intuitive. $x(t)$ real function is a [*band-limited almost periodic function*]{} with the period $P$ if $$\label{Qp}
x(t)\approx x(t+P), \ \ {\mathrm {or}} \ \ \vert x(t) - x(t+P) \vert <
\varepsilon \ \ \ \mathrm{for \ all \ } t,$$ where $$0 < \varepsilon \ll \Vert x \Vert = \sqrt {x^2}=
\sqrt{\lim_{\tau\to\infty}\frac{1}{\tau}\int_{-\tau /2 }^{\tau
/2}{x^2(t)dt}. }$$ Its Fourier representation is $$\label{F1}
x(t)=\frac{a_0(t)}{2} +
\sum_{n=1}^{\infty}{a_n(t) \sin \left[ 2\pi \int_0^{t} f_n (\tau)d
\tau + \varphi_n(0) \right]}.$$ Taking into account Eq. (\[F1\]) the instantaneous frequency is $$f_n(t)=n f_0 (t) + \frac{1}{2\pi} \varphi_{n}^{\prime}(t),$$ where $f_0(t)$, $a_n(t)$, $\varphi_n(t)$ are changing in time, but slowly. An important consequence of this expression is that the “harmonics” are not exactly harmonic: $f_n(t) \neq nf_0(t)$!
For the numerical tests we need a base light curve. Since this light curve should be strictly periodic we use the form Eq. (\[F0\]) for producing it. In practice, we chose Fourier parameters of a typical RR Lyrae star’s light curve and an artificial light curve prepared from them as it is shown in Fig. \[fig:modell\].
\[fig:modell\] {height="18em"}
Quasi-periodic Pulsation {#sec:quasi}
========================
First we made quasi-periodic signals using the light curve in Fig. \[fig:modell\] with the assumptions: $f_0(t_{i+1})=f_0(t_i)+k^{\mathrm f} e_i$; $a_n(t_{i+1})=a_n(t_i)+k^{\mathrm a}_n e_i$ and $\varphi_n(t_{i+1})=\varphi_n(t_i)+k^{\varphi}_n e_i$, where $e_i$ is the standard white noise process and $k$-s are constants. The indices $i=1, 2, \dots, $ mean the consecutive sampling points. In other words, we supposed the time variability to be an [*auto-regressive (AR) random process*]{}. This approximation tests stochastic signals (see Fig. \[fig:quasi\]). The simulated light curves show correlated amplitude and frequency changes, but the time scale of the variations support only the longest period Blazhko cycles. The corresponding Fourier spectra show harmonics surrounded by side peaks (even multiplets), but generally no peaks can be detected in the low frequency regime where the Blazhko frequency $f_{\mathrm m}$ can be found for observed stars. If we change the zero point strongly some peaks appear in this low frequency range, but we parallelly get a strange random walk in the average brightness that we never observed in real stars.
If the quasi-periodicity is caused by [*(multi)periodic variation*]{} of which has a characteristic period(s) $f^{\mathrm m}_j \ll f_0$ for all $j$, we are facing a general amplitude and frequency [*modulated signal*]{}. Here the modulation functions depend on the harmonics as it was introduced by [@szeidl12]. It is easy to verify: since $a_n(t)$, $f_0(t)$ and $\varphi_n(t)$ functions are periodic, they can be represented by Fourier series in a form of (\[F0\]). Substituting these forms into expression (\[F1\]) we get the equations of [@szeidl12]. If the modulation function always depends on the harmonics of the Blazhko stars as stated by [@szeidl12], our formalism suggests that even the simplest Blazhko stars have a quasi-periodic (at least a multi-periodic) nature. This consequence, however, needs a careful check and validation.
\[fig:quasi\] {width="31em"}
Simultaneous Pulsation and Modulation
=====================================
As we have seen in the previous section the multi-periodicity is in close connection with the modulation. For simplicity, we investigated the modulated signals with global (harmonic independent) modulation functions as they were used in [@benko11]. We constructed light curves using the strictly periodic unmodulated light curve in Fig. \[fig:modell\]. It is changed with random modulation, namely, the modulation functions were assumed as $f^{\mathrm m}(t_{i+1})=f^{\mathrm m}(t_i)+k^{\mathrm f} e_i$ where $e_i$ and $k^{\mathrm f}$ are the white noise process and a constant term, respectively. As an alternative we tested cases where the modulation functions were strictly periodic and the pulsation varied randomly as it is described in Sec. \[sec:quasi\].
\[fig:mod\]
{height="23em"}
Since the average pulsation and modulation are described exactly the consecutive runs demonstrate the possible light curve and Fourier spectra deviations. The simulated light curves have very similar characteristics to the observed ones (see left panels in Fig. \[fig:mod\]). They show changes of the envelope curves, changes in the Fourier amplitudes of the modulation frequency and the side peaks; sometimes additional peaks appear (right panels in Fig. \[fig:mod\]). All of these features are in the same magnitude as the observed effects.
Conclusions
===========
We checked two options and their sub-cases: almost periodic pulsation alone and simultaneous pulsation and modulation where (at least) one of them was almost periodic. Random pulsation yields light curves with amplitude and frequency variation but they show significant differences compared to the observed light curves. We showed that the strictly periodic sub-case results in a general modulation description.
Simultaneous modulation and pulsation where one of them has quasi-periodic behaviour yields completely analogous light curves and Fourier spectra with the observed data. The key question is whether the random perturbation is caused by a stochastic or a chaotic process. The best method for determining which one is the so-called “phase space reconstruction” which will be the next step in our future analysis.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '[**Many-body localization (MBL) describes a quantum phase where an isolated interacting system subject to sufficient disorder displays non-ergodic behavior, evading thermal equilibrium that occurs under its own dynamics. Previously, the thermalization-MBL transition has been largely characterized with the growth of disorder. Here, we explore a new axis, reporting on an energy resolved MBL transition using a 19-qubit programmable superconducting processor, which enables precise control and flexibility of both disorder strength and initial state preparations. We observe that the onset of localization occurs at different disorder strengths, with distinguishable energy scales, by measuring time-evolved observables and many-body wavefunctions related quantities. Our results open avenues for the experimental exploration of many-body mobility edges in MBL systems, whose existence is widely debated due to system size finiteness, and where exact simulations in classical computers become unfeasible.** ]{}'
author:
- Qiujiang Guo
- Chen Cheng
- 'Zheng-Hang Sun'
- Zixuan Song
- Hekang Li
- Zhen Wang
- Wenhui Ren
- Hang Dong
- Dongning Zheng
- 'Yu-Ran Zhang'
- Rubem Mondaini
- Heng Fan
- 'H. Wang'
bibliography:
- 'MainText\_bib.bib'
title: 'Observation of energy resolved many-body localization'
---
[^1]
[^2]
[^3]
The phenomenon of MBL represents one of the paradigmatic examples of a typical out-of-equilibrium quantum phase transition [@Nandkishore2015; @Altman2018; @Abanin2019; @Alet2018]. It goes beyond the standard ground-state classification of the quantum matter, and its associated low-lying excitations. Instead, it is described by a high-energy phase transition, inherently manifested via the unitary dynamics of an isolated system, wherein by tuning the strength of disorder, one is able to halt the onset of ergodic behavior and thermalization [@Deutsch1991; @Srednicki1994; @Rigol2008; @Alessio2016]. As a direct consequence, the ability MBL systems possess in retaining information, naturally paves the way for a quantum information storage device. A fundamental ingredient of the MBL is the interplay of disorder and interactions, in contrast with a typical non-interacting Anderson localization [@Anderson1958], and hence the intricate balance of these two knobs triggers the onset of the out-of-equilibrium quantum phase transition.
The recent high-precision experimental simulation of quantum many-body systems, engineered to operate quantum mechanically up to large time scales compared to the characteristic equilibration times, have demonstrated the capability of observation of this phenomenon, originally described by analytical [@Basko2006; @Imbrie2016], and numerical means [@Znidaric2008; @Luitz2015; @Mondaini2015; @Wahl2019]. As a typical dynamical phase transition, the experimental investigations largely rely on the preparation of low entropy initial states – often product states with high fidelity – whose time-evolution is followed after a quench protocol. This was used to investigate the MBL transition, set off by growing disorder amplitudes, in experiments using ultracold atoms trapped by quasi-disordered optical lattices [@Schreiber2015; @Bordia2016; @Luschen2017; @Kohlert2019; @Rispoli2019], or in disordered settings [@Choi2016], as in the case of ion chains [@Smith2016] and, more recently, on quantum processors, emulated via transmon superconducting qubits [@Xu2018; @Roushan2017; @Neill2018]. In common, they follow a similar protocol: by checking how a few-body observable, tailored to be easily quantified on the initial state, evolves in time, a distinction can be drawn between ergodic and localized behavior, naturally occuring below and above the critical disorder strength. However, experimental explorations of this transition have often neglected the energy dependence on the onset of localization, and the precise understanding of how the triggering of the MBL phase is influenced by this extra variable has direct consequences to potential technological applications on a quantum memory instrument. Figure \[fig:Fig0\] direct exemplifies a situation where information of the initial conditions can be or not preserved depending on the energy of the isolated system, evolved under unitary dynamics.
Here, by using a newly designed 20-qubit quantum processor and flexibly programming $N = 19$ of the qubits (see Fig. \[fig:Fig1\]A), we are able to have a large control of interactions and disorder, which, combined with a precise initial state preparation, enables a direct probe of the energy-disorder phase diagram of the many-body localized phase transition with remarkable accuracy. Our insight is to perceive that in a quench problem, the initial state fully encodes the total energy of the system (within the regime the system is yet isolated from perturbations from the environment), and, as such, would allow to probe when localization occurs with energy resolution. That is, by initializing a product (Fock) state $|\Psi_0\rangle$, the unitary time-evolved wave function $|\Psi_t\rangle = e^{-{\rm i}Ht}|\Psi_0\rangle$ preserves the total energy $\langle\Psi_t|H|\Psi_t\rangle=\langle\Psi_0|H|\Psi_0\rangle = E$, under the effective Hamiltonian of the superconducting quantum processor (see Supplementary Material) [@Song2017; @Xu2018; @Song2019], $$\begin{aligned}
\label{Ham}
\frac{H}{\hbar} &=& \sum_{\{m,n\}\in N} J_{mn}\left({\sigma^+_{m} \sigma_{n}^- + \sigma^-_{m} \sigma_{n}^+}\right) \nonumber \\
&& + \sum_{m} V_m \sigma^+_{m} \sigma_{m}^-,\end{aligned}$$ where $\sigma^-_{m}$ ($\sigma^+_{m}$) is the lowering (raising) operator for qubit $Q_m$, and the first term runs at pairs of qubits $Q_m$ and $Q_n$. The effective coupling strength $J_{mn}$ between pairs of qubits is schematically represented in Fig. \[fig:Fig1\]B, in which “nearest-neighbor” couplings are much larger. But the system possesses smaller long range couplings, resulting in a non-integrable Hamiltonian (Eq. 1) even in the absence of disorder [@Xu2018]. The second term in Eq. 1 is the disordered potential $V_m$ of the $m$-th qubit, which can be flexibly set by individually adjusting its resonant frequency, without noticeably altering the values of $J_{mn}$. Finally, to mimic a fully disordered system, we choose $V_m$ from a uniform random distribution $[-V,V]$.
The protocol we follow is: (i) With the $N=19$ qubits in $|0\rangle$, we prepare initial product states via $\pi$ pulses on individually selected $N_1 = 9$ qubits, producing generic states $|\Psi_0\rangle$. (ii) For a given disorder realization $\{V_m\}$, we compute the total energy of the system, where only diagonal terms in $H$ have a finite contribution: $E/\hbar=\sum\limits_{m \in N_1} V_{m}$, which sums over the 9 qubits initialized in $|1\rangle$. (iii) We estimate, for this $\{V_m\}$ realization, where the total energy $E$ lies in the energy density spectrum, $$\varepsilon = \frac{E-E_{\rm min}}{E_{\rm max}-E_{\rm min}}.$$ The extremal eigenvalues of $H$, $E_{\rm min}$ and $E_{\rm max}$, are easilly obtainable by numerical means, without resorting to full diagonalization of $H$. (iv) As a direct quantification of the preservation of the information encoded in the initial state, we measure a few-body (and local) observable, the generalized imbalance described as $${\cal I}_{\rm gen} = \sum\limits_{m=1}^{19} \beta_m\sigma^+_m \sigma^-_m,$$ where $\beta_m = 1/N_{1} \ (-1/N_{0})$ on the qubit $m$ initialized to $|1\rangle$ ($|0\rangle$); $N_1 =9$ and $N_0 = 10$ define the occupancy of our many-body system, where we select to excite 9 out of the $N=19$ qubits, in order to emulate the largest possible Hilbert space with our quantum processor, for a conserved total magnetization. The measurements are carried out for $k=20$ disorder realizations, together with carefully selecting the initial states, to fill a mesh of energy densities $\delta\varepsilon = 0.05$. This observable is similar to the charge imbalance often used in optical lattice experiments, when preparing a charge density wave [@Schreiber2015; @Bordia2016; @Luschen2017; @Kohlert2019], or a Néel state for trapped ions [@Smith2016], with the advantage of being specifically customized for whichever initial state with $\langle\Psi_0|{\cal I}_{\rm gen}|\Psi_0\rangle=1$.
The experimental pulse sequence to execute the abovementioned protocol is shown in Fig. \[fig:Fig1\]C, which consists of $\pi$ pulses for exciting qubits, square pulses for tuning qubit frequencies, and simultaneous readout for obtaining multiqubit probabilities. Our single-qubit $\pi$ pulses are calibrated to be around 0.996 in fidelity using randomized benchmarking, which ensures that the state preparation is accurate. The readout fidelity values are around 0.97 (0.92) on average for $|0\rangle$ ($|1\rangle$) as measured simultaneously for all 19 qubits, which are used to correct the multiqubit probabilities for elimination of the readout errors before further processing of the data. Additionally, we can reliably obtain the quantum unitary evolution up to times $t =1500$ ns, which are much shorter than the typical energy relaxation times of the qubits (in the range from 30 to 70 $\mu$s) and much longer than the average shortest tunneling times $\tau = 1/({J_{m,m+1}^{\rm ave}}) \approx 60$ ns, sufficient to observe equilibration before decoherence processes ultimately affect the results (See Supplementary Material).
Figure \[fig:Fig2\][A]{} displays the disorder averaged generalized imbalance ${\cal I}_{\rm gen}$ at $t=1000$ ns: A characteristic ‘$D$’-shape structure on the energy density-disorder amplitude ($\varepsilon$-$V$) diagram highlights that loss of memory of the initial product states is both disorder- and energy-dependent in our $19$-qubit quantum processor, for a Hilbert space size with 92,378 states. Closely matching results can be obtained from exact numerical simulation using experimentally measured qubit couplings, and render a very similar diagram for ${\cal I}_{\rm gen}$ at equivalent time scales.
From numerical simulation, many different quantities can characterize the onset of the MBL transition [@Luitz2015; @Mondaini2015]. The simplest is to look at the statistics of the gaps in the eigenenergy spectrum of $H$. Thermalization is intimately connected with the manifestation of quantum chaotic behavior, associated with eigenenergy level repulsion, whereas for MBL, disorder renders uncorrelated energies, resulting in Poissonian distribution of level spacings [@Alessio2016]. Experimentally, a spectroscopic analysis is rather elusive due to the exponentially dense energy spectra, but it was done on a 9-qubit device, when dealing with up to 45 states forming the Hilbert space [@Roushan2017]. One commonly tracks the average value of the ratio of adjacent gaps $r_\alpha=\min(\delta_\alpha,\delta_{\alpha+1})/\max(\delta_\alpha,\delta_{\alpha+1})$, with $\delta_\alpha$ a gap in the energy spectrum between consecutive levels $E_\alpha$ and $E_{\alpha+1}$; ergodic and localized regimes possess average values $\langle r\rangle$ approximately equal to 0.53 and 0.39, respectively [@Atas2013]. By restricting this analysis to different regions of the spectrum, we can construct the simulated MBL phase diagram for our case (Fig. \[fig:Fig2\]B), after using the experimentally measured $J_{mn}$. The striking similarity between the experimental and simulated phase diagrams, reassures the indication of energy-dependent localization transition. Focusing on the generalized imbalance at intermediate disorder ($V/2\pi= 16~{\rm MHz}$) in Fig. \[fig:Fig2\]C, experimental and numerical results exhibit a remarkable agreement, enabling us to infer that our finite system prepared at low energy densities ($\varepsilon = 0.15$ and $0.3$) will fail to thermalize $[{\cal I}_{\rm gen}(t\to\infty)\neq0]$, since infinite-time averages [@Alessio2016; @Mondaini2015; @SM] are readily available from numerics.
Further characterization of the transition can be done by noticing that subdiffusive equilibration dynamics precedes the onset of the localized phase, which can be traced to the formation of bottlenecks of rare regions in the disordered potential, thereby hampering energy transport. These are often referred as Griffiths regions [@Griffiths1969; @Potter2015; @Vosk2015; @Luitz2016], and have been experimentally shown to lead to a power-law decay in time of the imbalance, ${\cal I_{\rm gen}}(t)\propto t^{-\xi}$, with a disorder-dependent exponent $\xi$, that vanishes at the transition $V_c$ [@Luschen2017; @Kohlert2019]. Here, since we observe that the onset of localization depends on the energy of the system, the regime of manifestation of subdiffusive behavior is also $\varepsilon$-dependent. We further report in Fig. \[fig:Fig3\]A the disorder averaged imbalance at different times with energy resolution, using disorder strength $V/2\pi = 4, 16$ and $50\ {\rm MHz}$. Figure \[fig:Fig3\][B]{} displays similar results when focusing on the central part of the energy spectrum ($\varepsilon = 0.5$). A variety of MBL studies performed to date mostly focus on this general trend: the reduction of the decaying exponent $\xi$ under the increase of disorder. We extend this analysis to include energy dependence of this behavior. When checking the decay with time, subsequent to the initial dynamics displaying coherent oscillations (Figures \[fig:Fig2\][C]{} and \[fig:Fig3\][B]{}), we are able to extract the disorder controlled $\xi$ exponent, that is dependent on both energy density and disorder (Fig. \[fig:Fig3\][C]{}). The compilation of the values of $\xi$ (Fig. \[fig:Fig3\]C) shows that $V_c = V_c(\varepsilon)$: it is typically easier to trigger localization when the system is not at the regime of infinite temperatures.
The degree of control of our quantum processor enables direct determination of quantities intrinsically related to the time-evolved wavefunction, not previously investigated in experiments. As an example, we are able to trace the time evolution of the multiqubit probabilities related to each Fock state $|n\rangle$, $p_n(t) = |\langle n|\Psi_t\rangle|^2$, allowing the quantification of how fast $|\Psi_t\rangle$ spreads over the Hilbert space. A direct measure of this is provided by the participation ratio, ${\rm PR} = \left[\sum_n p_n^2(t)\right]^{-1}$, shown in Fig. \[fig:Fig4\]A, that compares $\varepsilon=0.15$ and $0.5$. The contrast is evident: The fast spread at the larger energy density is dictated by ergodic behavior, whereas at $\varepsilon=0.15$, the slower growth of PR suggests the onset of non-ergodicity, with an order of magnitude difference at the longest experimental time.
A distinctive aspect of the MBL phase is the characteristic slow growth of the entanglement entropy ($\propto \log t$), a result directly connected to the exponential localization of its emerging set of local integrals of motion, the $\ell$-bits [@Nandkishore2015; @Vosk2015; @Abanin2019]. This should be contrasted with the ballistic spreading of entanglement occurring in thermalizing systems, with diffusive energy transport [@Kim2013]. It has been tested numerically [@Bardarson2012], and experimentally measured with quantum state tomography (QST) [@Xu2018] and indirectly probed by $n$-point correlations [@Lukin2019; @Rispoli2019]. Since QST is extremely time consuming for a large number of qubits, here we implement the quantum Fisher information, defined as ${\cal F}_{Q}(t) = \langle{\cal I}_{\rm gen}^2(t)\rangle - \langle{\cal I}_{\rm gen}(t)\rangle^2$, for a lower bound of the entanglement entropy. After the initial transient, ${\cal F}_{Q}$ also grows logarithmically in time within a regime governed by the MBL phenomenon. Figure \[fig:Fig4\][B]{} compares the time evolution at intermediate disorder strengths ($V/2\pi=16\ {\rm MHz}$) with energy densities $\varepsilon =$ 0.15 and 0.5. In the latter, the fast growth of ${\cal F}_{Q}(t)$ approaching the regime of saturation indicates thermal behavior, whereas at smaller $\varepsilon$’s, the growth is logarithmically slow, suggesting localization.
By using a highly programmable quantum processor with 19 superconducting qubits, we provide a phase diagram of the MBL transition with unprecedented level of detail, resolving the energy dependence of the critical disorder. This is done by checking the evolution of a generalized imbalance, that encodes information on the initial state preparations. Exact numerical characterization of the emulated system is in excellent agreement with the experimental realization, suggesting the coexistence of localization and thermalization for the same disorder amplitude, albeit at different energy densities. Furthermore, our processor is scalable, with the potential to probe the yet unsolved question of the possible existence of mobility edges. The present experimental results suggest so, but a proper scaling with larger number of qubits in the regime where numerical simulation using classical computers becomes intractable, could possibly settle down this question.
Acknowledgments {#acknowledgments .unnumbered}
===============
RM thanks discussions with Rosario Fazio and Marcos Rigol. Devices were made at the Nanofabrication Facilities at Institute of Physics in Beijing and National Center for Nanoscience and Technology in Beijing. [**Funding:**]{} Supported by National Natural Science Foundation of China (Grants No. NSAF-U1930402, 11674021, 11974039, 11851110757, 11725419 11434008, and 11904145), and National Basic Research Program of China grants (Grants No. 2017YFA0304300 and 2016YFA0300600). [**Author contributions:**]{} C.C. and R.M. proposed the idea; C.C. and Z.-H.S. performed the numerical simulation; Q.G. conducted the experiment; H.L. and D.Z. fabricated the device; R.M., C.C., Q.G., Z.-H.S., H.F., and H.W. cowrote the manuscript; and all authors contributed to the experimental setup, discussions of the results, and development of the manuscript. [**Competing interests:**]{} Authors declare no competing interests. [**Data and materials availability:**]{} All data needed to evaluate the conclusions in the paper are present in the paper or the supplementary materials.
{width="85.00000%"}
{width="95.00000%"}
{width=".8\textwidth"}
{width="95.00000%"}
{width=".7\textwidth"}
[^1]: Those authors contributed equally to this work.
[^2]: Those authors contributed equally to this work.
[^3]: Those authors contributed equally to this work.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In the paper I study properties of random polynomials with respect to a general system of functions. Some lower bounds for the mathematical expectation of the uniform and recently introduced integral-uniform norms of random polynomials are established.\
[Key words and phrases:]{} Random polynomial, estimates for maximum of random process, integral-uniform norm.
author:
- 'P.G. Grigoriev'
title: ' Random linear combinations of functions from $L_1$'
---
[^1]
Introduction
============
In the article we research random polynomials (linear combinations of functions with random coefficients) of the type $$\label{polinom2}
F_n(\omega,x):=\sum_{i=1}^n \xi_i(\omega)f_i(x),$$ where $\{f_i\}_{1}^n$ is a system of functions on a measure space $(X,\mu)$ and $\{\xi_i\}_{i}^n$ is a set of independent random variables on a probability space $(\Omega,{\sf P})$ (“random coefficients”). Our interest will be concentrated on the properties of the random variable $\|F_n(\omega,\cdot)\|_{B(X)}$, where ${\|\cdot\|_B}$ is a norm in a certain space $B$ consisting of functions on $(X,\mu)$.
Note, when ${B=L_\infty}$, the formulated above research topic becomes the classical problem of estimating the expectation of the supremum of random polynomial (\[polinom2\]). In 1954 Salem and Zygmund [@sz] established that $$\label{salz}
{\sf E}\big\|\sum_{k=-n}^nr_k(\omega)e^{ikx}\big\|_\infty
\equiv \frac1{2^{2n+1}}
\hspace{-.6em}
\sum_{{\genfrac{}{}{0pt}{}{\text{over all choices}}{\text{of signs:}\
\varepsilon_k=\pm1}}}
\hspace{-.4em}
\Big\|\sum_{k=-n}^n\varepsilon_ke^{ikx} \Big\|_\infty\!
\asymp\sqrt{n\log n},$$ where $r_k(\omega)=\text{sign}\sin(2^{k+1}\pi\omega)$ are the Rademacher functions on $[0,1]$ and ${\sf E}$ denotes the mathematical expectation. However, by the virtue of Khinchin’s inequality it follows that $${\sf E}\big\|
\sum_{k=-n}^n r_k(\omega)e^{ikx}\big\|_p
\asymp_p\sqrt{n},
\quad\ 1\le p<\infty.$$ These facts demonstrate that the embeddings of the spaces of trigonometric polynomials into $L_\infty$ and $L_p$ with ${p<\infty}$ are qualitatively different.
By now a number of estimates for the uniform norm of random polynomials (\[polinom2\]) with various constraints on $\{\xi_i\}_1^n$ and $\{f_i\}_{1}^n$ have been established. Note while the upper estimates have many important applications in analysis and probability, the applications of the lower estimates are rather rare. Nevertheless, proofs of the lower estimates are more challenging and usually need involving some specific properties of systems $\{f_i\}$ and $\{\xi_i\}$. E.g. the proof of the lower bound in (\[salz\]) in [@sz] essentially relied on the facts that the trigonometric functions is a system of [*characters*]{} (i.e. formulae like ${\cos(\alpha+\beta)}={\cos\alpha\cdot\cos\beta
-\sin\alpha\cdot\sin\beta}$ were used) and that for distribution of linear combinations of Rademacher functions an exponential estimate holds. In [@ks] one can find a short proof of a sharp lower estimate for the random polynomial $\sum_{k}\xi_k\cos kx$, where $\{\xi_k\}$ are independent Gaussian variables, this proof relies on the fact that the matrix $\big(\cos \frac{2\pi}n jk\big)_{j,k=0}^{n-1}$ is a multiple of an orthogonal matrix and that uncorrelated Gaussian variables are independent. Proofs of the upper estimates normally use a simple averaging argument and need essentially weaker constraints on system $\{f_i\}$. Usually, if an upper estimate for the expectation of a norm of the random polynomial (\[polinom2\]) takes place, then essentially the same estimate holds if we substitute in (\[polinom2\]) the functions $f_i$ with their absolute values $|f_i|$. In this paper, continuing the work [@ja3], there is established a lower estimate for the uniform norm of random polynomials with respect to an abstract functional system $\{f_i\}_1^n$ from $L_1$, provided that the system satisfies a weak condition.
In monographs [@ledouxtal], [@marcuspis] one can find a deep theory which enables to estimate the expectation of the uniform norm of random polynomials (\[polinom2\]), provided $\{f_i\}$ is a system of characters of a locally compact Abelian group restricted to a compact neighborhood $V$ of the group identity element. The basic method for estimating the supremum of a random process in [@ledouxtal], [@marcuspis] (in particular, of a random polynomial (\[polinom2\])) is reduction of the problem to a problem of estimating the $\varepsilon$-entropy of $V$ with respect to a metric induced by the random process. This method was originally introduced by Dudley [@dud] and Sudakov [@sud1], [@sud2] and later was developed by Fernique, Marcus, Pisier, Talagrand and others (see [@ledouxtal], [@marcuspis], [@sud3], [@fern]). To obtain a lower estimate for the maximum of a random process using this method, one has to apply a variant of Slepian’s lemma [@sl] for Gaussian vectors. This lemma enables one to estimate from below the probability $${\sf P}\big\{\max_{1\le k\le m}
\sum_{i=1}^n\xi_if_i(x_k) \ge\alpha\big\},
\qquad\{x_k\}_1^m \text{\ is a net in\ }X,$$ provided that [**all**]{} normalized inner products (cosines of angles): $$\frac{(W_{x_j},W_{x_k})}{|W_{x_j}|\cdot|W_{x_k}|}
,\quad
{j\ne k},\quad~
\text{for vectors}\quad
W_{x_j}:=\big(f_i(x_j)\big)_{i=1}^n\in{\Bbb C}^n,
\quad~ j=1,\dots,m$$ are sufficiently small and $\{\xi_i\}$ are independent Gaussian variables. Demonstrating the existence of such a net $\{x_j\}$ is usually a separate non-trivial problem. Moreover, since Slepian’s lemma can be applied only to Gaussian vectors there arise some serious difficulties with transfer of the estimates for random polynomials with Gaussian coefficients to the case of non-Gaussian $\{\xi_i\}$.
In 1995 Kashin and Tzafriri [@kt1], [@kt2], [@kt3] introduced another method for obtaining lower estimates of the uniform norm of random polynomials. In particular, in [@kt1] it was shown that the lower estimate in the Salem-Zygmund theorem (\[salz\]) stays true for random polynomials with respect to an [**arbitrary**]{} orthonormal system, provided that the system is uniformly bounded in $L_3$. This approach relies on a version of the central limit theorem with an estimate of the error term. To apply this method it is not necessary to estimate [**all**]{} angles between the vectors $W_{x_j}$, instead of that it suffices to demonstrate that cosines of these angles are [**small on average**]{}.
In [@kt2] Kashin and Tzafriri defined the following norm $$\label{norma}
\|f\|_{m,\infty}:=\int_X...\int_X \max\{|f(x_1)|,...,|f(x_m)|\}
d\mu(x_1)...d\mu(x_m),$$ where $f$ is an integrable function on a measure space $(X,\mu)$, ${\mu(X)=1}$. As in author’s works [@ja2], [@ja3] let us call it [*integral-uniform norm*]{} for this norm. While obtaining the lower estimate for the expectation of the uniform norm of the random polynomial (\[polinom2\]) in [@kt1], [@kt3], a similar estimate for the [*integral-uniform*]{} norm with parameter ${m\asymp n^{1/2+\varepsilon}}$ was [*de facto*]{} proved. It is worth mentioning that estimates for the norm $\|f\|_{m,\infty}$ (or rather for the [*family*]{} of norms with ${m\ge1}$) are self-interesting, since the values of the norms carry quite full information about the distribution of function $f$: $\lambda_f(t)={\mu\{x\in X:|f(x)|\ge t\}}$. Indeed, it is easy to see that $$\label{norma2}
\|f\|_{m,\infty}=\int_0^\infty
\big(1-(1-\lambda_f(t))^m\big)dt$$ and ${\|f\|_1\equiv\|f\|_{1,\infty}}$${\le\|f\|_{m,\infty}\le\|f\|_\infty}$, moreover, ${\|f\|_{m,\infty}\!\to\|f\|_\infty}$ as ${m\to\infty}$. In Section \[auxiliary\] (Th. \[kee\]) it will be shown that $$\|f\|_{m,\infty}
\asymp
\sup_{{\genfrac{}{}{0pt}{}{\Delta\subset X}{\mu\Delta=1/m}}}
\Big\{m\int_\Delta|f|\,d\mu\Big\}$$ provided there exists a subset of $X$ of measure $1/m$.
Using a simple modification of the method from [@kt1], [@kt3], the author [@ja3] proved the following
[**Theorem A.**]{}\[genth\]
*Let $\{f_i\}_{i=1}^n$ and $\{\xi_i\}_{i=1}^n$ be systems of functions defined on probability spaces $(X,\mu)$ and $(\Omega,{\sf P})$ respectively, satisfying:*
- $\|f_i\|_2=1$ and $\|f_i\|_{2+\varepsilon}\le M$ for all $i=1,\ldots,n$ with some constants ${M>0}$, ${\varepsilon>0}$.
- \[bstar\] $
\big\|\sum\limits_{i=1}^n c_i f_i\big\|_2\le
M R^p
\big(\sum\limits_{i=1}^n|c_i|^2\big)^{1/2}$ for all sets of coefficients $\{c_i\}_{1}^n$ with some constants $p\in[0,\frac12)$, ${M>0}$, where $
R\equiv
R(\{a_i\}_1^n):=\frac{(\sum_{i=1}^n|a_i|^2)^2}{\sum_{i=1}^n|a_i|^4}$ and $\{a_i\}_{1}^n$ is a [**fixed**]{} set of complex coefficients.
- $\{\xi_i\}_1^n$ is a system of independent variables such that ${\sf E}\xi_i=0$, ${{\sf E}|\xi_i|^2=1}$ and ${({\sf E}|\xi_i|^{2+\varepsilon})\le M}$
Then there exist positive constants ${q=q(p)}$, ${C_j=C_j(p,M,\varepsilon)}$, ${j=1,2,3}$, such that[^2]$$\label{genth1}
{\sf P}\Big\{\omega\in\Omega :
\big\|\sum_{i=1}^n a_i\xi_i (\omega )f_i\big\|_{m,\infty}
\le C_1 \Big(\sum_{i=1}^n|a_i|^2\log P\Big)^{1/2}
\Big\}
\le \frac{C_2}{P^q},$$ where $P=\min(m,R)+1$ and $R=R(\{a_i\})$ are defined in condition [**(b)**]{}. It obviously implies $$\label{genth2}
{\sf E}\Big\|\sum_{i=1}^n a_i\xi_i f_i\Big\|_{m,\infty}
\ge C_3\Big(\sum_{i=1}^n|a_i|^2\log P\Big)^{1/2}.$$ (In the case of small values of $m$ or $R$ the inequality (\[genth2\]) easily follows from Khinchin’s and Hölder’s inequalities and the trivial estimate: $\|\cdot\|_{m,\infty}\ge \|\cdot\|_1$.)
Theorem A is a simple generalization of the results from [@kt1]–[@kt3], where the of random polynomials was estimated. There estimates of type (\[genth1\]), (\[genth2\]) for the uniform norm were proved in the case ${p=0}$ and was noted that if ${a_i\equiv1}$, these estimates could be generalized for the case ${p\in[0,1/2)}$. It was also noted that the established estimates stay true for the integral-uniform norm with parameter ${m\asymp R^{1/2}}$.
Roughly speaking, Theorem A implies that an estimate of type (\[genth2\]) for a random polynomial (\[polinom2\]) does [**not**]{} take place only if the system $\{f_i\}$ (of normalized in $L_2$ functions) is “significantly far” from an orthonormal one (e.g. if the functions of the system converge to a fixed function).
Note when ${a_i\equiv1}$ (${R(\{a_i\})=n}$), applying Lemma \[geomlem\] (see below), one can easily deduce the condition [**(b)**]{} for functions $\{f_i\}$, satisfying ${\|f_i\|_2=1}$, from the condition:
- $\|\sum_{i=1}^n \varepsilon_if_i\|_2\le
Mn^{\frac12+\beta}$ for all choices of signs $\varepsilon_i=\pm1$\
with some ${\beta\in[0,\frac12)}$, ${M>0}$.
In [@ja3] it was also shown that provided ${m\le R}$ and some additional constraints on $\{\xi_k\}_1^n$ the lower estimate (\[genth2\]) is sharp in sense of order. It was demonstrated by the following
[**Theorem B.**]{}\[upper\] [*Let $\{\xi_k\}_1^n$ be a system of independent variables such that the following exponential estimate holds: $$\label{exp}
{\sf P}\Big\{\big|\sum_{k=1}^n c_k\xi_k\big|> t\big(\sum_{k=1}^n
c_k^2\big)^{1/2}\Big\} \le C_4e^{-t^2C_5}$$ for all sets of coefficients $\{c_k\}_{1}^n$ with some constants $C_4$, ${C_5>0}$. Then there exists a constant $C_6=C_6(C_4,C_5)>0$ such that $$\label{upper1}
{\sf E}\big\|\sum_{k=1}^n\xi_k f_k\big\|_{m,\infty} \le
C_6\big\| \big(\sum_{k=1}^n|f_k|^2\big)^{1/2}\big\|_{m,\infty}
\cdot\sqrt{1+\log m}$$ [**for all**]{} systems of functions ${\{f_k\}_1^n\subset L_1(X,\mu)}$ (${\mu X=1}$) and for all ${m\ge1}$.* ]{}
Note that the condition imposed in Theorem A, which requires the functions $\{f_i\}$ to be [*uniformly bounded*]{} in $L_{2+\varepsilon}$, looks rather unnatural since the theorem provides [*lower*]{} estimates. Roughly speaking, the reason for necessity of such a condition is that the uniform boundness of system $\{f_i\}$ in $L_{2+\varepsilon}$, along with its orthonormality (or a weaker condition [**(b$'$)**]{} combined with ${\|f_i\|_2=1}$), ensures that the “essential” supports of functions $f_i$ mutually intersect “strongly enough”. To convince that the condition of uniform boundness cannot be simply omitted from Theorem A consider the example of the functional system ${f_i:=\sqrt{n}\chi_i}$ on $[0,1]$, where $\chi_i$ are the indicators of the intervals $\big(\frac{i-1}n,\frac in\big)$.
The main target of the article is to generalize Theorem A in the particular case ${a_i\equiv1}$ for the random polynomials of type (\[polinom2\]) with respect to a system of functions $\{f_i\}_1^n$, ${\|f_i\|_1=1}$, [*which are not necessarily bounded*]{} in $L_p$, ${p>1}$. To avoid extreme functional systems, such as one in the previous paragraph, the conditions [**(a)**]{} and [**(b$'$)**]{} are substituted for
- $\|f_i\|_1=1$ for all $i=1,\cdots,n$ and\
$\|\sum_{i=1}^n\theta_if_i\|_1\le
Mn^{\frac12 +p}$ for all choices of signs $\{\theta_i\}_1^n$, $\theta_i=\pm1$
with some constants $p\in[0,\frac1{12})$, $M>0$. The main result is the following
\[l1th\] Let $\{f_i\}_{i=1}^n$ be a system of functions on a probability space $(X,\mu)$ which satisfies the condition [**(d)**]{} with $p\in[0,\frac1{12})$. Let $\{\xi_i\}_{i=1}^n$ be independent variables defined on another probability space $(\Omega,{\sf P})$, satisfying ${{\sf E}\xi_i=0}$, ${{\sf E}|\xi_i|^2=1}$ and ${{\sf E}|\xi_i|^3\le M^3}$. Then there exist some constants $q'=q'(p)>0$, $C'_j=C'_j(p,M)>0$, $j=1,2,3$, such that whenever ${m\le n}$ $$\label{l1th1}
{\sf P}\Big\{\omega\in\Omega :
\| F_n(\omega,x)
\|_{m,\infty} \le C'_1
\sqrt{n\cdot(1+\log m)} \Big\} \le
{C'_2}{m^{-q'}}$$ and, consequently, $$\label{l1th2}
{\sf E}\|\sum_{i=1}^n \xi_i f_i\|_{m,\infty} \ge C'_3
\big(n\cdot(1+\log m)\big)^{1/2},$$ where $F_n$ is a random polynomial defined by (\[polinom2\]). (For small $m$ inequality (\[l1th\]) follows from Khinchin’s inequality.)
This result provides a new estimate not only for the integral-uniform norm but also for the uniform norm of random polynomials (\[polinom2\]):
\[l1uniform\] Let functions $\{f_i\}_1^n$ and random variables $\{\xi_i\}_1^n$ satisfy the conditions of Theorem \[l1th\]. Then for the [*uniform*]{} norm of the random polynomial (\[polinom2\]) the following estimate holds $${\sf P}\Big\{\omega\in\Omega :
\| F_n(\omega,x)\|_{\infty} \le C'_1
\sqrt{n\cdot(1+\log n)} \Big\} \le
{C'_2}{n^{-q'}}.$$
Note that Theorem \[l1th\], being stronger than Theorem A in some environments, is weaker than that in the two aspects: first, it cannot be applied to polynomials of type $\sum_k a_k\xi_k(\omega) f_k(x)$ with an arbitrary choice of non-random coefficients $\{a_i\}$; second, it imposes the condition [**(d)**]{} with parameter $p<\frac1{12}$, while in the corresponding condition for Theorem A (condition [**(b)**]{} provided ${a_i\equiv1}$ or [**(b$'$)**]{}) it suffices to have ${p<1/2}$. I think that the constraint ${p<\frac1{12}}$ can be relaxed[^3] (see Conjecture \[conja1\] below). In [@jadis] a conjecture about possible generalization of Theorem \[l1th\] for the case of polynomials $\sum_k a_k\xi_k f_k$ (with non-trivial $a_k$) is formulated.
The organization of the paper is as follows. In Section \[auxiliary\] we shall prove several auxiliary results which were explicitly or implicitly used in [@kt1], [@kt3], [@ja3]. We gather these results with the intention of making the method developed in [@kt1], [@kt3], [@ja3] easier to understand, apply and make appropriate alterations, e.g. for generalizing of Theorem A for the case of specific values of the power parameter $q$ (see Footnote \[powersnosk\] and Remark \[powerq\] below). In Section \[secl1\] we shall prove Theorem \[l1th\]. In Section \[applic\] Theorem \[l1th\] will be applied to give a partial solution for a Functional Analysis problem formulated by Montgomery-Smith and Semenov in [@sem] and to estimate the Marcinkiewicz norm of random polynomials. In Section \[applic\] we shall also formulate two hypotheses, concerning the potential generalizations of Theorem \[l1th\].
I would like to express my special gratitude to B.S. Kashin, whose advisement led me to establish the results of the paper, I thank also E.M. Semenov and A.M. Zubkov for valuable remarks and discussions.
Auxiliary Results {#auxiliary}
=================
[**Integral-Uniform Norm.**]{} Let us check that the definitions (\[norma\]) and (\[norma2\]) of the integral-uniform norm are identical. It is well-known that ${\|g\|_1=\int_0^\infty\lambda_g(t)\,dt}$ for every function ${g\in L_1(Y,\nu)}$, where $\lambda_g(t):={\nu\{y:|g(y)|\ge t\}}$ is the distribution of $g$. Thus, to prove equivalence of (\[norma\]) and (\[norma2\]) it suffices to notice that the function $$g(\bar x)=\max\big\{|f(x_1)|,...,|f(x_m)|\big\},
\ \quad \bar x=(x_j)_1^m\in X^m=:Y,$$ has the distribution: $\lambda_g(t)=1-(1-\lambda_f(t))^m$.
By inequality ${\max(|a|,|b|)\le|a|+|b|}$ it is easy to see that $$\label{qw4}
\|f\|_{m,\infty}\le m\|f\|_1\quad\mbox{and}\quad
\|f\|_{n,\infty}\le\frac{n+1}m\|f\|_{m,\infty}
\ \quad\mbox{for all $f\in L_1(X)$, $m\le n$}.$$ For the indicator $\chi_\Delta$ of a set $\Delta\subset X$ the identity (\[norma2\]) implies $
\|\chi_\Delta\|_{m,\infty}={1-(1-|\Delta|)^m}
$ (here and further we denote $|\Delta|\equiv\mu\Delta$). Thus, if $m\ge c|\Delta|^{-1}$, then $\|\chi_\Delta\|_{m,\infty}\ge C(c)$ with a constant [$C(c)>0$]{}.
For an integrable function $f\in L_1$ define the following “relative” norms: $$\begin{aligned}
\label{norma3}
\|f\|_m^* &:=
\sup_{\Delta\subset X}
\Big\{\frac{1-(1-|\Delta|)^m}{|\Delta|}
\int_\Delta|f|d\mu\Big\};\\
\label{norma4}
\|f\|'_m &:=
\sup_{{\genfrac{}{}{0pt}{}{\Delta\subset X}{\mu\Delta=1/m}}}
\Big\{m\int_\Delta|f|d\mu\Big\}.\end{aligned}$$ Right side of (\[norma4\]) is well-defined only if there exists an event $e\subset X$ such that $\mu (e)=1/m$. However, we will use this norm only if $X=[0,1]$ with standard Lebesque measure.
The ${\|\cdot\|^*_m}$- and ${\|\cdot\|'_m}$-norms are equivalent to ${\|\cdot\|_{m,\infty}}$-norm, this is proved in the following result, which was implicitly established by the author in [@ja3].
\[kee\] For all functions $f\in L_1[0,1]$ the following inequalities take place $$\label{kl}
(1-\frac1e)\|f\|'_m \le
\|f\|_{m}^* \le
\|f\|_{m,\infty} \le
2\|f\|'_m.$$
[**Proof.**]{} Let us show that $2\|f\|'_m\ge\|f\|_{m,\infty}$. Let $\Delta^*\subset [0,1]$ satisfy $|\Delta^*|=1/m$ and $$m\int_{\Delta^*}|f|=\|f\|'_m.$$ (It is easy to see that such a $\Delta^*$ exists though not necessarily unique). We have $$\|f\|_{m,\infty}\le
\|f\cdot\chi_{\Delta^*}\|_{m,\infty}+
\|f\cdot(1-\chi_{\Delta^*})\|_{m,\infty}.$$ By (\[qw4\]) we estimate $\|f\cdot\chi_{\Delta^*}\|_{m,\infty}
\le m\int_{\Delta^*}|f|=\|f\|'_m$. From the extremality of $\Delta^*$ we get $$\|f\cdot(1-\chi_{\Delta^*})\|_{m,\infty}
\le\|f\cdot(1-\chi_{\Delta^*})\|_\infty
\le\frac1{|\Delta^*|}\int_{\Delta^*}|f|=\|f\|'_m.$$ Thus, the inequality $\|f\|_{m,\infty}\le2\|f\|'_m$ is proved.
Let us check now that for every $\Delta\subset [0,1]$ the following inequality holds: $$\label{s11}
\|f\|_{m,\infty}\ge
\frac{1-(1-|\Delta|)^m}{|\Delta|}\int_\Delta|f|d\mu.$$ Obviously, it suffices to consider the case when $f$ vanishes outside $\Delta$ (${\text{supp}(f)\subset\Delta}$). By (\[norma2\]) we get $$\begin{aligned}
\|f\|_{k+1,\infty}-\|f\|_{k,\infty}
&=\int_0^\infty \!\big(1-(1-\lambda_f(t))^{k+1}\big)dt-
\int_0^\infty \!\big(1-(1-\lambda_f(t))^k\big)dt \\
&=\int_0^\infty \lambda_f(t)\big(1-\lambda_f(t)\big)^k dt \\
&\ge (1-|\Delta|)^k \int_0^\infty \!\lambda_f(t) dt=(1-|\Delta|)^k \|f\|_1.\end{aligned}$$ Summing these inequalities from $k=1$ to $k=m-1$, we get $$\|f\|_{m,\infty}-\|f\|_1
\equiv
\|f\|_{m,\infty}-\|f\|_{1,\infty}
\ge\|f\|_1\sum_{k=1}^{m-1}(1-|\Delta|)^k,$$ which implies $$\|f\|_{m,\infty}\ge\|f\|_1\frac{1-(1-|\Delta|)^m}{|\Delta|}.$$ Now, to prove (\[s11\]) it remains to notice that $\|f\|_1=\int_\Delta |f|$ (recall $\text{supp}f\subset\Delta$). Therefore, $\|f\|_{m,\infty}\ge\|f\|_m^*$.
The inequality ${\|f\|^*_m\ge(1-e^{-1})\|f\|'_m}$ obviously follows from the fact that ${(1-\frac1m)^m<e^{-1}}$. The proof of Theorem \[kee\] is completed.
[**A Fact from Geometry.**]{}
\[geomlem\] (See Lemma 1 in [@ja3] or Lemma 2.1 in [@jadis]). Let $\{w_i\}_{i=1}^n$ be a set of vectors in a linear space equipped with a norm $\|\cdot\|$ (or a semi-norm), satisfying $\|w_i\|=1$ and $$\label{geom}
\|\sum_{i=1}^n \theta_i w_i\|\le c\cdot n^{\frac12+\beta}
\nopagebreak$$ for all choices of signs $\{\theta_i\}_{i=1}^n$, $\theta_i=\pm1$, with some constants ${\beta\in[0,1/2)}$, ${c>0}$. Then $$\big\|\sum_{i=1}^n a_i w_i\big\|\le
C(c)\, n^{\frac14+\frac\beta2}
\big(\sum_{i=1}^n a_i^2\big)^{\frac12}
\eqno(\ref{geom}')$$ for all sets of coefficients $\{a_i\}_1^n$. This estimate is sharp, i.e. there exist vectors $\{w_i\}_1^n$, a norm ${\|\cdot\|}$, and coefficients $\{a_i\}_1^n$ such that (\[geom\]) takes place, but estimate (\[geom\]$'$) is sharp in sense of order.
Geometrically Lemma \[geomlem\] claims that the convex hull of the set $B^d_\infty\cup (n^{1/2+\beta}\cdot B^d_1)$ has the inscribed sphere with radius of order $n^{1/4+\beta/2}$, here $B^d_\infty$ denotes $d$- cube whose vertices have coordinates $\pm1$ and ${B^d_1:=\{(v_k)\in{\Bbb R}^d:\sum_1^d|v_k|\le1\}}$.
To prove the main result we need the following lemmas which enable us to transfer one property of the multidimensional normal distribution (Lemma \[lem2\]) to the case of an abstract multidimensional distribution.
\[lem1\] Let $\Omega_j\subset\Omega$, $j=1,\dots,m$, be events, satisfying $$(1-\kappa)
\sum_{j,k=1}^m{\sf P}\big({\Omega_j\cap \Omega_k}\big)
\le
\Big(\sum_{j=1}^m {\sf P}(\Omega_j)\Big)^{2}$$ with some $\kappa\in(0,1)$. Then ${\sf P}\big(\bigcup_{j=1}^m\Omega_j\big)
\ge1-\kappa.$
\[lem2\] Let $\{h_j\}_{j=1}^m$ be a set of Gaussian variables such that ${{\sf E}h_j=0}$, ${ {\sf E}h_j^2=D_j\ge r^2>0}$, ${{\sf E}h_jh_k=v_{j,k}}$ and each pare $(h_j,h_k)$, ${j\ne k}$ has a 2-dimensional normal distribution with density $$\phi_{0,V_{j,k}}(Y):=\frac1{2\pi(\det V_{j,k})^{1/2}}
\exp\big\{-\frac12(Y,V_{j,k}^{-1}Y)\big\},\qquad\ Y=(y_1,y_2),$$ $$\text{where}\qquad\
V_{j,k}=\left(
\begin{array}{cc}
D_j & v_{j,k}\\
v_{j,k} & D_k
\end{array}
\right)
\quad\text{is a covariance matrix.}$$ Assume also that there exist some constants $R\ge1$, $c_0$, ${\delta>0}$ such that $$\frac1{m^2}
\sum_{{\genfrac{}{}{0pt}{}{j,k=1}{j\ne k}}}^m |v_{j,k}|\le c_0R^{-\delta}r^2.$$ Then for arbitrary choice of ${\alpha<\delta^{1/2}}$ there exists an [*independent of $R$*]{} constant ${C_7=C_7(c_0,\alpha,\delta)}$ such that $$\label{lem22}
\sum_{{j,k=1}}^m
{\sf P}\big(\Psi_j\cap\Psi_k\big)
\le(1+C_7P^{-q_0}) \Big(\sum_{j=1}^m {\sf P}(\Psi_j)\Big)^2,$$ where ${P:=\min(R,m)+1}$, $
\Psi_j=\Psi_j(\alpha):={\big\{h_j>\alpha\sqrt{D_j\log P}\big\}}$ and $q_0=q_0(\alpha,\delta)=
{\min\big(\frac13(\delta-\alpha^2),\frac32\alpha^2\big)}>0$.
\[lem3tex\] Let $(\eta_j)_1^m$ and $(h_j)_1^m$ be random vectors with identical first and second moments: $${\sf E}\eta_j ={\sf E}h_j=0,\quad\
{\sf E}\eta_j\eta_k={\sf E}h_jh_k=v_{j,k},\quad\
D_j\equiv v_{j,j}\ge r^2>0.
$$ Moreover, let $(h_j)_1^m$ be a Gaussian vector whose covariance matrix satisfies the assumption of Lemma \[lem2\] with parameters $r$, $R$, $c_0$, ${\delta>0}$. Assume also that there exist some positive constants ${\delta_i\le1}$, $M_i$ ${i=1,2,3}$, and $\alpha<\alpha_0:=
\big(\min(\delta,\delta_1,\delta_2,\delta_3)\big)^{1/2}$ such that $$\begin{aligned}
\label{lem3usl1}
\big|{\sf P}\big(U_j(\alpha)\big)-
{\sf P}\big(\Psi_j(\alpha)\big)\big|
&\le M_1P^{-\delta_1},\quad \ 1\le j\le n; \\
\label{lem3usl2}
\big|{\sf P}\big(U_j(\alpha)\cap U_k(\alpha)\big)-
{\sf P}\big(\Psi_j(\alpha)\cap\Psi_k(\alpha)\big)\big|
&\le M_2P^{-\delta_2},\quad \ (j,k)\in\sigma,
$$ where $P=\min(R,m)+1$ and $$\begin{aligned}
\Psi_j=\Psi(\alpha)&:=
\big\{\omega:h_j(\omega)>\alpha\sqrt{D_j\log P}\big\};\\
U_j=U_j(\alpha)&:=\big\{\omega:\eta_j(\omega)>\alpha\sqrt{D_j\log P}\big\};\end{aligned}$$ and the index set $\sigma\subset\{(j,k):j\ne k,1\le j,k\le m\}$ satisfies ${|\sigma|\ge m^2(1-M_3P^{-\delta_3})}$.\
Then there exists a constant ${C_7'=C_7'(\alpha,c_0,\delta,\delta_i,M_i)}$, ${i=1,2,3}$, such that $$\label{lem33tex}
\sum_{j,k=1}^m
{\sf P}\big(U_j(\alpha)\cap U_k(\alpha)\big)
\le(1+C_7'P^{-q'})
\Big(\sum_{j=1}^m {\sf P}\big(U_j(\alpha)\big)\Big)^2,$$ where the power parameter $q'=
\min\big\{
\frac13(\delta-\alpha^2),\frac32\alpha^2,
\frac56\min\limits_{i=1,2,3}(\delta_i-\alpha^2)\big\}>0$.
[**Remark on Lemmas \[lem1\]–\[lem3tex\].**]{} Lemma \[lem1\], being a generalization of the Borel-Cantelli Lemma, is a known and important result. One can find similar statements, e.g. in [@cher], [@sprind], [@kt1]. Lemmas \[lem2\] and \[lem3tex\], as far as I am aware of, are formulated for the first time, however the ideas for their proofs have been thoroughly borrowed from [@kt1], [@kt3]. Lemma \[lem3tex\] serves to transfer the estimates of type (\[lem22\]) to the “non-Gaussian case.” Note that the main results of [@kt1], [@kt3], [@ja3] could be easier proved and perceived with the help of Lemmas \[lem1\] and \[lem3tex\].
Note also that from Lemmas \[lem1\] and \[lem2\] one can easily derive for Gaussian random variables $\{h_i\}_1^m$, satisfying the assumption of Lemma \[lem2\], the following estimate: $${\sf P}(\max\limits_{1\le j\le m}h_j > \alpha r\sqrt{\log m})
\ge{1-Cm^{-q_0}} \qquad \text{for}\quad m\le R,$$ where $R$ and $r$ are from the statement of Lemma \[lem2\] and $\alpha$, $q_0$, $C$ are some positive constants. This fact links Lemma \[lem2\] with the results of Slepian [@sl] and Šidak [@sidak1], [@sidak2], devoted to estimating the distribution of maximum of Gaussian vectors with a non-trivial covariance matrix.
[**Proof of Lemma \[lem1\].**]{} Let $\chi_j$ be the indicators of the events ${\Omega_j\subset\Omega}$ and let ${\zeta:=\sum_{j=1}^n\chi_j}$. Then ${\sf E}|\zeta|=\sum_{j=1}^m{\sf P}(\Omega_j)$, $
{\sf E}|\zeta|^2=\sum_{j,k=1}^m
{\sf P}\big({\Omega_j\cap \Omega_k}\big)
$ and ${\text{supp}\,\zeta}=\bigcup_{j=1}^m\Omega_j$. Applying the Cauchy-Schwarz inequality, we get $$(1-\kappa)^{1/2}\big({\sf E}|\zeta|^2\big)^{1/2}\le
{\sf E}|\zeta|\le\big({\sf E}|\zeta|^2\big)^{1/2}
\Big({\sf P}\Big(\bigcup_{j=1}^m\Omega_j\Big)\Big)^{1/2}.$$ This proves Lemma \[lem1\].
[**Proof of Lemma \[lem2\].**]{} Since the random variables $h_j$ are normal we have $$\label{lem2nomer1}
{\sf P}(\Psi_j)=\frac1{\sqrt{2\pi D_j}}
\int_{\alpha\sqrt{D_j\log P}}^\infty
e^{-{y^2}/({2D_j})}\,dy=
\frac1{\sqrt{2\pi}}\int_{\alpha\sqrt{\log P}}^\infty
e^{-{y^2}/2}\,dy.$$ Taking into account that $
\int_z^\infty e^{-t^2/2}\,dt\asymp z^{-1}
e^{-z^2/2}$ when $z\ge1$, we have $C_\alpha^{-1} (\sqrt{\log P})^{-1}P^{-\alpha^2/2} \le
{\sf P}(\Psi_j)\le
C_\alpha (\sqrt{\log P})^{-1}P^{-\alpha^2/2}$ with a constant $C_\alpha$ depending only on $\alpha$. Thus $$\sum_{j=1}^m{\sf P}(\Psi_j)
\asymp_\alpha \frac m{\alpha\sqrt{\log P}}P^{-\alpha^2/2}.\eqno(*)$$ Thus, to prove (\[lem22\]) it suffices to check that $$\sum_{{\genfrac{}{}{0pt}{}{j,k=1}{j\ne k}}}^m
{\sf P}\big(\Psi_j\cap\Psi_k\big)
\le(1+C'_7P^{-q_0}) \Big(\sum_{j=1}^m {\sf P}(\Psi_j)\Big)^2.
\eqno(\ref{lem22}')$$
Define the following index set: $$\sigma_1:=\big\{(j,k): 1\le j\ne k\le m,
|v_{j,k}|<\frac1{32}r^2\big\}.$$ Chebyshev’s inequality for the set ${\sigma_1^c=\{(j,k)\!:1\le j\ne k\le m\}\setminus\sigma_1}$ implies $$|\sigma_1^c|\le \frac{32}{r^2}\sum_{j\ne k}|v_{j,k}|
\le{32}c_0m^2R^{-\delta}.$$ Therefore, $$\sum_{(j,k)\in\sigma_1^c}{\sf P}(\Psi_j\cap\Psi_k)
\le {32}c_0m^2R^{-\delta}\max_{1\le j\le m}{\sf P}(\Psi_j).$$ Thus, taking into account (\[lem2nomer1\]), we get $$\sum_{(j,k)\in\sigma_1^c}{\sf P}(\Psi_j\cap\Psi_k)
\le 32c_0m^2R^{-\delta}\frac{\text{\rm const}}
{\alpha\sqrt{\log P}}P^{-\alpha^2/2}.$$ Applying (\*), we conclude that $$\sum_{(j,k)\in\sigma_1^c}
{\sf P}\big(\Psi_j\cap\Psi_k\big)
\le K_2R^{-\delta+\alpha^2}
\cdot
\Big(\sum_{j=1}^m {\sf P}(\Psi_j)\Big)^2
\eqno(**)$$ with a constant $K_2(\alpha,c_0)>0$. Thus, to prove $(\ref{lem22}')$ we can neglect the summation over $\sigma_1^c$ on the left-hand side.
Now, let us estimate the sum $$\Sigma_1:=\hspace{-.2em plus 1pt minus .7em}
\sum_{s=(j,k)\in\sigma_1}\hspace{-.2em plus 1pt minus .7em}
{\sf P}(\Psi_j\cap\Psi_k)
=\hspace{-.3em plus .1em minus 1.2em}
\int\limits_{\alpha\sqrt{D_j\log P}}^\infty
\int\limits_{\alpha\sqrt{D_k\log P}}^\infty
\sum_{s\in\sigma_1}
\frac{\exp\big\{\!-\frac12(Y,V_s^{-1}Y)\big\}}{2\pi\sqrt{\det V_s}}
dy_1dy_2.$$ Changing the integration variables $t_1=\frac{y_1}{\sqrt{D_j}}$, $t_2=\frac{y_2}{\sqrt{D_k}}$, we get $$\Sigma_1=
\int_{\alpha\sqrt{\log P}}^\infty
\int_{\alpha\sqrt{\log P}}^\infty
\sum_{s\in\sigma_1}
\frac{\sqrt{D_jD_k}}{2\pi\sqrt{\det V_s}}
e^{-\frac12Q(t_1,t_2)}
dt_1dt_2,$$ where $Q(t_1,t_2):=(Y,V_s^{-1}Y)$ is a quadratic form. Evaluating the determinant, we get ${\det V_s=D_jD_k-v_s^2}$ and $$V_s^{-1}= \frac1{D_jD_k-v_s^2}
\left(
\begin{array}{ccc}
D_k &-{v_s}\\
-{v_s} &{D_j}
\end{array} \right).$$ Thus, the coefficients of the quadratic form ${Q=a_st_1^2+a_st_2^2-2b_st_1t_2}$ are defined by $$a_s= \frac{D_jD_k}{D_jD_k-v_s^2};
\qquad\qquad
b_s= \frac{v_s\sqrt{D_jD_k}}{D_jD_k-v_s^2}.$$ Notice also that $|v_s|\le r^2/32$, provided $s\in\sigma_1$. Therefore, taking into account that $D_j\ge r^2$ for all $j=1,\dots,m$, we have $$a_s-b_s=\frac{\sqrt{D_jD_k}}{\sqrt{D_jD_k}+v_s}
\ge\frac12.$$ Thus, we can estimate the values of the quadratic form $Q$ as follows: $$Q(t_1,t_2)=a_st_1^2+a_st_2^2-2b_st_1t_2\ge
(a_s-b_s)(t_1^2+t_2^2)
\ge\frac12(t_1^2+t_2^2).$$
Now, for arbitrary ${L>1}$ we can estimate $$\begin{aligned}
J(L)&:=\frac1{2\pi}
\int_{L\alpha\sqrt{\log P}}^\infty
\int_{\alpha\sqrt{\log P}}^\infty
\sum_{s\in\sigma_1}
\Big( \frac{D_jD_k}{D_jD_k-v_s^2}\Big)^{\frac12}
e^{-\frac12Q(t_1,t_2)}
dt_1dt_2\\
&\le
\int_{L\alpha\sqrt{\log P}}^\infty
\int_{\alpha\sqrt{\log P}}^\infty
\sum_{s\in\sigma_1}
e^{-\frac1{4}(t_1^2+t_2^2)}dt_1dt_2\\
&\asymp_\alpha
\frac{|\sigma_1|}{L\log P}e^{-\frac{\alpha^2}{4}(L^2+1)\log P}
\le
\frac{K_3m^2}{L\log P}P^{-\frac{\alpha^2(L^2+1)}{4}},\end{aligned}$$ where $K_3=K_3(\alpha)$ is a constant. Choose $L=3$ and take into account (\*) to get $$J(3)<K_4 P^{-\frac32{\alpha^2}}
\Big(\sum_{j=1}^m{\sf P}(\Psi_j)\Big)^2\eqno(***)$$ with a constant $K_4=K_4(\alpha)>0$.
Notice that $$\Sigma_1\le 2 J(3)+
\frac1{2\pi}
\hspace{-.1em plus .2em minus .2em}
\int_{\alpha\sqrt{\log P}}^{3\alpha\sqrt{\log P}}
\int_{\alpha\sqrt{\log P}}^{3\alpha\sqrt{\log P}}
\sum_{s\in\sigma_1}
\sqrt{a_s} \cdot
e^{-\frac12Q(t_1,t_2)}
dt_1dt_2.$$ Now, in order to finish the proof of (\[lem22\]) it remains to compare the expression $$A:=\frac1{2\pi} \sum_{s\in\sigma_1}
\sqrt{a_s}
e^{-\frac12Q(t_1,t_2)}$$ with the expression $B:=(2\pi)^{-1}|\sigma_1|\exp\{-\frac12(t_1^2+t_2^2)\}$ on the square ${\alpha\sqrt{\log P}\le t_1,t_2\le 3\alpha\sqrt{\log P}}$. In fact, if we show that $A\le B(1+K_5 P^{-q_0})$ with a constant ${K_5(\alpha,\delta)>0}$, then integrating this inequality we get $$\begin{aligned}
\frac1{2\pi}
& \int_{\alpha\sqrt{\log P}}^{3\alpha\sqrt{\log P}}
\int_{\alpha\sqrt{\log P}}^{3\alpha\sqrt{\log P}}
\sum_{s\in\sigma_1}
\sqrt{a_s} \exp\big\{-\frac12Q(t_1,t_2)\big\}dt_1dt_2<\\
&<(1+{K_5}{P^{-q_0}})
\int_{\alpha\sqrt{\log P}}^\infty
\int_{\alpha\sqrt{\log P}}^\infty
\frac{|\sigma_1|}{2\pi}
\exp\big\{-\frac12(t_1^2+t_2^2)\big\} dt_1dt_2\\
&= (1+{K_5}{P^{-q_0}})
\sum_{s\in\sigma_1}{\sf P}(\Psi_j)\cdot{\sf P}(\Psi_k).\end{aligned}$$ This inequality combined with (\*\*) and (\*\*\*) would imply (\[lem22\]) and, thus, prove the lemma.
Split the index set $\sigma_1$ into the subsets $$\sigma_i:=
\big\{s\in\sigma_1:2^{-i}r^2\le|v_s|<2^{-i+1}r^2\big\},
\qquad \ \ i=6,7,\dots.$$ Clearly, $\sigma_1=\bigcup_{i\ge6}\sigma_i$. Chebyshev’s inequality for $|\sigma_i|$ implies $$|\sigma_i|2^{-i}r^2\le\sum_{s\in\sigma_1}|v_s|
\le c_0r^2m^2R^{-\delta}$$ and, consequently, $$|\sigma_i|\le\min\big(m^2,
2^ic_0m^2 R^{-\delta}\big).$$ Taking into account that $D_j\ge r^2$ for $s\in\sigma_i$ we have $$\begin{gathered}
\big|(a_s-1)(t_1^2+t_2^2)-2b_st_1t_2\big|
\le\big(|a_s-1|+|b_s|\big)(t_1^2+t_2^2)=\\
=\frac{|v_s|}{\sqrt{D_jD_k}-|v_s|}(t_1^2+t_2^2)
<2^{2-i}(t_1^2+t_2^2)\le2^{7-i}\alpha^2\log P\end{gathered}$$ in the domain $\alpha\sqrt{\log P}\le t_1,t_2\le 3\alpha\sqrt{\log P}$. Moreover, it is easy to see that if $s\in\sigma_i$ and $i\ge6$, then $$\sqrt{a_s}=
\Big(
\frac{{D_jD_k}}{D_jD_k-v_s^2}
\Big)^{1/2}
=\Big(1+
\frac{v_s^2}{D_jD_k-v_s^2}
\Big)^{1/2}
\le1+2^{3-2i}.$$
Gathering all these facts, we get $$\quad A\le\frac{B}{|\sigma_1|}S,$$ $$\text{where}\quad\
\hspace{0em plus 1em minus .4em}
S:=
\hspace{-.1em plus 1em minus .3em}
\sum_{i=6}^\infty |\sigma_i|(1+2^{3-2i})
e^{\frac12 2^{7-i}\alpha^2\log P}
=
\hspace{-.1em plus 1em minus .4em}
\sum_{i=6}^{[2q_0\log_2 P]}
\hspace{-.1em plus 1em minus .5em}
+
\hspace{-.2em plus 1em minus .5em}
\sum_{i>[2q_0\log_2 P]}^\infty
\hspace{-.3em plus 1em minus .5em}=:S_1+S_2.$$ Taking into account that $q_0=\min\big(\frac13(\delta-\alpha^2),\frac32\alpha^2\big)$, we estimate $$S_1 \le \hspace{-.3em plus 1em minus .6em}
\sum_{i=6}^{[2q_0\log_2 P]}
\hspace{-.3em plus 1em minus .6em}
2^ic_0m^2R^{-\delta}\!\cdot
(1+2^{-9})P^{\alpha^2}
\hspace{-.1em plus 1em minus .2em}
\le 4c_0 m^2 P^{\alpha^2+2q_0}R^{-\delta}
\hspace{-.1em plus 1em minus .2em}
\le K_6 R^{-q_0}|\sigma_1|,$$ where $K_6=K_6(c_0)>0$ is a constant.
To estimate $S_2$ notice that $P^{2^6\alpha^2P^{-2q_0}}<(1+K_{7}P^{-q_0})$, where $K_{7}(\alpha,q_0)>0$ is a constant. Thus, $$S_2\le
\sum_{i>[2q_0\log_2 P]}
|\sigma_i|(1+8P^{-4q_0})(1+K_7P^{-q_0})\le
|\sigma_1|(1+K_8 P^{-q_0}),$$ where $K_{8}=8+9K_7$. Therefore, $$S\le\big(1+(K_6+K_8)P^{-q_0}\big)|\sigma_1|$$ and $A\le B(1+K_5P^{-q_0})$. This completes the proof of the inequality (\[lem22\]) and Lemma \[lem2\].
\[remlem2\] In Lemma \[lem2\] the power parameter $q_0(\alpha,\delta)$ is not optimally chosen. It is not difficult to show that the inequality (\[lem22\]) stays true, provided $q_0={q_0(\alpha,\delta,\varepsilon)=
\min\big\{\frac12(\delta-\alpha^2),\frac32\alpha^2\big\}
-\varepsilon>0}$ for arbitrarily small $\varepsilon$. (In order to check this it suffices to draw a sharper estimate for the sum $S$). However, in this case the constant $C_7$ in the inequality (\[lem22\]) would depend also on ${\varepsilon>0}$. Moreover, somewhat more advanced modification of the proof enables to derive (\[lem22\]) with parameter ${q_0=\frac12(\delta-\alpha^2)-\varepsilon}$.
[**Proof of Lemma \[lem3tex\].**]{} Since the random vector $(h_j)_1^m$ satisfies the assumption of Lemma \[lem2\] it follows that the estimate (\[lem22\]) for the events $\{\Psi_j\}_{j=1}^m$ with the power parameter ${q_0=\frac13(\delta-\alpha^2)}$ holds. Thus, to prove (\[lem33tex\]) it suffices to show that for arbitrary $\alpha<\alpha_0$ the following inequalities take place: $$\begin{aligned}
\label{lem3t1}
\sum_{j=1}^m{\sf P}\big(\Psi_j(\alpha)\big)
& \le
\big(1+L_1P^{-q_1}\big)
\sum_{j=1}^m{\sf P}\big(U_j(\alpha)\big);\\
\label{lem3t2}
\sum_{j,k=1}^m\!{\sf P}\big(U_j(\alpha)\cap U_k(\alpha)\big)
& \!\le
\!\sum_{j,k=1}^m\!
{\sf P}\big(\Psi_j(\alpha)\cap \Psi_k(\alpha)\big)+
\frac{L_2}{P^{q_2}}
\Big(\sum_{j=1}^m\!{\sf P}\big(\Psi_j(\alpha)\big)\Big)^2,\end{aligned}$$ where the constants $q_1$, $q_2\ge q'$ and $L_1$, ${L_2>0}$ depend only on $\alpha$, $c_0$, $\delta$, $\delta_i$, $M_i$, ${i=1,2,3}$.
Notice that $$\label{psi4}
{\sf P}\big(\Psi_j(\alpha)\big)=
\frac1{2\pi D_j}
\int_{\alpha\sqrt{D_j\log P}}^\infty
e^{-\frac{y^2}{2D_j}}dy
\asymp_\alpha \frac{P^{-\alpha^2/2}}{\alpha\sqrt{\log P}}.$$ Thus, for all ${\alpha^2<\delta_1}$ we have (see (\[lem3usl1\])) $$\begin{aligned}
\label{lem3t3}
{\sf P}\big(U_j(\alpha)\big)
&\asymp_\alpha
{\sf P}\big(\Psi_j(\alpha)\big);\\
\notag
M_1P^{-\delta_1}
&\le L_3 P^{-\delta_1+\alpha^2}
\cdot {\sf P}\big(\Psi_j(\alpha)\big)\end{aligned}$$ with a constant $L_3(\alpha,\delta_1,M_1)$. Taking into account (\[lem3usl1\]), we get the inequality (\[lem3t1\]) with parameter ${q_1=\delta_1-\alpha^2}$.
In order to prove (\[lem3t2\]) let us recall that according to the Lemma assumption $$|\sigma^c|\equiv
\text{card}\big\{(j,k):j\ne k,(j,k)\notin\sigma\big\}
\le M_3m^2P^{-\delta_3}.$$ Thus, for all $\alpha^2<\delta_1$ we get (see (\[lem3t3\])) $$\sum_{(j,k)\in\sigma^c}{\sf P}(U_j\cap U_k)
\le M_3m^2P^{-\delta_3}\max_{1\le j\le m}{\sf P}(U_j)
\le L_4 m^2P^{-\delta_3}
\frac{P^{-\alpha^2/2}}{\sqrt{\log P}}$$ with a constant $L_4(\alpha,\delta_1,\delta_3,M_1,M_3)$. Using (\[psi4\]) and (\[lem3t3\]), for ${\alpha^2<\min\{\delta_1,\delta_3\}}$ we get $$\sum_{(j,k)\in\sigma^c}
{\sf P}\big(U_j(\alpha)\cap U_k(\alpha)\big)
\le L_5P^{-\delta_3+\alpha^2}
\Big( \sum_{j=1}^m{\sf P}\big(\Psi_j(\alpha)\big)\Big)^2.$$ Hence, to prove (\[lem3t2\]) we can neglect the summation over $\sigma^c$ on the left-hand side.
Notice, when ${\alpha^2<\delta_1}$ the estimate (\[lem3t3\]) implies $$\sum_{j=1}^m{\sf P}\big(U_j(\alpha)\big)
\le L_5 P^{-1+\alpha^2}
\Big(\sum_{j=1}^m{\sf P}\big(\Psi_j(\alpha)\big)\Big)^2$$ with a constant $L_5(\alpha,\delta_1,M_1)$. Therefore, to prove (\[lem3t2\]) we can also neglect summation over the pares ${\big\{(j,k):j=k\big\}}$ on the left-hand side.
Now, to prove (\[lem3t2\]) it remains to notice that the assumption (\[lem3usl2\]) implies $$\sum_{(j,k)\in\sigma}
{\sf P}\big(U_j(\alpha)\cap U_k(\alpha)\big)
\le\sum_{(j,k)\in\sigma}
{\sf P}\big(\Psi_j(\alpha)\cap\Psi_k(\alpha)\big)
+M_2m^2P^{-\delta_2}.$$ Taking into account (\[psi4\]), for $\alpha^2<\min\{\delta_1,\delta_2\}$ we can estimate the error term as follows: $$M_2m^2P^{-\delta_2}\le
L_6 P^{-\frac56({\delta_2-\alpha^2}) }
\Big(\sum_{j=1}^m{\sf P}\big(\Psi_j(\alpha)\big)\Big)^2,$$ where $L_6=L_6(\alpha,\delta_1,\delta_2,M_1,M_2)$ is a constant. Thus, the inequality (\[lem3t2\]) is proved with the constants ${L_2=L_4+L_5+L_6}$ and ${q_2=\min\{\frac56({\delta_2-\alpha^2}),\delta_3-\alpha^2\}}$.
The inequalities (\[lem3t1\]), (\[lem3t2\]), combined with (\[lem22\]), prove (\[lem33tex\]) with ${q'=\min\{q_0,q_1,q_2\}}$, where ${q_0=\min(\frac13(\delta-2\alpha^2),\frac32\alpha^2)}$ is the power parameter in (\[lem22\]). The proof of Lemma \[lem3tex\] is completed.
\[powerq\] A simple modification to the proof of Theorem A in [@ja3], which would involve Lemmas \[lem1\]–\[lem3tex\], enables one to establish the estimate (\[genth1\]) with the power parameter $q$ arbitrarily chosen from ${\big(0,\frac3{11}({1-2p})\big)}$. Moreover, if we took unto account Remark \[remlem2\] and made some simple refinements in Lemmas \[lem2\] and \[lem3tex\], then we could prove (\[genth1\]) with the parameter $q$ from the interval ${\big(0,({1-2p})/2\big)}$, however in this case the constants $C_1$, $C_2$ would depend on $q$. In [@jadis] one can find a detailed proof of Theorem A for the case ${q=(1-2p)/4}$.
[**Central Limit Theorem.**]{} For the proof of the main result we need to apply a version of 2-dimensional central limit theorem with an estimate of the error term. We shall use in one- and two-dimensional case the following result due to Rotar’ [@rotar] (or see Corollary 17.2 in [@bhat]):
\[rot\] Let $\{X_i\}_{i=1}^N$ be a set of independent random vectors in ${\Bbb R}^d$, satisfying ${\sf E}X_i=0$, $1\le i\le N$, then $$\sup_{A\in {\mathcal C}} |P_N(A)-\Phi_{0,V}(A)|\le K_1(d)
N^{-1/2}m_3\lambda^{-3/2},$$ where $P_N(A)$ is the probability of the event that $N^{-1/2}\sum_{i=1}^NX_i$ belongs to the set $A$, ${\mathcal C}$ denotes the class of all Borel convex sets in ${\Bbb R}^d$, $K_1(d)<\infty$ is a constant, $$m_3:=\frac1N\sum_{i=1}^N{\sf E} |X_i|^3,$$ $\lambda$ is the smallest eigenvalue of the matrix $V=N^{-1}\sum_{i=1}^N \text{\rm cov} (X_i)$, $\text{\rm cov} (X_i)$ denotes the covariance matrix of vector $X_i$, finally, $\Phi_{0,V}$ denotes the Gaussian measure on ${\Bbb R}^d$ with the density $$\phi_{0,V}(Y):=(2\pi)^{-d/2}(\det V)^{-1/2}
\exp\big\{-\frac12(Y,V^{-1}Y)\big\},\qquad \ \ Y\in {\Bbb R}^d.$$
Proof of Theorem \[l1th\] {#secl1}
=========================
Theorem \[l1th\] is a direct corollary of more general Theorem \[l1th\]$'$ and Khinchin’s inequality.
[**Theorem [\[l1th\]]{}$'$.**]{}
*Let $\{f_i\}_{i=1}^n$ be a system of functions on a probability space $(X,\mu)$, satisfying*
- $\|f_i\|_1=1$ for all $i=1,\cdots,n$;\
$\big\|\sum_{i=1}^n \theta_i f_i\big\|_1\le Mn^{\frac 12+p_1}$ for all choices of signs $\theta_i=\pm1$;\
$\big\|(\sum_{i=1}^n|f_i|^2)^{1/2}\big\|_1\le Mn^{\frac12+p_2}$, where $M$, ${p_1, p_2\ge0}$ are some constants, satisfying $p_1+2p_2<\frac12$ and ${p_2<\frac1{12}}$.
Let $\{\xi_i\}_{i=1}^n$ be a system of independent random variables on another probability space $(\Omega,{\sf P})$, satisfying ${{\sf E}\xi_i=0}$, ${{\sf E}|\xi_i|^2=1}$ and ${{\sf E}|\xi_i|^3\le M^3}$. Then whenever ${m\le n}$ for random polynomial (\[polinom2\]) the estimates (\[l1th1\]) and (\[l1th2\]) hold with some constants $q=q(p_1,p_2)>0$, $C_j=C_j(p_1,p_2,M)>0$, ${j=1,2,3}$. (These constants, of course, are not the same as the constants from the statement of Theorem A.)
To deduce Theorem \[l1th\] from Theorem \[l1th\]$'$ it suffices to notice that validity of condition [**(d)**]{}, combined with integrated over ${x\in X}$ Khinchin’s inequality (e.g. see [@ks]) for the sum ${\sum r_i(\omega) f_i(x)}$ with fixed $x$, where $r_i$ are the Rademacher functions, implies $$\big\|\Big(\sum_{i=1}^n |f_i|^2\Big)^{\frac12}\big\|_1
\le \text{const}\cdot{\sf E}\int_X
\big|\sum_{i=1}^n r_i(\omega) f_i(x)\big|d\mu(x)
\le M'(M)n^{\frac12+p}.$$ Thus, if condition [**([d]{})**]{} holds with parameter $p=p_1$, the condition [**([d]{}$'$)**]{} holds with $p_2\le p_1$ and, consequently, if ${p_1<\frac1{12}}$, then ${p_1+2p_2<1/2}$ automatically.
Theorem \[l1th\]$'$, like Theorem A, is based on the central limit theorem (Proposition \[rot\]), however its proof requires essentially subtler preparatory work. Roughly speaking, the reason for this is that while the basis functions $\{f_i\}$ were supposed to be [*uniformly*]{} bounded in $L_{2+\varepsilon}$ it was possible to find a “sufficiently large” set ${E\subset X}$ such that $$\label{tridva}
\frac{\sum_{i=1}^n|f_i(x)|^3}{\big(\sum_{i=1}^n|f_i(x)|^2\big)^{3/2}}
\le\text{\rm{const}}\cdot n^{-\epsilon},\qquad\ x\in E,$$ with some $\epsilon>0$. This inequality is needed to estimate the error term after the application of the central limit theorem. Such a trick was used to prove (\[genth1\]), (\[genth2\]) in [@ja3] (and before that it had been uses in [@kt1], [@kt3] to prove similar estimates for the uniform norm).
However, in the assumptions of Theorem \[l1th\]$'$ there may be no point ${x\in X}$ such that (\[tridva\]) holds. As a corresponding example take the functions ${f_i:=r_i+n^{q}\chi_i}$, where $r_i$ are the Rademacher functions, $\chi_i$ are the characteristic functions of the intervals $\big(\frac{i-1}n,\frac in\big)$ and ${\frac13<q<\frac7{12}}$. (You can normalize $f_i$ in $L_1$ to make the example more appropriate for Theorem \[l1th\]$'$.) For such functions the condition [**(d)**]{} holds with $p=\max\{0,q-\frac12\}<\frac1{12}$, however the inequality (\[tridva\]) fails a.e. on $[0,1]$. Thus, one cannot directly apply Proposition \[rot\], as well as [*the other*]{} versions of the central limit theorem, for the sums $\sum_1^n\xi_i f_i(x)$. Nevertheless, the condition [**(d$'$)**]{} reserves a possibility for “sufficiently large” set of points $x\in X$ to pick out “long enough” subsum $\sum_{i\in{{\mathcal}I}_x}\xi_if_i(x)$ for which the estimate (\[tridva\]) holds and, consequently, the central limit theorem can be applied. It turns out that the index set ${{\mathcal}I}_x$ may depend on $x$, for this reason there arise some difficulties with the transfer of estimates for subsums to the case of original polynomial.
[**Proof of Theorem \[l1th\]$'$.**]{} Without loss of generality assume ${2\le m\le n^{\varepsilon_1}-1}$ with some fixed ${\varepsilon_1=\varepsilon_1(p_1,p_2)\in(0,\frac12)}$, whose value will be explicitly set later. Moreover, since ${\|f\|_{m,\infty}}$ depends only on the distribution of $f$ (see (\[norma2\])) and $L_1[0,1]$ contains equimeasurable copy of vector $(f_j)_1^n$ we can suppose that $X=[0,1]$ with standard Lebesque measure. The last means that we can use Theorem \[kee\].
[**Step 1.**]{} For each $x\in X$ define the index set $${{\mathcal}A}(x)\equiv {{\mathcal}A}_x:=\Big\{k:1\le k\le n,
\; |f_k(x)|\le n^{-\frac12-\varepsilon_1}\sum_{j=1}^n|f_j(x)|\Big\}.$$ By Chebyshev’s inequality we get $|{{\mathcal}A}_x^c|\le n^{\frac12+\varepsilon_1}$ and $|{{\mathcal}A}_x|\ge {n-n^{\frac12+\varepsilon_1}}$. For each $x\in X$ define a smaller index set: ${\Lambda(x)\equiv\Lambda_x\subset {{\mathcal}A}_x}$ such that ${|\Lambda_x|=n-[n^{\frac12+2\varepsilon_1}]}$ and at every $x$ the set $\Lambda_x$ indexes the ${n-[n^{\frac12+2\varepsilon_1}]}$ least values of $|f_k(x)|$. In order to define the set $\Lambda_x$ formally and provide it with an additional property that for all $k\in\{1,\cdots,n\}$ the set ${\{x:k\in\Lambda(x)\}}$ is $\mu$-measurable we use the following inductive procedure. Let $$\begin{aligned}
k_1(x) &:=
\min\big\{ k\in\{1,\dots,n\}: |f_k(x)|\ge|f_i(x)|~
\forall i=1,\dots,n\big\};\\
{{\mathcal}K}_1(x) &:=\big\{1,\dots,n\big\}\setminus\{k_1(x)\}.\end{aligned}$$ Assume that $k_l(x)$, ${{\mathcal}K}_l(x)$ for $l=1,\dots,j-1$ are defined and set $$\begin{aligned}
k_j(x) &:=
\min\big\{ k\in{{\mathcal}K}_{j-1}(x): |f_k(x)|\ge|f_i(x)|~
\forall i\in{{\mathcal}K}_{j-1}(x)\big\};\\
{{\mathcal}K}_j(x) &:={{\mathcal}K}_{j-1}(x)\setminus\{k_j(x)\}.\end{aligned}$$ It is easy to show that the indices $k_j(x)$, ${j=1,\dots,n}$, are measurable functions of $x$. Set $
\Lambda_x :=\big\{1,\dots,n\big\}\setminus
{{\mathcal}K}_{[n^{1/2+2\varepsilon_1}]}(x).
$ Clearly, $|\Lambda_x|={n-[n^{\frac12+2\varepsilon_1}]}$, ${\Lambda_x\subset{{\mathcal}A}_x}$ and $$\int_X\sum_{k\in \Lambda_x^c}|f_k(x)|d\mu(x) \le\int_X
\sqrt{|\Lambda_x^c|}\cdot
\Big(\sum_{k=1}^n|f_k|^2\Big)^{1/2}d\mu(x) \le
Mn^{\frac34+p_2+\varepsilon_1}.$$ and, consequently, $$\int_X\sum_{k\in \Lambda_x}|f_k(x)|d\mu(x)\ge
n-Mn^{\frac34+p_2+\varepsilon_1}.$$ Non-triviality of this estimate will be ensured by the choice of $\varepsilon_1$, satisfying ${\frac34+p_2+\varepsilon_1<1}$.
[**Step 2.**]{} Set $$~X':=\Big\{x\in X:
\sum_{k\in\Lambda_x}|f_k(x)|\ge \frac13\sum_{k=1}^n|f_k(x)|\Big\}$$ and notice that $
\int_{X\setminus X'}\sum_{k\in\Lambda_x}|f_k|
\le\frac13\int_X\sum_{k=1}^n|f_k|
=n/3$. Therefore, for sufficiently large[^4] $n\ge n_0(\varepsilon_1+p_2,M)$ we have $$\int_{X'}
\sum_{\Lambda_x}|f_k|\ge n-Mn^{\frac34+p_2+\varepsilon_1}-\frac
n3\ge\frac n2.$$ Hölder’s inequality for the function $$F_2(x):=\Big(\sum_{k\in\Lambda_x}|f_k(x)|^2\Big)^{1/2}$$ implies $$\int_{X'}F_2(x)\,d\mu(x)\ge
\int_{X'}|\Lambda_x|^{-\frac12}\cdot\sum_{k\in \Lambda_x}|f_k|\,d\mu(x)
\ge \frac{\sqrt n}2.$$
Notice, if $x\in X'$ and an index set $I_x\subset {{\mathcal}A}_x$ satisfies $|I_x|\ge|\Lambda_x|$, then by the definitions of $\Lambda_x$ and ${{\mathcal}A}_x$ we get $$\begin{aligned}
\sum_{k\in I_x}|f_k(x)|^3 &\le \Big(\sum_{k\in I_x}|f_k(x)|^2\Big)
\cdot\max_{k\in I_x}\big\{|f_k(x)|\big\}
\le \Big(\sum_{k\in I_x}|f_k(x)|^2\Big)
\cdot n^{-\frac12-\varepsilon_1}\sum_{k=1}^n|f_k(x)|\\
&\le n^{-\frac12-\varepsilon_1}\Big(\sum_{k\in I_x}|f_k(x)|^2\Big)
\cdot 3\Big(\sum_{k\in\Lambda_x}|f_k(x)|\Big)\\
&\le 3n^{-\frac12-\varepsilon_1}
|\Lambda_x|^{\frac12}\Big(\sum_{k\in I_x}|f_k(x)|^2\Big)^{3/2}
\le 3n^{-\varepsilon_1} \Big(\sum_{k\in I_x}|f_k(x)|^2\Big)^{3/2}.\end{aligned}$$ Thus $$\label{tridva2}
\frac{\sum_{k\in I_x}|f_k(x)|^3}
{\big(\sum_{k\in I_x}|f_k(x)|^2\big)^{3/2}}\le 3n^{-\varepsilon_1}.
$$ We shall need this inequality to estimate the error term in the central limit theorem.
[**Step 3.**]{} Define the sets $$E_\ell:=\Big\{x\in X':\frac{\sqrt n}4 2^{\ell-1}
\le F_2(x)<\frac{\sqrt n}4 2^\ell \Big\},
\qquad\ \ell=1,2,\dots.$$ Assuming $n$ sufficiently large to ensure that $$\frac{\sqrt n}2\le\int_{X'}F_2\le Mn^{\frac12+p_2},$$ for arbitrary $\varepsilon_2>0$ we have $$\label{st3}
\mu\Big\{x\in X':F_2(x)\ge\frac{\sqrt n}4 2^{n^{\varepsilon_2}} \Big\}
\le\frac4{\sqrt n} 2^{-n^{\varepsilon_2}}\int_{X'}F_2\,d\mu
<
4Mn^{p_2} 2^{-n^{\varepsilon_2}}.$$ Notice $$\sum_{\ell=1}^\infty\int_{E_\ell}F_2\,d\mu=
\int\limits_{\{F_2\ge \frac{\sqrt n}4\}\cap X'}
\!\!\!F_2 \,d\mu
\ge\frac{\sqrt n}4$$ so that at least one of the following cases takes place:
- $\sum\limits_{1\le \ell< n^{\varepsilon_2}}\,
\int\limits_{E_\ell}F_2\ge {\sqrt n}/8$;
- $\sum\limits_{\ell\ge n^{\varepsilon_2}}\,
\int\limits_{E_\ell}F_2> {\sqrt n}/8$.
Assume first that [**(i)**]{} holds. Define the following index set $${{\mathcal}L}:=\Big\{1\le
\ell<n^{\varepsilon_2}:
\int_{E_\ell}F_2\ge\frac{n^{\frac12-\varepsilon_2}}{16}\Big\}$$ and notice that $\sum\limits_{\ell\in{{\mathcal}L}}\int_{E_\ell}F_2\ge\frac{\sqrt n}{16}$. Denote $$\mu_\ell:=\mu E_\ell;\qquad\qquad \rho_\ell:=\frac{\sqrt n}4 2^{\ell-1}.$$ It is easy to see that in the case [**(i)**]{} we have $$\sum_{\ell\in{{\mathcal}L}}\mu_\ell\rho_\ell\ge\frac{\sqrt n}{32};
\qquad\qquad \mu_\ell\rho_\ell\ge\frac{n^{\frac12-\varepsilon_2}}{32}
\qquad\ \text{for all~}\ell\in{{\mathcal}L}.$$
Further, on steps 5–11 we shall deal with the case [**(i)**]{} only. The case [**(ii)**]{} is simpler and we shall consider it on the final step 12.
[**Step 4.**]{} Assume ${{\mathcal}J}_x\subset \{1,\dots,n\}$ is an index set (which may depend on $x\in X$), satisfying ${|{{\mathcal}J}_x|\ge n-n^{\frac12+2\epsilon}}$ (${\epsilon<1/4}$), and for each $k_0$ the set ${\{x\in X: k_0\in{{\mathcal}J}_x\}}$ is $\mu$-measurable. Then for an arbitrary set of signs ${\theta_k=\pm1}$, $k=1,\dots,n$, it follows that $$\begin{aligned}
\int_X \big|\sum_{k\in{{\mathcal}J}_x}\theta_kf_k(x)\big|d\mu(x)&\le
\int_X\big|\sum_{k=1}^n\theta_kf_k(x)\big|d\mu(x)
+\int_X \big|\!\sum_{k\in{{\mathcal}J}^c_x}\theta_kf_k(x)\big|d\mu(x)\\
&\le Mn^{\frac12+p_1}+
\int_X\sqrt{|{{\mathcal}J}^c_x|}\cdot\Big(\sum_{k=1}^n|f_k(x)|^2\Big)^{1/2}d\mu(x)\\
&\le Mn^{\frac12+p_1}+Mn^{\frac34+p_2+\epsilon}
\!\le2Mn^{\frac12+\max\{p_1,p_2+\frac14+\epsilon\}}\!.\end{aligned}$$ Therefore, by Lemma \[geomlem\] for arbitrary coefficients $\{a_k\}_1^n$ we have $$\int_X\big|\sum_{k\in{{\mathcal}J}_x}a_kf_k\big|d\mu(x)\le
Cn^{\frac14+\frac12\max\{p_1,p_2+\frac14+\epsilon\}}
\Big(\sum_{k=1}^n |a_k|^2\Big)^{\frac12}$$ with a constant ${C(M)>0}$.
[**Step 5.**]{} Let us choose the constants $\varepsilon_1,\varepsilon_2,\varepsilon_3>0$ such that the following inequalities take place: $$\label{epsdelta}
\begin{array}{rl}
\vphantom{\sum\limits_k}
2\varepsilon_2+\varepsilon_3 &\le \frac14(1-2p_1-4p_2);\\
\frac{\varepsilon_1}2+2\varepsilon_2+\varepsilon_3
&\le\frac32\big(\frac1{12}-p_2\big).
\end{array}$$ The reasons for such a choice will be clear soon. To be definite we could set $$2\varepsilon_2=\frac{\varepsilon_1}2=\varepsilon_3:=
\min\big\{\frac12\big(\frac1{12}-p_2\big),\frac18(1-2p_1-4p_2)
\big\}.$$ The constraints imposed on $p_1$ and $p_2$ ensure that $\varepsilon_1$, $\varepsilon_2$ and $\varepsilon_3$ are positive.
Consider some $E_\ell$, $\ell\in{{\mathcal}L}$. For ${\bar{x}=(x_i)_1^m\in (E_\ell)^m}$ set $$\begin{aligned}
{{\mathcal}I}(x_1,\dots,x_m)&:=\bigcap_{j=1}^m{{\mathcal}A}_{x_j};\\
{\varphi}(x_1,\dots,x_m)&:=
\frac1{m^2}\sum_{{\genfrac{}{}{0pt}{}{j,k=1}{j\ne k}}}^m
\Big|\sum_{i\in {{\mathcal}I}(\bar{x})}f_i(x_j)f_i(x_k)\Big|.\end{aligned}$$ Note that $|{{\mathcal}I}(\bar{x})|\ge n-mn^{\frac12+\varepsilon_1}
\ge n-n^{\frac12+2\varepsilon_1}$. Using the fact that the set ${{\mathcal}I}(x_1,\dots,x_m)$ and the function $\varphi(x_1,\dots,x_m)$ are invariant with respect to permutations of the coordinates $(x_1,\ldots,x_m)$, we get $$\begin{gathered}
\frac1{\mu_\ell^m}
\int\limits_{(E_\ell)^m}{\varphi}(\bar{x})d\mu^m(\bar{x})
=\frac{m(m-1)}{m^2\mu_\ell^m} \int\limits_{(E_\ell)^m}
\Big|\! \sum_{k\in {{\mathcal}I}(\bar{x})}\! f_k(x_1)f_k(x_2)\Big|
d\mu^{m}(\bar x)
\le\\
\le\frac1{\mu_\ell^m}\!\!
\int\limits_{(E_\ell)^{m-2}}\!\!\!\Big\{ \int\limits_X\!\int\limits_X
\Big|\! \sum_{k\in {{\mathcal}I}(\bar{x})}\! f_k(x_1)f_k(x_2)\Big|
d\mu(x_1)d\mu(x_2)\Big\}d\mu^{m-2}(x_3,\dots,x_m).\end{gathered}$$ Applying the arguments of Step 4 for the set ${{\mathcal}I}(\bar{x})$ с $\epsilon=\varepsilon_1$, we get $$\begin{gathered}
\int_X\int_X \Big|\sum_{k\in {{\mathcal}I}(\bar{x})}
f_k(x_1)f_k(x_2)\Big| d\mu(x_1)d\mu(x_2)\le\\
\le
Cn^{\frac14+\frac12\max\{p_1,p_2+\frac14+\varepsilon_1\}}
\cdot\int_X\Big(\sum_{k=1}^n|f_k(x_1)|^2\Big)^{1/2}d\mu(x_1),\end{gathered}$$ which in turn implies $$\label{st5}
\frac1{\mu_\ell^m} \int_{(E_\ell)^m}
{\varphi}(\bar{x})d\mu(x_1)\dots d\mu(x_m) \le
\frac{CM}{\mu_\ell^2}
n^{\frac34+p_2+\frac12\max\{p_1,p_2+\frac14+\varepsilon_1\}}.$$
We are going to show that there exists a set $G_\ell\subset(E_\ell)^m$ such that ${\mu^mG_\ell\ge\frac12\mu_\ell^m}$ and $${\varphi}(\bar{x})\le C_0\rho_\ell^2n^{-\varepsilon_3}$$ for all ${\bar{x}=(x_i)_1^m\in G_\ell}$ with some constants $\varepsilon_3$, $C_0(p_1,p_2,M)>0$. Indeed, whenever ${\ell\in{{\mathcal}L}}$ we have $\mu_\ell\rho_\ell\ge2^{-5}n^{\frac12-\varepsilon_2}$ which, combined with (\[st5\]), implies $$\frac1{\mu_\ell^m} \int_{(E_\ell)^m}{\varphi}(\bar{x})d\mu^m(\bar{x})
\le
CM2^{10}{\rho_\ell^2}n^{2\varepsilon_2-1}
n^{\frac34+p_2+\frac12\max\{p_1,p_2+\frac14+\varepsilon_1\}}
\le CM2^{10}\rho_\ell^2n^{-\varepsilon_3},$$ where the constants $\varepsilon_1$, $\varepsilon_2$, $\varepsilon_3$ chosen positive and satisfying the inequalities (\[epsdelta\]). Thus, by Chebyshev’s inequality for the set $$G_\ell:=\big\{\bar x\in E_\ell:
\varphi(\bar x)\le CM2^{11}\rho_\ell^2n^{-\varepsilon_3}\big\}
,\qquad\ell\in{{\mathcal}L},$$ we have $\mu^m
G_\ell\ge\frac12\mu_\ell^m$.
[**Step 6.**]{} For every $\ell\in{{\mathcal}L}$ and $(x_j)_1^m\in G_\ell$ define the random vector ${\eta=(\eta_j)_1^m}$ by $$\eta_j(\omega):= |{{\mathcal}I}(\bar x)|^{-1/2}\!\!
\sum_{i\in {{\mathcal}I}(x_1,\dots,x_m)}\!
\xi_i(\omega)f_i(x_j).$$ We are going to show that $$\label{st6}
\sup_{v\in{\Bbb R}^m}{\sf P}
\Big\{\max_{1\le j\le m}
|\eta_j+v_j|\le
\alpha\rho_\ell\sqrt{\frac{\log m}{|{{\mathcal}I}(\bar x)|}}\Big\}
\le\frac{K_0}{m^q}$$ with some positive constants $\alpha$, $K_0$ and $q$ which depend only on $p_1$, $p_2$, $M$. Obviously, inequality (\[st6\]) holds if $$\min_{\theta_j=\pm1}
{\sf P}\Big(\bigcup_{j=1}^m\Big\{\theta_j\eta_j\ge
\alpha\rho_\ell\Big(\frac{\log m}{|{{\mathcal}I}(\bar x)|}\Big)^{1/2}
\Big\}\Big)
\ge 1-\frac{K_0}{m^q}$$ (to see that it suffices to take ${\theta_j=-\text{sign}(v_j)}$). To prove the latter inequality we show that $$\min_{\theta_j=\pm1}
{\sf P}\Big(\bigcup_{j=1}^m\Big\{\theta_j\eta_j
\ge \alpha\sqrt{d_{j,\ell}\log m}
\Big\}\Big) \ge 1-\frac{K_0}{m^q},
\eqno(\ref{st6}')$$ $$\text{where}\quad
d_{j,\ell}:={\sf D}(\eta_j)=
\frac1{|{{\mathcal}I}(\bar x)|}\sum_{i\in {{\mathcal}I}(\bar x)}\!|f_i(x_j)|^2\ge
\frac1{|{{\mathcal}I}(\bar x)|}\!\sum_{i\in \Lambda(x_j)}\!\!|f_i(x_j)|^2
\ge\frac{\rho_\ell^2}{|{{\mathcal}I}(\bar x)|}.$$ (While estimating $d_{j,\ell}$ we used the definition of $\Lambda_x$ and the fact that $|{{\mathcal}I}(\bar x)|\ge {n-m\cdot n^{\frac12+2\varepsilon_1}\ge|\Lambda_x|}$.)
For a fixed set of signs $\{\theta_j\}_1^m$ ($\theta_j=\pm1$) denote $$\begin{aligned}
U_j
&=U_j(\alpha):=
\big\{\omega\in\Omega:\theta_j\eta_j\ge \alpha\sqrt{d_{j,\ell}
\log m}\big\};\\
v_{j,k} &:=\frac{\theta_j\theta_k}{|{{\mathcal}I}(\bar x)|}
\sum_{i\in {{\mathcal}I}(\bar x)}f_i(x_j)f_i(x_k),\end{aligned}$$ and note that by the definition of the set $G_\ell$ $$\label{st6nv}
\frac1{m^2}\sum_{{\genfrac{}{}{0pt}{}{j,k=1}{j\ne k}}}^m|v_{j,k}|\le
C_0\rho_\ell^2n^{-\varepsilon_3}|{{\mathcal}I}(\bar x)|^{-1}.$$
In what follows we are going to demonstrate that the random variables $\eta_j$ are “almost” Gaussian and, moreover, “almost” parewise Gaussian. Then we apply Lemma \[lem3tex\] (with parameters ${R=n}$, $P=m$) and establish for the events $U_j$ the following inequality: $$\label{l1star2}
\sum_{j,k=1}^m
{\sf P}\big(U_j(\alpha)\cap U_k(\alpha)\big)
\le (1+L_0 m^{-q}) \Big(\sum_{j=1}^m{\sf P}
\big(U_j(\alpha)\big)\Big)^2$$ with some constants $L_0(p_1,p_2,M)$, $q(p_1,p_2)>0$, this inequality with help of Lemma \[lem1\] implies (\[st6\]$'$). Thus, to prove (\[st6\]) it suffices to prove (\[l1star2\]). Steps 7 and 8 are devoted to the proof of (\[l1star2\]).
[**Step 7.**]{} Let $(h_j)_1^m$ denote a Gaussian vector with zero mean and the covariances: ${{\sf E}h_jh_k=v_{j,k}}$ and ${\Psi_j=\Psi_j(\alpha)
:=\{{\omega:h_j>\alpha\sqrt{d_{j,\ell}\log m}}\}}$. Note that $${{\sf P}(\Psi_j)=(2\pi d_{j,\ell})^{-1/2}
\int_{\alpha\sqrt{d_{j,\ell}\log m}}^\infty
\exp\big(-\frac{y^2}{2d_{j,\ell}}\big)dy}.
$$ We are going to apply Lemma \[lem3tex\] to the vectors $(\eta_j)_1^m$ and $(h_j)_1^m$. In order to check the conditions (\[lem3usl1\]), (\[lem3usl2\]) we shall use Proposition \[rot\].
Given fixed $\bar x=(x_j)_1^m\in G_\ell$ and a set of signs $\{\theta_j\}_1^m$, ${\theta_j=\pm1}$, let us apply Proposition \[rot\] to the set of random variables ${\big\{\theta_jf_i(x_j)\xi_i(\omega)\big\}_{i\in {{\mathcal}I}(\bar x)}}$ for each ${1\le j\le m}$. Here ${N=|{{\mathcal}I}(\bar x)|}$ and $$\begin{aligned}
m_3 &=
\frac1{|{{\mathcal}I}({\bar x})|}
\sum_{k\in {{\mathcal}I}({\bar x})}{\sf E}|\xi_k|^3|f_k(x_j)|^3
\le \frac{M^3}{|{{\mathcal}I}(\bar x)|} \sum_{k\in {{\mathcal}I}(\bar x)}|f_k(x_j)|^3\\
\lambda &=V=
\frac1{|{{\mathcal}I}(\bar x)|}\sum_{k\in {{\mathcal}I}({\bar x})} {\sf E} |\xi_k|^2 |f_k(x_j)|^2
=d_{j,\ell}.
\end{aligned}$$ Taking into account (\[tridva2\]), we get $$|{{\mathcal}I}({\bar x})|^{-1/2}m_3 \lambda^{-3/2}\le 3M^3n^{-\varepsilon_1}$$ and, consequently, by Proposition \[rot\] applied for the variable $\theta_j\eta_j$ it follows that $$\begin{aligned}
|{\sf P}(U_j(\alpha))- {\sf P}(\Psi_j(\alpha))|
&\equiv
\big|{\sf P}
(U_j(\alpha))-\Big(\frac{1}{2\pi d_{j,\ell}}\Big)^{1/2}
\hspace{-.1em plus 0em minus .1em}
\int_{\alpha\sqrt{d_{j,\ell}\log m}}^\infty
\hspace{-.2em plus 0em minus .2em}
\exp\Big\{\frac{-y^2}{2d_{j,\ell}}\Big\}dy\big|
\notag\\
\label{l1tex1}
&\le3K_1(1)M^3n^{-\varepsilon_1}.\end{aligned}$$
[**Step 8.**]{} As in the proof of Lemma \[lem2\] set $$\sigma_1:=\big\{(j,k): 1\le j\ne k\le m,
|v_{j,k}|<\frac{\rho_\ell^2}{8|{{\mathcal}I}(\bar x)|}\big\}.$$ By Chebyshev’s inequality it follows that (see (\[st6nv\])) $$\label{l1tex3}
|\sigma_1^c|\le
\frac{8|{{\mathcal}I}(\bar x)|}{\rho_\ell^2}
\sum_{{\genfrac{}{}{0pt}{}{j,k=1}{j\ne k}}}^m|v_{j,k}|
\le {8C_0m^2n^{-\varepsilon_3}}
\le 8C_0m^{2-{\varepsilon_3}/{\varepsilon_1}}.$$ (We used the assumption that $m\le n^{\varepsilon_1}$).
To apply Lemma \[lem3tex\] and, thus, to prove (\[l1star2\]) it remains to demonstrate that an estimate of type (\[lem3usl2\]) holds for all ${(j,k)\in\sigma_1}$.
For a fixed pare $s=(j,k)\in \sigma_1$ consider the following set of random vectors in ${\Bbb R}^2$: $$\Big\{
\Big(\frac{\theta_j\xi_i(\omega)f_i(x_j)}{\sqrt{d_{j,\ell}}},
\frac{\theta_k\xi_i(\omega)f_i(x_k)}{\sqrt{d_{k,\ell}}}
\Big)\Big\}_{i\in {{\mathcal}I}(x_1,\dots,x_m)}.$$ Apply Proposition \[rot\] to this set. Here $$\begin{aligned}
m_3^s
&=\!\frac1{|{{\mathcal}I}(\bar x)|} \sum_{i\in {{\mathcal}I}(\bar x)}\!\!
{\sf E}|\xi_i|^3
\Big(\frac{|f_i(x_j)|^2}{d_{j,\ell}}
+\frac{|f_i(x_k)|^2}{d_{k,\ell}}
\Big)^{3/2};\\
V^s
&=\frac1{|{{\mathcal}I}(\bar x)|}
\left(\!
\mbox{$
\begin{array}{l}
\sum\limits_{i\in {{\mathcal}I}(\bar x)\vphantom{\frac1I}}
\frac{|f_i(x_j)|^2}{d_{j,\ell}}\\
\frac{\theta_j\theta_k}{\sqrt{d_{j,\ell}d_{k,\ell}}}
\sum\limits_{i\in {{\mathcal}I}(\bar x)}\!\!\!f_i\!(x_j)f_i(x_k)
\end{array}
$}
\hspace{-1em}
\mbox{$
\begin{array}{r}
\frac{\theta_j\theta_k}{\sqrt{d_{j,\ell}d_{k,\ell}}}
\sum\limits_{i\in {{\mathcal}I}(\bar x)\vphantom{\frac1I}}\!\!\! f_i(x_j)f_i(x_k)\\
\sum\limits_{i\in {{\mathcal}I}(\bar x)}
\frac{|f_i(x_k)|^2}{d_{k,\ell}}
\end{array}
$}
\right).\end{aligned}$$ Therefore, $$V^s=
\left(
\begin{array}{cc}
1 & \frac{v_s}{\sqrt{d_{j,\ell}d_{k,\ell}}} \\
\frac{v_s}{\sqrt{d_{j,\ell}d_{k,\ell}}} & 1
\end{array}\right).$$ Since $s\in\sigma_1$, we can estimate: $$\det V^s=1-\frac{v_s^2}{d_{j,\ell}d_{k,\ell}}
\ge 1-\frac{1}{64}
\ge\frac12.$$ Since the matrix $V^s$ is positive definite both its eigenvalues: ${\lambda_2\ge\lambda_1}$ are positive. Therefore, taking into account that ${\lambda_1+\lambda_2}=\text{\rm{trace\ }} V^s=2$, we get $$\frac12\le\det V^s=\lambda_2\lambda_1
\le 2\lambda_1$$ so that $\lambda_1>1/4$. Taking into account (\[tridva2\]) and the definition of $d_{j,\ell}$, ${{\mathcal}I}(\bar x)$, we estimate $$m^s_3 \le 8M^3|{{\mathcal}I}(\bar x)|^{1/2}\cdot 3n^{-\varepsilon_1}.$$
Thus, by Proposition \[rot\] we have $$\begin{aligned}
\Big|&{\sf P}( U_j\cap U_k)-
\frac1{2\pi\sqrt{\det V^s}}\!
\hspace{-.2em plus 1pt minus .5em}
\int\limits_{\alpha\sqrt{d_{j,\ell}\log m}}^\infty\,
\int\limits_{\alpha\sqrt{d_{k,\ell}\log m}}^\infty
\hspace{-.4em }
e^{-\frac12(Y,(V^s)^{-1}Y)}dy_1dy_2
\Big|\equiv \notag\\
&\label{l1tex2}
\equiv
\big|{\sf P} ( U_j\cap U_k)-{\sf P}(\Psi_j\cap \Psi_k)\big|
\le \frac{K_1(2)m_3^s\lambda_1^{-3/2}}{|{{\mathcal}I}(\bar x)|^{1/2}}
< \frac{200M^3K_1(2)}{n^{\varepsilon_1}}.\end{aligned}$$
Now, equipped with the estimates (\[l1tex1\]), (\[l1tex3\]) and (\[l1tex2\]) we can apply Lemma \[lem3tex\] with parameters ${R=n}$, ${P=m}$, ${\delta_1=\delta_2=\varepsilon_1}$ and ${\delta=\delta_3={\varepsilon_3}/{\varepsilon_1}}$ to derive (\[l1star2\]) and, thus, to prove (\[st6\]) with some[^5] fixed ${\alpha>0}$, ${q(p_1,p_2)>0}$.
[**Step 9.**]{} The aim of this step is to prove that $$\label{l1st11}
{\sf P}\Big(\bigcup_{j=1}^m
\big\{\omega:\big|\sum_{i=1}^n\xi_i(\omega)f_i(x_j)\big|
\le\alpha\rho_\ell\sqrt{\log m}\big\}\Big)
\le K_0m^{-q}
$$ for $(x_j)_1^m\in G_\ell$, $\ell\in{{\mathcal}L}$. This inequality will easily follow from (\[st6\]) and the following
\[tver\] Let $\eta=(\eta_j)_1^m$ and $\eta^c=(\eta^c_j)_1^m$ be independent random vectors and let ${B\subset {\Bbb R}^m}$ be an open or closed set, satisfying $$\sup_{v\in{\Bbb R}^m}
{\sf P}\big\{\eta+v\in B\big\}\le p$$ with some fixed ${p\in(0,1)}$. Then $
{\sf P}\big\{\eta+\eta^c\in B\big\}\le p.
$
[**Proof of Lemma \[tver\].**]{} Let $\chi(v,w)$ denote the characteristic function of the set ${\{(v,w):v+w\in B\}\subset{\Bbb R}^{2m}}$, then $${\sf P}\big\{\eta+\eta^c\in B\big\}
=\int_{{\Bbb R}^m}\int_{{\Bbb R}^m}
\chi(v,w)\, dF_{\eta|\,\eta^c}(v|\,w)dF_{\eta^c}(w),$$ where $F_{\eta|\,\eta^c}(v|\,w)$ is the conditional distribution of vector $\eta$ given $\eta^c$ (see [@shir] for the definition), $F_{\eta^c}(w)$ is the distribution of $\eta^c$. Since the vectors $\eta$ and $\eta^c$ are independent, we have ${F_{\eta |\,\eta^c}(v|\,w)=F_{\eta}(v)}$ and $$\begin{aligned}
{\sf P}\big\{\eta+\eta^c\in B\big\}
&= \int_{{\Bbb R}^m}\int_{{\Bbb R}^m}
\chi(v,w)\, dF_{\eta}(v)dF_{\eta^c}(w) \\
&=\int_{{\Bbb R}^m} {\sf P}\big\{\eta\in B-w\big\}
\,dF_{\eta^c}(w)\le p\end{aligned}$$ Lemma \[tver\] is proved.$\Box$
Now, to deduce (\[l1st11\])[^6] from (\[st6\]) it suffices to notice that the random vectors, defined by $$\eta_j:=
\frac1{\sqrt{|{{\mathcal}I}(\bar x)|}}\sum_{i\in {{\mathcal}I}(\bar x)} \xi_if_i(x_j),
\qquad\
\eta^c_j:=\frac1{\sqrt{|{{\mathcal}I}(\bar x)|}}\sum_{i=1}^n
\xi_if_i(x_j)-\eta_j,$$ are independent and apply Lemma \[tver\] to these vectors with parameter ${p=K_0m^{-q}}$ and ${B:=\{(y_j)_1^m\in{\Bbb R}^m: |y_j|\le\alpha\rho_\ell
\sqrt{\frac{\log m}{|{{\mathcal}I}(
\bar x)|}}\}}$.
[**Step 10.**]{} For $\ell\in{{\mathcal}{L}}$ set $$\Omega_\ell:=\Big\{\omega\in\Omega:
\mu^m\big\{\bar x\in G_\ell:
\!\max_{1\le j\le m}\big|
F_n(\omega,x_j)\big|
\ge\alpha\rho_\ell\sqrt{\log m}\big\}
\ge\frac{\mu^mG_\ell}2\Big\}.$$ From (\[l1st11\]) it follows that $$\gamma:=
{\sf P}\!\times\!\mu^m
\Big\{(\omega,\bar{x})\in\Omega\times G_\ell:
\!\max_{1\le j\le m}\big|
F_n(\omega,x_j)\big|
\ge\alpha\rho_\ell\sqrt{\log m} \Big\}
\ge\big(1-\frac{K_0}{m^{q}}\big)\mu^mG_\ell.$$ Thus, we have $${\sf P}(\Omega_\ell)\mu^mG_\ell
+ \big(1-{\sf P}(\Omega_\ell)\big)\frac{\mu^mG_\ell}2
\ge \gamma \ge (1-{K_0}{m^{-q}})\mu^mG_\ell$$ and, therefore ${\sf P}(\Omega_\ell)\ge1-2K_0m^{-q}$.
We shall need the following
\[tver2\] Let $T_\ell\ge0$ be numbers, satisfying $\sum_{\ell=1}^L T_\ell=T$ ($L$ may be infinite), and let $\Omega_\ell$ be events, satisfying ${\sf P}(\Omega_\ell)\ge1-p$. Then $${\sf P}\big\{\sum_{\ell=1}^LT_\ell I_{\Omega_\ell}(\omega)
\le\frac T2\big\}\le2p,$$ where $I_{\Omega_\ell}$ are the indicators of $\Omega_\ell$.
[**Proof of Lemma \[tver2\].**]{} Set $q={\sf P}(\sum_{\ell=1}^LT_\ell I_{\Omega_\ell}\le\frac T2)$, then $$q\frac T2+(1-q)T\ge
{\sf E}\sum_{\ell=1}^L T_\ell I_{\Omega_\ell}\ge T(1-p).$$ Therefore, $q\le2p$. Lemma \[tver2\] is proved.$\Box$
Let us apply Lemma \[tver2\] to the numbers $T_\ell:=\mu_\ell\rho_\ell$ and the events $\Omega_\ell$ for ${\ell\in{{\mathcal}L}}$ with the parameter ${p=2K_0m^{-q}}$. We derive that there exists an event $\Omega_0$ and a subset ${{\mathcal}{L}}_0(\omega)\subset{{\mathcal}{L}}$ such that $\Omega_0\subset\bigcap_{\ell\in{{\mathcal}{L}}_0}\Omega_\ell$, $${\sf P}(\Omega_0)\ge1-4K_0m^{-q}$$ and (see Step 3) $$\label{l1st12}
\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}\mu_\ell\rho_\ell
\ge\frac12\sum_{\ell\in{{\mathcal}{L}}}\mu_\ell\rho_\ell
\ge\frac{\sqrt n}{64}
\qquad\text{a.s. on}\quad \Omega_0$$
[**Step 11.**]{} For $\ell\in{{\mathcal}{L}}$ set $$\Delta_\ell(\omega):=
\big\{x\in E_\ell:
|F_n(\omega,x)|\ge\alpha\rho_\ell\sqrt{\log m}\big\}.$$ Almost surely on $\Omega_\ell$ we have $$\begin{aligned}
(\mu E_\ell)^m-\big(\mu E_\ell-\mu\Delta_\ell\big)^m
&=\mu^m\big\{
\bar{x}\in E_\ell^m:\max_{1\le j\le m}
|F_n(\omega,x_j)|\ge\alpha\rho_\ell\sqrt{\log
m} \big\} \\
&\ge\frac{\mu^mG_\ell}2\ge\frac{(\mu E_\ell)^m}4
\equiv\frac{\mu_\ell^m}4.\end{aligned}$$ Therefore, $1\ge\frac14+\big(1-\frac{\mu\Delta_\ell}{\mu_\ell}\big)^m$ and $$\frac{\mu\Delta_\ell}{\mu_\ell}\ge1-\Big(\frac34\Big)^{1/m}
\ge\frac1{4m}.$$ There exist some subsets $\Delta'_\ell(\omega)\subset\Delta_\ell(\omega)$ such that ${\mu\Delta'_\ell=\frac{\mu_\ell}{4m}}$. Set ${E'=\bigcup_{\ell\in{{\mathcal}{L}}_0}\Delta'_\ell(\omega)}$, ${\mu E'=\frac1{4m}\sum_{\ell\in{{\mathcal}{L}}_0}\mu_\ell}$.
Now, for almost all $\omega\in\Omega_0$ ($\Omega_0$ and ${{\mathcal}{L}}_0$ defined on the previous step) by Theorem \[kee\] we can estimate the integral-uniform norm of the random polynomial $F_n(\omega,x)$ of the type (\[polinom2\]), using its “relative” norm ${\|\cdot\|_m^*}$ (see (\[norma3\])). We have $$\begin{aligned}
\|F_n(\omega,\cdot)\|_{m}^*
&\ge
\frac1{\mu E'(\omega)}
\Big(
\int_{E'(\omega)}
|F_n(\omega,x)|d\mu(x)\Big)
\Big(1-\big(1-
\mu E'(\omega)\big)^m\Big)\ge\\
&\ge
\frac1{\mu E'(\omega)}
\Big(\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}
\frac{\mu_\ell}{4m}\alpha\rho_\ell\sqrt{\log m}\Big)
\Big(1-\big(1-\mu E'(\omega)\big)^m\Big)=\\
&=
\alpha\sqrt{\log m}\Big(
\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}\mu_\ell\rho_\ell\Big)
\cdot t_0^{-1}\Big(1-\big(1-\frac{t_0}{4m}\big)^m\Big),\end{aligned}$$ where $t_0=\sum_{\ell\in{{\mathcal}{L}}_0}\mu_\ell$. Taking into account (\[l1st12\]) and that ${1-(1-\frac t{4m})^m\ge t/8}$ for $t\in(0,1)$ and $m\ge1$, by Theorem \[kee\] (see (\[kl\])) we get $$\|F_n(\omega,\cdot)\|_{m,\infty}\ge\|F_n(\omega,\cdot)\|_{m}^*
\ge 8^{-3}\alpha\sqrt{n\log m}
\qquad\text{a.s. on}~\Omega_0.$$ This, combined with ${\sf P}(\Omega_0)\ge1-{4K_0}{m^{-q}}$, proves (\[l1th1\]) if the case[** (i)**]{} takes place (see Step 3).
[**Step 12.**]{} In order to finish the proof of the theorem, it remains to consider the case[** (ii)**]{} from Step 3. Notice, since the inequality (\[tridva2\]) holds a.e. on $X'$ (for ${I_x={{\mathcal}A}_x}$), we can apply Proposition \[rot\] (in the one-dimensional case) for the sum of random variables ${\eta_x:={|{{\mathcal}A}_x|}^{-1/2}
\sum_{i\in {{\mathcal}A}_x}\xi_if_i(x)}$. We get $$\sup_{v\in{\Bbb R}}
{\sf P}\Big\{|\eta_x+v|\le\gamma\sqrt{d(x)}\Big\}=
(2\pi)^{-1/2}\int_{-\gamma}^{\gamma}
e^{-y^2/2}dy
+O(n^{-\varepsilon_1}),$$ where $d(x):=|{{\mathcal}A}_x|^{-1}\sum_{i\in {{\mathcal}A}_x}|f_i(x)|^2$. Set $\gamma=m^{-\varepsilon_1}$, then for almost all ${x\in E_\ell}$, ${\ell\ge1}$, we get $$\sup_{v\in{\Bbb R}}
{\sf P}\Big\{|\eta_x+v|\le\sqrt{d(x)}
m^{-\varepsilon_1}\Big\}
\le K'm^{-\varepsilon_1}$$ with a constant $K'(p_1,p_2,M)>0$. Taking into account that ${|{{\mathcal}A}_x|d(x)\ge\rho_\ell^2}$ for ${x\in E_\ell}$, by Lemma \[tver\] we get $${\sf P}\Big\{\big|\sum_{i=1}^n\xi_i f_i(x)\big| \le
\rho_\ell m^{-\varepsilon_1}\Big\}
\le K'm^{-\varepsilon_1}
\qquad
\mbox{for all $\ell\ge1$ a.e. on $E_\ell$}.$$
For each $\ell\ge n^{\varepsilon_2}$ define the sets: $$\Omega'_\ell:=\Big\{\omega\in\Omega:
\mu\big\{x\in E_\ell:
\big| F_n(\omega,x)\big|
\ge \rho_\ell m^{-\varepsilon_1}\big\}
\ge\frac{\mu E_\ell}2\Big\}.$$ As on Step 10, it is easy to prove that there exist events $\Omega'_0$ and an index set ${{{\mathcal}{L}}_0(\omega)\subset\{\ell:\ell\ge n^{\varepsilon_2}\}}$ such that ${\Omega'_0\subset\bigcap_{\ell\in{{\mathcal}{L}}_0}\Omega'_\ell}$, ${{\sf P}(\Omega'_0)\ge1-{4K'}{m^{-\varepsilon_1}}}$ and $$\label{l1last}
\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}\mu_\ell\rho_\ell
\ge\frac1{2}\sum_{\ell\ge n^{\varepsilon_2}}\mu_\ell\rho_\ell
\ge\frac{\sqrt n}{32}$$ a.s. on $\Omega'_0$. As on Step 11 set $$\Delta_\ell(\omega):=\big\{x\in E_\ell:|F_n(\omega,x)|\ge
m^{-\varepsilon_1}\rho_\ell\big\},
\quad\ \ell\ge n^{\varepsilon_2}.$$ By definition of $\Omega'_\ell$ we have $\mu\Delta_\ell\ge{\mu E_\ell}/2\equiv{\mu_\ell}/2$ a.s. on $\Omega'_\ell$, so there exist subsets ${\Delta'_\ell(\omega)\subset\Delta_\ell(\omega)}$ such that ${\mu\Delta'_\ell=\mu_\ell/2}$. Set ${E':=\bigcup_{\ell\in{{\mathcal}{L}}_0}\Delta'_\ell(\omega)}$. Estimate the ${\|\cdot\|_m^*}$-norm of random polynomial (\[polinom2\]) (see (\[norma3\])) a.s. on $\Omega'_0$ as follows $$\begin{aligned}
\|F_n(\omega,\cdot)\|_{m}^*
&\ge
\frac1{\mu{E'(\omega)}}
\Big(
\int_{E'(\omega)}
|F_n(\omega,x)|d\mu(x)\Big)
\Big(1-\big(1-\mu{E'(\omega)}\big)^m\Big)\\
&\ge
\frac{2}{\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}\!\mu_\ell}
\Big(\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}
\frac{\mu_\ell}{2}\rho_\ell m^{-\varepsilon_1}\Big)
\Big(1-\Big(1-\!\!\!
\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}\frac{\mu_\ell}{2}\Big)^m\Big)\\
&=
m^{-\varepsilon_1}\Big(
\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}\mu_\ell\rho_\ell\Big)
\cdot\frac{1-(1-t)^m}{2t},\end{aligned}$$ where $t=\sum_{\ell\in{{\mathcal}{L}}_0(\omega)}{\mu_\ell}/2
<Mn^{p_2}2^{-n^{\varepsilon_2}}$ (see (\[st3\])). Notice that ${(1-y)^m\le1-\frac m2 y}$ for ${y\in(0,\frac1m)}$, so provided sufficiently large ${n>n_0(p_1,p_2,M)}$ we have ${t^{-1}(1-(1-t)^m)>m/2}$. Therefore, combining (\[l1last\]) and Theorem \[kee\], we obtain $$\|F_n(\omega,\cdot)\|_{m,\infty}>128^{-1}m^{1-\varepsilon_1}\sqrt{n}
\qquad \ \ \text{a.s. on\ } \Omega'_0.$$ The proof of Theorem \[l1th\]$'$ is completed.
\[remgenth2\] (An analog of a remark from [@kt1]). In the statement of Theorem \[l1th\] the condition of uniform boundness of the third moments ${{\sf E}|\xi_i|^3\le M^3}$ can be relaxed to a weaker condition: ${{\sf E}|\xi_i|^{2+\varepsilon}\le M}$ with some ${\varepsilon>0}$. In this case the constants in (\[l1th1\]) would depend also on $\varepsilon$. In order to prove such a statement it suffices to apply instead of Proposition \[rot\] with a more precise version of the central limit theorem (Corollary 18.3 in [@bhat]).
Applications and open problems {#applic}
==============================
[**Applications of the Integral-Uniform Norm.**]{} In [@sem] Montgomery and Semenov in connection with their research of strictly singular embeddings of rearrangement invariant spaces in $L_1[0,1]$ (i.e. the spaces whose norms are invariant with respect to measure invariant changes of variable) put forward a hypothesis which we formulate in somewhat simplified form:
\[consem2\] For an arbitrary set of functions $\{f_i\}_{i=1}^n$ from $L_1[0,1]$ such that ${\|f_i\|_1=1}$, there exist a set of signs $\{\theta_i\}_{i=1}^n$, ${\theta_i=\pm1}$, and a constant ${c_0>0}$ such that $$\label{se0}
\big\|\sum_{i=1}^n \theta_i f_i\big\|'_{2^k}\equiv
\sup_{{\genfrac{}{}{0pt}{}{\Delta\subset [0,1]}{\mu\Delta=2^{-k}}}}
\Big\{ 2^k
\int_\Delta \big|\sum_{i=1}^n \theta_i f_i(x)\big| d\mu(x)\Big\}
\ge c_0 \sqrt{nk}$$ for all ${k=1,\ldots,n}$ (see (\[norma4\])).
Using Theorem \[l1th\] we can show that the assertion of Conjecture \[consem2\] is true, at least if parameter $k$ in (\[se0\]) varies only in ${1,\ldots,[\log n]}$:
\[thsem\] For an arbitrary set of functions $\{f_i\}_{i=1}^n\subset L_1[0,1]$ such that ${\|f_i\|_1=1}$, there exist a sequence of signs $\{\theta_i\}_{i=1}^n$, $\theta_i=\pm 1$, and a constant ${c_0>0}$ such that (\[se0\]) holds for all ${k=1,\dots,[\log n]}$.
The proof of Theorem \[thsem\] almost coincide with the proof of Theorem 5 in [@ja3]. The only difference is that instead of Theorem A, formulated in the Introduction, we need to apply Theorem \[l1th\].
Conjecture \[consem2\] in the general form stays open. Note that for ${k\asymp n^\sigma}$ the inequality (\[se0\]) cannot be proved by a random signs argument, since in most cases the order of the uniform norm (and, thus the integral-uniform one) of random polynomials (\[polinom2\]) is bounded from above by $\sqrt{n\log n}$, e.g. this follows from the Salem-Zygmund estimate (\[salz\]) (see also Th. B from Introduction and Th. 4.3 in [@kah]).
Now, let us show that from an estimate of the integral-uniform norm it is possible to get one for the Marcinkiewicz norm. Let ${\varphi:[0,1]\to[0,1]}$ be an increasing concave function such that ${\varphi(0)=0}$ and ${\varphi(1)=1}$. Then the Marcinkiewicz space is defined as a space of functions on $[0,1]$, equipped with the norm (e.g. see [@sem]): $$\|f\|_{M(\varphi)}
:=\sup_{0<t<1}\Big\{\frac1{\varphi(t)}\int_0^t f^*(s)\,ds\Big\},$$ where $f^*$ is the [*decreasing rearrangement*]{} of $f$, defined by $$f^*(s):=\inf\Big\{\tau>0: \mu\big\{x\in[0,1]: f(x)\ge\tau
\big\}<s\Big\}.$$
As a direct corollary of (\[genth2\]) or (\[l1th2\]) and the equivalence of the norms ${\|\cdot\|'_{m}}$ and ${\|\cdot\|_{m,\infty}}$ (see Th. \[kee\]) one can get an estimate for the Marcinkiewicz norm:
Assume that for a random polynomial $F_n(\omega,x)$ of type (\[polinom2\]) the estimate (\[l1th2\]) takes place for all ${m\le n}$. And let ${\varphi:[0,1]\to[0,1]}$ be an increasing concave function such that $\varphi(0)=0$, $\varphi(1)=1$. Then the Marcinkiewicz norm of $F_n(\cdot,x)$ can be estimated as follows: $${\sf E}\|F_n(\omega,x)\|_{M(\varphi)}\ge
A\sqrt{n}\max_{m=2,\dots,n}
\Big\{\frac{\sqrt{\log m}}{m\varphi(1/m)}\Big\}$$ with a constant $A>0$.
\[conja1\] The conclusion of Theorem \[l1th\] stays true for random polynomials (\[polinom2\]) with respect to functional systems $\{f_i\}_1^n$ which satisfy condition [**(d)**]{} with parameter ${p\in[0,1/2)}$ (in Theorem \[l1th\] it is assumed that ${p<1/12}$).
A somewhat weaker form of Conjecture \[conja1\] could be proved if we could prove a statement of the following type:
\[conja3\] [For systems of functions $\{f_i\}_1^n$ which satisfy [**(d)**]{} with parameter ${p\in[0,1/2)}$ there exist some constants ${\varepsilon_0\in(0,\frac12)}$, ${L>0}$ such that $$\int_X
\sum_{1\le k\le n^{1/2+\varepsilon}}f_{k}^*(x)\,d\mu(x)
\le \text{\rm const}\cdot n^{p+L\varepsilon},$$ provided ${\varepsilon\in(0,\varepsilon_0)}$, here $f_k^*(x)$ denote the decreasing ordered values $|f_j(x)|$ at a fixed point ${x\in X}$. ]{}
By the assertion of Conjecture \[conja3\] one could make some estimates in Step 4 in the proof of Theorem \[l1th\]$'$ more efficient. It is “optimization” of Step 4 that is most promising for prospective generalizations of Theorem \[l1th\] in the direction of Conjecture \[conja1\].
[99]{}
R.N. Bhattacharya and R. Ranga Rao, [*Normal Approximation and Asymptotic Expantions,*]{} J. Wiley, 1976.
K.L. Chang and P. Erdös, [*On the application of the Borel-Cantelli lemma*]{}// Trans. Amer. Math. Soc. 1952. v. 72. pp. 179–186.
R.M. Dudley, [*The series of compact subsets of Hilbert space and continuity of Gaussian processes*]{}// J. Funct. Analysis. 1967. v. 1. N. 3. pp. 290–330.
X. Fernique, [*Régularié des trajectoires des foction aléatoires gaussiennes*]{}// Lectures Notes in Mathematics. v. 480. Springer. Berlin. 1975. Heidelberg. pp. 1–96.
P.G. Grigoriev, [*Estimates for norms of random polynomials and their application*]{}// Math. Notes. 2001. [v. 69. N 6.]{} pp. 868–872.
P.G. Grigoriev, [*Estimates for norms of random polynomials*]{}// East J. Approx. 2001. [v. 7. N. 4.]{} pp. 445–469.
P.G. Grigoriev, [*Random and Special Polynomials with respect to a General System of Functions*]{}. Ph.D. Dissertation ([*kandidatskaya*]{}), Steklov Institute of Math., Moscow, 2002, (in Russian).
J.-P. Kahane, [*Some Random Series of Functions.*]{} Heath math. monographs. Lexington, Mass., 1968.
B.S. Kashin and A.A. Saakyan, [*Orthogonal Series*]{}. Amer. Math. Soc., Providence, 1989.
B. Kashin and L. Tzafriri, [*Lower estimates for the supremum of some random processes*]{}// East J. Approx. 1995. [v. 1. N. 1]{}. pp. 125–139.
B. Kashin and L. Tzafriri, [*Lower estimates for the supremum of some random processes, II *]{}// East J. Approx. 1995. [v. 1. N. 3.]{} pp. 373–377.
B. Kashin and L. Tzafriri, [*Lower estimates for the supremum of some random processes, II*]{}. Preprint, MPIM, n. 242 95-85. Max-Plank Institut für Mathematik, Bonn, 1995.
M. Ledoux and M. Talagrand, [*Probability in Banach Spaces*]{}. [Springer Verlag,]{} 1991.
M. Marcus and G. Pisier, [*Random Fourier Series with Applications to Harmonic Analysis*]{}. [Princeton Univ. Press,]{} 1981.
S. Montgomery-Smith and E.M. Semenov, [*Embeddings of rearrangement invariant spaces that are not strictly singular*]{}// Positivity. 2000. [v. 4]{}. pp. 397–402.
V.I. Rotar’, [*Nonuniform estimate of the rate of convergence in multidimensional central limit theorem*]{}// Theory of Probab. and Appl. 1970. v. 15. pp. 647–665.
R. Salem and A. Zygmund, [*Some properties of trigonometric series whose terms have random signs*]{}// [Acta Math.]{} 1954. [v. 91.]{} pp. 245–301.
A.N. Shiryaev. [*Probability*]{}. Springer Verlag, 1996.
D. Slepian, [*The one-sided barrier problem for Gaussian noise*]{}// Bell. System Tech. J. 1962. v. 41. pp. 463–501.
Z. Šidak, [*Rectangular confidence regions for the means of multivariate normal distributions*]{}// J. Amer. Statist. Assoc. 1967. v. 62. N. 318. pp. 626–633.
Z. Šidak., [*On multivariate normal probabilities of rectangles*]{}// Ann. Math. Statist. 1968. v. 39. pp. 1425–1434.
V.G. Sprindzhuk, [*Metric Theory of Diophant Approximations*]{}, Moscow. Nauka, 1977, (in Russian).
V.N. Sudakov, [*Gaussian measures, Cauchy measures and $\varepsilon$-entropy*]{}// Soviet Math. Dokl. 1969. v. 10. pp. 310–313.
V.N. Sudakov, [*Gaussian processes and measures of solid angles in Hilbert space*]{}// Soviet Math. Dokl. 1971. v. 12. pp. 412–415.
V.N. Sudakov, [*Geometrical Problems in the Theory of Infinite-Dimensional Probability Distributions*]{}// Trudy MIAN. 1976. v. 141. (in Russian). English transl. in Proc. Steklov Inst. Math., 1979. 2.
[^1]: This work was supported by grants “Leading Scientific Schools of the Russian Federation” 00-15-96047 and RFFI 02-01-00315.\
This paper is to be published in Mathematical Notes (*Matematicheskie zametki).*
[^2]: \[powersnosk\]It is possible to show that in (\[genth1\]) one can choose the power parameter $q$ arbitrarily from the interval $\big(0,{(1-2p)/2}\big)$, in this case the constants ${C_1}$, $C_2$ may depend on the choice of $q$. One can find a detailed proof of (\[genth1\]) with ${q=(1-2p)/4}$ in [@jadis].
[^3]: The condition [**(d)**]{} with parameter ${p=1/2}$ is obviously satisfied for [**all**]{} functional systems $\{f_i\}_1^n$, ${\|f_i\|_1=1}$. In particular, it holds for the trivial system ${f_i\equiv1}$ for which by Khinchin’s inequality the $L_\infty$-norm of random polynomial with respect to that system is of order $\sqrt n$. Thus, it makes sense to think about possible validity of estimates of type (\[l1th1\]), (\[l1th2\]) only in the case ${p<1/2}$.
[^4]: As usual, the case of small $n$ can be dealt by increasing or reducing some constants.
[^5]: In this work we do not try to choose the power parameter $q$ optimally. However, note if ${\alpha^2=\frac1{4}(\frac1{12}-p_1)}$ than Lemma \[lem3tex\] ensures (\[l1star2\]) and (\[st6\]) with ${q=\frac14(\frac1{12}-p_1)}$.
[^6]: If we were only interested in estimates for the norm of random polynomials (Corollary \[l1uniform\]), then practically we could finish the proof on this step. Indeed, inequality (\[l1st11\]) with $m=[n^{\varepsilon_1}]$ proves Corollary \[l1uniform\] for the case [**(i)**]{} from Step 3. If the case [**(ii)**]{} takes place, then it suffices to apply Proposition \[rot\] to $\{f_i(x)\xi_i\}_{i\in {{\mathcal}A}(x)}$ at a single point ${x\in E_\ell}$, ${\ell>n^{\varepsilon_1}}$, and, by Lemma \[tver\], derive the desired estimate with a “great reserve.”
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The link between a modified Higgs self-coupling and the strong first-order phase transition necessary for baryogenesis is well explored for polynomial extensions of the Higgs potential. We broaden this argument beyond leading polynomial expansions of the Higgs potential to higher polynomial terms and to non-polynomial Higgs potentials. For our quantitative analysis we resort to the functional renormalization group, which allows us to evolve the full Higgs potential to higher scales and finite temperature. In all cases we find that a strong first-order phase transition manifests itself in an enhancement of the Higgs self-coupling by at least 50%, implying that such modified Higgs potentials should be accessible at the LHC.'
author:
- 'M. Reichert'
- 'A. Eichhorn'
- 'H. Gies'
- 'J. M. Pawlowski'
- 'T. Plehn'
- 'M. M. Scherer'
title: 'Probing Baryogenesis through the Higgs Self-Coupling'
---
Introduction {#sec:introduction}
============
The existence of a scalar Higgs potential is the most fundamental insight from the LHC to date. It is based on the observation of a likely fundamental Higgs scalar in combination with measurements of the massive electroweak bosons, fixing the infrared theory and its model parameters after electroweak symmetry breaking to high precision. The one remaining parameter is the Higgs self-coupling and its relation to the Higgs mass, defining a standard benchmark measurement for current and future colliders. This in itself very interesting measurement may also be related to more fundamental physics questions. A prime candidate for such a question is electroweak baryogenesis, specifically the nature of the electroweak phase transition.
For the single Higgs boson of the renormalizable Standard Model we can test the electroweak phase transition through the Higgs mass. Here, electroweak baryogenesis [@ew_phase; @review_ew] requires a Higgs mass well below the observed value of 125 GeV [@misha_higgs; @ew_higgs; @Shaposhnikov:1991cu]. Only then will the electroweak phase transition be strongly first order. If we consider the Standard Model an effective field theory (EFT), a sizable dimension-6 contribution to the Higgs potential, $(\phi^\dag \phi)^3/\Lambda^2$, is known to circumvent this bound [@eft1; @eft2; @Noble; @christophe_geraldine]. In principle, this scenario can be tested through a measurement of the Higgs self-coupling at colliders [@uli1; @eft2; @Noble; @christophe_geraldine]. The problem with this link is that the new-physics scale required by a first-order phase transition is typically not large, $\Lambda \gtrsim v = 246$ GeV. If LHC data should indeed point to a dimension-6 Lagrangian with a low new-physics scale, we will see this in many other channels long before we will actually measure the Higgs self-coupling [@legacy]. As a matter of fact, a global analysis of the effective Higgs Lagrangian including $(\phi^\dag \phi)^3/\Lambda^2$ might never probe the required values of the Higgs self-coupling once we take into account all operators and all uncertainties, so it hardly serves as a motivation to measure a SM-like Higgs self-coupling.
In this paper we take a slightly different approach. First, we assume that the new physics responsible for the strongly first-order electroweak phase transition only appears in the Higgs sector. In the EFT framework we would consider, for example, the operator $(\phi^\dag
\phi)^3/\Lambda^2$ [@christophe_geraldine; @eft2; @Noble]. While this approach systematically includes higher-dimensional operators in a power-counting expansion, it is not at all guaranteed that such an expansion is appropriate for the underlying new physics. Furthermore, a description of first-order phase transitions requires one to extract global information about the effective potential. Again, a simple polynomial expansion around a vanishing Higgs field might not be sufficient to resolve the fluctuation-driven competition between different minima of the effective potential that induce a first-order phase transition.
A simple global approximation to the effective potential is provided by mean-field theory, which works remarkably well for Standard Model parameters [@mean-field; @Gies:2013fua; @Borchardt:2016xju; @Sondenheimer:2017jin] because of the dominance of the top quark. Depending, however, on the strength of the bosonic and order-parameter fluctuations in the new physics model, mean-field approaches may become unreliable. We demonstrate this explicitly using a simple example case in this paper. This situation calls for non-perturbative methods. Recently, lattice simulations have been used to study the possibility of first-order phase transition in the presence of the operator $(\phi^\dag\phi)^3/\Lambda^2$, both in a Higgs-Yukawa model [@Akerlund:2015fya] and in a gauged-Higgs system [@Akerlund:2015gfy].
Here we use the functional renormalization group (FRG) [@christof_eq] as a non-perturbative tool, for reviews see, [*e.g.*]{}, [@rg_reviews]. It is able to provide global information about the Higgs potential, bridge a wide range of scales, include fluctuations of bosonic and fermionic matter fields as well as gauge bosons and deal with extended classes of Higgs potentials. The two questions which will guide us are:
1. Do extended Higgs potentials help with electroweak baryogenesis?
2. Can they be systematically tested by measuring the Higgs self-coupling?
We study the influence of operators or functions of operators in the Higgs sector on the electroweak phase transition using several representative examples. We determine the consequences for the Higgs self-coupling for suitable extended Higgs potentials supporting electroweak baryogenesis and being compatible with the standard-model mass spectrum.
The global properties of the Higgs potential are also intimately related to the questions of vacuum stability and Higgs mass bounds [@Krive:1976sg; @Lindner:1985uk; @Buttazzo:2013uya]. In fact, higher-dimensional operators can also increase the stability regime of the vacuum [@Branchina:2005tu; @Gies:2013fua; @eichhorn_scherer; @our_paper; @Akerlund:2015fya; @stable_frg; @Jakovac:2015kka]. The example Higgs potentials studied in this paper suggest new-physics scales well below a possible instability scale of $10^{10\cdots12}$ GeV of the Standard Model. While vacuum instability is therefore not an issue for our study, extended potentials generally do have the potential to both support electroweak baryogenesis and stabilize the Higgs vacuum. A measurement of the Higgs self-coupling can therefore be indicative for both aspects.
Electroweak phase transition {#sec:phase}
----------------------------
The asymmetry between the matter and anti-matter contents in the Universe is one of the great mysteries in cosmology and particle physics. Experimentally, the effective absence of anti-matter in the Universe has been proven in many different ways [@antimatter_meas]. A quantitative measurement is given by the baryon-to-photon ratio $n_B/n_\gamma \approx 6 \cdot 10^{-10}$, which is many orders of magnitude larger than what we would expect from the thermal history in the presence of anti-matter. It can be explained by a small initial asymmetry in the number of baryons and anti-baryons which leads to a finite density of baryons after essentially all anti-baryons have annihilated away.
Theoretically, the mechanisms behind the baryon asymmetry are well understood. Most notably, it can be shown that the presence of an asymmetry is equivalent to the three Sakharov conditions for our fundamental theory [@ew_phase]: baryon number violation, C as well as CP violation, and departure from thermal equilibrium. The first two conditions can be probed by precision measurements of the Lagrangian of the Standard Model and its extensions. The third condition can in principle be achieved at the time of the electroweak phase transition, where it then requires a strong first-order phase transition. The nature of the electroweak phase transition can be read off from the scalar potential in or beyond the Standard Model.
The strength of the phase transition which occurs at the critical temperature $T_c$ is measured by the ratio $\phi_c/T_c$, where $\phi_c=\langle \phi\rangle_{T_c}$ is the expectation value of the Higgs at the critical temperature. The critical temperature describes the transition where for small temperatures $T <T_c$ the potential exhibits a single, non-trivial minimum for some value of the scalar field $\phi$. The field value at the minimum is temperature dependent, approaching $v=246$ GeV for $T \rightarrow 0$. With increasing temperature, a second minimum at zero field value and with an unbroken electroweak symmetry appears in a first-order scenario. At the critical temperature $T_c$, the two minima of the potential, [*i.e.*]{}the one at finite field value and the one at vanishing field value are degenerate, and the system undergoes a phase transition from the symmetry-broken regime with a finite Higgs expectation value to the symmetric regime.
The field value at the minimum constitutes an order parameter. For $\phi_c \neq 0$ the transition is of first order, [*i.e.*]{}the vacuum does not evolve continuously through the phase transition. For electroweak baryogenesis, the transition has to be a strong first-order one, $$\frac{\phi_c}{T_c} \gtrsim 1 \,,$$ otherwise the baryon asymmetry is washed out [@Shaposhnikov:1991cu].
Higgs self-coupling measurement {#sec:self}
-------------------------------
![Feynman diagrams contributing to Higgs pair production at the LHC. Figure from Ref. [@uli1].[]{data-label="fig:feynman"}](feyn_hh){width="65.00000%"}
At energy scales relevant for the LHC, the self-interaction of the Higgs boson is described by the infrared (IR) Higgs potential in the broken phase. In the renormalizable Standard Model, and ignoring Goldstone modes, it reads at tree level $$\begin{aligned}
V = \frac{\mu^2}{2} \, (v+H)^2 + \frac{\lambda_4}{4} \, (v+H)^4 \; ,
\label{eq:pot_ir}\end{aligned}$$ where $H$ is the physical Higgs field. The two parameters describing the SM-Higgs potential in the IR, $\mu$ and $\lambda_4$, can be traded for the vacuum expectations value $v$ and the Higgs mass $m_H$ [@lecture] $$\begin{aligned}
v = \sqrt{ \frac{\mu^2}{2 \lambda_4} } = 246~{\text{GeV}}\,, {\qquad \qquad \qquad}m_H = \sqrt{2 \lambda_4} \, v = 125~{\text{GeV}}\; .
\label{eq:ir_values}\end{aligned}$$ The interaction between three and four physical Higgs bosons in the Standard Model is then given by $$\begin{aligned}
\lambda_{H^3,0} = \frac{3 m_H^2}{v}\,, {\qquad \qquad \qquad}\lambda_{H^4,0} = \frac{3 m_H^2}{v^2} \; .
\label{eq:ir_selfs}\end{aligned}$$ In the limit of heavy top quarks, $2 m_t > m_H$, an effective Higgs–gluon Lagrangian [@low_energy] $$\begin{aligned}
{\mathcal{L}}_{ggH}
=\frac{\alpha_s}{12\pi} \; G^{\mu\nu}G_{\mu\nu} \;
\log \left(1+\frac{H}{v} \right)
= \frac{\alpha_s}{12 \pi} \; G^{\mu\nu}G_{\mu\nu} \; \frac{1}{v} \;
\left( H - \frac{H^2}{2v} + \ldots \right)\,,
\label{eq:higgs_eff}\end{aligned}$$ with the gluon field strength tensor $G_{\mu\nu}$ and the strong coupling $\alpha_s$, can be used to describe many relevant LHC observables.
When we include new physics contributions in the Higgs potential, the relations in Eq. change. It is instructive to follow the simple example of the modified Higgs potential [@lecture] $$\begin{aligned}
V = \frac{\mu^2}{2} \, (v+H)^2
+ \frac{\lambda_4}{4} \, (v+H)^4
+ \frac{\lambda_6}{\Lambda^2} \, (v+H)^6 \; .
\label{eq:mod-Higgs-pot}\end{aligned}$$ The modified relations between the observables become $$\begin{aligned}
m_H &= \sqrt{2 \lambda_4} \, v \left( 1 +12 \frac{\lambda_6 v^2}{
\lambda_4 \Lambda^2} \right) \,,\notag \\[1ex]
\lambda_{H^3} &= \frac{3 m_H^2}{v} \left( 1 + \frac{16 \lambda_6 v^4}{
m_H^2 \Lambda^2} \right)
\equiv \lambda_{H^3,0} \left( 1 + \frac{16
\lambda_6 v^4}{ m_H^2 \Lambda^2} \right)\,, \notag \\[1ex]
\lambda_{H^4} &= \frac{3 m_H^2}{v^2} \left( 1 +
\frac{96 \lambda_6 v^4}{m_H^2 \Lambda^2} \right)
\equiv \lambda_{H^4,0} \left( 1 +
\frac{96 \lambda_6 v^4}{m_H^2 \Lambda^2} \right) \; .
\label{eq:ir_shifts}\end{aligned}$$ Because $m_H$ and $v$ have to keep their measured values, we need to adjust $\lambda_4$ to compensate for the effect of $\lambda_6$ on the Higgs mass. This shift has to be accounted for in the expressions for the Higgs self-couplings as a function of $m_H$ and $v$. The reference couplings $\lambda_{H^n,0}$ keep their Standard Model values in terms of the unchanged parameters $m_H$ and $v$, but the physical Higgs couplings $\lambda_{H^n}$ change.
The standard channel to measure $\lambda_{H^3}$ at the LHC is Higgs pair production in gluon fusion, as illustrated in Fig. \[fig:feynman\], [@orig; @spirix; @uli1; @uli2; @uli3; @review_hh]. Its production rate is known including NLO [@nlo] and NNLO [@nnlo]. One of the problems with such a measurement is that the link between the total di-Higgs production rate and the Higgs self-coupling requires us to know the top Yukawa coupling. An appropriate framework is the global Higgs analysis [@legacy; @hh_d6], which is expected to give at best a 10% measurement of the top Yukawa coupling. A model-independent precision measurement of the top Yukawa coupling at the percent level will only be possible at a 100 TeV collider [@nimatron_yt].
The experimental situation improves once we include kinematic information in the di-Higgs production process. Two kinematic regimes are well known to carry information on the Higgs self-coupling, both exploiting the (largely) destructive interference between the two graphs shown in Fig. \[fig:feynman\]. While the continuum contribution dominates over most of the phase space, the two diagrams become comparable close to threshold [@spirix; @uli1]. The low-energy theory of Eq. gives us for the combined di-Higgs amplitude $$\begin{aligned}
\mathcal{A} \propto
\frac{\alpha_s}{12 \pi v} \;
\left( \frac{\lambda_{H^3}}{s-m_H^2} - \frac{1}{v} \right)
\stackrel{\lambda_{H^3} = \lambda_{H^3,0}}{\longrightarrow}
\frac{\alpha_s}{12 \pi v^2} \;
\left( \frac{3m_H^2}{3m_H^2} -1 \right) = 0
\qquad \text{for} \quad
m_{HH} \to 2 m_H \; ,
\label{eq:higgs_pair}\end{aligned}$$ where $m_{HH}$ is the invariant di-Higgs mass. An exact cancellation occurs in the Standard Model. Whereas the heavy-top approximation is known for giving completely wrong kinematic distributions for Higgs pair production [@uli1], it does correctly predict this threshold behavior. Note, that the momenta of the outgoing particles in such processes are typically small compared to the Higgs mass and the low-energy regime of the theory is probed. In the analysis in Sec. \[sec:main\], we thus read off the Higgs self-couplings from the low-energy effective potential.
The second relevant kinematic regime is boosted Higgs pair production [@boosted], because of top threshold contributions to the triangle diagram around $m_{HH} = 2 m_t$. In terms of the transverse momentum this happens around $p_{T,H} \approx 100$ GeV, where the combined amplitude develops a minimum for large Higgs self-couplings.
At the LHC, we define di-Higgs signatures simply based on Higgs decay combinations. The most promising channel is the $b\bar{b}\gamma\gamma$ final state [@uli3; @madmax; @vernon; @atlas], where we can easily reconstruct one of the two Higgs bosons and measure the continuum background in the side bands. We can also use the $b\bar{b} \tau
\tau$ final state [@uli2; @boosted], assuming very efficient tau-tagging. The combination $b\bar{b}WW$ [@bbww] requires an efficient suppression of the $t\bar{t}$ background, while the $4b$ [@uli2; @bbbb] and $4W$ [@uli1; @wwww] signatures are unlikely to work for SM-like Higgs bosons. Finally, the $b\bar{b} \mu
\mu$ is in many ways similar for the $b\bar{b} \gamma \gamma$ channel [@uli3], but with a much lower rate in the Standard Model.
To get an idea of what to expect, we quote the optimal reach of the high-luminosity LHC run with $3~{\text{ab}^{-1}}$, based on the Neyman-Pearson theorem applied to the $b\bar{b} \gamma \gamma$ channel for self-couplings relatively close to the Standard Model [@madmax], $$\begin{aligned}
{9}
\frac{\lambda_{H^3}}{\lambda_{H^3,0}} &= \phantom{-} 0.4~...~1.7 \qquad \text{at 68\% CL,}
\label{eq:lhc}\end{aligned}$$ so any value for $\lambda_{H^3}/\lambda_{H^3,0}$ outside the range given above will not be compatible with the vanishing di-Higgs amplitude in Eq.. This reach will be improved when we combine several Higgs decay channels, but will also suffer from systematic uncertainties. In addition, it assumes a perfect knowledge of the top Yukawa coupling. This implies that models which predict a change in the Higgs self-coupling by less than 50% will not be testable at the LHC.
Modified Higgs potentials {#sec:model}
=========================
Similar to the EFT approach we assume that beyond an ultraviolet (UV) scale or cutoff scale $\Lambda$ new physics exists and modifies the form of the Higgs potential. As the additional degrees of freedom are heavy, their effects below $\Lambda$ can be parametrized by additional terms in the Higgs potential, without modifying the propagating degrees of freedom. The details of the new physics are encoded in the initial condition for the RG flow of the Standard Model at $k =\Lambda$. Exploring different higher-order terms thus provides access to large classes of high-scale physics scenarios, for which we do not have to investigate the detailed matching of the additional terms in the Higgs potential and the underlying high-scale degrees of freedom at $k = \Lambda$.
Our system features three relevant energy scales. First, the RG scale $k$ ranges between $k=0$, where all quantum fluctuations are taken into account, and $k=\Lambda$, where we initialize the flow. Second, the temperature $T$ defines the external physics scale with which we probe our system. Third, the field value $\phi$ defines an additional, internal energy scale of our system. As is usual in EFT analyses, it is important to clearly disentangle these three scales, even though $\phi$ and $T$ can in principle act similarly to the RG scale $k$ in that they suppress IR quantum fluctuations [@our_paper]. We employ a method that can straightforwardly account for the RG flow in the presence of these different scales, namely the functional renormalization group. In this setting, quantum fluctuations in the presence of further internal and external scales are taken into account by a functional differential equation that is structurally one-loop, without being restricted to a weak-coupling regime. This provides access to classes of non-perturbative microscopic models with a manageable computational effort. Most importantly, the functional RG approach enables us to keep track of the separate dependence of the potential on the RG scale $k$, the temperature and the field value even in cases with non-perturbative UV potentials, where, [*e.g.*]{}a mean-field approach breaks down.
For our study, we concentrate on that part of the Standard Model which is relevant for the RG flow of the Higgs potential using the framework developed in [@our_paper]. Here, we follow that framework by implementing the effects of weak gauge bosons through a fiducial coupling, and upgrade our treatment by including a thermal mass generated by the corresponding fluctuations as their leading contribution instead of implementing a fully-fledged dynamical treatment of that sector, see App. \[app:der-flows\] for details. Similarly, would-be Goldstone modes do not need to be considered explicitly, such that it suffices to concentrate on a real scalar field $\phi$, which after electroweak symmetry breaking can be described in terms of the physical Higgs field $H$ as $\phi = H + v$. At the UV scale $k=\Lambda$, the Higgs-potential is parametrized as $$\begin{aligned}
V_{k=\Lambda} = \frac{\mu^2}{2} \, \phi^2 +
\frac{\lambda_4}{4} \, \phi^4 + \Delta V \,,
\label{eq:poly-start-pot}\end{aligned}$$ where $\Delta V$ contains the contribution of some higher dimensional operator. In principle, higher-order modifications of the Yukawa sector could also be included, cf. [@Pawlowski:2014zaa; @deVries:2017ncy; @Gies:2017zwf]. We investigate three classes of modifications to the SM-Higgs potential:
1. additional $\phi^6$ or $\phi^8$ terms, which cover the leading-order terms in an effective-field theory approach and have been extensively studied in the literature [@eft1; @christophe_geraldine; @eft2; @Noble];
2. a logarithmic dependence on the Higgs-field, inspired by Coleman-Weinberg potentials. It does not allow for a Taylor expansion around $\phi=0$. Logarithmic modifications are naturally generated by functional determinants, [*i.e.*]{}by integrating out heavy scalars or fermions.
3. a simple example of non-perturbative contributions of the form $\exp(-1/\phi^2)$, [*i.e.*]{}an exponential dependence on the inverse field, consequently not admitting a Taylor expansion in the field around $\phi=0$. This is inspired by semiclassical contributions to the path integral with $\phi$ reminiscent to a moduli parameter of an underlying model.
We denote these modifications of the potential by $$\begin{aligned}
{9}
\Delta V_6 &= \lambda_6 \, \frac{\phi^6}{\Lambda^2}\,, {\qquad \qquad \qquad}&
\Delta V_8 &= \lambda_6 \, \frac{\phi^6}{\Lambda^2}
+ \lambda_8 \, \frac{\phi^8}{\Lambda^4} \,, \notag\\[2ex]
\Delta V_{\ln,2} &= -\lambda_{\ln,2} \, \frac{\phi^2\Lambda^2}{100} \;
\ln \frac{\phi^2}{2\Lambda^2}\,, {\qquad \qquad \qquad}&
\Delta V_{\ln,4} &= \lambda_{\ln,4} \, \frac{\phi^4}{10} \; \ln \frac{\phi^2}{2\Lambda^2}
\,, \notag\\[2ex]
\Delta V_{\exp,4} &= \lambda_{\exp,4} \phi^4 \exp\left(-
\frac{2\Lambda^2}{\phi^2}\right)\,, {\qquad \qquad \qquad}&
\Delta V_{\exp,6} &= \lambda_{\exp,6} \frac{\phi^6}{\Lambda^2}
\exp\left(-\frac{2\Lambda^2}{\phi^2} \right)\,.
\label{eq:pots}\end{aligned}$$ In all these potentials $\Lambda$ describes a new physics scale, which absorbs the mass dimension of the Higgs field. The case of $\phi^6/\Lambda^2$ has been explored in the literature [@eft2; @Noble; @christophe_geraldine] and serves as a test of our method, as discussed in the Appendix. Neither the logarithmic nor the exponential potentials can be expanded around $\phi=0$, so they cannot be treated in an EFT framework. Similar bare potentials have been suggested in [@Sondenheimer:2017jin] in the context of Higgs mass bounds and vacuum stability. Instead, all potentials that can be expanded around $\phi=0$ can be approximated by the power-ordered, first kind of potentials. As expected by canonical power counting, terms of higher order in $\phi$ can only play a role for very low values of $\Lambda/v$, unless their prefactors are non-perturbatively large. From a more general viewpoint, the set of power law, logarithmic and exponential potential functions does not only reflect the physics structures arising from local vertex expansions, one-loop determinants or semiclassical approximations. It also includes the set of functions to be expected on mathematical grounds if the effective potential permits a potentially resurgent transseries expansion [@Dunne:2012ae].
To investigate the different classes of modifications, a variety of tools appears to be at our disposal, a priori ranging from mean-field techniques to non-perturbative lattice tools and functional methods. It turns out that the former are only applicable to a restricted class of potentials, not allowing us to adequately explore the full range of possible UV potentials corresponding to diverse underlying microscopic models. This is displayed in Fig. \[fig:mean-field\] where the $\phi^6$- modification of the Higgs potentials shows the expected physical behavior as the strength of the first-order phase transition is decreasing with an increasing cutoff. The logarithmic modifications on the other hand show a rather unphysical behavior as the strength of the first-order phase transition remains constant or even increases with the UV scale. This indicates that scalar order-parameter fluctuations are important, which are ignored in simple mean-field theory. Therefore we make use of powerful functional techniques, which treat bosonic and fermionic fluctuations on the same footing.
![Mean-field results for $\phi_c/T_c$ as a function of the cutoff for different modifications of the Higgs potential. Second-order and weak first-order phase transitions are excluded from the plot. The results of the $\phi^6$ modification are reasonable, while the results for the $\phi^2\ln\phi^2$ and the $\phi^4\ln\phi^2$ modifications are clearly unphysical, see explanation in the text. More elaborate methods than mean-field are needed. []{data-label="fig:mean-field"}](Mean-Field){width="50.00000%"}
When allowing for modifications of the Higgs potential, we need to ensure that at $T=0$ the IR-values for $\mu$, $\lambda_4$, and the top-Yukawa-coupling $y_t$ are such that the measured observables do not change. We adjust the corresponding masses to $$\begin{aligned}
v = 246 \,{\text{GeV}}\,, {\qquad \qquad \qquad}m_H = 125 \,{\text{GeV}}\,, {\qquad \qquad \qquad}m_t = 173 \,{\text{GeV}}\,.
\label{eq:ir_data}\end{aligned}$$ Within our numerical analysis, we require $v$ and $m_t$ to be reproduced to an accuracy of $\pm$0.5 GeV. The Higgs mass is adjusted within a somewhat larger numerical band of $\pm 1.5$ GeV. Since it is related to the second derivative (curvature) of the potential at the minimum, a higher precision is numerically more expensive, see App. \[app:grid-code\] for details. Moreover, it is expected that the curvature mass used here shows small deviations from the Higgs pole mass $m_H$, see [@Helmboldt:2014iya], and the above band also contains an estimate of this systematic error. In the symmetry broken regime, the potential given in Eq. can be expanded in powers of $(\phi^2 - v^2)$. In the decoupling region in the deep IR, we use the parametrization $$\begin{aligned}
V_{k\ll v} &= \frac{\lambda_{4,\,\rm IR}}{4} (\phi^2 - v^2)^2
+ \frac{\lambda_{6,\,\rm IR}}{8 v^2} (\phi^2 - v^2)^3
+ \frac{\lambda_{8,\,\rm IR}}{16v^4} (\phi^2 - v^2)^4 + \cdots\notag \\
&= \lambda_{4,\,\rm IR} v^2 H^2
+ \left( \lambda_{4,\,\rm IR} + \lambda_{6,\,\rm IR} \right) v H^3
+ \frac{1}{4} \left( \lambda_{4,\,\rm IR} + 6 \lambda_{6,\,\rm IR} +
4 \lambda_{8,\,\rm IR} \right) H^4
+ \cdots\,.\end{aligned}$$ Note that this is the full effective potential in the IR, differing from the tree-level potential in Eq.. In particular, higher-order terms, encoded in $\lambda_{6,\, \rm IR }$ are generated by quantum fluctuations even if the tree-level potential is quartic. At tree level, the Higgs potential is described by two parameters, [*i.e.*]{}$\lambda_{6,\,\rm IR} = \lambda_{8,\,\rm IR} = \ldots =0$. If we allow higher-order terms, all measurable parameters are affected, in close analogy to Eq.. As described in Sec. \[sec:self\] the vacuum expectation value $v$ and the Higgs mass $m_H^2/(2v^2) \equiv \lambda_4$ are known very precisely from collider measurements and thus we have to keep them fixed. The physical Higgs self-couplings change from the values given in Eq. to the more general form $$\begin{aligned}
\lambda_{H^3} = \frac{\delta^3}{\delta H^3} V_{k=0}
= 6 v(\lambda_{4,\,\rm IR} + \lambda_{6,\,\rm IR})\,, \qquad
\lambda_{H^4} = \frac{\delta^4}{\delta H^4} V_{k=0}
= 6 (\lambda_{4,\,\rm IR} + 6 \lambda_{6,\,\rm IR} + 4 \lambda_{8,\,\rm IR}) \; .\end{aligned}$$ The first terms are precisely the couplings $\lambda_{H^3,0}=6 v\lambda_{4,\,\rm IR}$ and $\lambda_{H^4,0}=6 \lambda_{4,\,\rm IR}$ familiar from the tree-level structure. With the present setup we can compute the Higgs self-couplings in the pure Standard Model including higher-order terms generated by quantum fluctuations by initializing the flow at some high cutoff scale without any modifications of the Higgs potential. As long as the cutoff is not too close to the electroweak scale the results will be largely independent of the cutoff choice. For our level of numerical precision, a cutoff $\Lambda=2$ TeV is sufficient. The Higgs self-couplings are given by $$\begin{aligned}
\frac{\lambda_{H^3}}{\lambda_{H^3,0}} \approx 0.92\,, {\qquad \qquad \qquad}\frac{\lambda_{H^4}}{\lambda_{H^4,0}} \approx 0.68 \;.
\label{eq:Higgs-self-SM}\end{aligned}$$ These values are equivalent to computations of the Higgs potential with Coleman-Weinberg corrections. We then go beyond the pure Standard Model by adjusting a combination of the coefficients $\lambda_j$ and the new physics scale $\Lambda$ in Eq.. These can now be used to adjust $\phi_c/T_c$ such that we obtain a strong first-order phase transition.
Phase transition {#sec:main}
================
For the modified Higgs potentials defined in Eq. we need to explore which values of the UV scale $\Lambda$ and the coefficients $\lambda_j$ lead to a sufficiently strong first-order transition. Simultaneously, we monitor whether this leads to a measurable modification of the Higgs self-couplings in the IR.
First-order phase transition {#sec:pt}
----------------------------
![Temperature evolution of the potentials of the type $\phi^4 \ln\phi^2$ (solid) and $\phi^4 \exp(-1/\phi^2)$ (dashed) for fixed $\phi_c/T_c\approx1$. We plot the temperatures $T=25$ GeV (violet), $T=50$ GeV (red), $T=75$ GeV (blue), $T=100$ GeV (green) and $T=T_c$ (orange). Note that $T_c^{\ln,4} = 116.4\;\mathrm{GeV} > T_c^{\exp,4} = 110.5$ GeV and thus one curve overtakes the other. A magnification of the curves at $T=T_c$ is displayed in Fig. \[fig:IR-plots\] []{data-label="fig:T-comp"}](T-comp){width="60.00000%"}
![Effective potentials at $T=0$ (left) and $T=T_c$ (right). We show all modified Higgs potentials from Eq. with $\Lambda=2$ TeV. The values of the coefficients at the UV scale $\Lambda$ are fixed by the requirement $\phi_c \approx T_c$, leading to $\lambda_6=1.2$, $\lambda_6=1$ with $\lambda_8=1.4$, $\lambda_{\ln,4}=0.89$, $\lambda_{\ln,2}=0.27$, $\lambda_{\exp,4}=23.3$, and $\lambda_{\exp,6}=27.5$.[]{data-label="fig:IR-plots"}](IR-Plots){width="\textwidth"}
In Fig. \[fig:T-comp\] we show the evolution of two example potentials from Eq. from zero temperature to $T_c$, where the latter is defined as the temperature at which the two competing minima become degenerate. The latter is not distinctly apparent in Fig. \[fig:T-comp\], but becomes visible in the magnification in the right panel of Fig. \[fig:IR-plots\]. We also require the second minimum to be at $\phi_c= T_c$, to guarantee a sufficiently strong first-order phase transition. This way, the $\phi$ dependence of the two cases becomes comparable. A key feature already visible in this figure is that the potential with the deeper minimum at small temperature turns into the steeper potential at $T_c$. This is achieved by a larger value of $T_c$ for the potential with the deeper minimum. Note that the potentials in Fig. \[fig:T-comp\] and \[fig:IR-plots\] are read off at the RG scale $k_\text{IR}$, which is an infrared scale where the Higgs potential and all observables are frozen out. Below this scale only convexity generating processes take place. The freeze out occurs once fluctuations of fields decouple from the RG flow because the RG scale $k$ crosses their mass-threshold. This decoupling is built into the FRG setup. We choose $k_\text{IR}$ to be smaller than the masses of the model, such that the exact choice of $k_\text{IR}$ does not matter.
In Fig. \[fig:IR-plots\] we illustrate the behavior of all our modified Higgs potentials in the IR at vanishing temperature (left panel) and at the critical temperature (right panel), respectively. Note the different scales on the vertical axes. The UV scale $\Lambda$ and the respective coefficients $\lambda_j (\Lambda)$ are chosen such that they result in a strong first-order phase transition, $\phi_c/T_c=1$. The different potentials at zero temperature are similar to that of the Standard Model, as expected from the fact that we fix the Higgs vacuum expectation value and mass to their observed values. In particular, the minima all appear at $v = 246$ GeV, and the second derivatives have to reproduce the measured Higgs mass. Nevertheless, if we fix $V_{k=k_\text{IR}}(0) = 0$, an imprint of modified UV physics remains visible.
In the left panel of Fig. \[fig:IR-plots\] we see that up to $\phi \approx 300$ GeV, all modifications we consider lead to a very similar form of the zero-temperature IR potential, if their coefficients are fixed such that $\phi_c/T_c$ is the same for all our potentials. At higher field values the different UV modifications lead to distinct field-dependence of the potential. The sizable impact of the modified microscopic action on the IR potential is due to the finite UV scale $\Lambda = 2$ TeV. This is not sufficiently far above the electroweak scale for the contributions $\Delta V$ to be washed out by the RG flow.
At finite temperature, we see in the right panel of Fig. \[fig:IR-plots\] that the potentials show significant deviations and the six different modifications fall into three distinct forms of the IR potential at $T_c$. The Standard Model is not displayed, since it exhibits a second-order phase transition with $\phi_c=0$. The other potentials show different sizes of the bump that separates the minima at $\phi=0$ and $\phi=\phi_c$. The exponential modifications show the smallest bump, while logarithmic modifications show the largest bump. The third class is given by the polynomial UV potentials, which fall in between the two other classes.\
It is worth noting that the resulting IR modifications almost coincide *within* each class of UV potentials, [*i.e.*]{}, the polynomial, logarithmic, and exponential class. Although there are manifestly different UV modifications within each class, like for instance $\phi^4\exp(-1/\phi^2)$ vs $\phi^6\exp(-1/\phi^2)$, the resulting IR behavior appears to be dominated by the exponential dependence, and accordingly is nearly the same for the two cases – as stressed before, the two exponential cases differ from the two logarithmic cases, which are within a separate class of their own.
Comparing the two panels we observe that zero-temperature potentials with a steeper increase at larger field values turn into more shallow potentials for finite temperature near the broken vacuum. The latter corresponds to a lower barrier between the two minima. The reason for this link is that the phase transition occurs once positive thermal corrections to the mass parameter are large enough to change the extremum at $\phi=0$ from a maximum to a minimum, which then becomes degenerate with the minimum at a finite field value. For potentials with a lower zero-temperature depth — and correspondingly a more substantial slope at large $\phi$ — the corresponding critical temperature $T_c$ is lower. Therefore, the steepest increase towards large $\phi$ in the left panel in Fig. \[fig:IR-plots\] corresponds to the smallest bump in the right panel of Fig. \[fig:IR-plots\]. Phrased differently: for potentials with a flatter inner region, scalar fluctuations are quantitatively more relevant. At the same time, the phase transition turns first order as soon as the scalar fluctuations dominate over the fermionic ones. This connection will become important when evaluating the prospects of the different cases with regards to detectability at the LHC.
Scale of new physics {#sec:valid}
--------------------
Given a particular microscopic model containing additional degrees of freedom, the UV scale or cutoff $\Lambda$ is typically identified with the mass scale of those additional fields, below which their fluctuations are suppressed. From an EFT point of view, one correspondingly associates $\Lambda$ with the energy scale, above which new physics can appear as on-shell excitations. In turn, below $\Lambda$ the effect of new physics is only visible indirectly. Such an indirect effect would be a deviation of the Higgs potential from its form in the renormalizable Standard Model. A key aspect of this kind of approach is that an EFT description by definition comes with a region of validity, above which we will be sensitive to the actual UV completion. Hence, before we use our modified Higgs potential to link a strong first-order phase transition to the Higgs self-coupling we need to study the validity range of our description.
![Coefficient $\lambda_j(\Lambda)$ of the dimension-6 operator $\phi^6/\Lambda^2$ (left), the modification $\Lambda^2
\phi^2 \ln \phi^2/\Lambda^2$ (center), and the modification $\phi^4
\ln \phi^2/\Lambda^2$ (right) as a function of the cutoff, requiring $\phi_c/T_c = 1 \pm 0.05$.[]{data-label="fig:pctc1"}](pctc1-all){width="\textwidth"}
Following Eq. we see that an indirect measurement using an EFT-like approach is only sensitive to a combination of the scale $\Lambda$ and the (Wilson) coefficients $\lambda_j$. In Fig. \[fig:pctc1\] we show the correlation between $\Lambda$ and the corresponding $\lambda_j$ evaluated at the UV scale $\Lambda$ for a set of modified Higgs potentials, assuming a strong first-order phase transition with $\phi_c/T_c = 1$. We can interpret these results as lines of constant IR physics: the running coefficient $\lambda_j(\Lambda)$ then describes a family of effective models defined at different scales $\Lambda$, all yielding the same IR observables. Without new physics effects, $\Delta V = 0$, this corresponds to fixing $v$, $m_H$ and $m_t$ in the IR and simply evolving them toward the UV with their known RG equations. In our extended setup, the additional coefficients measure the strength of the new physics contribution, that we initialize at the UV scale $\Lambda$. We then use a corresponding parameter $\lambda_j$ to fix $\phi_c/T_c$ to a value of our choice. Doing so for different UV scales $\Lambda$, the coefficient $\lambda_j$ becomes a function of $\Lambda$.
Without running effects for the coefficients $\lambda_j$ the correlation between the coefficient and the UV scale would be simple. For instance, the dimension-6 Wilson coefficient would follow a parabola, $\lambda_6 \propto \Lambda^2$. However, the condition on $\phi_c/T_c$ for the strong first-order phase transition is defined at energies around the Higgs VEV, while the shown values of $\lambda_j$ are defined in the UV. The complete correlation is well-described by a quadratic polynomial. In the case of $\lambda_6$, this reflects the quadratic running due to the canonical dimension. While the normalization of $\Delta V$ can be adjusted at will and the absolute values of the coefficients $\lambda_j$ do not carry any physical significance, the growth of these coefficients towards the ultraviolet suggests the possible onset of a strongly coupled regime.
To investigate the onset of this strongly coupled regime we fit the correlation between $\lambda_j$ and $\Lambda$ to a broken rational polynomial. A motivation for the particular choice of fit function in Fig. \[fig:pctc1\] is given by an approach to a power-like Landau-pole singularity. Indeed, this ansatz fits our numerical results well for the given range of UV scales. From the broken polynomial we can estimate the critical scales, where the respective models might become strongly coupled, $$\begin{aligned}
\Lambda_6^\text{crit} = 7.0\,{\text{TeV}}, {\qquad \qquad}\Lambda_{\ln,2}^\text{crit} = 10 \,{\text{TeV}}, {\qquad \qquad}\Lambda_{\ln,4}^\text{crit} = 6.8 \,{\text{TeV}}\,.\end{aligned}$$ These critical scales should be viewed as conservative estimates of the validity scale up to which our field-theory description using purely Standard-Model degrees of freedom is applicable. These estimates are of the same order of magnitude as maximum values of $\Lambda$ that lead to a first-order phase transition in studies based on mean-field arguments, see [*e.g.*]{}[@christophe_geraldine].
Baryogenesis vs Higgs self-coupling {#sec:baryogenesis}
-----------------------------------
![Modification of the self-coupling $\lambda_{H^3}/\lambda_{H^3,0}$ as a function of $\phi_c$ (left) and $1/T_c$ (right) for the UV potentials given in Eq.. The asterisk in both plots represents the Standard Model expectation, including Coleman-Weinberg corrections, cf. Eq..[]{data-label="fig:Higgs-self-sep"}](Higgs-self-sep){width="\textwidth"}
After showing how a modified Higgs potential can lead to a strong first-order phase transition in Sec. \[sec:pt\] and confirming that our approach is consistent in Sec. \[sec:valid\], we can now explore the link between the strong first-order phase transition and the observable Higgs self-coupling. As laid out in the Introduction, the crucial question is as to whether modifications of the Higgs potential that lead to a sufficiently strong first-order phase transition for electroweak baryogenesis can be tested through the Higgs self-coupling measurement at the LHC.
Following the above discussion, the remaining question is how a value $\phi_c/T_c \approx 1$ due to the potentials given in Eq. is reflected in shifted physical Higgs self-couplings $\lambda_{H^3}$ and $\lambda_{H^4}$. All new physics models are adjusted to reproduce the low-energy measurements in Eq.. First, we can separate the two parameters $1/T_c$ and $\phi_c$ and show their individual effects on the physical Higgs self-couplings. In Fig. \[fig:Higgs-self-sep\] we first see that the two parameters contribute roughly similar amounts to an increase in the Higgs self-couplings, if we push the model towards a strong first-order phase transition. Second, we see that the individual potentials in the general class of power-series, logarithmic, and exponential potentials give essentially degenerate results. Finally, the effect on the self-couplings is the weakest for the logarithmic potential, slightly stronger for the power-law modification, and the strongest for the exponential modification.
![Modification of the self-couplings $\lambda_{H^3}/\lambda_{H^3,0}$ (left) and $\lambda_{H^4}/\lambda_{H^4,0}$ (right) as a function of $\phi_c/T_c$ for the UV potentials given in Eq.. The asterisk in the lower left of both plots represents the Standard Model expectation, including Coleman-Weinberg corrections, cf. Eq.. []{data-label="fig:Higgs-self"}](Higgs-self){width="\textwidth"}
As already observed in Sec. \[sec:pt\], a steeper zero-temperature potential at large field values can be linked to a decrease in $T_c$. On the other hand, a steeper increase at large field values will be tied directly to larger values of the cubic and quartic Higgs self-coupling. This dependence is confirmed by Fig. \[fig:Higgs-self-sep\], where potentials with smaller $T_c$ feature larger $\lambda_{H^3}$. This feature holds both within each class of potentials where we can decrease $T_c$ by enhancing $\Delta
V$, and between different classes of potentials. This trend should be generic in that additions $\Delta V$ leading to a strong first-order transition at low $T_c$ will be easier to detect at the LHC.
Given that we do not see any striking effects from the individual dependence on $1/T_c$ and $\phi_c$, we study the dependence of the different Higgs potentials on the physically relevant ratio $\phi_c/T_c$. In Fig. \[fig:Higgs-self\], we show the modifications of both Higgs self-couplings as a function of $\phi_c/T_c$. The free model parameter along the shown line is an appropriate combination of new-physics scale $\Lambda$ and the new-physics coefficient $\lambda_j$. For $\phi_c/T_c \gtrsim 1$ we find a strong first-order phase transition, suitable for electroweak baryogenesis. From the location of the Standard Model point it is clear that there exists a range of modified self-couplings where the electroweak phase transition remains second order. Only for $$\begin{aligned}
\frac{\lambda_{H^3}}{\lambda_{H^3,0}} \gtrsim 1.5
{\qquad \qquad}\text{or} {\qquad \qquad}\frac{\lambda_{H^4}}{\lambda_{H^4,0}} \gtrsim 4 \,,\end{aligned}$$ we have a chance to generate a first-order phase transition. This number should be compared to the LHC reach given in Eq.. We conclude that the prospects of a detectable imprint appear to be good for all models that we have studied. A strong first-order phase transition corresponding to $\phi_c/T_c >1$ can in all scenarios be achieved by further increasing the new physics contributions and thereby increasing the Higgs self-couplings. In particular, we observe that the non-perturbative modifications $\exp(-1/\phi^2)$ lead to a significantly higher value of the Higgs self couplings at fixed $\phi_c/T_c$ and are thus easier to detect. Given that for example exponential potentials feature a minimum value of $\lambda_{H^3}$ significantly larger than the simple $\phi^6$ extension, the LHC measurement might even allow first clues to the nature of new physics, even if the corresponding scale $\Lambda$ remains out of direct reach at the LHC.
Because the curves in Fig. \[fig:Higgs-self\] connect an IR observable with a UV property we can link the two regimes and make two observations. First, we can start in the IR and fix $\lambda_{H^3}$ for different UV potentials. Here, we find that an increase in $\phi_c/T_c$ or decrease in $T_c$ leads to a decrease in $\lambda_{H^4}$ for constant $\lambda_{H^3}$. Alternatively, we can fix $\phi_c/T_c$ for different UV potentials and find that a decrease in $\lambda_{H^3}$ corresponds to a decrease also in $\lambda_{H^4}$ or an increase in $T_c$.
Finally, Fig. \[fig:Higgs-self-coeff\] explicitly shows the connection between the strength of the observable effect at LHC scales, measured by $\lambda_{H^3}/\lambda_{H^3,0}$ and the size of the new physics contribution $\Delta V$ at the microscopic scale $\Lambda$, measured by the value of the dimensionless coefficients $\lambda_j$. The nature of the electroweak phase transition is encoded in the coloring of the lines. The onset of the first-order phase transition is at values that can also be read off from Fig. \[fig:Higgs-self\]: for logarithmic modifications we find the lowest value of $\lambda_{H^3}/\lambda_{H^3,0} \approx 1.4$, for the $\phi^6$ modification $\lambda_{H^3}/\lambda_{H^3,0} \approx 1.5$, and for exponential modifications $\lambda_{H^3}/\lambda_{H^3,0} \approx 1.9$. This size of all modifications can be probed in the high-luminosity run at the LHC. Importantly, the Higgs self-couplings grow continuously as a function of $\lambda_j$ while $\phi_c/T_c$ remains zero till the onset of the first-order phase transition and only then starts to grow continuously.
![Modification of the self-coupling $\lambda_{H^3}/\lambda_{H^3,0}$ as a function of the coefficients $\lambda_j$ from the different UV potentials given in Eq.. Blue lines represent first-order phase transitions and red dotted lines second-order phase transitions. The cutoff is $\Lambda = 2$ TeV.[]{data-label="fig:Higgs-self-coeff"}](Higgs-self-coeff){width="\textwidth"}
Outlook {#sec:outlook}
=======
Higgs pair production or the measurement of the Higgs self coupling is an extraordinarily interesting LHC analysis. We find that it is well motivated by modified Higgs potentials which allow for a strong first-order electroweak phase transition and hence an explanation of the observed matter vs anti-matter asymmetry. We have studied a wide range of such modifications to the Higgs potential, especially potentials that cannot be expanded as an effective field theory. We used the functional renormalization group to describe the dependence on the field value $\phi$ and on the temperature $T$. For all classes of potentials considered here, there exists an appropriate choice of model parameters, for which the phase transition is of first order and sufficiently strong, $\phi_c/T_c \gtrsim 1$.
Our numerical analysis indicates that the requirement $\phi_c/T_c=1$ corresponds to a critical scale of the order of 10 TeV for all our potentials, where the potentials become strongly coupled. Below this scale we can rely on our assumed potentials to describe LHC signals. We then found that a strong first-order phase transition universally predicts an enhancement of the Higgs self-couplings $\lambda_{H^3} \gtrsim 1.5 \lambda_{H^3,0}$ and $\lambda_{H^4} \gtrsim
4 \lambda_{H^4,0}$. Extending earlier studies, we systematically established this connection between a first-order transition and a measurable deviation of the Higgs self couplings, employing a method that can describe systems with multiple physical scales in a controlled manner. While it might be possible that a new physics model features a strong first-order transition with all effects on $\lambda_{H^{3/4}}$ canceling accidentally [@Noble], none of our examples falls into this class. We conclude that a measurement of the Higgs self-couplings at the LHC indeed serves as an indirect probe of a first-order phase transition and thus of electroweak baryogenesis in generic setups.
On the other hand, we observed that it is possible to obtain large deviations in the Higgs self-interactions for our class of non-perturbative potentials without the condition $\phi_c/T_c\geq 1$ being fulfilled. For example with an exponential modification of the Higgs potential the physical Higgs self-coupling reaches $\lambda_{H^3}
\approx 1.9 \lambda_{H^3,0}$ already significantly below $\phi_c/T_c=1$. On the theoretical side, a quantitative upgrade of our analysis includes, but is not limited to, a full treatment of the weak gauge sector as well as improvements in our treatment of the Yukawa sector, which might result in quantitative changes of the order of 10 %, cf. [@Gies:2017zwf]. An as precise as possible measurement of the triple-Higgs interaction is clearly desirable. For instance a $20\%$ measurement of a relatively small modification of $\lambda_{H^3}/\lambda_{H^3,0}$ could exclude such exponential potentials as sources of electroweak baryogenesis. Such an actual measurement could therefore provide valuable hints guiding theoretical studies of interesting extended Higgs models.
**Acknowledgments**
We thank R. Sondenheimer for insightful discussions. MR acknowledges funding from IMPRS-PTFS and is grateful to the DFG research training group GRK 1523 at TPI Jena for hospitality. JMP is supported by the Helmholtz Alliance HA216/EMMI and by ERC-AdG-290623. AE is supported by an Emmy-Noether grant of the DFG under Ei-1037/1 and an Emmy-Noether visiting fellowship at the Perimeter Institute for Theoretical Physics. This work is part of and supported by the DFG Collaborative Research Centre “SFB 1225 (ISOQUANT)”.
Flow equations {#app:der-flows}
==============
The set of couplings in our setup consists of the $SU(3)$ coupling $g_3$, a fiducial coupling $g_F$ that simulates the $SU(2)$ and the $U(1)$ sector, the top-Yukawa coupling $y_t$, and the full Higgs potential $V(\phi)$ [@our_paper]. For the $SU(3)$ coupling it suffices to consider one-loop running, since higher-order or threshold corrections have little impact on the phase transition. The one-loop beta function is given by $$\begin{aligned}
\beta_{g_3} = -\frac{g_3^3}{(4 \pi)^2} \left(11 - \frac{2}{3} n_f\right) \,,\end{aligned}$$ with $n_f=6$. We fix the $SU(3)$ coupling through $g_3(1\,\text{TeV})=1.06$, so the scale-dependent $SU(3)$ coupling is known analytically. We approximate its temperature dependence by replacing $k \rightarrow \sqrt{k^2+\pi\, T^2}$, $$\begin{aligned}
g_3(k, T) = \left( \frac7{8 \pi ^2} \ln \frac{\sqrt{k^2+\pi T^2}}{1\,\text{TeV}}
+\frac{1}{1.06^2}\right)^{-1/2} \,.\end{aligned}$$ The logarithmic running of the $U(1)$ and $SU(2)$ couplings is sufficiently slow to be negligible for our purpose [@our_paper]. We model it as a fiducial coupling $g_F$ that is a constant as a function of the RG scale and thus also a constant as a function of the temperature. At finite temperature, this simplified treatment must be ameliorated by a thermal mass generated by fluctuations from the electroweak sector. According to the high-$T$ expansion of the one-loop thermal potential it is given by $$\begin{aligned}
V_{\text{thermal mass}} (\phi, T ) = \frac{1}{16}\left( 3g^2 + {g'}^{2} \right)
\; \frac{T^2 \phi^2}{2} \,,\end{aligned}$$ where $g=0.65$ and $g'=0.36$ are the $SU(2)$ and $U(1)$ gauge couplings, respectively.
To derive beta functions for the Higgs potential and the top-Yukawa coupling we introduce the renormalized dimensionless field $\rho$ and the dimensionless potential $u$ $$\begin{aligned}
\rho = \frac{\phi^2}{2 k^2 Z_\phi}
{\qquad \qquad \qquad}u(\rho) = \frac{V(\phi(\rho))}{k^4} \,.\end{aligned}$$ The wave function renormalizations of the fields appear in the beta functions only via their anomalous dimension $$\begin{aligned}
\eta_\phi = - \frac{\mathrm d \log Z_\phi }{\mathrm d \log k}
{\qquad \qquad \qquad}\eta_\psi = - \frac{\mathrm d \log Z_\psi }{\mathrm d \log k} \,.\end{aligned}$$ Written in terms of threshold functions, the beta function for the top Yukawa coupling agrees with that from Refs. [@our_paper; @Gies:2013fua], see, [*e.g.*]{} Eq.(C8) of Ref. [@our_paper]. However, we use a spatial regulator as described below and temperature-dependent threshold functions. The spatial regulator changes some prefactors, which is compensated by the different definition of the threshold functions. The beta function is given by $$\begin{aligned}
\label{eq:flow-eq-yuk} \frac{\mathrm d y_t^2}{\mathrm d
\log k} ={}& y_t^2 \left(\eta_\phi+2 \eta_\psi\right) -
\frac{y_t^4}{\pi^2} \left(3 \kappa u''(\kappa) + 2 \kappa^2
u^{(3)}(\kappa)\right) l_{1,2}^{(FB)4}\left(\kappa
y_t^2,u'(\kappa) + 2 \kappa
u''(\kappa);\eta_\psi,\eta_\phi;T\right)
\notag \\
&+ \frac{y_t^4}{2\pi^2}
\left(l_{1,1}^{(FB)4}\left(\kappa y_t^2,u'(\kappa) +
2 \kappa u''(\kappa);\eta_\psi,\eta_\phi;T\right)
\right. \notag \\
&\left.\hspace{2cm}-2 \kappa y_t^2
\,l_{2,1}^{(FB)4}\left(\kappa y_t^2,u'(\kappa) + 2
\kappa u''(\kappa);\eta_\psi,\eta_\phi;T\right)
\right)
\notag \\
&+\frac{3}{\pi^2} \frac{\left(N_c^2-1\right)}{2N_c}
g_3^2 y_t^2 \left(2 \kappa y_t^2\,
l_{2,1}^{(FB)4}(\kappa y_t^2,0;\eta_\psi,\eta_A;T)
- l_{1,1}^{(FB)4}(\kappa y_t^2,0;\eta_\psi,\eta_A;T)\right) \notag\\
&-\frac{c_y g_F^2 y_t^2}{16 \pi ^2 \left(1+
\left(\frac{80}{246}\right)^2\kappa\right)}\,,\end{aligned}$$ where $c_y=97/30$ and $N_c=3$. It depends on the position of the renormalized dimensionless minimum $\kappa$ of the potential, the anomalous dimensions of the fields, as well as on regulator-dependent threshold functions specified below. Here, we have employed the same projection scheme onto the Yukawa flow as in [@our_paper] for reasons of comparison. In principle, there exists an improved scheme [@Pawlowski:2014zaa] more adequately capturing higher-order contributions to the Yukawa flow for the present model [@Gies:2017zwf], possibly improving the fixing of initial conditions on the 5% level. In either case, working in the symmetric regime with $\kappa=0$ and neglecting the additional $\eta$ dependence in the threshold functions reproduces the universal one-loop beta functions, as it should.
The beta function for the Higgs potential at vanishing temperature has been computed in Ref. [@our_paper; @Gies:2013fua], see, [*e.g.*]{} Eq.(E1) of Ref. [@our_paper]. As for the beta function of the Yukawa coupling, the present finite temperature beta function for the Higgs potential agrees with the $T=0$ one in terms of the threshold functions $$\begin{aligned}
\label{eq:flow-eq-pot} \frac{\mathrm d u (\rho)}{\mathrm
d \log k}={}& -4 u(\rho) +(2+\eta_\phi) \rho\, u'(\rho) \notag
\\&+\frac{1}{4\pi^2}\left( l_0^{(B)4}\left(u'(\rho)+2 \rho\,
u''(\rho);\eta_\phi;T\right) - 4 N_c l_0^{(F)4}(y_t^2\rho
;\eta_\psi;T)\right) +\frac {c_l}{2 \pi ^2 \left(1 + \frac{g_F^2
\rho}{2}\right)} \,,\end{aligned}$$ where $c_l=9/16$ and again $N_c=3$.
Finally, we need expressions for the anomalous dimensions of the Higgs field and the top-quark: the first two terms in Eq. are integral parts of the universal one-loop contribution. In terms of the threshold functions the anomalous dimension of the top quark agrees with the $T=0$ one in Eq.(C8) of Ref. [@our_paper], and the anomalous dimension of the scalar field has the same form as in Eq.(16) of Ref. [@Gies:2013fua]. With the thermal threshold functions of the present work this means $$\begin{aligned}
\label{eq:anom-dim} \eta_\phi ={}& \frac{2}{3\pi^2} N_c
y_t^2 \left(m_4^{(F)4}(\kappa y_t^2;\eta_\psi;T)
- \kappa y_t^2 m_2^{(F)4}(\kappa y_t^2;\eta_\psi;T) \right) \notag\\
&+ \frac{1}{3\pi^2} \kappa \left(3 u''(\kappa)
+ 2 \kappa u^{(3)}(\kappa)\right)^2 m_4^{(B)4}\left(u'(\kappa)
+ 2 \kappa u''(\kappa);\eta_\phi;T\right) \,,\notag\\
\eta_\psi ={}& \frac{1}{6\pi^2} y_t^2 \,
m_{1,2}^{(FB)4}\left(\kappa y_t^2,u'(\kappa) + 2 \kappa u''(\kappa);\eta_\phi;T\right)\notag\\
&+ \frac{1}{2\pi^2} \frac{\left(N_c^2-1\right)}{ 2N_c} g_3^2
\left(m_{1,2}^{(FB)4}(\kappa y_t^2,0;\eta_\psi,0;T)
-\tilde{m}_{1,1}^{(FB)4}(\kappa y_t^2,0;\eta_\psi,0;T)\right) \,.\end{aligned}$$ The beta functions found above are expressed in terms of regulator-dependent and temperature-dependent threshold functions. Here we provide explicit analytic results for these threshold functions for one specific regulator. The analyticity of the threshold function is rooted in the use of a Litim-type regulator [@flat-reg] that only regularizes the spatial momenta. The dimensionless bosonic and fermionic propagators are regularized as $$\begin{aligned}
G_\phi(\omega_n^2,{\vec p\,}^2 , m_\phi^2) =
\left(\omega_n^2 + {\vec p\,}^2/k^2 (1+r_B({\vec p\,}^2/k^2))+m_\phi^2\right)^{-1}\,,\notag\\
G_\psi(\nu_n^2,{\vec p\,}^2 , m_\psi^2) =
\left(\nu_n^2 + {\vec p\,}^2/k^2 (1+r_F({\vec p\,}^2/k^2))+m_\psi^2\right)^{-1}\,,\end{aligned}$$ with the bosonic Matsubara frequency $\omega_n=2\pi n T/k$ and the fermionic Matsubara frequency $\nu_n = 2\pi (n+\frac12)T/k$. Note that $m_\phi$ and $m_\psi$ are dimensionless mass-like arguments. The bosonic and fermionic regulator shape functions read [@flat-reg] $$\begin{aligned}
r_B(x) = \left(x^{-1}-1\right)\Theta(1-x)\,,{\qquad \qquad}r_F(x) = \left(x^{-1/2}-1\right)\Theta(1-x)\,,
\label{eq:flat-reg}\end{aligned}$$ where $x={\vec p\,}^2/k^2$. In the following, we express the threshold functions in terms of the bosonic and fermionic distribution functions, $$\begin{aligned}
n_{F,B}(m_{\psi,\phi}^2,T) = \left(\exp\left(\frac{k}{T} \sqrt{1+m_{\psi,\phi}^2}\right)\mp 1\right)^{-1}\; .\end{aligned}$$ The set of threshold functions we need in our calculation includes $$\begin{aligned}
l_0^{(B)d}(m_\phi^2;\eta_\phi;T) ={}& \frac{2}{d-1} \left(1-\frac{\eta_\phi}{d+1}\right) \mathcal{B}_{(1)} (m_\phi^2;T) \,,\notag \\
l_0^{(F)d}(m_\psi^2;\eta_\psi;T) ={}& \frac{2}{d-1} \left(1-\frac{\eta_\psi}{d}\right) \mathcal{F}_{(1)} (m_\psi^2;T) \,,\notag \\
l_{n,m}^{(FB)d}(m_\psi^2,m_\phi^2;\eta_\psi,\eta_\phi;T) ={}& \frac{2}{d-1}
\left(n \left( 1 - \frac{\eta_\psi}d \right) \mathcal{FB}_{(n+1,m)}(m_\psi^2,m_\phi^2;T)\right.
\notag \\&\left.\hspace{1.7cm}
+ m \left( 1 - \frac{\eta_\phi}{d+1} \right) \mathcal{FB}_{(n,m+1)}(m_\psi^2,m_\phi^2;T) \right) \,,\notag
\end{aligned}$$ $$\begin{aligned}
m_4^{(B)d}\left(m_\phi^2;\eta_\phi;T\right) &= \mathcal{B}_{(4)}(m_\phi^2;T) \,,\notag \\
m_{2}^{(F)d}(m_\psi^2;T) &= \mathcal{F}_{(4)}(m_\psi^2;T) \,,\notag \\
m_{4}^{(F)d}(m_\psi^2;\eta_\psi;T) &= \mathcal{F}_{(4)}(m_\psi^2;T)
+ \frac{1- \eta_\psi}{d-3} \mathcal{F}_{(3)}(m_\psi^2;T)
- \frac12\left(\frac{1- \eta_\psi}{ d - 3} + \frac{1}{2} \right) \mathcal{F}_{(2)}(m_\psi^2;T) \,,\notag \\
m_{1,2}^{(FB)d}(m_\psi^2,m_\phi^2;\eta_\psi,\eta_\phi;T) &=
\left(1- \frac{\eta_\phi}d\right) \mathcal{FB}_{(1,2)}(m_\psi^2,m_\phi^2;T) \,,\notag \\
\tilde{m}_{1,1}^{(FB)d}(m_\psi^2,m_\phi^2;\eta_\psi,\eta_\phi;T) &=
\frac{2}{d-2} \left( \left(1- \frac{\eta_\phi}d\right) \mathcal{FB}_{(1,2)}(m_\psi^2,m_\phi^2;T) \right. \notag \\
&{\qquad \qquad}+ \left(1- \frac{\eta_\psi}{d-1}\right) \mathcal{FB}_{(2,1)}(m_\psi^2,m_\phi^2;T) \notag \\
&{\qquad \qquad}\left.- \frac{1}{2} \left(1- \frac{\eta_\psi}{d-1}\right) \mathcal{FB}_{(1,1)}(m_\psi^2,m_\phi^2;T) \right) \,.\end{aligned}$$ All threshold functions are expressed in terms of $$\begin{aligned}
\mathcal{F}_{(1)} (m_\psi^2;T) &= \frac{T}{k} {\sum_{n\in\mathbb Z}}G_\psi (\nu_n,m_\psi^2) \,,\notag \\
\mathcal{B}_{(1)} (m_\phi^2;T) &= \frac{T}{k} {\sum_{n\in\mathbb Z}}G_\phi (\omega_n,m_\phi^2) \,,\notag \\
\mathcal{FB}_{(1,1)}(m_\psi^2,m_\phi^2;T) &= \frac{T}{k}
{\sum_{n\in\mathbb Z}}G_\psi(\nu_n,m_\psi^2) G_\phi(\omega_n,m_\phi^2) \,.\end{aligned}$$ At finite temperature for the flat regulators in Eq. they are given by $$\begin{aligned}
\mathcal{F}_{(1)} (m_\psi^2;T) ={}& \frac{1}{\sqrt{ 1 + m_\psi^2}} \left( \012 - n_F ( m_\psi^2 , T ) \right)\,,\notag \\
\mathcal{B}_{(1)} (m_\phi^2;T) ={}& \frac{1}{\sqrt{ 1 + m_\phi^2}} \left( \012 + n_B ( m_\phi^2 , T ) \right)\,,\notag \\
\mathcal{FB}_{(1,1)}(m_\psi^2,m_\phi^2;T) ={}&
\Bigg[ \frac{\frac{1}{2} + n_B(m_\phi^2 , T) }{2\sqrt{1+m_\phi^2}}
\left( \left(m_\psi^2 +1 - \left(i\pi T /k + \sqrt{1+m_\phi^2} \right)^2\right)^{-1}\right.\notag \\
&\hspace{3.5cm}\left.+\left(m_\psi^2 +1 -
\left(i\pi T /k - \sqrt{1+m_\phi^2} \right)^2\right)^{-1} \right) \notag \\
&\,\,+\frac{ \frac{1}{2} - n_F( m_\psi^2 , T )}{2\sqrt{1+m_\psi^2}}
\left( \left(m_\phi^2 +1 - \left(i\pi T /k + \sqrt{1+m_\psi^2} \right)^2\right)^{-1}\right.\notag \\
&\hspace{3.5cm} +\left. \left(m_\phi^2 +1 -
\left(i\pi T /k - \sqrt{1+m_\psi^2} \right)^2\right)^{-1} \right) \Bigg]\,.\end{aligned}$$ They obey the relations $$\begin{aligned}
{9}
\frac{\partial\mathcal{F}_{(n)} }{\partial m_\psi^2} &= -n \mathcal{F}_{(n+1)}\,, {\qquad \qquad \qquad}&
\frac{\partial\mathcal{B}_{(n)} }{\partial m_\phi^2} &= -n \mathcal{B}_{(n+1)} \,,\notag \\[1ex]
\frac{\partial\mathcal{FB}_{(m,n)}}{\partial m_\psi^2} &= -m \mathcal{FB}_{(m+1,n)}\,, {\qquad \qquad \qquad}&
\frac{\partial\mathcal{FB}_{(m,n)}}{\partial m_\phi^2} &= -n \mathcal{FB}_{(m,n+1)}\,.\end{aligned}$$ The notation and the threshold functions agree with Ref. [@Pawlowski:2014zaa]. Note, that the $T\to0$ limit of the threshold functions does not agree with the ones given in Ref. [@Gies:2013fua], since we use a spatial regulator while Ref. [@Gies:2013fua] uses a covariant regulator. This concludes the list of threshold functions and relations necessary in order to numerically evaluate the previously given beta functions.
Grid approach and benchmarking {#app:grid-code}
==============================
We solve the functional differential equation for the Higgs potential, Eq., using a grid code. This means that the potential $u(\rho)$ and its derivative $u^\prime(\rho)$ are discretized on a grid in the field invariant $\rho$. The discretization converts the partial differential equation for $u(\rho)$ into a large set of coupled ordinary differential equations. The grid code has to manage a numerical integration from $k=\Lambda$, where we initialize the flow, down to $k=k_\text{IR} \approx 100$ GeV. At this IR value all physical relevant quantities are frozen out and only convexity-generating processes take place.
The grid code also has to cover a large range of values in the scalar field $ 0 \leq \phi \leq c \Lambda$, where we typically choose $c =
\mathcal{O}(1\dots 10)$. To resolve both, large field values and the minimum of the potential at small field values, we employ an exponential distribution of the grid points $\rho_i=\phi_i^2/2$ with $i\in{0,\ldots,N-1}$ according to $$\begin{aligned}
\rho_i = \rho_\text{a} + \frac{\exp\left(\frac{i}{c_\text{grid}}\right)-1}{\exp\left(\frac{(N-1)}{c_\text{grid}}\right)}\rho_\text{b}\,,\end{aligned}$$ where $N$ is the number of grid points, $c_\text{grid}$ a grid parameter that governs the distributions of the grid points, and $\rho_\text{a}$ and $\rho_\text{b}$ the smallest and largest included field value, respectively.
We introduce a grid for the potential $u(\rho_i)$ as well as for the derivative of the potential $u'(\rho_i)$, and we match the second and third derivative of the potential in between the grid points [@grid-code]. This is augmented by a differential equation for the top-Yukawa coupling, while the $SU(3)$ coupling is already integrated out and the fiducial coupling for $SU(2)$ and $U(1)$ remains constant. Consequently, we obtain a system of $2N + 1$ coupled differential equations for a grid consisting of $N$ points, which is solved with an iterative Runge-Kutta-Fehlberg method with an adaptive step size.
At the IR scale and at vanishing temperature, we match the output of the grid code with the physically known observables, see Eq.. This is implemented on the level of the variables of the grid code and in particular we demand that the errors fulfill $\Delta \rho_\text{min} \leq 20$ GeV$^2$, $\Delta \lambda_4
\leq0.002$ and $\Delta y_t \leq 0.0014$. Expressed in the quantities of Eq. these errors correspond to $\Delta v \leq
0.08$ GeV, $\Delta m_H \leq 0.28$ GeV, and $\Delta m_t \leq 0.23$ GeV. It is important to determine the vacuum expectation value more precisely since its error directly influences the error on the Higgs and the top mass.
To achieve this precision we tune the parameters $\mu$, $\lambda_4$ and $y_t$ at the UV scale, which is done by a secant method in $\mu$ and a two-dimensional bisection method in $\lambda_4$ and $y_t$ . The grid code might exhibit other systematic errors and in particular the measurement of the Higgs mass is challenging since it is related to the second derivative of the potential. Hence we conservatively estimate the total accuracy of the IR values with $$\begin{aligned}
\Delta v \leq 0.2 \ {\text{GeV}}\,, {\qquad \qquad}\Delta m_H \leq 1.5 \ {\text{GeV}}\,, {\qquad \qquad}\Delta m_t \leq 0.5 \ {\text{GeV}}\,.\end{aligned}$$ The tuning process is performed at vanishing temperature and the tuned initial values are subsequently used as initial values for all finite-temperature computations. For each temperature we initialize the flow in this way and determine the position of the minimum at the IR scale $k_\text{IR}$. The critical temperature is obtained with a bisection method where we demand an accuracy of $\Delta T_c \leq 0.2$ MeV. This high accuracy is necessary for a precise value of $\phi_c$, which is in turn given by the position of the minimum at the temperature just below $T_c$. From the grid code, it is difficult to get a clear signature distinguishing between second-order phase transitions and weak first-order phase transitions. Within our numerical accuracy, a reliable distinguishing signature is not available for $\phi_c\lesssim20$ GeV. For finite temperature computations we slightly increase the number of grid points, since the exponential functions in the bosonic and fermionic distribution functions make these computations technically more challenging.
![Modification of the self-couplings $\lambda_{H^3}/\lambda_{H^3,0}$ (left) and $\lambda_{H^4}/\lambda_{H^4,0}$ (right) as a function of $\phi_c/T_c$ for polynomial and logarithmic modifications of the UV potentials, cf. Eq.. We compare results for $N=70$ and $N=90$ grid points.[]{data-label="fig:convergence"}](Higgs-self-convergence){width="\textwidth"}
![Modification of the self-couplings $\lambda_{H^3}/\lambda_{H^3,0}$ (left) and $\lambda_{H^4}/\lambda_{H^4,0}$ (right) as a function of $\phi_c/T_c$ for exponential modifications of the UV potentials, cf. Eq.. We compare results for $N=130$ and $N=150$ grid points.[]{data-label="fig:convergence-exp"}](Higgs-self-convergence-exp){width="\textwidth"}
We test our numerical results by first comparing the observables for two different numbers of grid points. The necessary number varies with our choice of cutoff and the modification of the Higgs potential. For example, more grid points are necessary for the exponential modifications of the potential. For polynomial and logarithmic modifications and a cutoff $\Lambda=2$ TeV, we use typically $N=90$ grid points, while for exponential modifications with the same cutoff we use $N=150$ grid points. In Fig. \[fig:convergence\] we display results for polynomial and logarithmic modifications. In particular we show the correlation between the strength of first-order phase transition and the Higgs-self couplings. In Fig. \[fig:convergence-exp\] we show the same correlation but for exponential modifications and for $N=130$ and for $N=150$ grid points. The results for $N=90$ and for $N=150$ are identical with those displayed in Fig. \[fig:Higgs-self\].
![Relative change of $\lambda_{H^3}/\lambda_{H^3,0}$ with different numbers of grid points as a function of $\phi_c/T_c$ for polynomial and logarithmic modifications of the UV potentials, cf. Eq.. In the regime of interest of $\phi_c/T_c
\geq 1$, the relative difference between $N=70$ and $N=90$ is in the sub-percent regime, $\leq 0.5 \%$.[]{data-label="fig:relative-error"}](Higgs-self-relative-error){width=".6\textwidth"}
To make our analysis more quantitative we also display the relative change of the correlation for polynomial and logarithmic modifications in Fig. \[fig:relative-error\]. The results do not change significantly when we increase the number of grid points. In case of polynomial and logarithmic modifications the amount of wiggles in the region of a weak first-order phase transition, which originates from numerical uncertainties, is further reduced. In the region of a weak first-order phase transition we have a relative change of less than 2%, while in the region of a strong first-order phase transition we have a relative change of less than 0.5%. This is sufficient for our analysis, since we are only interested in the latter case. In case of the exponential modifications the change is hardly visible. The relative change is globally less than 0.02%. These results illustrate that our findings are indeed numerically stable.
Finally, we can compare our functional renormalization group results to other methods, for instance to the mean-field-like methods of Ref. [@christophe_geraldine]. To perform a meaningful comparison, we have to take into account the slightly different setup: while we modify the microscopic potential, Ref. [@christophe_geraldine] implements the modifications directly at the level of the effective potential. This means that in our setup a $\phi^6$ modification of the microscopic potential generates finite higher-order modifications through quantum fluctuations, which in the weak coupling regime are similar to the one-loop determinant. These additional terms do not appear in Ref. [@christophe_geraldine].
For our comparison we therefore adjust the parameter $\lambda_6$ such that the $T=0$ effective potentials of both setups agree. Due to the impact of quantum fluctuations, different values of $\Lambda$ require slightly different initial conditions for $\lambda_6$ in our setup. With a cutoff $\Lambda=1$ TeV it turns out that this is the case for $\lambda_6\approx 0.21$, while for a cutoff $\Lambda=0.6$ TeV we find $\lambda_6\approx 0.19$. The difference in values of $\lambda_6$ is accounted for by the RG flow between the two choices of cutoff scale. With these values we can then compare $T_c$ and $\phi_c/T_c$. As expected, we indeed find good qualitative agreement. For instance, for $\Lambda=0.6$ TeV we find $\phi_c/T_c=2.7$ and $T_c=83$ GeV vs $\phi_c/T_c=2.8$ and $T_c=75$ GeV from Ref. [@christophe_geraldine]. We emphasize that a more precise agreement cannot be expected: the modification of the microscopic and the effective Higgs potential are necessarily different, as our setup accounts for quantum fluctuations, in particular affecting $\lambda_6$ between the microscopic scale and the IR.
[99]{}
A. D. Sakharov, Pisma Zh. Eksp. Teor. Fiz. [**5**]{}, 32 (1967) \[JETP Lett. [**5**]{}, 24 (1967)\] \[Sov. Phys. Usp. [**34**]{}, 392 (1991)\] \[Usp. Fiz. Nauk [**161**]{}, 61 (1991)\]; V. A. Kuzmin, V. A. Rubakov and M. E. Shaposhnikov, Phys. Lett. [**155B**]{}, 36 (1985); V. A. Rubakov and M. E. Shaposhnikov, Usp. Fiz. Nauk [**166**]{}, 493 (1996) \[Phys. Usp. [**39**]{}, 461 (1996)\] \[hep-ph/9603208\]; M. Trodden, Rev. Mod. Phys. [**71**]{}, 1463 (1999) \[hep-ph/9803479\]. M. Sher, Phys. Rept. [**179**]{}, 273 (1989); A. G. Cohen, D. B. Kaplan and A. E. Nelson, Ann. Rev. Nucl. Part. Sci. [**43**]{}, 27 (1993) \[hep-ph/9302210\]; J. M. Cline, \[hep-ph/0609145\]; D. E. Morrissey and M. J. Ramsey-Musolf, New J. Phys. [**14**]{}, 125003 (2012) \[arXiv:1206.2942 \[hep-ph\]\]; T. Konstandin, Phys. Usp. [**56**]{}, 747 (2013) \[Usp. Fiz. Nauk [**183**]{}, 785 (2013)\] \[arXiv:1302.6713 \[hep-ph\]\]. M. E. Shaposhnikov, Nucl. Phys. B [**287**]{}, 757 (1987); Nucl. Phys. B [**299**]{}, 797 (1988). Z. Fodor, J. Hein, K. Jansen, A. Jaster and I. Montvay, Nucl. Phys. B [**439**]{}, 147 (1995) \[hep-lat/9409017\]; W. Buchmuller, Z. Fodor and A. Hebecker, Nucl. Phys. B [**447**]{}, 317 (1995) \[hep-ph/9502321\]; K. Kajantie, M. Laine, K. Rummukainen and M. E. Shaposhnikov, Phys. Rev. Lett. [**77**]{}, 2887 (1996) \[hep-ph/9605288\]; K. Rummukainen, M. Tsypin, K. Kajantie, M. Laine and M. E. Shaposhnikov, Nucl. Phys. B [**532**]{}, 283 (1998) \[hep-lat/9805013\]; F. Csikor, Z. Fodor and J. Heitger, Phys. Rev. Lett. [**82**]{}, 21 (1999) \[hep-ph/9809291\]. M. E. Shaposhnikov, Phys. Lett. B [**277**]{}, 324 (1992) Erratum: \[Phys. Lett. B [**282**]{}, 483 (1992)\]. X. m. Zhang, Phys. Rev. D [**47**]{}, 3065 (1993) \[hep-ph/9301277\]; X. Zhang, B. L. Young and S. K. Lee, Phys. Rev. D [**51**]{}, 5327 (1995) \[hep-ph/9406322\].
C. Grojean, G. Servant and J. D. Wells, Phys. Rev. D [**71**]{}, 036001 (2005) \[hep-ph/0407019\].
S. Kanemura, Y. Okada and E. Senaha, Phys. Lett. B [**606**]{}, 361 (2005) \[hep-ph/0411354\]; F. P. Huang, P. H. Gu, P. F. Yin, Z. H. Yu and X. Zhang, Phys. Rev. D [**93**]{}, no. 10, 103515 (2016) \[arXiv:1511.03969 \[hep-ph\]\]; P. Huang, A. Joglekar, B. Li and C. E. M. Wagner, Phys. Rev. D [**93**]{}, no. 5, 055049 (2016) \[arXiv:1512.00068 \[hep-ph\]\]; A. Kobakhidze, L. Wu and J. Yue, JHEP [**1604**]{}, 011 (2016) \[arXiv:1512.08922 \[hep-ph\]\]; Q. H. Cao, G. Li, B. Yan, D. M. Zhang and H. Zhang, Phys. Rev. D [**96**]{}, no. 9, 095031 (2017) \[arXiv:1611.09336 \[hep-ph\]\]; D. Curtin, P. Meade and H. Ramani, arXiv:1612.00466 \[hep-ph\]; X. Gan, A. J. Long and L. T. Wang, Phys. Rev. D [**96**]{}, no. 11, 115018 (2017) \[arXiv:1708.03061 \[hep-ph\]\]; Q. H. Cao, F. P. Huang, K. P. Xie and X. Zhang, Chin. Phys. C [**42**]{}, no. 2, 023103 (2018) \[arXiv:1708.04737 \[hep-ph\]\]; B. Jain, S. J. Lee and M. Son, arXiv:1709.03232 \[hep-ph\]. A. Noble and M. Perelstein, Phys. Rev. D [**78**]{}, 063518 (2008) \[arXiv:0711.3018 \[hep-ph\]\]. U. Baur, T. Plehn and D. L. Rainwater, Phys. Rev. Lett. [**89**]{}, 151801 (2002) \[hep-ph/0206024\]; Phys. Rev. D [**67**]{}, 033003 (2003) \[hep-ph/0211224\]. for an analysis of the Higgs-gauge sector based on the full Run I data set see A. Butter, O. J. P. Eboli, J. Gonzalez-Fraile, M. C. Gonzalez-Garcia, T. Plehn and M. Rauch, JHEP [**1607**]{}, 152 (2016) \[arXiv:1604.03105 \[hep-ph\]\]. K. Holland and J. Kuti, Nucl. Phys. Proc. Suppl. [**129**]{}, 765 (2004) \[hep-lat/0308020\]; K. Holland, Nucl. Phys. Proc. Suppl. [**140**]{}, 155 (2005) \[hep-lat/0409112\]; J. R. Espinosa, T. Konstandin and F. Riva, Nucl. Phys. B [**854**]{}, 592 (2012) \[arXiv:1107.5441 \[hep-ph\]\]. H. Gies, C. Gneiting and R. Sondenheimer, Phys. Rev. D [**89**]{}, no. 4, 045012 (2014) \[arXiv:1308.5075 \[hep-ph\]\]. H. Gies and R. Sondenheimer, Eur. Phys. J. C [**75**]{}, no. 2, 68 (2015) \[arXiv:1407.8124 \[hep-ph\]\]; Phil. Trans. Roy. Soc. Lond. [**376**]{}, no. 2114, 20170120 (2018) \[arXiv:1708.04305 \[hep-ph\]\]. J. Borchardt, H. Gies and R. Sondenheimer, Eur. Phys. J. C [**76**]{}, no. 8, 472 (2016) \[arXiv:1603.05861 \[hep-ph\]\].
R. Sondenheimer, arXiv:1711.00065 \[hep-ph\]. O. Akerlund and P. de Forcrand, Phys. Rev. D [**93**]{}, no. 3, 035015 (2016) \[arXiv:1508.07959 \[hep-lat\]\]; D. Y.-J. Chu, K. Jansen, B. Knippschild and C.-J. D. Lin, EPJ Web Conf. [**175**]{}, 08017 (2018) \[arXiv:1710.09737 \[hep-lat\]\]. O. Akerlund, P. de Forcrand and J. Steinbauer, PoS LATTICE [**2015**]{}, 229 (2016) \[arXiv:1511.03867 \[hep-lat\]\]. C. Wetterich, Phys. Lett. B [**301**]{}, 90 (1993). for an overview see [*e.g.*]{} J. Berges, N. Tetradis and C. Wetterich, Phys. Rept. [**363**]{}, 223 (2002); J. M. Pawlowski, Annals Phys. [**322**]{}, 2831 (2007); H. Gies, Lect. Notes Phys. [**852**]{}, 287 (2012); J. Braun, J. Phys. G [**39**]{}, 033001 (2012). I. V. Krive and A. D. Linde, Nucl. Phys. B [**117**]{}, 265 (1976). M. Lindner, Z. Phys. C [**31**]{}, 295 (1986). D. Buttazzo, G. Degrassi, P. P. Giardino, G. F. Giudice, F. Sala, A. Salvio and A. Strumia, JHEP [**1312**]{}, 089 (2013) \[arXiv:1307.3536 \[hep-ph\]\]. A. Eichhorn, H. Gies, J. Jaeckel, T. Plehn, M. M. Scherer and R. Sondenheimer, JHEP [**1504**]{}, 022 (2015) \[arXiv:1501.02812 \[hep-ph\]\]. Z. Fodor, K. Holland, J. Kuti, D. Nogradi and C. Schroeder, PoS LAT [**2007**]{}, 056 (2007) \[arXiv:0710.3151 \[hep-lat\]\]; P. Hegde, K. Jansen, C.-J. D. Lin and A. Nagy, PoS LATTICE [**2013**]{}, 058 (2014) \[arXiv:1310.6260 \[hep-lat\]\]; D. Y.-J. Chu, K. Jansen, B. Knippschild, C.-J. D. Lin, K. I. Nagai and A. Nagy, PoS LATTICE [**2014**]{}, 278 (2014) \[arXiv:1501.00306 \[hep-lat\]\]. A. Jakovac, I. Kaposvari and A. Patkos, Mod. Phys. Lett. A [**32**]{}, no. 02, 1750011 (2016) \[arXiv:1508.06774 \[hep-th\]\]; Int. J. Mod. Phys. A [**31**]{}, no. 28n29, 1645042 (2016) \[arXiv:1510.05782 \[hep-th\]\]. A. Eichhorn and M. M. Scherer, Phys. Rev. D [**90**]{} (2014) 025023 \[arXiv:1404.5962 \[hep-ph\]\]. V. Branchina and H. Faivre, Phys. Rev. D [**72**]{}, 065017 (2005) \[hep-th/0503188\]; V. Branchina, H. Faivre and V. Pangon, J. Phys. G [**36**]{}, 015006 (2009) \[arXiv:0802.4423 \[hep-ph\]\]. P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], Astron. Astrophys. [**594**]{}, A13 (2016) \[arXiv:1502.01589 \[astro-ph.CO\]\]. with the exception of a different normalization of $\lambda_6$ we use the conventions of T. Plehn, Lect. Notes Phys. [**886**]{} (2015). <http://www.thphys.uni-heidelberg.de/~plehn/>
M. A. Shifman, A. I. Vainshtein, M. B. Voloshin and V. I. Zakharov, Sov. J. Nucl. Phys. [**30**]{}, 711 (1979) \[Yad. Fiz. [**30**]{}, 1368 (1979)\]; B. A. Kniehl and M. Spira, Z. Phys. C [**69**]{}, 77 (1995) \[hep-ph/9505225\]; M. Spira, JHEP [**1610**]{}, 026 (2016) \[arXiv:1607.05548 \[hep-ph\]\]. O. J. P. Eboli, G. C. Marques, S. F. Novaes and A. A. Natale, Phys. Lett. B [**197**]{}, 269 (1987); D. A. Dicus, C. Kao and S. S. D. Willenbrock, Phys. Lett. B [**203**]{}, 457 (1988); E. W. N. Glover and J. J. van der Bij, Nucl. Phys. B [**309**]{}, 282 (1988). T. Plehn, M. Spira and P. M. Zerwas, Nucl. Phys. B [**479**]{}, 46 (1996) Erratum: \[Nucl. Phys. B [**531**]{}, 655 (1998)\] \[hep-ph/9603205\]; A. Djouadi, W. Kilian, M. Muhlleitner and P. M. Zerwas, Eur. Phys. J. C [**10**]{}, 45 (1999) \[hep-ph/9904287\]; X. Li and M. B. Voloshin, Phys. Rev. D [**89**]{}, no. 1, 013012 (2014) \[arXiv:1311.5156 \[hep-ph\]\]. U. Baur, T. Plehn and D. L. Rainwater, Phys. Rev. D [**68**]{}, 033001 (2003) \[hep-ph/0304015\]. U. Baur, T. Plehn and D. L. Rainwater, Phys. Rev. D [**69**]{}, 053004 (2004) \[hep-ph/0310056\]. J. Baglio, A. Djouadi, R. Gröber, M. M. Mühlleitner, J. Quevillon and M. Spira, JHEP [**1304**]{}, 151 (2013) \[arXiv:1212.5581 \[hep-ph\]\]. S. Dawson, S. Dittmaier and M. Spira, Phys. Rev. D [**58**]{}, 115012 (1998) \[hep-ph/9805244\]; J. Grigo, J. Hoff, K. Melnikov and M. Steinhauser, Nucl. Phys. B [**875**]{}, 1 (2013) \[arXiv:1305.7340 \[hep-ph\]\]; F. Maltoni, E. Vryonidou and M. Zaro, JHEP [**1411**]{}, 079 (2014) \[arXiv:1408.6542 \[hep-ph\]\]. D. de Florian and J. Mazzitelli, JHEP [**1509**]{}, 053 (2015) \[arXiv:1505.07122 \[hep-ph\]\]; J. Grigo, J. Hoff and M. Steinhauser, Nucl. Phys. B [**900**]{}, 412 (2015) \[arXiv:1508.00909 \[hep-ph\]\]; S. Borowka, N. Greiner, G. Heinrich, S. P. Jones, M. Kerner, J. Schlenk, U. Schubert and T. Zirke, Phys. Rev. Lett. [**117**]{}, no. 1, 012001 (2016) \[arXiv:1604.06447 \[hep-ph\]\]; D. de Florian, M. Grazzini, C. Hanga, S. Kallweit, J. M. Lindert, P. Maierhöfer, J. Mazzitelli and D. Rathlev, JHEP [**1609**]{}, 151 (2016) \[arXiv:1606.09519 \[hep-ph\]\]. F. Goertz, A. Papaefstathiou, L. L. Yang and J. Zurita, JHEP [**1504**]{}, 167 (2015) \[arXiv:1410.3471 \[hep-ph\]\]; Q. H. Cao, B. Yan, D. M. Zhang and H. Zhang, Phys. Lett. B [**752**]{}, 285 (2016) \[arXiv:1508.06512 \[hep-ph\]\]; M. Gorbahn and U. Haisch, JHEP [**1610**]{}, 094 (2016) \[arXiv:1607.03773 \[hep-ph\]\]; W. Bizon, M. Gorbahn, U. Haisch and G. Zanderighi, JHEP [**1707**]{}, 083 (2017) \[arXiv:1610.05771 \[hep-ph\]\]; G. Degrassi, P. P. Giardino, F. Maltoni and D. Pagani, JHEP [**1612**]{}, 080 (2016) \[arXiv:1607.04251 \[hep-ph\]\]; S. Di Vita, C. Grojean, G. Panico, M. Riembau and T. Vantalon, JHEP [**1709**]{}, 069 (2017) \[arXiv:1704.01953 \[hep-ph\]\]. M. L. Mangano, T. Plehn, P. Reimitz, T. Schell and H. S. Shao, J. Phys. G [**43**]{}, no. 3, 035001 (2016) \[arXiv:1507.08169 \[hep-ph\]\]; R. Contino [*et al.*]{}, CERN Yellow Report, no. 3, 255 (2017) \[arXiv:1606.09408 \[hep-ph\]\]. M. J. Dolan, C. Englert and M. Spannowsky, JHEP [**1210**]{}, 112 (2012) \[arXiv:1206.5001 \[hep-ph\]\]; A. J. Barr, M. J. Dolan, C. Englert and M. Spannowsky, Phys. Lett. B [**728**]{}, 308 (2014) \[arXiv:1309.6318 \[hep-ph\]\]. F. Kling, T. Plehn and P. Schichtel, Phys. Rev. D [**95**]{}, no. 3, 035026 (2017) \[arXiv:1607.07441 \[hep-ph\]\]. V. Barger, L. L. Everett, C. B. Jackson, A. D. Peterson and G. Shaughnessy, Phys. Rev. D [**90**]{}, no. 9, 095006 (2014) \[arXiv:1408.2525 \[hep-ph\]\]; A. Alves, T. Ghosh and K. Sinha, Phys. Rev. D [**96**]{}, no. 3, 035022 (2017) \[arXiv:1704.07395 \[hep-ph\]\]. see [*e.g.*]{}G. Aad [*et al.*]{} \[ATLAS Collaboration\], Phys. Rev. D [**92**]{}, 092004 (2015) \[arXiv:1509.04670 \[hep-ex\]\]; V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Rev. D [**94**]{}, no. 5, 052012 (2016) \[arXiv:1603.06896 \[hep-ex\]\]. A. Papaefstathiou, L. L. Yang and J. Zurita, Phys. Rev. D [**87**]{}, no. 1, 011301 (2013) \[arXiv:1209.1489 \[hep-ph\]\]. D. E. Ferreira de Lima, A. Papaefstathiou and M. Spannowsky, JHEP [**1408**]{}, 030 (2014) \[arXiv:1404.7139 \[hep-ph\]\]; D. Wardrope, E. Jansen, N. Konstantinidis, B. Cooper, R. Falla and N. Norjoharuddeen, Eur. Phys. J. C [**75**]{}, no. 5, 219 (2015) \[arXiv:1410.2794 \[hep-ph\]\]; J. K. Behr, D. Bortoletto, J. A. Frost, N. P. Hartland, C. Issever and J. Rojo, Eur. Phys. J. C [**76**]{}, no. 7, 386 (2016) \[arXiv:1512.08928 \[hep-ph\]\]. Q. Li, Z. Li, Q. S. Yan and X. Zhao, Phys. Rev. D [**92**]{}, no. 1, 014015 (2015) \[arXiv:1503.07611 \[hep-ph\]\]. J. M. Pawlowski and F. Rennecke, Phys. Rev. D [**90**]{}, no. 7, 076002 (2014) \[arXiv:1403.1179 \[hep-ph\]\].
A. Jakovac, I. Kaposvari and A. Patkos, Mod. Phys. Lett. A [**32**]{}, no. 02, 1750011 (2016) \[arXiv:1508.06774 \[hep-th\]\]; J. de Vries, M. Postma, J. van de Vis and G. White, JHEP [**1801**]{}, 089 (2018) \[arXiv:1710.04061 \[hep-ph\]\]. H. Gies, R. Sondenheimer and M. Warschinke, Eur. Phys. J. C [**77**]{}, no. 11, 743 (2017) \[arXiv:1707.04394 \[hep-ph\]\]. A. J. Helmboldt, J. M. Pawlowski and N. Strodthoff, Phys. Rev. D [**91**]{} (2015) no.5, 054010 \[arXiv:1409.8414 \[hep-ph\]\]. G. V. Dunne and M. Unsal, JHEP [**1211**]{}, 170 (2012) \[arXiv:1210.2423 \[hep-th\]\]. D. F. Litim, Phys. Rev. D [**64**]{}, 105007 (2001) \[hep-th/0103195\]. J. A. Adams, J. Berges, S. Bornholdt, F. Freire, N. Tetradis and C. Wetterich, Mod. Phys. Lett. A [**10**]{}, 2367 (1995) \[hep-th/9507093\].
|
{
"pile_set_name": "ArXiv"
}
|
‘=11 makefntext\#1[ to 3.2pt [-.9pt $^{{\ninerm\@thefnmark}}$]{}\#1]{} makefnmark[to 0pt[$^{\@thefnmark}$]{}]{} PS. @myheadings[mkbothgobbletwo oddhead[ ]{} oddfootevenheadevenfoot \#\#1\#\#1]{}
\[appendixc\] \[subappendixc\]
\#1
=1.5pc
citex\[\#1\]\#2[@fileswauxout citeacite[forciteb:=\#2]{}[\#1]{}]{}
@cghi cite\#1\#2[[$\null^{#1}$@tempswa ]{}]{}
=cmbx10 scaled1 =cmr10 scaled1 =cmti10 scaled1 =cmbxti10 scaled=cmbx10 scaled=cmr10 scaled=cmti10 scaled=cmbxti10 =cmbx10 =cmr10 =cmti10 =cmbx9 =cmr9 =cmti9 =cmbx8 =cmr8 =cmti8 ł ø 6.0in 8.6in -0.25truein 0.30truein 0.30truein =1.5pc
**STRINGS IN COSMOLOGICAL SPACETIMES AND THEIR BACK-REACTION**
H. J. DE VEGA
*Laboratoire de Physique Théorique et Hautes Energies,*
*Université Pierre et Marie Curie (Paris VI) et Université Denis Diderot (Paris VII),*
*Tour 16, 1er. étage, 4, Place Jussieu*
*75252 Paris, cedex 05, France. Laboratoire Associé au CNRS URA280.*
Lecture delivered at STRINGS ’95, Future Perspectives in String Theory, USC, Los Angeles, march 13 - 18, 1995.
Strings and Quantum Gravity
===========================
The construction of a sensible quantum theory of gravitation is probably the greatest challenge in today’s theoretical physics. Deep conceptual problems (as the lost of quantum coherence) arise when one tries to combine (second) quantization concepts with General Relativity. That is, it may be very well that a quantum theory of gravitation needs new concepts and ideas. Of course, this future theory must have the today’s General Relativity and Quantum Mechanics (and QFT) as limiting cases. In some sense, what everybody is doing in this domain (including string theories approach) may be to the real theory what the old quantum theory in the 10’s was compared with quantum mechanics [@eri92].
The main drawback to develop a quantum theory of gravitation is clearly the total lack of experimental guides for the theoretical development. The physical effects combining gravitation and quantum mechanics are relevant only at energies of the order of $M_{Planck} = \hbar c / G = 1.22 \, 10^{16} $Tev. Such energies were available in the Universe at times $ t < t_{Planck}
= 5.4 \, 10^{-44} $sec. Anyway, as a question of principle, to construct a quantum theory of gravitation is a problem of fundamental relevance for theoretical physics . In addition, one cannot rule out completely the possibility of some “low energy” ($E \ll M_{Planck}$) physical effect that could be experimentally tested.
Since $M_{Planck}$ is the heaviest possible particle scale, a theory valid there (necessarily involving quantum gravitation) will also be valid at any lower energy scale. One may ignore higher energy phenomena in a low energy theory, but not the opposite. In other words, it will be a ‘theory of everything’. We think that this is the [**key point**]{} on the quantization of gravity. A theory that holds till the Planck scale must describe [**all**]{} what happens at lower energies including all known particle physics as well as what we do not know yet (that is, beyond the standard model). Notice that this conclusion is totally independent of the use of string models. A direct important consequence of this conclusion, is that it does not make physical sense to quantize [**pure gravity**]{}. A physically sensible quantum theory cannot contain only gravitons. To give an example, a theoretical prediction for graviton-graviton scattering at energies of the order of $M_{Planck}$ must include all particles produced in a real experiment. That is, in practice, all existing particles in nature, since gravity couples to all matter[@eri92].
String theory is a serious candidate for a quantum description of gravity since it provides a unified model of all interactions overcoming at the same time the nonrenormalizable character of quantum fields theories of gravity.
As a first step on the understanding of quantum gravitational phenomena in a string framework, we started in 1987 a programme of string quantization on curved spacetimes [@dvs87; @negro].
The investigation of strings in curved spacetimes is currently the best framework to study the physics of gravitation in the context of string theory since it provides essential clues about the physics in this context but is clearly not the end of the story. The next step beyond the investigation of [**test**]{} strings, consist in finding [**self-consistently**]{} the geometry from the strings as matter sources for the Einstein equations or better the string effective equations (beta functions). This goal is achieved in [@cos] for cosmological spacetimes at the classical level. Namely, we used the energy-momentum tensor for a gas of strings as source for the Einstein equations and we solved them self-consistently. \[For more detailed reviews see [@eri92; @eri94]\].
Let us consider bosonic strings (open or closed) propagating in a curved D-dimensional spacetime defined by a metric $G_{AB}(X),
0 \leq A,B \leq D-1$. The action can be written as $$\label{accion}
S = {{1}\o{2 \pi \a'}} \int d\s d\tau \sqrt{g}\, g_{\a\b}(\s,\tau) \;
G_{AB}(X) \,
\partial^{\a}X^A(\s,\tau) \, \partial^{\b}X^B(\s,\tau)$$ Here $ g_{\a\b}(\s,\tau)$ ( $0 \leq \a, \b \leq 1$ ) is the metric in the worldsheet, $\a'$ stands for the string tension. As in flat spacetime, $\a'
\sim (M_{Planck})^{-2} \sim ( l_{Planck})^2$ fixes the scale in the theory.
We will start considering given gravitational backgrounds $G_{AB}(X)$. That is, we start to investigate [*test*]{} strings propagating on a given spacetime. In section 3, the back reaction problem will be studied. That is, how the strings may act as source of the geometry.
String propagation in massless backgrounds other than gravitational (dilaton, antisymmetric tensor) can be investigated analogously.
The string equations of motion and constraints follow by extremizing eq.(\[accion\]) with respect to $X^{A}(\s,\tau)$ and $
g_{\a\b}(\s,\tau)$, respectively. In the conformal gauge, they take the form: $$\begin{aligned}
\label{conouno}
\partial_{-+}X^{A}(\s,\tau) + \Gamma^{A}_{BC}(X)\, \partial_{+}
X^{B}(\s,\tau) \, \partial_{-}X^{C}(\s,\tau) = 0~, \quad
0 \le A \le D-1 ,\end{aligned}$$ $$\begin{aligned}
\label{conodos}
T_{\pm\pm} \equiv G_{AB}(X) \, \partial_{\pm}X^{A}(\s,\tau) \,
\partial_{\pm}X^{B}(\s,\tau)= 0\, , \quad T_{+-} \equiv T_{-+} \equiv 0
\end{aligned}$$ where we introduce light-cone variables $x_{\pm} \equiv \s \pm \tau $ on the world-sheet and where $\Gamma^{A}_{BC}(X)$ stand for the connections (Christoffel symbols) associated to the metric $ G_{AB}(X)$.
Notice that these equations in the conformal gauge are still invariant under the conformal reparametrizations: $$\label{conftr}
\s + \tau \to \s' + \tau' = f(\s+ \tau) \qquad , \qquad
\s - \tau \to \s' - \tau' =g(\s-\tau)$$ Here $f(x)$ and $g(x)$ are arbitrary functions.
The string boundary conditions in curved spacetimes are identical to those in Minkowski spacetime. That is, $$\begin{aligned}
\label{condc}
X^{A}(\s + 2 \pi,\tau) = \, X^{A}(\s,\tau) \quad & {\rm closed
\,strings} \cr \cr
\partial_{\s}X^{A}(0,\tau)=\, \partial_{\s} X^{A}(\pi,\tau) = \,0
\quad & {\rm open \,strings. }
\end{aligned}$$
We shall consider, as usual, that only four space-time dimensions are uncompactified. That is, we shall consider the strings as living on the tensor product of a curved four dimensional space-time with lorentzian signature and a compact space which is there to cancel the anomalies. Therefore, from now on strings will propagate in the curved (physical) four dimensional space-time. However, we will find instructive to study the case where this curved space-time has dimensionality $D$, where $D$ may be 2, 3 or arbitrary.
Strings Falling into Spacetime Singularities: nonlinear plane waves.
=====================================================================
Let us first consider strings propagating in gravitational plane-wave space-times [@ondpl; @ondpl2]. In this geometry the full non-linear string equations (\[conouno\]) and constraints (\[conodos\]) can be exactly solved in closed form. The plane-wave space-times are described by the metric $$\label{opla}
(ds)^2 = - dU dV + \sum_{i=1}^{D-2}(dX^{i})^2 - \left[ W_{1}(U)\,
(X^2 - Y^2 )
+ 2 \,W_{2}(U) \,X Y \right]\, (dU)^2$$ where $ X \equiv X^1$ , $Y \equiv X^2$ . These space-times are exact solutions of the vacuum Einstein equations for any choice of the profile functions $ W_{1}(U)$ and $W_{2}(U)$. In addition they are exact string vacua [@guven]. The case when $W_{2}(U) = 0$ describes waves of constant polarization. When both $ W_{1}(U) \neq 0$ and $W_{2}(U)\neq 0$ , eq.(\[opla\]) describes waves with arbitrary polarization. If $W_{1}(U)$ and/or $W_{2}(U)$ are singular functions, space-time singularities will be present. The singularities will be located on the null plane $U =$ constant. We consider profiles which are nonzero only on a finite interval $-T < U < T$ , and which have power-type singularities [@ondpl; @ondpl2], $$\label{singu}
W_{1}(U) \buildrel{U \to 0}\over= { {\a_{1}} \o {|U|^{\beta_{1}}}}
\quad , \quad
W_{2}(U) \buildrel{U \to 0}\over= { {\a_{2}} \o {|U|^{\beta_{2}}}}$$ The spacetimes (\[opla\]) share many properties with the shockwaves [@camas; @horo]. In particular, $U(\s,\tau)$ obeys the d’Alembert equation and we can choose the light-cone gauge $$\label{gauge}
U = 2 \, \a' p^U \tau \; .$$ The string equations of motion (\[conouno\]) become then in the metric (\[opla\]) : $$\begin{aligned}
\label{equac}
V'' - \ddot V ~ + ~(2 \a' p^U )^2 \left[ \partial_{U} W_{1}\,(X^2 - Y^2 )
~ +~2 \,\partial_{U}W_{2} \,X Y \right]\cr
+ ~8 \a' p^U \left[ W_{1} ( X \dot X - Y \dot Y ) +W_{2} (X \dot Y +
\dot Y X )
\right] = \,0 \cr
X'' - \ddot X ~ +~ (2 \a' p^U )^2 \left[ W_{1} X -W_{2} Y \right] = \,0
\cr Y'' - \ddot Y ~ +~ (2 \a' p^U )^2 \left[ W_{2} X -W_{1} Y \right] = \, 0\end{aligned}$$ and the constraints (\[conodos\]) take the form: $$\label{vinop}
\pm \partial_{\pm}V_{<} = {1 \o {\a' p^{U}}}
\left[ (\partial_{\pm}X)^2 +
(\partial_{\pm}Y)^2 + \sum_{i=3}^{D-2}(\partial_{\pm}X^{i})^2
\right] + \a' p^{U} \left[ W_{1}\, (X^2 - Y^2 )
+ 2 \,W_{2} \,X Y \right]$$ Let us analyze now the solutions of the string equations (\[equac\]) and (\[vinop\]) for a closed string. The transverse coordinates obey the d’Alembert equation, with the solution $$\begin{aligned}
\label{stra}
X^{i}(\s,\tau) = q^i + 2 p^i \a' \tau + i \sqrt{\a'} \sum_{n \neq 0}
\{ \a^{i}_{n} \exp[in(\s - \tau)]
+ \tilde \a^{i}_{n} \exp[-in(\s + \tau )] \}/n , \cr ~ 3 \leq i
\leq D-2 . \quad\end{aligned}$$ For the $X$ and $Y$ components it is convenient to Fourier expand as $$X(\s,\tau) = \sum_{n=-\infty}^{+\infty} \exp(in\s) ~ X_{n}(\tau)\quad , \quad
Y(\s,\tau) = \sum_{n=-\infty}^{+\infty} \exp(in\s) ~ Y_{n}(\tau)$$ Then, eqs.(\[equac\]) for $X$ and $Y$ yield $$\begin{aligned}
\label{schr}
\ddot X_{n} + n^2 X_{n}-(2 \a' p^U )^2
\left[ W_{1} X_{n} - W_{2} Y_{n} \right]= ~ 0 \cr
\ddot Y_{n} + n^2 Y_{n}-(2 \a' p^U )^{2}
\left[ W_{2} X_{n} - W_{1} Y_{n}\right]= ~ 0\end{aligned}$$ where we consistently set $ U = 2 \a' p^U \tau$. Formally, these are two coupled one dimensional Schrödinger-like equations with $\tau$ playing the rôle of a spatial coordinate.
We study now the interaction of the string with the gravitational wave. For $2\a' p^{U} \tau < -T$, $W_{1,2}(\tau) = 0$ and therefore $X, Y$ are given by the usual flat-space expansions $$X(\s,\tau) =~ q^X_{<} + 2 p^X_{<} \a' \tau + i \sqrt{\a'} \sum_{n \neq 0}
\{ \a^{X}_{n<} \exp[-in\tau)]
- \tilde \a^{X}_{-n<} \exp[in\tau )] \} \exp[in\s]/n$$ $$Y(\s,\tau) =~ q^Y_{<} + 2 p^Y_{<} \a' \tau + i \sqrt{\a'} \sum_{n \neq 0}
\{ \a^{Y}_{n<} \exp[-in\tau)]
- \tilde \a^{Y}_{-n<} \exp[in\tau )] \} \exp[in\s]/n
$$ These solutions define the initial conditions for the string propagation in $\tau \geq -\tau_{0} \equiv -{T \o {2\a' p^{U}}}$. In the language of the Schrödinger-like equations we have a two channel potential in the interval $-\tau_{0} < \tau < +\tau_{0}$. We consider the propagation of the string when it approaches the singularity at $U = 0 = \tau$ from $\tau < 0$.
The general case when $W_1 \neq 0 \neq W_2$ is solved in [@ondpl2]. Let us concentrate here on the case $W_2 \equiv 0$, $W_1(U) = \a \left [ |U|^{-\b} - |T|^{-\b} \right]$ for $ |U| < T$.
Eq.(\[schr\]) can be approximated near $\tau = 0^-$ as $$\ddot X_{n}-{{(2 \a' p^U )^{2-\beta}}\o{|\tau|^{\beta}}}
\; \a \; X_{n} = 0$$ $$\ddot Y_{n}+{{(2 \a' p^U )^{2-\beta}}\o{|\tau|^{\beta}}}
\; \a \; Y_{n}= 0$$ The behaviour of the solutions $X_{n}(\tau)$ and $Y_{n}(\tau)$ for $\tau \to 0$ depends crucially on the value range of $\beta$. Namely, i) $\beta > 2$, ii) $\b = 2$, iii) $\beta < 2$.
When $\beta < 2$ the solution for $\tau\to 0^{-}$ behaves as $$X(\s, \tau)\buildrel{\tau\to 0^{-}}\over = B^X(\s) + A^X(\s)~ \tau +
O(|\tau|^{2-\beta}) \; , \;$$ $$Y(\s,\tau)\buildrel{\tau\to 0^{-}}\over = B^Y(\s) + A^Y(\s)~ \tau +
O(|\tau|^{2-\beta})~~,~~\beta \neq 1$$ \[In the special case $\beta = 1$ one should add a term $0 ( \tau
\ln|\tau|)$\]. Here and in what follows, $ B^X(\s), A^X(\s), B^Y(\s)$ and $ A^Y(\s)$ are arbitrary functions depending on the initial data.
For $\beta < 2$, the string coordinates $X, Y$ are always regular indicating that the string propagates smoothly through the gravitational-wave singularity U = 0. (Nevertheless, the velocities $\dot X$ and $\dot Y$ diverge at $\tau = 0$ when $1 \leq
\beta < 2$).
For the case $\beta = 2$ the solution is [@ondpl]: $$\begin{aligned}
\label{beta2}
X(\s,\tau)\buildrel{\tau\to 0^{-}}\over = B^X(\s)
|\tau|^{(1-\sqrt{1+4\a})/2} + A^X(\s)
|\tau|^{(1+\sqrt{1+4\a})/2} \cr
Y(\s,\tau)\buildrel{\tau\to 0^{-}}\over = B^Y(\s)
|\tau|^{(1-\sqrt{1-4\a})/2} + A^Y(\s)
|\tau|^{(1+\sqrt{1-4\a})/2}\end{aligned}$$
Let us now consider the case $\beta > 2$. We have [@ondpl] $$\begin{aligned}
\label{betaG}
X(\s,\tau)&\buildrel{\tau\to 0^{-}}\over=& B^X(\s)
|\tau|^{{\beta}\o 4} ~ \exp[K|\tau|^{1-\beta/2}] +
A^X(\s)
|\tau|^{{\beta}\o 4} ~ \exp[-K|\tau|^{1-\beta/2}] \quad ,
\quad \cr
Y(\s,\tau)&\buildrel{\tau\to 0^{-}}\over=& A^Y(\s)
|\tau|^{{\beta}\o 4} ~ \cos{\left[K |\tau|^{1-\beta/2} + C^Y(\s)
\right]} \; .\end{aligned}$$ where $$K = \,{{(2 \a' p^{U} )^{1-\beta/2}}\o
{\beta/2 - 1}}\sqrt{\a} > 0 \quad , \quad$$
Let us now analyze the string behaviour near the singularity $\tau \to 0^- $ for $\beta \geq 2$. We see that for strong enough singularities ($\b \leq 2$) the transverse coordinate $X$ tends to infinity when the string approaches the singularity $\tau
\to 0, U\to 0$. This means that the string does not cross the gravitational wave, since it does not reach the $U>0$ region. The $Y$ coordinate tends to zero oscillating, when $\tau \to 0$.
The string goes off to $X=\infty$, grazing the singularity plane $U = 0$ (therefore never crossing it). At the same time, the string oscillates in the $Y$ direction, with an amplitude vanishing for $\tau \to 0^- $.
The spatial string coordinates $X^{i}(\s,\tau)~ [3 \leq i \leq D-2]$ behave freely \[eq.(\[stra\])\]. The longitudinal coordinate $V(\s,\tau)$ follows from the constraint eqs.(\[vinop\]) and the solutions (\[beta2\],\[betaG\]) for $X(\s,\tau),
Y(\s,\tau)$ and $X^{j}(\s,\tau)~ [3 \leq j \leq D-2]$. We see that for $\tau \to 0^- $ , $V(\s,\tau)$ diverges as the square of the singular solutions (\[beta2\],\[betaG\]).
Let us consider the spatial length element of the string, i.e. the length at fixed $U = 2 \a'p^{U} \tau $ , between two points $(\s,\tau)$ and $(\s+d\s,\tau)$, $$ds^2 = dX^2 + dY^2 + \sum_{j=3}^{D-2} (dX^{j})^2$$ For $\tau\to 0^{-}$ eqs.(\[beta2\],\[betaG\]) yield $$\begin{aligned}
\label{explo}
ds^2\buildrel{\tau\to 0^{-}}\over =
~ \left[ {B^{X}}(\s)'\right]^{2} d\s^2 ~~ |\tau|^{1-\sqrt{1+4\a}}
\quad {\rm for} ~
\beta=2 \, , \cr
ds^2\buildrel{\tau\to 0^{-}}\over =
~ \left[ {B^{X}}(\s)'\right]^{2}
d\s^2 ~~ |\tau|^{\beta/2} ~ \exp\left[K|\tau|^{1-\beta/2}\right]
\quad {\rm for} ~ \beta \geq 2.\end{aligned}$$ That is, the proper length between $(\s_{0},\tau)$ and $(\s_{1},\tau)$ is given by $$\begin{aligned}
\label{strech}
\Delta s \buildrel{\tau\to 0^{-}}\over =&
[B^X(\s_{1}) - ~B^X(\s_{2})]~ \sqrt{|\tau|}^
{1-\sqrt{1+4\tilde\a}}\to \infty \quad {\rm for} ~ \beta=2 \, , \cr
~\Delta s \buildrel{\tau\to 0^{-}}\over =&
[B^X (\s_{1}) - ~B^X(\s_{2})]~~ |\tau|^{\beta/4} ~
\exp\left[K|\tau|^{1-\beta/2}\right] \to \infty
~{\rm for} ~ \beta \geq 2.\end{aligned}$$ We see that $\Delta s \to \infty$ for $\tau\to 0^{-}$ . That is, the string stretches infinitely when it approaches the singularity plane. This stretching of the string proper size also occurs for $\tau \to 0$ in the inflationary cosmological backgrounds as we shall see below.
Another consequence of eqs. (\[beta2\],\[betaG\]) is that the string reaches infinity in a finite time $\tau$. In particular, for $\s$-independent coefficients, eqs. (\[beta2\],\[betaG\]) describe geodesic trajectories. The fact that for $\beta \geq 2$, a point particle (as well as a string) goes off to infinity in a finite $\tau$ indicates that the space-time is singular.
Finally, we would like to remark that the string evolution near the space-time singularity is a [**collective motion**]{} governed by the nature of the gravitational field. The (initial) state of the string fixes the overall $\s$-dependent coefficients $A^X(\s),B^X(\s),$ $ A^Y(\s), B^Y(\s)$ \[see eqs. (\[beta2\],\[betaG\])\], whereas the $\tau$-dependence is fully determined by the space-time geometry. In other words, the $\tau$-dependence is the same for all modes $n$. In some directions, the string collective propagation turns to be an infinite motion (the escape direction $X$), whereas in the orthogonal direction ($Y$), the motion is oscillatory, but with a fixed ($n$-independent) frequency. In fact, these features are not restricted to singular gravitational waves, but [**are generic**]{} to strings in strong gravitational fields \[see sec.(6) and refs.([@camas; @eri92])\].
For sufficiently weak spacetime singularities ($\beta_{1} < 2$ and $\beta_{2} < 2$), the string crosses the singularity and reaches the region $U > 0$. Therefore, outgoing scattering states and outgoing operators can be defined in the region $U > 0$. We explicitly found in [@ondpl; @ondpl2] the transformation relating the ingoing and outgoing string mode operators. For the particles described by the quantum string states, this relation implies two types of effects as described in [@trans; @eri92] for generic asymptotically flat spacetimes: (i) rotation of spin polarization in the $(X,Y)$ plane, and (ii) [**transmutation**]{} between different particles. We computed in [@ondpl; @ondpl2] the expectation values of the outgoing mass $M_{>}^{2}$ operator and of the mode-number operator $N_{>}$, in the ingoing ground state $|O_{<}>$. As for shockwaves (see [@camas] ) , $M_{>}^{2}$ and $N_{>}$ have different expectation values than $M_{<}^{2}$ and $N_{<}$ . This difference is due to the excitation of the string modes after crossing the space-time singularity. In other words, the string state is not an eigenstate of $M_{>}^{2}$, but an infinity superposition of one-particle states with different masses. This is a consequence of the particle transmutation which allows particle masses different from the initial one.
Strings and Multistrings in Cosmological Spacetimes and the Self-consistent string cosmology
============================================================================================
Recently, several interesting progresses in the understanding of string propagation in cosmological spacetimes have been made [@prd]${}^-$[@ijm]. The classical string equations of motion plus the string constraints were shown to be exactly integrable in D-dimensional de Sitter spacetime, and equivalent to a Toda-type model with a potential unbounded from below. In 2+1 dimensions, the string dynamics in de Sitter spacetime is exactly described by the sinh-Gordon equation.
[**Exact**]{} string solutions were systematically found by soliton methods using the linear system associated to the problem (the so-called dressing method in soliton theory) [@dms; @cdms]. In addition, exact circular string solutions were found in terms of elliptic functions[@dls]. All these solutions describe one string, several strings or even an infinite number of different and independent strings. A single world-sheet simultaneously describes many different strings. This is a new feature appearing as a consequence of the interaction of the strings with the spacetime geometry. Here, interaction among the strings (like splitting and merging) is neglected, the only interaction is with the curved background. Different types of behaviour appear in the multistring solutions. For some of them the energy and proper size are bounded (‘stable strings’) while for many others the energy and size blow up for large radius of the universe ($R \to \infty$, ‘unstable strings’). In addition, such stable and unstable string behaviours are exhibited by the ring solutions found in [@din] for Friedmann-Robertson-Walker (FRW) universes and for power type inflationary backgrounds. In all these works, strings were considered as [*test*]{} objects propagating on the given [*fixed*]{} backgrounds.
We report here the recent results [@cos] further in the investigation of the physical properties of the string solutions above mentioned. We compute the energy-momentum tensor of these strings and we use it to find the back reaction effect on the spacetime. That is, we investigate whether these classical strings can sustain the corresponding cosmological background. This is achieved by considering [**self-consistently**]{}, the strings as matter sources for the Einstein (general relativity) equations (without the dilaton field), as well as for the string effective equations (beta functions) including the dilaton, the dilaton potential and the central charge term.
In spatially homogeneous and isotropic universes, $$ds^2 = (dT)^2 - R(T)^2 \sum_{i=1}^{D-1}(dX^i)^2
\label{metA}$$ the string energy-momentum tensor $T_A^B(X) \; , (A,B=1,\ldots D) $ for our string solutions takes the fluid form, allowing us to define the string pressure $p$ through $\, -\delta_i^k~p = T_i^k \, $ and the string energy density as $ \rho = T_0^0 $. The continuity equation $D^A\;T_A^B = 0 $ takes then the form $$\dot{\rho} + (D-1)\, H\, (p + \rho ) = 0 ,
\label{contA}$$ where $ H \equiv {{d\log R}\over{dT}} $. We consider $D=1+1, D=2+1$, and generic $D-$dimensional universes.
In $1+1$ cosmological spacetimes we find the general solution of the string equations of motion and constraints for arbitrary expansion factor $ R $ . It consists of two families: one depends on two arbitrary functions $ f_{\pm}(\sigma \pm \tau) $ and has [*constant*]{} energy density $ \rho $ and [*negative*]{} pressure $
p = - \rho $. That is, a perfect fluid relation holds $$p = (\gamma - 1 ) \rho
\label{flup}$$ with $\gamma = 0$ in $ D = 1 + 1 $ dimensions. The other family of solutions depends on [*two*]{} arbitrary constants and describes a massless point particle (the string center of mass). This second solution has $ p = \rho = u \; R^{-2} > 0 $ . This is a perfect fluid type relation with $\gamma = 2 $. These behaviours fulfil the continuity equation (\[contA\]) in $ D = 2$.
In $2+1$ dimensions and for any factor $ R $ , we find that circular strings exhibit three different asymptotic behaviours :
- \(i) [**unstable**]{} behaviour for $ R \to \infty $ in inflationary universes (this corresponds to conformal time $ \eta \sim \tau \to \tau_0 $ with finite $ \tau_0 $ and proper string size $ S \sim R \to \infty $), for which the string energy $ E_u
\sim R \to \infty $ and the string pressure $ p_u \simeq -E_u/2 \to -\infty
$ is [**negative**]{} . This behavior dominates for $ R \to \infty $ in inflationary universes.
- \(ii) [**Dual**]{} to unstable behaviour for $ R \to 0 $. This corresponds to $ \eta \sim (\tau - \tau_0)^{-1} \to +\infty $ for finite $ \tau \to \tau_0 $, $ S \sim R \to 0 $ (except for de Sitter spacetime where $ S \to 1/H $ ), for which the string energy $ E_d \sim 1/R \to
\infty $ and the string pressure $ p_d \simeq E/2 \to + \infty $ , is [**positive**]{}.
- \(iii) [**Stable**]{} for $ R \to \infty $, (corresponding to $\eta
\to \infty , \tau\to \infty, S = $ constant ), for which the string energy is $ E_s = $ constant and the string pressure vanishes $ p_s = 0 $ .
Here the indices $(u,d,s)$ stand for ‘unstable’, ‘dual’ and ‘stable’ respectively. The behaviours (i) and (ii) are related by the duality transformation $ R \leftrightarrow 1/R $ , the case (ii) being invariant under duality. In the three cases, we find perfect fluid relations (\[flup\]) with the values of $\gamma$ : $$\gamma_u = 1/2 \quad , \quad \gamma_d = 3/2 \quad , \quad\gamma_s = 1~.
\label{glup}$$ For a perfect gas of strings on a comoving volume $ R^2 $, the energy density $ \rho $ is proportional to $ E / R^2 $, which yields the scaling $ \rho_u = u \, R^{-1} , \rho_d = d \, R^{-3} , \rho_s = s\, R^{-2}$. All densities and pressures obey the continuity equation (\[contA\]) as it must be.
The $ 1+1 $ and $ 2+1 $ string solutions here described exist in any spacetime dimension. Embedded in D-dimensional universes, the $ 1+1 $ and $ 2+1 $ solutions describe straight strings and circular strings, respectively. In D-dimensional spacetime, strings may spread in $ D - 1 $ spatial dimensions. Their treatment has been done asymptotically in [@gsv]. We have three general asymptotic behaviours:
- \(i) [**unstable**]{} for $ R \to \infty $ in inflationary universes with $ \rho_u = u \, R^{2-D} ,\; p_u = -\rho_u/(D-1) < 0 $
- \(ii) [**Dual**]{} to unstable for $ R \to 0 $ with $ \rho_d = d \, R^{-D},\; p_d = \rho_d/(D-1) > 0 $ .
- \(iii) [**Stable**]{} for $ R \to \infty $, with $\rho_s = s \, R^{1-D} , \; p_s = 0 $ .
We find perfect fluid relations with the factors $$\gamma_u = {{D-2}\over {D-1}} \quad , \quad \gamma_d =
{D\over {D-1}} \quad , \quad\gamma_s = 1~.
\label{glud}$$ This reproduces the two dimensional and three dimensional results for $ D = 2 $ and $ D = 3 $, respectively. The stable regime is absent for $ D = 2 $ due to the lack of string transverse modes there.
The dual strings behave as [*radiation*]{} (massless particles) and the stable strings are similar to [*cold matter*]{}. The unstable strings correspond to the critical case of the so called [*coasting universe*]{}. That is, classical strings provide a [*concrete*]{} realization of such cosmological models.
Strings continuously evolve from one type of behaviour to another, as is explicitly shown by our solutions [@prd; @dls]. For intermediate values of $ R $, the string equation of state is clearly more complicated. We propose a formula of the type: $$\rho = \left( u_R \; R + {{d} \over R} + s \right) {1 \over
{R^{D-1}}} ~~~,~~~
p = {1 \over {D-1}} \left( {d \over R} - u_R \; R\right) {1 \over
{R^{D-1}}} \label{rope}$$ where $$\begin{aligned}
\lim_{R\to\infty} u_R = \cases{ 0 \quad & {\rm FRW } \cr
u_{\infty} \neq 0 & {\rm Inflationary } \cr}\end{aligned}$$ is qualitatively correct for all $ R $ and becomes exact for $ R \to 0 $ and $ R \to
\infty $. The parameter $u_R$ varies smoothly with $R$ and tends to the constant $ u_{\infty}$ for $ R \to \infty$.
We stress here that we obtained the string equation of state from the exact string evolution in cosmological spacetimes.
Inserting the equation of state (\[rope\]) in the Einstein-Friedmann equations of general relativity, we obtain a self-consistent solution for $ R $ as a monotonically increasing function of the cosmic time $
T $ $$T = \sqrt{{(D-1)(D-2)}\over 2}~
\int_0^R dR \; {{R^{D/2-1}} ~\over {\sqrt{ u_R \; R^2 + d + s\; R}}}
\label{inte}$$ where we set $R(0) = 0$. This string dominated universe starts at $
T = 0 $ with a radiation dominated regime $
R(T) \buildrel{T \to 0}\over \simeq C_D \, (T)^{2 \over D}$, then the universe expands for large $ T $ as $ R(T) \buildrel{T \to \infty}\over \simeq C'_D ~
(T)^{2 \over {D-1}}$, as (cold) matter dominated universes. For example, at $ D = 4 ,\; R $ grows as $ \, R \sim (T)^{2 \over 3}$.
It must be noticed that an universe dominated by unstable strings ($u_R$) would yield $ R(T) \buildrel{T \to \infty}\over \simeq C'_D ~
(T)^{2 \over {D-2}}$, which is faster than (cold) matter dominated universes. However, this is not a self-consistent solution of the Einstein-Friedmann equations plus the string equations of motion, as shown in [@cos].
The unstable string solutions are called in this way since their energy and invariant length grow as $R$ for large $R$. However, it must be clear that as [*classical*]{} string solutions they [**never decay**]{}.
Our self-consistent solution $R(T)$ yields the realistic matter behaviour $R(T) \buildrel{T \to \infty}\over \sim (T)^{2 \over
{D-1}}$ .
The [*stable*]{} strings (which behave as cold matter) are those dominating for $R \to \infty$. The ‘dual’ strings give $R(T) \buildrel{T \to 0}\over \simeq C_D \, (T)^{2 \over D}$, the radiation type behaviour. For intermediate $R$, the three types of string behaviours (unstable, dual and stable) are present. Their cosmological implications as well as those associated with string decay deserve investigation. For a thermodynamical gas of strings the temperature $T$ as a function of $R$, scales as $1/R$ for small $R$ (the usual radiation behaviour).
For the sake of completeness we analyze the effective string equations in [@cos]. These equations have been extensively treated in the literature and they are not our central aim.
It must be noticed that there is no satisfactory derivation of inflation in the context of the effective string equations. This does not mean that string theory is not compatible with inflation, but that the effective string action approach [*is not enough*]{} to describe inflation. The effective string equations are a low energy field theory approximation to string theory containing only the [*massless*]{} string modes. The vacuum energy scales to start inflation are typically of the order of the Planck mass where the effective string action approximation breaks down. One must also consider the [*massive*]{} string modes (which are absent from the effective string action) in order to properly get the cosmological condensate yielding inflation. De Sitter inflation does not emerge as a solution of the the effective string equations.
Conclusions
===========
Strings in curved spacetimes show a rich variety of new behaviours unknown in flat spacetimes. The most spectacular effect is clearly given by the unstable strings with size and energy tending to infinity. This phenomenon appears both for singular plane wave spacetimes and for non-singular de Sitter spacetimes and for all other inflationary universes. We think that it is a generic feature for strings in strong gravitational fields.
Comparison of the energy-momentum behaviours in inflationary universes and singular plane-waves show interesting differences. We find for the fastest growing energy-momentum components in singular plane-waves ($\b \geq 2$) [@ondpl]: $$\begin{aligned}
T^{VV}(\s,\tau) \buildrel{\tau \to 0}\over= C_V \, [\xi(\tau)]^2
\int_0^{2\pi} [B^X(\s)]^2 \to \infty \cr
T^{VX}(\s,\tau) \buildrel{\tau \to 0}\over= C_X \, \xi(\tau)
\int_0^{2\pi} B^X(\s) \to \infty\end{aligned}$$ where $$\xi(\tau) \equiv |\tau|^{-\b/4}\; \exp[K|\tau|^{1-\b/2}] \to \infty$$ and $ C_V $ and $ C_X $ are constants.
A typical unstable string behaviour on an inflationary spacetime with scale factor $R(T) = a \; T^{{k \o {k+2}}} \; (k < 0)$ is as follows [@eri94]. $$\begin{aligned}
E \buildrel{\tau \to 0}\over= {C \o {\a'}} \; \tau^{k/2} = {R \o
{\a'}}\to +\infty \cr
P \buildrel{\tau \to 0}\over= -E/2 \to -\infty \cr
S \buildrel{\tau \to 0}\over=\tau^{k/2}\to +\infty\end{aligned}$$ where $C$ is a constant and we considered a ring solution for simplicity.
The main difference is that in singular plane wave spacetimes only the null energy (conjugated to the null variable $V$) diverges. In the inflationary case, both the energy ($T^0_0$) and the pressure ($-T^i_i$, not summed) diverge and they blow up at the same rate. Obviously the kind of spacetime singularity is very different.
[11]{} H.J. de Vega and N. Sánchez in “String Quantum Gravity and the Physics at the Planck Scale”, Proceedings of the Erice Workshop held in June 1992. Edited by N. Sánchez, World Scientific, 1993. Pages 73-185, and references given therein.
H.J. de Vega and N. Sánchez in “Current Topics in Astrofundamental Physics”, Proceedings of the Chalonge School, Erice 4-16 September 1994, Proceedings edited by N. Sánchez and A. Zichichi.
H. J. de Vega and N. Sánchez, Phys. Lett. [**B 197**]{}, 320 (1987).
H. J. de Vega and N. Sánchez, Nucl. Phys. [**B 309**]{}, 552 and 577 (1988).
H. J. de Vega and N. Sánchez, Phys. Rev. [**D50**]{}, 7202 (1994).
H. J. de Vega and N. Sánchez, Phys. Rev. [**D 45**]{} , 2783 (1992).
H. J. de Vega, M. Ramón Medrano and N. Sánchez,
Class. and Quantum Grav. [**10**]{}, 2007 (1993).
H. J. de Vega and N. Sánchez, Nucl. Phys. [**B 317**]{}, 706 and 731 (1989).
H. J. de Vega and N. Sánchez, Phys. Rev. Lett. (C), [**65**]{}, 1517 (1990).
G. Horowitz and A.R. Steif, Phys. Rev. Lett. [**64**]{}, 260 (1990)
and Phys. Rev. [**D 42**]{} , 1950 (1990).
R. Güven, Phys. Lett. [**B 191**]{}, 275 (1987).
D. Amati and K. Klimĉik, Phys. Lett. [**B 210**]{} , 92 (1988).
H. J. de Vega, M. Ramón Medrano and N. Sánchez,
Nucl. Phys. [**B 351**]{}, 277 (1991), Nucl. Phys. [**B 374**]{}, 425 (1992)
and Phys. Lett. [**B 285**]{}, 206 (1992).
H. J. de Vega and N. Sánchez, Phys. Rev. [**D47**]{}, 3394 (1993).
H. J. de Vega, A. V. Mikhailov and N. Sánchez,
Teor. Mat. Fiz. [**94**]{} (1993) 232.
F. Combes, H. J. de Vega, A. V. Mikhailov and N. Sánchez,
Phys. Rev. [**D50**]{}, 2754 (1994).
H. J. de Vega, A. L. Larsen and N. Sánchez, Nucl. Phys. [**B 427**]{}, 643 (1994).
I. Krichever, Funct. Anal. and Appl. [**28**]{}, 21 (1994),
\[Funkts. Anal. Prilozhen. [**28**]{}, 26 (1994)\].
H. J. de Vega and I. L. Egusquiza, Phys. Rev. [**D 49**]{}, 763 (1994).
A. L. Larsen and N. Sánchez, Phys. Rev. [**D 50**]{}, 7493 (1994).
N. Sánchez and G. Veneziano, Nucl. Phys. [**B 333**]{}, 253 (1990).
M. Gasperini, N. Sánchez and G. Veneziano,
Int. J. Mod. Phys. [**A 6**]{}, 3853 (1991) and Nucl. Phys. [**B364**]{}, 365 (1991).
H. J. de Vega and N. Sánchez, Int. J. Mod. Phys. [**A 7**]{}, 3043 (1992).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'With the rapid growth of online fashion market, demand for effective fashion recommendation systems has never been greater. In fashion recommendation, the ability to find items that goes well with a few other items based on style is more important than picking a single item based on the user’s entire purchase history. Since the same user may have purchased dress suits in one month and casual denims in another, it is impossible to learn the latent style features of those items using only the user ratings. If we were able to represent the style features of fashion items in a reasonable way, we will be able to recommend new items that conform to some small subset of pre-purchased items that make up a coherent style set. We propose *Style2Vec*, a vector representation model for fashion items. Based on the intuition of distributional semantics used in word embeddings, Style2Vec learns the representation of a fashion item using other items in matching outfits as context. Two different convolutional neural networks are trained to maximize the probability of item co-occurrences. For evaluation, a fashion analogy test is conducted to show that the resulting representation connotes diverse fashion related semantics like shapes, colors, patterns and even latent styles. We also perform style classification using Style2Vec features and show that our method outperforms other baselines.'
author:
- |
Hanbit Lee, Jinseok Seol, Sang-goo Lee\
Department of Computer Science and Engineering\
Seoul National University\
[{skcheon, jamie, sglee}@europa.snu.ac.kr]{}
bibliography:
- 'egbib.bib'
title: 'Style2Vec: Representation Learning for Fashion Items from Style Sets'
---
Introduction
============
[0.32]{} ![Examples of style sets in (a) formal, (b) punk and (c) business types. A style set is collage of coherent items from distinct categories.\[fig:example\]](images/formal.jpg "fig:"){width="95.00000%"}
[0.32]{} ![Examples of style sets in (a) formal, (b) punk and (c) business types. A style set is collage of coherent items from distinct categories.\[fig:example\]](images/punk.jpg "fig:"){width="95.00000%"}
[0.32]{} ![Examples of style sets in (a) formal, (b) punk and (c) business types. A style set is collage of coherent items from distinct categories.\[fig:example\]](images/business.jpg "fig:"){width="95.00000%"}
In fashion recommendation, the ability to find items that *goes well* with a few of other items based on style is more important than recommending a single item without context. Traditional recommendation approaches use user purchase data (generally referred to as *user ratings*) to learn the latent features of item sets. However, since the same user may have purchased both dress suits and casual denims (within a certain time span), it is impossible to learn specific features of individual styles using such user ratings. Therefore, the ability to identify detailed style features of an item is quite challenging. A straightforward approach is to use item categories and hand-crafted attributes to represent an item. However, it is almost impossible to define a set of fine-grained attributes exquisite enough to characterize the subtle nuisance of fashion. Recent approaches use visual images of items to obtain more rich latent item features. However, individual fashion features cannot effectively explain whether one item matches another or a set of items make up a stylish outfit.
In this paper, we propose Style2Vec, a distributed representation of fashion items that is learned from a large dataset of user created “style sets”. We define a *style set* to be a collection of garments and accessories that make up a single outfit. Figure \[fig:example\] shows three examples of style sets in different types. As we can see, a style set is composed of items belong to different categories and our network attempts to learn latent style shared by items in distinct shapes.
To devise a model that can learn styles from set of item images, we borrow the intuition of distributional semantics from the natural language processing community. Word2Vec [@mikolov2013distributed] and other distributed word representation models learn the representation by training shallow neural network to maximize the probability of word co-occurrences within a context window (or sentence). For Style2Vec, a word is a single fashion item, (e.g., a jacket, a shirt, a hat, etc.) while a sentence corresponds to a style set. The key idea of our approach is to use the member items of a style set as the context items. While Word2Vec model directly updates word embedding vectors, but for fashion items, it is impossible to learn embedding vectors directly since each item appears in small number of style sets. To overcome such context sparseness, we use convolutional neural network (CNN) a as projection network from image to embedding vector since CNN clumps similar images. This utilizes even rare items as meaningful vectors, which was difficult in traditional Word2Vec models.
Total 297,083 user created style sets have been collected from a popular fashion website as our corpus for training and testing. We use two different convolutional neural networks to map an image to its latent features, one for the input item and the other for the context items. The former network, after training, is used as the embedding network for extracting the latent style feature vector for each item.
For evaluation, a fashion analogy test is conducted to show that the resulting representation connotes diverse fashion related semantics like shapes, colors, patterns and even latent styles. We also perform style classification using Style2Vec and show that our method outperforms other baselines.
Related Work
============
There are several works trying to uncover visual relationship between pair of compatible fashion items. McAuley et al. [@mcauley2015image] use parameterized distance metric (i.e. Mahalanobis distance) to learn relationships between co-purchased item pairs. He et al. [@he2016learning] extends [@mcauley2015image] to allow multiple notions of ‘relatedness’ with ensemble of $k$ Mahalanobis transforms. Both of these approaches only use image features extracted from the network that is trained for image classification of general images. So, the item features they use are not fashion-centric nor context-sensitive.
The closest work to ours is the work by Veit et al. [@veit2015learning]. They use Siamese convolutional neural network to learn dyadic item co-occurrences. Since they train network to minimize the $l_2$ difference between co-purchased pair of item features, it is impossible for them to learn features from set of arbitrary number of items. In addition, they use co-purchase dataset from Amazon[^1] to learn latent features of fashion items. However, as mentioned before, we claim that co-purchase data is not suitable for identifying specific style relationships among well-matched style set.
[1.0-3em]{} ![Model architecture of Style2Vec. (a) Structure of VGGNet [@simonyan2014very] that is used as input and context projection networks. (b) Process of computing the loss function given a style set. CNN colored in gray is VGGNet for the input item, and the yellow one is for the context items.\[fig:model\]](images/vgg.jpg "fig:"){width="95.00000%"}
[1.0-3em]{} ![Model architecture of Style2Vec. (a) Structure of VGGNet [@simonyan2014very] that is used as input and context projection networks. (b) Process of computing the loss function given a style set. CNN colored in gray is VGGNet for the input item, and the yellow one is for the context items.\[fig:model\]](images/model.jpg "fig:"){width="95.00000%"}
Another research stream we need to consider is unsupervised representation learning from images. In this area, generative adversarial networks [@goodfellow2014generative] are receiving enormous attention as they can generate highly realistic images from latent vectors and they have been shown to learn good feature representation from images [@donahue2016adversarial; @dumoulin2016adversarially; @radford2015unsupervised]. The most popular architecture is DCGAN [@radford2015unsupervised], which uses transposed convolutional network for generator and convolutional network for discriminator that are trained in adversarial fashion. However, these approaches learn representation from standalone image, not in the context of relationship among items in a set.
Meanwhile, recently, noteworthy fashion recommender systems are being proposed, which try to recommend items that go well with other items. Lee et al. [@lee2015style] use heterogeneous graph to link fashion items those make up a stylish outfit, as well as to link items to their attributes. Hu et al. [@hu2015collaborative] propose a tensor factorization approach to recommend a set of fashion items. They do not learn item features based on sets, but use discrete item attributes or low level image features.
In summary, our approach is distinct from the above in that we aim to learn latent styles of fashion items from style sets consist of multiple items. Our approach captures fashion semantics on the style space we learn, which can be effectively utilized by fashion recommender systems.
Style2Vec model
===============
Our goal is to implement a representation learning framework for fashion items where learned representation contains latent styles shared by items in a style set. The latent styles of an item are highly determined by the other items in the same style set as well as by features of item itself. In order to utilize features of the context items, we borrow the concept of distributional hypothesis from linguistics, meaning that the words used and occur in the similar contexts tend to purport similar meanings. We assume that fashion items in the similar style set share a coherent style, as words in the similar contexts have similar meanings.
Many researches in natural language processing have tried to effectively represent semantic of a word using the words in the same context under this hypothesis. Most successful model is Word2Vec model [@mikolov2013linguistic] which learns a simple neural network to reconstruct linguistic contexts of words. The network has one hidden layer and there are two types of parameters, one for projecting input word to its latent vector on the hidden layer and another for multiplying weight vector to the latent vector so as to predict distribution of context words.
Inspired by Word2Vec model, we devise our model *Style2Vec* to use two types of neural network, one for projecting a target fashion item to its representation vector and another for mapping the other context items in the same style set to their weight vectors. The inner product between target item vector and each context weight vector is made to probabilities with softmax function and entire networks are trained to maximize the sum of the log probabilities. Figure \[fig:model\] shows our entire model architecture.
Formally, let the image of an item $i$ be $I_i$. We parameterize two types of networks $f_{\text{in}} : I_i \mapsto u_i$ for the input item image $I_i$ and $f_{\text{cxt}} : I_c \mapsto v_c$ for the context item image $I_c$, both from the image space $\mathcal{I}$ to a latent feature space $\mathbb{R}^d$. We use the same VGGNet [@simonyan2014very] structure for both networks, except that they do not share parameters since those two networks have different purposes. Our goal is to learn $f_{\text{in}}$ and $f_{\text{cxt}}$ that maximize the log probability of the item co-occurrence in a style set.
Let $F$ be a set of all fashion items. We assume that each of *style set* $S$ in our training dataset $D \subset \mathcal{P}(F)$ has coherent style. We iterate $i$ over $S$ as an input, with items in $S \setminus \{i\}$ as a context. The overall objective of Style2Vec model is to maximize the average log probability for each $S$ in $D$:
$$\frac{1}{|S|} \sum_{i \in S} \sum_{c \in S \setminus \{i\}} \log \left( \frac{\exp(u_i^\intercal v_c)}{\sum_{j \in F} \exp(u_i^\intercal v_j)} \right)$$
where $u_i = f_{\text{in}}(I_i)$, $v_c = f_{\text{ctx}}(I_c)$ and $v_j = f_{\text{ctx}}(I_j)$.
Since computing the softmax is very expensive, we apply negative sampling approach [@mikolov2013distributed]. We randomly pick $k$ items, different from the items in the training sample $S$ to form a negative sample set $N_{S, i, c} \subset F \setminus S$ for each $(i, c)$. Now our objective becomes
$$\frac{1}{|S|} \sum_{i \in S} \sum_{c \in S \setminus \{i\}} \left( \log \sigma (u_i^\intercal v_c) + \sum_{j \in N_{S, i, c}} \log \sigma (-u_i^\intercal v_j) \right)$$
for each $S$ in $D$.
{width="\textwidth"}
It is noteworthy that Style2Vec is capable of learning any size of a style set, since we use shared parameters for projecting all the context item images to their context features.
Experiment
==========
To evaluate the item representation of Style2Vec, we conducted several tasks using the item vectors learned by Style2Vec. Visualization of Style2Vec feature space shows how items with similar style gather together. Also we investigate the characteristics of item representation by performing a fashion analogy test. Experiment of style type classification demonstrates that our item vectors can be effectively used for more general fashion related tasks.
Dataset and training
--------------------
We collected 297,083 sets from popular fashion web service Polyvore[^2]. Each of set is composed with 2 to 4 categorically distinct items from a pool of 53,460 tops, 43,180 bottoms, 31,199 outers, 77,981 pairs of shoes, and 30,852 dresses, 236,672 unique items in total. We use VGGNet with 16 layers including 4 max pooling layers and one average pooling layer, which maps an image to 1,024 dimensional vector. The whole network is trained using mini-batch gradient descent through Adam optimizer [@kingma2014adam] with batch normalization [@ioffe2015batch].
Visualization of style space
----------------------------
The t-SNE [@maaten2008visualizing] algorithm is used to embed 1,024 dimensional item features into a 2D space. Figure \[fig:tsne\] shows visualized style space of our model. Since the convolutional neural network tend to map visually similar items to close points in the latent space, we can observe that the items of similar shape and color (e.g., black mini-skirt) are projected nearby. However, even when images are not visually similar, they might lie near together if they have been matched with similar context items. We can observe that the pastel tone items those go well with each other are located close to each other (upper right box). Also, sneakers and casual flat shoes gather together, since those items make good matches with other casual items (lower right box). It is remarkable that formal dress and heels are clustered, as well as the punky items of distinct categories are located nearby (upper left box & lower left box). This visual map gives us a glimpse of our model and justifies that our model transforms the images into a reasonable style space.
[1-2em]{} ![Results from analogy tests. Style factors can be calculated using basic arithmetics. Each test checks the transfer of following fashion factor: (a) color (b) pattern (c) adding punk style (d) removing punk style (e) adding formal style (f) adding punk style.\[fig:analogy\]](images/a.jpg "fig:"){width="\textwidth"}
[1-2em]{} ![Results from analogy tests. Style factors can be calculated using basic arithmetics. Each test checks the transfer of following fashion factor: (a) color (b) pattern (c) adding punk style (d) removing punk style (e) adding formal style (f) adding punk style.\[fig:analogy\]](images/b.jpg "fig:"){width="\textwidth"}
[1-2em]{} ![Results from analogy tests. Style factors can be calculated using basic arithmetics. Each test checks the transfer of following fashion factor: (a) color (b) pattern (c) adding punk style (d) removing punk style (e) adding formal style (f) adding punk style.\[fig:analogy\]](images/c.jpg "fig:"){width="\textwidth"}
[1-2em]{} ![Results from analogy tests. Style factors can be calculated using basic arithmetics. Each test checks the transfer of following fashion factor: (a) color (b) pattern (c) adding punk style (d) removing punk style (e) adding formal style (f) adding punk style.\[fig:analogy\]](images/d.jpg "fig:"){width="\textwidth"}
[1-2em]{} ![Results from analogy tests. Style factors can be calculated using basic arithmetics. Each test checks the transfer of following fashion factor: (a) color (b) pattern (c) adding punk style (d) removing punk style (e) adding formal style (f) adding punk style.\[fig:analogy\]](images/e.jpg "fig:"){width="\textwidth"}
[1-2em]{} ![Results from analogy tests. Style factors can be calculated using basic arithmetics. Each test checks the transfer of following fashion factor: (a) color (b) pattern (c) adding punk style (d) removing punk style (e) adding formal style (f) adding punk style.\[fig:analogy\]](images/f.jpg "fig:"){width="\textwidth"}
Analogy test
------------
In order to investigate the inherent characteristics of latent style features, we performed a fashion analogy test. Each test question has a form of “$x$ is to $y$ as $z$ is to $\rule{0.2cm}{0.15mm}$” which asks the system to find an item that fills the blank, in the context of fashion semantics. For example, a question “punk boots is to punk jacket, as formal heels is to $\rule{0.2cm}{0.15mm}$” (answer might be a formal blazer) can test whether there exist some regularities related to latent styles. We manually created 288 test questions. Specifically, we selected two items $x$ and $y$ from a style set, where $x$ and $y$ belong to different categories but sharing a concrete style (e.g., punk boots and punk jacket). For $z$, we choose a random item that belongs to same category as $x$ (e.g., formal heels). To fill the blank, we first obtain the item feature vectors of $x$, $y$, and $z$ with our trained model. Then we find the answer item by calculating the closest item vector to the vector $u_y - u_x + u_z$, where $u_x$, $u_y$, $u_z$ are feature vectors of item $x$, $y$ and $z$ [@mikolov2013efficient]. The answer item found by the system is evaluated by two human evaluators only if the answer item belongs to the same category as $y$. If the category of the answer item is different from that of $y$, then we mark that question as failed one. Furthermore, we mark as *acceptable* only if both of two human evaluators agree together that $y$ and the answer item go well enough to form a style set. Total 199 out of 288 questions are marked as acceptable, which is about 69.1%. This result quantitatively supports that our style features possess latent styles shared within a style set.
Among the analogy tests, we could find remarkable results which show that the feature vectors of our model possess basic characteristics such as color, pattern, shape as well as latent styles. Moreover, those latent factors can be transfered via simple vector arithmetic. Figure \[fig:analogy\] shows some of the results where the first three items of each row correspond to items $y$, $x$, $z$ in order, and the rest item in the same row corresponds to the answer item found by the system. Cases of (a) and (b) presents test results related to color and pattern factor. In case of (a), $x$ and $y$ are items with same category and detailed shape, but only different in color (briefly, a black mini skirt and a pink mini skirt), while $z$ is an item in a different category but in same color as $x$ (which is a black blazer). The model finds the pink blazer, which implies that the color factor is transferred via simple vector arithmetic. Similar regularity was found in case of pattern factor from (b). Cases of (c) and (d) show style transfer between punk style and casual style. In case of (e), we can observe transfer of formal style from formal dress to high-heels via casual dress and flat shoes. Moreover, we could find blue high heels those match well with given blue formal dress, as if it were a recommendation.
The result of analogy tests implies that our embedding model can learn fashion features of color, shape, pattern, and even latent styles, which are considered as the most ambiguous property in the fashion industry. We claim that our latent style features are derived from process of learning items in a style set simultaneously.
Set size Casual Punk Formal Business Total
---------- -------- ------- -------- ---------- --------
2 978 943 773 1,000 3,694
3 1,000 1,000 1,000 961 3,961
4 759 886 1,000 311 2,956
total 2,737 2,829 2,773 2,272 10,611
: Distribution of style set data used for style classification. For fare comparison, we collected sets of four style types uniformly.\[tab:data\]
Style classification for style sets
-----------------------------------
We try to classify style sets into explicit style types using item representation learned from large dataset of style sets. Style features of a style set is obtained by averaging the style vectors of items in the set. The style vector of each fashion item is achieved from three different models: Siamese CNN [@veit2015learning], DCGAN [@radford2015unsupervised], and Style2Vec.
All the models are trained on the same dataset with controlled number of parameters. However, in case of Siamese CNN, we transformed dataset by getting pairwise combinations from all the style sets, since Siamese CNN can be trained only with pairs of items. To generate negative samples, we randomly selected items from different categories to make item pairs. We also made another version of Style2Vec model by training the network on this pairwise dataset to investigate the difference between learning pairwise relationship and relationship among sets with more than two items. For DCGAN, the features are extracted from penultimate layer of discriminator CNN. So eventually item representation from four models are evaluated: Siamese CNN, DCGAN, Styl2Vec(Pair) and Style2Vec.
Since the original dataset has no style type labels on the style sets, we made a dataset for style classification task. We first choose four distinguishable style types: casual, punk, formal, and business. Then we collected style sets with specific style type by querying each style keyword to Polyvore site. Data distribution of style sets for each style type are shown in Table \[tab:data\].
Overall process is as follows: four models are trained on the original large dataset and the style features of style sets in the new dataset are obtained from the trained models. 90% of the new dataset is used for training the multi-layer perceptron classifier which classify style sets into four style types and the rest is used as a test set.
Model Accuracy
--------------------- ------------
Siamese CNN 51.14%
DCGAN 54.33%
Style2Vec(pairwise) 54.99%
**Style2Vec** **61.13**%
: Style classification accuracy of Siamese CNN, DCGAN, Style2Vec(Pairwise), and Style2Vec. Style2Vec(Pairwise) is a version which is trained on pairwise dataset. Multi-layer perceptron classifier is used to classify style sets to their style types. \[tab:style-clf\]
Table \[tab:style-clf\] shows the accuracy of each model. As we can see, Style2Vec outperforms all the other models. Style2Vec performs better than DCGAN even if DCGAN is the state-of-the-art representation learning network, because DCGAN learns item representation from standalone image without considering style factors shared within a style set. Siamese CNN achieves even lower accuracy than DCGAN although Siamese CNN learns relationship of compatible item pairs. We believe there are two reasons. One reason is that, with only pairwise relationship, it is insufficient to learn the shared style type among items in a set. Another is the objective function that Siamese CNN uses is not effective enough. It is clear when we compare the results of Siamese CNN, Style2Vec trained on pairwise dataset and Style2Vec trained on original dataset. Style2Vec with pairwise dataset performs worse than the original version, which shows that it is crucial to learn relationship within a set of more than two items simultaneously. Style2Vec achieves higher accuracy than Siamese CNN even if they both use the same pairwise dataset. It shows that the objective function and the structure of Style2Vec is more effective than those of Siamese CNN.
Conclusion
==========
In this work, we propose a novel representation learning framework, Style2Vec that can learn latent style features from large collection of user created style sets. Visualization and fashion analogy tests reveal that these style features meaningfully possess diverse semantics of shape, color, pattern and latent style. Moreover, we apply our style features to style classification task and show our approach outperforms the-state-of-the-art baselines. We believe this is a significant step towards utilizing latent style features of fashion items in various tasks including visual search, article tagging, and recommendation. As a future work, we plan to implement recommender system that recommends fashion items those go well with items in user’s closet using Style2Vec item representation. We also plan to further investigate and interpret the relationships between items in a style set.
[^1]: https://www.amazon.com/
[^2]: http://www.polyvore.com/
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We review the current state of results about the half-wave maps equation on the domain ${\mathbb{R}}^d$ with target ${\mathbb{S}}^2$. In particular, we focus on the energy-critical case $d=1$, where we discuss the classification of traveling solitary waves and a Lax pair structure together with its implications (e.g. invariance of rational solutions and infinitely many conservation laws on a scale of homogeneous Besov spaces). Furthermore, we also comment on the one-dimensional space-periodic case. Finally, we list some open problem for future research.'
address: |
University of Basel\
Department of Mathematics\
Spiegelgasse 1, CH-4051 Basel, Switzerland
---
Introduction
============
This expository note is intended to give an overview on results about the **half-wave maps equation** posed in ${\mathbb{R}}^d$ with target ${\mathbb{S}}^2$ (embedded in ${\mathbb{R}}^3$). The corresponding equation reads $$\tag{HWM}
{\partial}_t {\mathbf{u}}= {\mathbf{u}}\times {|\nabla|}{\mathbf{u}}$$ for the function ${\mathbf{u}}: [0,T) \times {\mathbb{R}}^d \to {\mathbb{S}}^2$. Here $\times$ denotes the standard vector product in ${\mathbb{R}}^3$ and the operator $|\nabla|$ is defined via the Fourier transform as $\widehat{(|\nabla| {\mathbf{u}}})(\xi) = |\xi| \widehat{{\mathbf{u}}}(\xi)$. Below we also address the one-dimensional periodic setting when ${\mathbb{R}}$ is replaced by ${\mathbb{T}}$.
In what follows, we will mainly deal with $d=1$ space dimensions. As a matter of fact, this case displays a very rich list of interesting analytical phenomena such as [*energy-criticality, conformal invariance,*]{} and a [*Lax pair structure*]{}. We try to highlight the relevant mathematical results that have been found so far, followed by an outline of future open problems.
Let us also mention that, from a physical perspective, the one-dimensional half-wave maps equation is also of significant interest, since it (formally) arises as a semi-classical and continuum limit of [*Haldane-Shastry (HS) spin chains*]{} and [*classical spin systems of Calogero-Moser (CM) type*]{}. Models of (HS) and (CM) are exactly solvable and completely integrable quantum and classical systems, respectively. For more details on this, we refer the reader to the appendix of [@LeSc-18].
Some Basics Facts
=================
We start with collecting some fundamental properties of the geometric evolution equation given by (HWM).
Conservation Laws and Hamiltonian Structure
-------------------------------------------
A moment’s reflection shows that (HWM) exhibits (formal) conservation of the energy $$E[{\mathbf{u}}] = \frac 1 2 \int_{{\mathbb{R}}^d} {\mathbf{u}}\cdot |\nabla| {\mathbf{u}}\, dx = c_d \iint_{{\mathbb{R}}^d \times {\mathbb{R}}^d} \frac{|{\mathbf{u}}(x)-{\mathbf{u}}(y)|^2}{|x-y|^2} \, dx \, dy,$$ with some constant $c_d > 0$ (where the last identity is a classical fact from harmonic analysis). From this observation, we see that the corresponding energy space is found to be the homogeneous Sobolev space $$\dot{H}^{\frac 1 2}({\mathbb{R}}^d; {\mathbb{S}}^2) = \{ {\mathbf{u}}: {\mathbb{R}}^d \to {\mathbb{R}}^3 \mid \mbox{$E[{\mathbf{u}}] < +\infty$ and $|{\mathbf{u}}| = 1$ a.\,e.} \}.$$ It is not hard to see that (HWM) can indeed be written as a Hamiltonian equation of motion ${\partial}_t {\mathbf{u}}= \{ {\mathbf{u}}, E \}$ with the canonical Poisson bracket for ${\mathbb{S}}^2$-valued functions defined as $$\{ u_i(x), u_j(y) \} = {\varepsilon}_{ijk} u_k(x) \delta(x-y),$$ where ${\varepsilon}_{ijk}$ is the anti-symmetric Levi-Cività symbol. The energy space $\dot{H}^{\frac 1 2}({\mathbb{R}}^d; {\mathbb{S}}^2)$ can then be seen as a phase space for the infinite-dimensional Hamiltonian system whose equation of motions are given by the half-wave maps equation.
Another type of (formally) conserved quantities for (HWM) are found to be $$\mathbf{S}[{\mathbf{u}}] = \int_{{\mathbb{R}}^d} ({\mathbf{u}}- \mathbf{P}) \, dx, \quad M[{\mathbf{u}}] = \int_{{\mathbb{R}}^d} |{\mathbf{u}}- \mathbf{P}|^2 \, dx,$$ where we require that ${\mathbf{u}}- \mathbf{P} \in L^1({\mathbb{R}}^d)$ or ${\mathbf{u}}- \mathbf{P} \in L^2({\mathbb{R}}^d)$, respectively, for some constant point $\mathbf{P}$ on ${\mathbb{S}}^2$. In physical terms, the quantity $\mathbf{S}[{\mathbf{u}}]$ is the total spin of the system represented by the field ${\mathbf{u}}$. Of course, the rotational symmetry on the domain ${\mathbb{R}}^d$ (i.e. the continuous action of the group $\mathrm{SO}(d)$ on the domain ${\mathbb{R}}^d$ for $d \geq 2$) also induces formal conservation laws by Noether’s theorem corresponding to the conservation of [*angular momentum*]{}. However, we omit any further details here because the conservation of angular momentum does not play any role in what follows. Another – and more interesting conservation law with respect to analysis – is related to the invariance under spatial translations in the domain ${\mathbb{R}}^d$. For example when $d=1$, we obtain the (formal) conversation of the quantity $$P[{\mathbf{u}}] = \int_{{\mathbb{R}}} \frac{u_2 {\partial}_x u_1 - u_1 {\partial}_x u_2}{1-u_3} \, dx ,$$ which can be regarded as a [*linear momentum*]{}. For more details and the geometric meaning of $P[{\mathbf{u}}]$, we refer to [@LeSc-18].
With regard to conserved quantities, we finally mention that the one-dimensional (HWM) in fact possesses [*infinitely many conservation laws*]{} due to the existence of a [*Lax pair*]{}; see Section \[sec:Lax\] for more on this remarkable feature that indicate some kind of complete integrability in the one-dimensional setting.
Criticality and Conformal Symmetry
----------------------------------
For a constant $\lambda >0$, we readily check that rescaling $${\mathbf{u}}(t,x) \mapsto {\mathbf{u}}_\lambda(t,x) = {\mathbf{u}}(\lambda t, \lambda x)$$ maps solutions of (HWM) into solutions defined on the time interval $[0, \lambda^{-1} T)$. Likewise, the energy transforms as $$E[{\mathbf{u}}_\lambda] = \lambda^{d-1} E[{\mathbf{u}}].$$ Hence, we see that $E[{\mathbf{u}}_\lambda] = E[{\mathbf{u}}]$ in $d=1$ space dimensions, which implies that the evolution problem is [*energy-critical*]{} in one space dimension. In higher dimensions $d \geq 2$, the equation (HWM) becomes [*energy-supercritical*]{}.
Another feature of the energy-critical case when $d=1$ is the [*conformal invariance*]{} of the energy. More precisely, let ${\mathbf{u}}\in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$ be given and consider its harmonic extension ${\mathbf{u}}^e : {\mathbb{R}}_+^2 \to {\mathbb{R}}^3$ to the upper half-plane ${\mathbb{R}}^2_+$. Then it is a classical fact that $$E[{\mathbf{u}}] = \frac 1 2 \int_{{\mathbb{R}}} {\mathbf{u}}(x) \cdot |\nabla| {\mathbf{u}}(x) \, dx = \frac 1 2 \iint_{{\mathbb{R}}^2_+} |\nabla_{(x,y)} {\mathbf{u}}(x,y)|^2 \, dx \, dy.$$ Now the right-hand side is invariant under conformal transformations $\phi : {\mathbb{R}}_+^2 \to {\mathbb{R}}_+^2$ preserving the upper half-plane ${\mathbb{R}}_2^+ \simeq {\mathbb{C}}_+$, which is known to be the Möbius subgroup $\mathrm{PSL}(2, {\mathbb{R}})$. An elementary calculation yields that $$E[{\mathbf{u}}\circ \phi] = E[{\mathbf{u}}] \quad \mbox{for all $\phi \in \mathrm{PSL}(2, {\mathbb{R}})$}.$$ This conformal invariance of $E[{\mathbf{u}}]$ will play a role in the classification of traveling solitary waves (in addition to another action of the Möbius group on the target ${\mathbb{S}}^2$).
Geometric Formulation and Variants
----------------------------------
In analogy to the Schrödinger maps equation $\partial_t {\mathbf{u}}= {\mathbf{u}}\times \Delta {\mathbf{u}}$ for ${\mathbb{S}}^2$-valued maps, we can recast the half-wave maps equation into a more geometric form by exploiting that ${\mathbb{S}}^2$ is a Kähler manifold. Indeed, we can write (HWM) as $$\label{eq:HWM_kaehler}
{\partial}_t {\mathbf{u}}= J_{\mathbf{u}}P_{\mathbf{u}}{|\nabla|}{\mathbf{u}}$$ where $J_{\mathbf{u}}= {\mathbf{u}}\times : T_{\mathbf{u}}{\mathbb{S}}^2 \to T_{\mathbf{u}}{\mathbb{S}}^2$ is the standard complex structure on ${\mathbb{S}}^2$ and $P_{\mathbf{u}}: {\mathbb{R}}^3 \to T_{\mathbf{u}}{\mathbb{S}}^2$ denotes the projection onto the tangent space $T_{\mathbf{u}}{\mathbb{S}}^2$. So far, this reformulation of (HWM) has not been proven to be fruitful in the analysis yet. However, from this point of view, a natural modification of (HWM) arises when the compact Kähler manifold ${\mathbb{S}}^2$ is replaced by the non-compact Kähler $\mathbb{H}^2$ (hyperbolic plane). For more details on this, see Section \[sec:h2\] below.
Furthermore, in analogy to the Landau-Lifshitz-Gilbert equation (which is a combination of Schrödinger and harmonic maps heat flow), we can generalise the equation by adding a parabolic term, i.e., $${\partial}_t {\mathbf{u}}= \alpha J_{\mathbf{u}}P_{\mathbf{u}}{|\nabla|}u - \beta P_{\mathbf{u}}{|\nabla|}u$$ with constants $\alpha, \beta \geq 0$. For $\alpha = 0$, this is [*half-harmonic maps heat flow*]{} with target ${\mathbb{S}}^2$. See [@SiWeZh-18b] for infinite-time blow-up solutions for the purely parabolic evolution problem (when $\alpha=0$) with target ${\mathbb{S}}^1$.
Cauchy Problem
--------------
Up to now, the Cauchy problem for (HWM) lacks a full-fledged well-posedness theory for initial data in the energy space $\dot{H}^{\frac 1 2}({\mathbb{R}}^d; {\mathbb{S}}^2)$. In high dimensions $d \geq 5$ and for small energy initial data, the recent work in [@KrSi-18] establishes global-in-time existence. However, these techniques do not seem to be applicable for low dimensions $d \leq 3$; in particular, the energy-critical case $d=1$ seems out of scope by these techniques.
Let us mention that short-time existence of solutions for initial data in $\dot{H}^{k}$ with $k > \frac{d}{2} + 1$ can be established by standard arguments. However, the global-in-time existence or finite-time blowup is a completely open question. Also, by standard approximation arguments (e.g. parabolic regularisation), we can deduce existence of global weak solutions for finite energy data (where uniqueness of weak solutions is of course an open problem). See, e.g., [@PuGu-13] for such results on the domain $\mathbb{T}^d$ with $d=1,2,3$, but the arguments there can be carried over to domain ${\mathbb{R}}^d$ with $d=1,2,3$.
Traveling Solitary Waves {#sec:solitary}
========================
We will now turn out attention to the energy-critical (HWM) in $d=1$ space dimension. We seek special solutions that are given by [*traveling solitary waves*]{}. By definition, these solutions are of the form $${\mathbf{u}}(t,x) = {\mathbf{Q}}_v(x-vt),$$ where $v \in {\mathbb{R}}$ is a given constant velocity. The profile function ${\mathbf{Q}}_v : {\mathbb{R}}\to {\mathbb{S}}^2$ is seen to satisfy the nonlinear equation $$\label{eq:Qv}
{\mathbf{Q}}_v \times |\nabla| {\mathbf{Q}}_v + v {\partial}_x {\mathbf{Q}}_v = 0.$$ Indeed, we shall be interested in finite-energy solutions and hence we shall assume that ${\mathbf{Q}}_v \in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$. Other non-trivial solutions ${\mathbf{Q}}_v$ of with infinite energy are also given below.
In the special case $v=0$, corresponding to static solutions of (HWM), leads to the [*half-harmonic maps equation*]{} $$\label{eq:Qv_0}
{\mathbf{Q}}\times {|\nabla|}{\mathbf{Q}}= 0$$ for maps ${\mathbf{Q}}\in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$. In fact, this equation was introduced in [@DaLioRi-11] as a model problem to study higher regularity for a nonlocal elliptic energy-critical problem with conformal symmetry (in analogy to harmonic maps from ${\mathbb{R}}^2$ into ${\mathbb{S}}^2$). See also [@MiSi-15].
Complete Classification of Traveling Solitary Waves
---------------------------------------------------
The following classification result obtained in [@LeSc-18] now extends the known classification result for half-harmonic maps (see e.g. [@MiSi-15]). However, the presence of the term $v {\partial}_x {\mathbf{Q}}_v$ in calls for some further ideas, since ${\mathbf{Q}}_v$ with $v \neq 0$ do not correspond technically speaking to free boundary minimal disks but to a more general case. From [@LeSc-18] we recall the following result.
\[thm:Qv\_class\] Let $v \in {\mathbb{R}}$ and suppose ${\mathbf{Q}}_v \in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$ solves . Then the following holds.
1. If $|v| < 1$, then ${\mathbf{Q}}_v$ must be of the form[^1] $${\mathbf{Q}}_v(x) = R \left ( \sqrt{1-v^2} \, \mathrm{Re} \, \mathcal{B}(x), \mp \sqrt{1-v^2} \, \mathrm{Im} \, \mathcal{B}(x), \mp v \right )$$ with some fixed rotation $R \in \mathrm{SO}(3)$ and $\mathcal{B} : \overline{{\mathbb{C}}}_+ \to {\mathbb{C}}$ is a finite Blaschke product of degree $m \in \mathbb{N}_0$, i.e., $$B(z) = \prod_{k=1}^m \frac{z-z_k}{z-z_k^*}$$ with arbitrary points $z_1, \ldots, z_m$ in the complex upper half-plane ${\mathbb{C}}_+$.
2. If $|v| \geq 1$, then ${\mathbf{Q}}_v(x) \equiv \mathbf{P}$ for some constant $\mathbf{P} \in {\mathbb{S}}^2$.
*1. In (i), the trivial case ${\mathbf{Q}}_v \equiv \mbox{const}.$ corresponds to degree $m=0$.*
2\. The energy of the maps ${\mathbf{Q}}_v \in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$ for $|v| < 1$ is found to be $$E[{\mathbf{Q}}_v] = (1-v^2) \cdot m \pi.$$ In particular, we see that $E[{\mathbf{Q}}_v] \to 0$ as $|v| \to 1^-$. As an interesting consequence, we can construct traveling solitary waves with arbitrarily small energy. This is in stark contrast to other energy-critical evolution PDEs (e.g., energy-critical wave maps, Schrödinger maps, NLW and NLS equations) where small energy data lead to scattering to free solutions.
3\. For degree $m=1$, the profiles ${\mathbf{Q}}_v$ can regarded as [**ground states**]{}, i.e. as the nontrivial solutions with smallest energy. For example if we take $z_1=+{{\mathrm{i}}}$ and $v=0$, then $${\mathbf{Q}}(x) = \left ( \frac{x^2-1}{x^2+1}, \frac{2x}{x^2+1}, 0 \right )$$ is a half-harmonic map, i.e., a finite-energy solutions of with $v=0$. More generally, we see that all half-harmonic maps are (up to rotations on the sphere) are given by rational parametrisations of the equator, where the integer $m \geq 1$ is the winding number.
4\. In fact, one can view the map ${\mathbf{Q}}_{v=0} \mapsto {\mathbf{Q}}_{v \neq 0}$ as [*Möbius transform*]{} (conformal transformation) on the target ${\mathbb{S}}^2$. See [@LeSc-18] for more details and why this transformation can be seen as a [*Lorentz boost*]{} implemented by the Möbius group acting on ${\mathbb{S}}^2$. The presence of the factor $\sqrt{1-v^2}$ can be viewed as a kind of [*Lorentz contraction*]{} occurring for relativistic Lorentz boosts.
5\. For $|v| < 1$ and integer $m \geq 1$, the maps $$\tilde{{\mathbf{Q}}}_v(x) = (\sqrt{1-v^2} \cos (mx), \sqrt{1-v^2} \sin(mx), -v ) \in \dot{H}^{\frac 1 2}_{\mathrm{loc}}({\mathbb{R}}; {\mathbb{S}}^2)$$ are also (distributional) solutions of . However, these maps do not belong to $\dot{H}^{\frac 1 2}$ and thus the corresponding solutions ${\mathbf{u}}(t,x)= \tilde{Q}_v(x-vt)$ are traveling waves for (HWM) with infinite energy. It seems an interesting open problem to show that $\tilde{Q}_v$ provide all infinite-energy traveling waves for (HWM) up to symmetries.
6\. See also [@ZhSt-15] for a discussion of traveling solitary waves for the half-wave maps equation (HWM) in the physics literature.
Proving cases (i) and (ii) are of very different nature.
We briefly summarise the arguments in [@LeSc-18] used to obtain (i). Let $|v| < 1$ be given. By a delicate bootstrap argument (building upon the regularity theory in [@DaLioRi-11] for half-harmonic maps), we first show that any ${\mathbf{Q}}_v \in \dot{H}^{\frac 1 2}$ solving belongs to $C^\infty \cap \dot{H}^2$. For this regularity argument, the assumption $|v| < 1$ is crucial.
Once the regularity of ${\mathbf{Q}}_v$ is improved (actually, some higher Hölder continuity would suffice), we show as a next step that – after a suitable rotation on ${\mathbb{S}}^2$ – we have $$\label{eq:Qv_plane}
{\mathbf{Q}}_v(x) = ( \sqrt{1-v^2} \, f(x), \sqrt{1-v^2} \, g(x), \pm v )$$ with some smooth function $f, g: {\mathbb{R}}\to {\mathbb{R}}$ such that $f^2 + g^2 \equiv 1$. In geometric terms, equation means that the image of ${\mathbf{Q}}_v$ lies in a plane $E$ that is parallel to the equatorial plane $\{z=0\}$ and $E$ has distance $|v|$ to the plane $\{ z= 0 \}$. Furthermore, since we have ${\mathbf{Q}}_v(-\infty) = {\mathbf{Q}}_v(+\infty)$ by the finite energy condition, we see that $x \mapsto {\mathbf{Q}}_v(x)$ parametrises a circle on ${\mathbb{S}}^2$ with radius $\sqrt{1-v^2}$.
But how to prove above? To attack this problem, we consider ${\mathbf{Q}}_v : {\mathbb{R}}\to {\mathbb{S}}^2$ as the boundary curve of a minimal surface $\Sigma$ inside the the unit ball $B_1(0) \subset {\mathbb{R}}^3$. Thanks to the finite energy assumption ${\mathbf{Q}}_v \in \dot{H}^{\frac 1 2}$, we find that ${\mathbf{Q}}_v(-\infty) = {\mathbf{Q}}_v(+\infty)$ showing that $Q_v$ yields a closed curve on the unit sphere ${\mathbb{S}}^2$.
Next we let ${\mathbf{Q}}_v^e : {\mathbb{R}}^2_+ \to {\mathbb{R}}^3$ denote the (unique) bounded harmonic extension of ${\mathbf{Q}}_v$ to the upper half-plane ${\mathbb{R}}_+^2$. By the strong maximum principle, it follows that ${\mathbf{Q}}_v^e({\mathbb{R}}_+^2) \subset B_1(0)$ holds, unless ${\mathbf{Q}}_v$ is constant (which is a trivial case that we exclude here). Now, we apply the techniques of [*Hopf differentials*]{} to prove that the map ${\mathbf{Q}}_v^e: {\mathbb{R}}_+^2 \to {\mathbb{R}}^3$ traces out a [*minimal surface*]{} $\Sigma \subset B_1(0)$ whose boundary ${\partial}\Sigma$ given by the parametrised curve ${\mathbf{Q}}_v : {\mathbb{R}}\to {\mathbb{S}}^2$. To carry out this argument, we identify ${\mathbb{R}}_+^2$ with the complex upper half-plane ${\mathbb{C}}_+$ via $z = x+ {{\mathrm{i}}}y$. Next, we define the functions $\mathrm{\tt Hopf}_{\nu} : {\mathbb{C}}_+ \to {\mathbb{C}}$ with $\nu=1,2$ by setting $$\mathrm{\tt Hopf}_1(z) = {\partial}_z {\mathbf{Q}}_v^e \cdot {\partial}_z {\mathbf{Q}}_v^e \quad \mbox{and} \quad \mathrm{\tt Hopf}_2(z) = {\partial}_z^2 {\mathbf{Q}}_v^e \cdot {\partial}_{z}^2 {\mathbf{Q}}_v^e,$$ where ${\partial}_z = \frac 1 2 ({\partial}_x - {{\mathrm{i}}}{\partial}_y)$ and $U \cdot V = \sum_{j=1}^3 U_j V_j$ for $U, V \in {\mathbb{C}}^3$. From the harmonicity $\Delta {\mathbf{Q}}_v^e = 0$, we directly infer that $\mathrm{\tt Hopf}_1$ and $\mathrm{\tt Hopf}_2$ are both holomorphic on ${\mathbb{C}}_+$. Now by using the equation satisfied by ${\mathbf{Q}}_v = {\mathbf{Q}}_v^e |_{{\partial}{\mathbb{R}}_+^2}$, we deduce that $\mathrm{\tt Hopf}_1(z) \equiv 0$ vanishes identically. But this shows that ${\mathbf{Q}}_v^e$ is a conformal map from ${\mathbb{R}}^2_+$ into $B_1(0)$, i.e., we have $|{\partial}_x {\mathbf{Q}}_v^e | = |{\partial}_y {\mathbf{Q}}_v^e |$ and ${\partial}_x {\mathbf{Q}}_v \perp {\partial}_y {\mathbf{Q}}_v^e$. By its harmonicity, this means that ${\mathbf{Q}}_v^e$ traces out a minimal surface $\Sigma \subset B_1(0)$. Finally, we prove that the second Hopf differential $\mathrm{\tt Hopf}_2(z) \equiv 0$ also vanishes identically. From this fact we can deduce that $\Sigma$ is indeed a [*flat disk*]{}, which implies that its boundary $\partial \Sigma$ is a circle on ${\mathbb{S}}^2$. This completes the sketch of the proof of .
The rest of the proof of case (i) now boils down to a problem in complex analysis. By the previous discussion, the harmonic extension ${\mathbf{Q}}_v^e$ can be identified with a bounded holomorphic function $F : {\mathbb{C}}_+ \to C$ satisfying the boundary condition $|F|^2 \equiv 1$ on ${\partial}{\mathbb{C}}_+ \simeq {\mathbb{R}}\times \{0\}$. Thus $F$ is an inner function and hence has the canonical factorisation $$F(z) = \lambda e^{{{\mathrm{i}}}\alpha z} B(z)S(z),$$ where $\lambda \in {\mathbb{C}}$, $|\lambda| = 1$ and $\alpha \geq 0$. Here $B(z)$ is (a possibly infinite) Blaschke product with zeros on ${\mathbb{C}}_+$ and $S$ is the so-called singular inner part. However, by the regularity estimates for $F$ on the boundary ${\partial}{\mathbb{C}}_+$, we can show that $S(z) \equiv 1$ is trivial. Furthermore, by the finite energy property, we have $$-\int_{{\partial}{\mathbb{C}}_+} \overline{F} {\partial}_y F |_{y=0} \,dx = \int_{{\mathbb{R}}} {\mathbf{Q}}_v \cdot |\nabla| {\mathbf{Q}}_v \, dx < +\infty,$$ we find that $\alpha = 0$ and that $B(z)$ must be a finite Blaschke product, i.e., we have $$B(z) = \prod_{k=1}^m \frac{z-z_k}{z-z_k^*}$$ with some $z_1, \ldots, z_m \in {\mathbb{C}}_+$ and $m \in \mathbb{N}$. The interested reader may consult [@LeSc-18] for details.
In the case $|v| \geq 1$, no regularity method is known to improve the regularity of maps ${\mathbf{Q}}_v \in \dot{H}^{\frac 1 2}$ solving . However, we can build up an ‘Pohozaev-type’ argument to show that finite-energy solutions of must be constants whenever $|v| \geq 1$ holds. Luckily, no higher regularity is needed to carry out this argument.
Indeed, let ${\mathbf{Q}}_v^e : {\mathbb{R}}^2_+ \to {\mathbb{R}}^3$ be the harmonic extension of ${\mathbf{Q}}_v : {\mathbb{R}}\to {\mathbb{S}}^2$. Since ${\mathbf{Q}}_v \in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$, we have that ${\mathbf{Q}}_v^e \in\dot{H}^1({\mathbb{R}}_+^2;{\mathbb{R}}^3)$. By testing the equation against the Hilbert transform[^2] $H({\mathbf{Q}}_v) \in \dot{H}^{\frac 1 2}$, we find after some calculations the following identity: $$\label{eq:poho}
v \iint_{{\mathbb{R}}_+^2} (|{\partial}_x {\mathbf{Q}}_v^e|^2 + |{\partial}_y {\mathbf{Q}}_v^e|^2 ) \, dx \, dy = 2 \iint_{{\mathbb{R}}^2_+} ( {\mathbf{Q}}_v^e \times {\partial}_x {\mathbf{Q}}_v^e) \cdot {\partial}_y {\mathbf{Q}}_v^e \, dx \, dy.$$ Since $|{\mathbf{Q}}_v^2| \leq 1$ on ${\mathbb{R}}^2_+$ by the maximum principle, we can apply the Cauchy-Schwarz inequality in ${\mathbb{R}}^3$ to find $$2|( {\mathbf{Q}}_v^e \times {\partial}_x {\mathbf{Q}}_v^e) \cdot {\partial}_y {\mathbf{Q}}_v^e| \leq (|{\partial}_x {\mathbf{Q}}_v^e|^2 + |{\partial}_y {\mathbf{Q}}_v^e|^2).$$ Thus we conclude that $|{\partial}_x {\mathbf{Q}}_v^e| = |{\partial}_y {\mathbf{Q}}_v^e| \equiv 0$ from if $|v| \geq 1$ (where a little refinement is needed in the limiting case $|v|=1$.) Hence ${\mathbf{Q}}_v^e$ and therefore ${\mathbf{Q}}_v$ must be constant. Again, we refer the reader to [@FrSc-18] for details of this argument.
The completes our sketch of the proof of Theorem \[thm:Qv\_class\].
Spectral Analysis of the Linearised Operator
--------------------------------------------
We shall now briefly review the results in [@LeSc-18] on the spectrum of the linearised operator around traveling solitary waves for (HWM) in the energy-critical dimension $d=1$. More specifically, by following [@LeSc-18], we consider the static case with vanishing velocity $v=0$, i.e., half-harmonic maps from ${\mathbb{R}}$ to ${\mathbb{S}}^2$. Moreover, we focus on the special case when the profile ${\mathbf{Q}}\in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$ is given by Blaschke factor of degree $m \geq 1$ having identical factors, i.e., the zeros $z_1 = \ldots = z_m \in {\mathbb{C}}_+$ in Theorem \[thm:Qv\_class\] coincide. (For $m=1$ this is no loss of generality, whereas for $m > 1$ this assumption is non-trivial.) Henceforth we assume that $z_k = +{{\mathrm{i}}}$ for all $k=1, \ldots, m$. Thus we deal with profiles of the form $$\label{eq:pure_Blaschke}
{\mathbf{Q}}(x) = {\mathbf{Q}}_m(x) = \left ( \mathrm{Re}\, Q_m(x), \mathrm{Im} \, Q_m(x), 0 \right ) \quad \mbox{with} \quad Q_m(z) = \left ( \frac{z-{{\mathrm{i}}}}{z+{{\mathrm{i}}}} \right )^m.$$ To express perturbations around ${\mathbf{Q}}: {\mathbb{R}}\to {\mathbb{S}}^2$, we use the frame $\{ \mathbf{e}, J \mathbf{e} \}$ for vectors in the tangent space $T_{\mathbf{Q}}{\mathbb{S}}^2$, where $\mathbf{e}=(0,0,1)$ and $J \mathbf{e} = {\mathbf{Q}}\times \mathbf{e}$. In this frame, the linearized equation (HWM) around ${\mathbf{Q}}$ is found to be $$\label{eq:linearized_hwm}
{\partial}_t \left [ \begin{array}{c} h_1 \\ h_2 \end{array} \right ] = J L \left [ \begin{array}{c} h_1 \\ h_2 \end{array} \right ] + O(\mathbf{h}^2)$$ for the perturbation $\mathbf{h} = h_1 \mathbf{e} + h_2 J \mathbf{e}$. Here $O(\mathbf{h}^2)$ stands for quadratic terms in $\mathbf{h}$ and $|\nabla| \mathbf{h}$. The linearised operator is found to be $$J L = \left [ \begin{array}{cc} 0 & -1 \\ 1 & 0 \end{array} \right ] \left [ \begin{array}{cc} L_+ & 0 \\ 0 & L_- \end{array} \right ] = \left [ \begin{array}{cc} 0 & - L_- \\ L_+ & 0 \end{array} \right ].$$ Here $L_+$ and $L_-$ are the scalar operators given by $$L_+ = |\nabla| - | {\partial}_x {\mathbf{Q}}_m | = |\nabla| - \frac{2m}{1+x^2} \quad \mbox{and} \quad L_- = L_+ + R ,$$ and $R$ denotes the integral operator of the form $$(R f)(x) = \frac{1}{2 \pi} \int_{{\mathbb{R}}} \frac{|{\mathbf{Q}}_m(x)- {\mathbf{Q}}_m(y)|^2}{|x-y|^2} f(y) \, dy.$$
Let us first consider the [*nullspaces*]{} of the operators $L_+$ and $L_-$, which are defined as $$\mathcal{N}(L_+) = \{ f \in \dot{H}^{\frac 1 2}({\mathbb{R}}) \mid L_+ f = 0 \} \quad \mbox{and} \quad \mathcal{N}(L_-) = \{ f \in \dot{H}^{\frac 1 2}({\mathbb{R}}) \mid L_- f =0 \}.$$ Note that also (non-trivial) constant functions belong to $\dot{H}^{\frac 1 2}({\mathbb{R}})$. Elements $f \in \mathcal{N}(L_\pm) \setminus L^2({\mathbb{R}})$ will be called [*resonances*]{} of $L_+$ and $L_-$, respectively. From [@LeSc-18] we recall the following complete description of the nullspaces.
\[thm:spec1\] For any degree $m \geq 1$, it holds that $$\mathcal{N}(L_+) = \mathrm{span} \, \{ f_m , g_m \} \quad \mbox{and} \quad \mathcal{N}(L_-) = \mathrm{span} \, \{ 1, f_1, \ldots, f_m, g_1, \ldots, g_m \},$$ where the functions $\{ f_k \}_{k=1}^m$ and $\{g_k \}_{k=1}^m$ are known in closed form. In particular, we have that $$\dim \mathcal{N}(L_+) = 2 \quad \mbox{and} \quad \dim \mathcal{N}(L_-) = 2m+1$$ and hence [**nondegeneracy**]{} holds for the linearised operator around half-harmonic maps ${\mathbf{Q}}= {\mathbf{Q}}_m \in \dot{H}^{\frac 1 2}({\mathbb{R}};{\mathbb{S}}^2)$ for any degree $m \geq 1$.
Moreover, the operators $L_+$ and $L_-$ have a common resonance $\varphi \in \dot{H}^{\frac 1 2}({\mathbb{R}}) \setminus L^2({\mathbb{R}})$, which is given by $$ =
f\_m & if $m$ is odd,\
g\_m & if $m$ is even.
$$
*1. The idea of the proof will be sketched below.*
2\. The notion of nondegeneracy means that all the elements in the nullspaces of $L_+$ and $L_-$ are entirely generated by the continuous symmetries of the family of half-harmonic maps, i.e., due to rotations, translations and changes of the $m$ zeros in the corresponding Blaschke product; see [@LeSc-18] for details.
3\. For the special case $m=1$, the nondegeneracy of half-harmonic maps was also shown in [@SiWeZh-18].
The spectral analysis of the operators $L_+$ and $L_-$ can be extended much further, as done in [@LeSc-18]. To this end, we recall that the point spectrum of the unbounded self-adjoint operators $L_+$ and $L_-$ acting on $L^2({\mathbb{R}})$ is defined as the set of eigenvalues in $L^2({\mathbb{R}})$. In [@LeSc-18], the following collection of results is established.
\[thm:spec2\] For any degree $m \geq 1$, we have the following.
- [**$L^2$-Eigenvalues of $L_+$:**]{} The point spectrum of $L_+$ consists of exactly $2m$ eigenvalues, i.e., $$\sigma_\mathrm{p}(L_+) = \left \{ E_{0}, E_1, \ldots, E_{2m-1} \right \}.$$ Moreover, each eigenvalue $E_k$ is [**simple**]{} and we have the inequalities $$E_0 < E_1 < \ldots < E_{2m-2} < E_{2m-1} = 0.$$
- [**$L^2$-Eigenvalues of $L_-$:**]{} The point spectrum of $L_-$ only contains zero, i.e., $$\sigma_{\mathrm{p}}(L_-) = \{ 0 \}.$$ Moreover, the eigenvalue $E=0$ is exactly $2m$-fold degenerate.
- Both operators $L_+$ and $L_-$ have [**no embedded**]{} eigenvalue $E > 0$ inside $\sigma_{\mathrm{ess}}(L_+) = \sigma_{\mathrm{ess}}(L_-) = [0, \infty)$.
*1. A particular corollary of the preceding result is a [*coercivity estimate*]{} for $L_+$ and $L_-$ under certain (natural) orthogonality conditions. Such a result is typically needed in the modulational analysis of solutions close to ${\mathbf{Q}}$.*
2\. It can be shown that the [*purely absolutely continuous spectrum*]{} $\sigma_{\mathrm{ac}}(L_+) = \sigma_{\mathrm{ac}}(L_-)$ equals $[0, \infty)$. For details on an explicit unitary transform establishing this fact, see [@LeSc-18].
3\. For the spectral analysis of the matrix operator $\mathcal{L} = JL$, we refer again to [@LeSc-18].
The proofs of Theorems \[thm:spec1\] and \[thm:spec2\] both essentially depend on the use of the stereographic projection $\Pi$ from the unit circle ${\mathbb{S}}$ to the projective real line $\hat{{\mathbb{R}}} = {\mathbb{R}}\cup \{ \infty \}$. To briefly sketch the main ideas, we note that, by applying $\Pi$, we can recast the spectral analysis for the operators $$L_+= |\nabla| - \frac{2m}{1+x^2} \quad \mbox{and} \quad L_-= |\nabla| - \frac{2m}{1+x^2} + R$$ in terms of the unbounded operators given by $$J = (1-\sin \theta) ( |\nabla|_{{\mathbb{S}}} - m \mathds{1}) \quad \mbox{and} \quad H = (1- \sin \theta) ( |\nabla|_{{\mathbb{S}}} - m \mathds{1} + \widetilde{R} )$$ acting on $L^2({\mathbb{S}})$. Here $|\nabla|_{{\mathbb{S}}}$ is the square root of the Laplacian on ${\mathbb{S}}$ and $\widetilde{R}$ is some integral operator. Clearly, the operators $J \neq J^*$ and $H\neq H^*$ fail to be self-adjoint on $L^2({\mathbb{S}})$. Hence their detailed spectral analysis seems to be a hopeless enterprise at first sight. However, the benefit of this approach is that $J$ and $H$ are both seen to be [**Jacobi operators**]{}. By this, we mean that the corresponding (infinite) matrices have a tridiagonal structure with respect to the standard Fourier basis $\{ e^{{{\mathrm{i}}}k \theta} \}_{k \in \mathbb{Z}}$ of $L^2({\mathbb{S}})$. For instance, the matrix for $J$ is of the form $$[J_{kl}] = \left [ \begin{array}{cccccc} \ddots & \ddots & \ddots & & & 0 \\ & a_{n} & b_{n} & c_{n} \\ & & a_{n+1} & b_{n+1} & c_{n+1} \\ 0 & & & \ddots & \ddots & \ddots \end{array} \right ]$$ with certain sequences $(a_n)$, $(b_n)$, $(c_n)$ and $n \in \mathbb{Z}$. Roughly speaking, the degree $m \in \mathbb{N}$ of the half-harmonic maps ${\mathbf{Q}}= {\mathbf{Q}}_m$ plays a central role by splitting the frequencies $k \in \mathbb{Z}$ on the unit circle with $|k| \leq m$ and $|k| \geq m$. More precisely, the analysis of the Jacobi operators given by $J$ and $H$ then shows the following.
- The [*bound states*]{} of $L_+$ and $L_-$ are determined via the actions of $J$ and $H$ on the $2m+1$-dimensional subspace $\mathrm{span} \, \{ e^{{{\mathrm{i}}}k \theta} : |k| \leq m \}$ in $L^2({\mathbb{S}})$.
- The [*scattering states*]{} of $L_+$ and $L_-$ can be analysed in detail via the action of $J$ and $H$ on the infinite-dimensional subspace $\mathrm{span} \, \{ e^{{{\mathrm{i}}}k \theta} : |k| \geq m \}$ in $L^2({\mathbb{S}})$.
A distinguished part will be played by the two-dimensional subspace spanned by $\{ e^{{{\mathrm{i}}}k \theta} : k = \pm m \}$, which lies at the interface between the bound states and scattering states (i.e. the continuous spectrum), and which yields the two linearly independent solutions of $L_+ \varphi = L_-\varphi = 0$ given by the $L^2$-zero mode and the zero-energy resonance.
Lax Pair Structure and Rational Solutions {#sec:Lax}
=========================================
In the recent work [@GeLe-18], a Lax pair was found for one-dimensional (HWM). This indicates that some sort of complete integrability is present in this case. To formulate this result (and some of its applications), we need to introduce some notation as follows. Let $\{ \sigma_1, \sigma_2, \sigma_3 \}$ denote the standard Pauli matrices. For notational convenience, we make use of the Pauli vector $\bm{\sigma}=(\sigma_1, \sigma_2, \sigma_3)$ in the following. Given a vector $\mathbf{x} \in {\mathbb{R}}^3$, we define the complex $2 \times 2$-matrix by setting $${\mathbf{X}}= \mathbf{x} \cdot \bm{\sigma} = \sum_{j=1}^3 x_j \sigma_j = \left [ \begin{array}{cc} x_3 & x_1 - {{\mathrm{i}}}x_2 \\ x_1 + {{\mathrm{i}}}x_2 & -x_3 \end{array} \right ] .$$ Clearly, we have that ${\mathbf{X}}= {\mathbf{X}}^*$ is Hermitian matrix with zero trace $\mathrm{Tr} \,{\mathbf{X}}= 0$. From elementary algebra for the Pauli matrices we get the matrix identity $$\label{eq:nice}
( \mathbf{x} \cdot \bm{\sigma})(\mathbf{y} \cdot \bm{\sigma}) = (\mathbf{x} \cdot \mathbf{y}) \mathds{1} + {{\mathrm{i}}}( \mathbf{x} \times \mathbf{y}) \cdot \bm{\sigma}.$$ for all vectors $\mathbf{x}, \mathbf{y} \in {\mathbb{R}}^3$, where $\mathds{1}$ denotes the $2 \times 2$-unit matrix. As a consequence of this identity, we find that (HWM) can be written in the form $${\partial}_t {\mathbf{U}}= -\frac{{{\mathrm{i}}}}{2} \left [ {\mathbf{U}}, {|\nabla|}{\mathbf{U}}\right ],$$ where $[A,B] = A B-BA$ is the commutator of matrices in ${\mathbb{C}}^{2 \times 2}$ and ${\mathbf{U}}= {\mathbf{u}}\cdot \bm{\sigma}$. In abstract terms, the ${\mathbb{S}}^2$-valued map ${\mathbf{u}}$ can be expressed in terms of the map ${\mathbf{U}}$ that takes values in the real Lie algebra $\mathfrak{su}(2)$ spanned by the Pauli matrices. Note that ${\mathbf{U}}^2 = |{\mathbf{u}}|^2 \mathds{1} = \mathds{1}$ holds as an immediate consequence of .
Given a map ${\mathbf{u}}: [0,T) \times {\mathbb{R}}\to {\mathbb{S}}^2$ (not necessarily a solution of (HWM)), we set ${\mathbf{U}}= {\mathbf{u}}\cdot \bm{\sigma}$ and define the operators $$\label{def:LB}
L _{\mathbf{u}}= [H, {\mathbf{U}}] \quad \mbox{and} \quad B_{\mathbf{u}}= -\frac{{{\mathrm{i}}}}{2} \left ( {\mathbf{U}}\circ {|\nabla|}+ {|\nabla|}\circ {\mathbf{U}}\right ) + \frac{{{\mathrm{i}}}}{2} {|\nabla|}{\mathbf{U}}$$ acting on $L^2({\mathbb{R}}; {\mathbb{C}}^2)$. Here $H$ denotes [*Hilbert transform*]{} on $L^2({\mathbb{R}})$ defined in Fourier space by $\widehat{(H f)}(\xi) =-{{\mathrm{i}}}(\mathrm{sgn} \, \xi) \widehat{f}(\xi)$. For ${\mathbb{C}}^2$-valued functions $f \in L^2({\mathbb{R}}; {\mathbb{C}}^2)$, we let $H$ act componentwise via $Hf = (H f_1, Hf_2)$ with a slight abuse of notation.
In the definition of $L$ and $B$, the matrix-valued functions ${\mathbf{U}}$ and $|\nabla|{\mathbf{U}}$ are understood as multiplication operators on $L^2({\mathbb{R}}; {\mathbb{C}}^2)$. By the skew-symmetry of the Hilbert transform $H=-H^*$, we deduce the formal properties $$L_{\mathbf{u}}=(L_{\mathbf{u}})^* \quad \mbox{and} \quad B_{\mathbf{u}}=-(B_{\mathbf{u}})^*.$$ Indeed, as we will see below, the operator $L_{\mathbf{u}}$ is of Hilbert-Schmidt class (and hence compact) if and only if ${\mathbf{u}}$ has finite energy $E[{\mathbf{u}}] < \infty$. Of course, the operator $B_{\mathbf{u}}$ is always an unbounded operator on $L^2({\mathbb{R}}; {\mathbb{C}}^2)$ due to the presence of the pseudo-differential operator $|\nabla|$.
From [@GeLe-18] we recall the following result.
\[thm:Lax\] Let ${\mathbf{u}}: [0,T) \times {\mathbb{R}}\to {\mathbb{S}}^2$ be a sufficiently regular solution of (HWM) in $d=1$ dimension. Then the following Lax equation holds true: $$\frac{d}{dt} L_{\mathbf{u}}= [B_{\mathbf{u}}, L_{\mathbf{u}}],$$ where the operators $L_{\mathbf{u}}$ and $B_{\mathbf{u}}$ are defined in above with ${\mathbf{U}}={\mathbf{u}}\cdot \bm{\sigma}=\sum_{j=1}^3 u_j \sigma_j$.
The proof of Theorem \[thm:Lax\] follows from a tricky calculation involving the fact that $|\nabla| = H \frac{d}{dx}$ holds combined with [*Cotlar’s product identity*]{} for the Hilbert transform: $$H(fg) = (Hf) g + f(Hg) + H((Hf)(Hg))$$ for all $f,g \in \mathcal{S}({\mathbb{R}})$. Also, the normalisation condition ${\mathbf{U}}^2 = |{\mathbf{u}}|^2 \mathds{1} = \mathds{1}$ enters in essential way into the arguments. We refer to [@GeLe-18] for the complete proof of Theorem \[thm:Lax\].
From the Lax identity above, we can obtain two fundamental corollaries about (HWM) in one space dimension:
1. Infinite number of conservation laws.
2. Rational initial data give rise to rational solutions.
To elaborate on these claims, we note that the spectrum of $L_{\mathbf{u}}$ stays constant in time thanks to the Lax equation in Theorem \[thm:Lax\]. As an immediate consequence, we obtain the following result on an infinite number of (formal) conservation laws for the one-dimensional (HWM).
\[cor:L\_trace\] For ${\mathbf{u}}: [0,T) \times {\mathbb{R}}\to {\mathbb{S}}^2$ as above, we have the (formal) conservation laws $$\mathrm{Tr} ( |L_{\mathbf{u}}|^p ) = \mathrm{const}.$$ for any $0 < p < \infty$.
Let us make some few comments on this result. If we take $p=2$ above, we obtain the conservation of energy again. Indeed, the integral kernel of the commutator $L_{\mathbf{u}}= [H, {\mathbf{U}}]$ is found to be $$K(x,y) = \frac{1}{\pi} \frac{{\mathbf{U}}(x)- {\mathbf{U}}(y)}{x-y} \in {\mathbb{C}}^{2 \times 2},$$ where we omit the $t$-dependence for notational convenience. Using that $\mathrm{Tr}_{{\mathbb{C}}^2} ({\mathbf{U}}^2) = \mathrm{Tr}_{{\mathbb{C}}^2}\, (\mathds{1} |u|^2) = 2 |{\mathbf{u}}|^2$, we find that the square of the Hilbert-Schmidt norm of $L_{\mathbf{u}}$ is $$\mathrm{Tr} \, ( |L_{\mathbf{u}}|^2) = \frac{2}{\pi^2} \iint_{{\mathbb{R}}\times {\mathbb{R}}} \frac{|{\mathbf{u}}(x)-{\mathbf{u}}(y)|^2}{|x-y|^2} \, dx \,dy = \frac{8}{\pi} E[{\mathbf{u}}].$$ In particular, we see that $L_{\mathbf{u}}$ is of Hilbert-Schmidt class if and only if ${\mathbf{u}}\in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$.
For general $0 < p < \infty$, we can extend the previous observation by exploiting the norm equivalence of [*Schatten norms*]{} of the operator $L_{\mathbf{u}}$ with certain [*homogeneous Besov norms*]{} of function ${\mathbf{u}}$. More precisely, let us recall that the family of Schatten norms is given by $$\| L_{\mathbf{u}}\|_{\mathfrak{A}_p} = (\mathrm{Tr} \, ( |L_{\mathbf{u}}|^p) )^{\frac{1}{p}} \quad \mbox{for} \quad 0 < p < \infty.$$ For $p \geq 1$, this in fact a norm, whereas for $0 < p < 1$ we only obtain a quasi-norm. By adapting a classical result due to V. Peller on Hankel operators (see e.g. [@Pe-04] and by using the splitting $L^2({\mathbb{R}}) = L_+^2({\mathbb{R}}) \oplus L_-^2({\mathbb{R}})$ into positive and negative frequencies, we obtain the equivalence $$\| L_{\mathbf{u}}\|_{\mathfrak{A}_p} \sim_p \| {\mathbf{u}}\|_{\dot{B}_{p,p}^{1/p}} \quad \mbox{for} \quad 0 < p < \infty.$$ Note that $\| \cdot \|_{\dot{B}_{p,p}^{1/p}}$ is a semi-norm if and only if $p \geq 1$, where it is only a quasi semi-norm when $0 < p < 1$. By Corollary \[cor:L\_trace\] and classical facts about Besov space, we deduce the a-priori bound $$\sup_{t \in [0,T)} \left ( \int_{{\mathbb{R}}} |{\partial}_x {\mathbf{u}}(t,x)| \, dx \right ) \lesssim \sup_{t \in [0,T)} \| {\partial}_x {\mathbf{u}}(t) \|_{B^0_{1,1}} \lesssim \| {\mathbf{u}}(0) \|_{\dot{B}^1_{1,1}}.$$ In geometric terms, this a-priori estimate implies that the [*length of the curve*]{} parametrised by ${\mathbf{u}}(t) : {\mathbb{R}}\to {\mathbb{S}}^2$ is uniformly bounded for all $t \in [0,T)$. However, it remains a very interesting open problem to show that the conservation laws obtained from $L_{\mathbf{u}}$ are strong enough to deduce global well-posedness result for the one-dimensional (HWM). Furthermore, it is possible that although a smooth solution ${\mathbf{u}}={\mathbf{u}}(t,x)$ exists for all times $t \geq 0$, higher Sobolev norms $\| {\mathbf{u}}(t) \|_{\dot{H}^{\frac{1}{2}+ {\varepsilon}}}$ could growth drastically in $t$, showing some sort of turbulent behaviour. Such a turbulence phenomenon was recently constructed in [@GeLePoRa-18] for the focusing cubic half-wave equation in ${\mathbb{R}}$, exploiting the existence of small traveling solitary waves. It may be conjectured that an analogous behaviour can happen for (HWM) building upon the existence of small traveling solitary waves, see Section \[sec:solitary\] above.
Another main corollary from Theorem \[thm:Lax\] follows from an adaptation of [*Kronecker’s theorem*]{} for Hankel operators, which says that $L_{\mathbf{u}}$ has finite rank if and only if ${\mathbf{u}}={\mathbf{u}}(t,x)$ is a rational function of $x$. Indeed, we have the following.
Let ${\mathbf{u}}: [0,T) \times {\mathbb{R}}\to {\mathbb{S}}^2$ solve (HWM) in $d=1$ dimension. If the initial datum ${\mathbf{u}}(0,x)$ is a rational function of $x$, then ${\mathbf{u}}(t,x)$ is a rational function of $x$ for all $t \in [0,T)$.
The idea of the proof involves showing that $L_{\mathbf{u}}=[H, {\mathbf{U}}]$ is of finite rank if and only if ${\mathbf{u}}(t,x)$ is rational in $x$. In fact, we can relate the rank of $L_{\mathbf{u}}$ with the number of poles of $x \mapsto {\mathbf{u}}(t,x)$ in a precise way: $$\mathrm{rank} \, L_{\mathbf{u}}= \mbox{number of poles of ${\mathbf{u}}(x)$ in ${\mathbb{C}}\setminus {\mathbb{R}}$}.$$ By the Lax type evolution stated in Theorem \[thm:Lax\], the rank of $L_{\mathbf{u}}$ stays constant in time and hence the result follows. Again, it is a very interesting open problem to show global-in-time existence in the restricted class of rational solutions to (HWM). Even in this setting, it not clear yet how to exploit the conservation laws to deduce global-in-time existence.
Note that a special subclass of rational solutions is given by the traveling solitary waves ${\mathbf{u}}(t,x) = \mathbf{Q}_v(x-vt)$ discussed in Section \[sec:solitary\] above. Furthermore, in the class of rational solutions for (HWM), there exist other interesting explicit solutions such as [*time-periodic solutions*]{}. For instance, the function (which is rational in $x$) given by $${\mathbf{u}}_{\mathrm{per}}(t,x) = \left ( \cos \left ( \frac{t}{\sqrt{2}} \right) \frac{2x^2}{x^4+1}, \sin \left ( \frac{t}{\sqrt{2}} \right ) \frac{2x^2}{x^4+1}, \frac{x^4-1}{x^4+1} \right )$$ is a time-periodic solution of the one-dimensional (HWM). At each time $t$ fixed, the map $x \mapsto {\mathbf{u}}_{\mathrm{per}}(t,x)$ parametrises a half-circle on ${\mathbb{S}}^2$ starting from ${\mathbf{u}}_{\mathrm{per}}(t,-\infty) = (0,0,1)$ and going back to ${\mathbf{u}}_{\mathrm{per}}(t,+\infty)=(0,0,1)$. Its time-dependence corresponds to a rotation of this half-circle around the $x_3$-axis with constant angular velocity $\omega = 1/\sqrt{2}$. It is an interesting open problem to study the stability properties of ${\mathbf{u}}_{\mathrm{per}}(t,x)$ under rational (or more general) perturbations.
Periodic Case and Hyperbolic Target $\mathbb{H}^2$ {#sec:h2}
==================================================
A natural variant of (HWM) arises in the one-dimensional periodic setting when ${\mathbb{R}}$ is replaced by $\mathbb{T} = {\mathbb{R}}/ 2 \pi \mathbb{Z}$. That is, we consider maps ${\mathbf{u}}: [0,T) \times \mathbb{T} \to {\mathbb{S}}^2$ solving $${\partial}_t {\mathbf{u}}= {\mathbf{u}}\times |\nabla| {\mathbf{u}}$$ where $|\nabla| f = \sum_{n \in \mathbb{Z}} |n| \widehat{f}_n e^{{{\mathrm{i}}}n \theta}$ now denotes the square root of the Laplacian on $\mathbb{T}$. By means of the stereographic projection $\Pi : {\mathbb{S}}\to {\mathbb{R}}\cup \{ \infty \}$ it can be shown that the traveling solitary wave profiles ${\mathbf{Q}}_v : {\mathbb{R}}\to {\mathbb{S}}^2$ on the real line give rise to traveling solitary waves profile $\tilde{{\mathbf{Q}}}_v : \mathbb{T} \to {\mathbb{S}}^2$ for (HWM) on $\mathbb{T}$ and vice versa, by the simple relation $\tilde{{\mathbf{Q}}}_v = {\mathbf{Q}}_v \circ \Pi$. This fact also reflects the conformal invariance of the energy functional $E[{\mathbf{u}}]$ as sketched above; see [@LeSc-18]. In particular, the classification result in Theorem \[thm:Qv\_class\] applies mutatis mutandis to the one-dimensional periodic (HWM). Furthermore, the existence and the formula for the Lax pair $(L_{\mathbf{u}}, B_{\mathbf{u}})$ both carry over to periodic case in a direct fashion; see [@GeLe-18].
A further natural variation of (HWM) occurs when the compact target ${\mathbb{S}}^2$ is replaced by the non-compact Kähler manifold $\mathbb{H}^2$, i.e., the two-dimensional hyperbolic plane. Again, we refer to [@GeLe-18] for more details and the construction of the Lax pair for the target $\mathbb{H}^2$.
Conclusion and (Some) Open Problems
===================================
In this short note, we have given a brief overview on the current state of affairs concerning the half-wave maps equation (HWM). This geometric evolution equation exhibits a set of analytically intriguing properties. In the energy-critical case of $d=1$ space dimensions, we obtain: complete classification of traveling solitary waves, detailed spectral theory of linearised operators, time-periodic solutions, the existence of a Lax pair, infinitely many conservation laws involving the theory of Hankel operators, invariance of rational solutions under the flow, etc. Yet, it is fair to say that we are still at the beginning of understanding the dynamical properties of solutions. In the author’s opinion, the following three circles of open problems are of main interest for a future study of the energy-critical (HWM):
1. Global well-posedness for (HWM) in energy space (or counterexamples by singularity formation).
2. Stability (or instability) of traveling solitary waves or other special solutions (e.g. the time-periodic rational solutions).
3. Develop a notion of complete integrability building on the Lax pair.
Any progress concerning (1)–(3) would be highly desirable.
### Acknowledgments {#acknowledgments .unnumbered}
The author is grateful to Patrick Gérard and Armin Schikorra for fruitful collaborations on (HWM).
[99]{}
F. Da Lio and T. Rivi[è]{}re, Sub-criticality of non-local Schrödinger systems with antisymmetric potentials and applications to half-harmonic maps, [*Advances in Mathematics*]{} 227 (2011), 1300 – 1348.
P. G[é]{}rard and E. Lenzmann, A Lax pair structure for the half-wave maps equation, [*Lett. Math. Phys.*]{} 108 (2018), no. 7, 1635–1648.
P. G[é]{}rard, E. Lenzmann, O. Pocovnicu, and P. Rapha[ë]{}l, A two-soliton with transient turbulent regime for the cubic half-wave equation on the real line, [*Annals of PDE*]{} 4 (2018), no.1, 166pp.
A. Fraser and R. Schoen, Uniqueness theorems for free boundary minimal disks in space forms, [*Int. Math. Res. Not. IMRN*]{} 17 (2015), 8268–8274.
J. Krieger and Y. Sire, Small data global regularity for half-wave maps, [*Anal. PDE*]{} 11 (2018), no. 3, 661–682.
E. Lenzmann and A. Schikorra, On energy-critical half-wave maps into $\mathbb{S}^2$, [*Invent. Math.*]{} 213 (2018), no. 1, 1–82.
V. Millot and Y. Sire, On a fractional Ginzburg-Landau equation and 1/2-harmonic maps into sphere, [*Arch. Rat. Mech. Anal.*]{} 215 (2015), no. 1, 125–210.
P. Mironescu and A. Pisante, A variational problem with lack of compactness for [$H^{1/2}(S^1;S^1)$]{} maps of prescribed degree, [*J. Funct. Anal.*]{} 217 (2004), no. 2, 249–279.
V. Peller, Hankel operators and their applications, Springer, New York (2003).
X. Pu and B. Guo, Well-posedness for the fractional Landau-Lifshitz equation without Gilbert damping, [*Calc. Var. PDE*]{} 46 (2013), no. 3–4, 441–460.
Y. Sire, J. Wei, and Y. Zheng, Nondegeneracy of half-harmonic maps from $\mathbb{R}$ into $\mathbb{S}^1$, [*Proc. Amer. Math. Soc.*]{} 146 (2018), no. 12, 5263–5268.
Y. Sire, J. Wei, and Y. Zhang, Infinite time blow-up for half-harmonic map flow from $\mathbb{R}$ into $\mathbb{S}^1$, [*arXiv:1711.05387*]{}.
T. Zhou and M. Stone, Solitons in a continuous classical [Haldane–Shastry]{} spin chain [*Phys. Lett. A*]{} 379 (2015), 2817–2825.
[^1]: Taking $-v$ here corresponds to $+v$ in [@LeSc-18], where a different sign convention for (HWM) is used.
[^2]: To be understood modulo constants. More precisely, for ${\mathbf{Q}}_v \in \dot{H}^{\frac 1 2}({\mathbb{R}}; {\mathbb{S}}^2)$, there exists some constant $\mathbf{P} \in {\mathbb{S}}^2$ such that ${\mathbf{Q}}_v - \mathbf{P} \in D^{\frac 1 2}({\mathbb{R}})$ = closure of $C^\infty_c({\mathbb{R}})$ with respect to $\| \cdot \|_{\dot{H}^{\frac 1 2}}$. Thus $H({\mathbf{Q}}_v)$ means actually $H({\mathbf{Q}}_v-\mathbf{P})$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study quantum imaging in a triangular quantum corral that is embedded in a superconducting host system with $s$-wave symmetry. We show that the corral acts as a [*quantum copying machine*]{} by creating multiple images of a [*quantum candle*]{}. We obtain new selection rules for the formation of quantum images that arise from the interplay of the corral’s geometry and the location of quantum candles. In more complex corral structures, we show that quantum images can be projected “around the corner".'
author:
- 'Nikolaos A. Stavropoulos and Dirk K. Morr'
title: Quantum Imaging and Selection Rules in Triangular Quantum Corrals
---
Over the last few years, a growing number of exciting quantum phenomena has been observed [@Man00; @exp; @Bra02; @Pie04; @Repp04] that arise from the interplay between the geometry and quantum properties of nanoscale atomic structures and their coupling to a fermionic quantum many-body systems. Among these phenomena are the formation of [*quantum images*]{} (also referred to as [*quantum mirages*]{}) in elliptical quantum corrals [@Man00], electronic lifetime effects in triangular quantum corrals [@Bra02], and magnetization effects in triangular Co islands [@Pie04]. While the investigation of these phenomena is of general fundamental interest, it could potentially lead to important applications in the field of spin electronics and quantum information technology [@spinQC]. Theoretically, much progress in understanding the interaction between nanostructures and quantum many-body systems has been made by studying the formation of a [*single*]{} quantum image in quantum corrals that are embedded in metallic [@theory; @review] or superconducting (SC) [@Morr04] host systems.
In this Letter, we argue that complex nanoscale structures are prototype systems for the observation of novel quantum phenomena. In particular, we demonstrate that a triangular quantum corral can be used as a [*quantum copying machine*]{} that creates multiple quantum images of characteristic features (so-called [*quantum candles*]{}) in the host system’s density-of-states (DOS). As a quantum candle, we employ the spectroscopic signature of a fermionic bound state induced by a magnetic impurity in a superconducting host system with $s$-wave symmetry. We show that the formation of quantum images inside a triangular corral consisting of [*non-magnetic*]{} impurities is determined by a set of selection rules that arises from the interplay between the corral’s geometry and the location of quantum candles. Moreover, we demonstrate that such a corral can suppress the formation of fermionic bound states, leading to the important result that [*non-magnetic impurities*]{} can reverse the pair-breaking effect of a [*magnetic defect*]{}. Finally, we show that [*double triangular corrals*]{} allow the projection of quantum images “around the corner", opening the interesting possibility to custom design the imaging properties of quantum corrals.
In order to study novel quantum effects arising from the interaction of a triangular quantum corral with a superconducting host system, we employ a generalized scattering $\hat{T}$-matrix theory [@Morr04; @Morr03a; @Shiba68]. The host system’s local Greens function (in Nambu-notation) in the presence of the corral is given by $$\begin{aligned}
\hat{G}({\bf r},{\bf r'},\omega_n)&=&\hat{G}_0({\bf r},{\bf r'},\omega_n) \nonumber \\
& & \hspace{-2cm} +\sum_{i,j=1}^N \hat{G}_0({\bf r},{\bf
r}_i,\omega_n)\hat{T}({\bf r}_i,{\bf r}_j,\omega_n)\hat{G}_0({\bf
r}_j,{\bf r'},\omega_n) \ , \label{Ghat}\end{aligned}$$ where the sum runs over the locations ${\bf r}_i \ (i=1,..,N)$ of the $N$ impurities forming the corral. The ${\hat T}$-matrix follows from the Bethe-Salpeter equation $$\begin{aligned}
\hat{T}({\bf r}_i,{\bf r}_j,\omega_n)&=&\hat{V}_i \, \delta_{i,j}
\nonumber \\
& & \hspace{-1.5cm} +\hat{V}_i \, \sum_{l=1}^N \hat{G}_0({\bf
r}_i,{\bf r}_l,\omega_n)\hat{T}({\bf r}_l,{\bf r}_j,\omega_n) \ ; \\
\hat{V}_i&=&\frac{1}{2} \left(U_i \sigma_0 + J_iS \sigma_3
\right)\tau_3 \ , \label{Tmatrix}\end{aligned}$$ and the electronic Greens function of the unperturbed (clean) system in momentum space is $$\hat{G}^{-1}_0({\bf k},i\omega_n)=\left[ i\omega_n \tau_0 -
\epsilon_{\bf k} \tau_3 \right] \sigma_0 + \Delta_0 \tau_2
\sigma_2 \ . \label{G0}$$ $U_i (J_i)$ is the potential (magnetic) scattering strength of the impurity at site ${\bf r}_i$, $S$ is the spin of a magnetic impurity, and ${\bf \sigma}$, ${\bf \tau}$ are the Pauli-matrices in spin and Nambu space, respectively. $\Delta_{0}$ is the superconducting $s$-wave gap, and $\epsilon_{\bf k}$ is the host system’s normal state dispersion. In this approach, magnetic impurities are treated as classical variables [@Shiba68] (corresponding to a large-$S$ limit) since $J$ is taken to be smaller than the critical value $J_c$ necessary for a Kondo-effect to occur in an s-wave superconductor [@Sat92], in full agreement with experiment [@Yaz97]. Finally, the local DOS, $N({\bf r},\omega)$, shown below, is obtained numerically from Eqs.(\[Ghat\])-(\[G0\]) with $N({\bf
r},\omega)=A_{11}+A_{22}$, $A_{ii}({\bf r},\omega)=-{\rm Im}\,
\hat{G}_{ii}({\bf r},\omega+i\delta)/ \pi$ and $\delta=0.1$ meV.
Motivated by recent experiments [@Bra02], we first study the eigenmode spectrum of a triangular equilateral quantum corral embedded in a normal host system. To facilitate comparison with experiment, we consider a two-dimensional host system with a triangular lattice (lattice constant $a_0=1$) and $\epsilon_{\bf
k}= k^2/2m-\mu$ ($\hbar=1$) where $\mu=-65$ meV is the chemical potential and $k_F=0.24$ the Fermi wave-vector (qualitatively similar results to the ones shown below are expected for a 3D host system). In Fig. \[DOSTriangle\](a)-(h) we present a spatial DOS plot of the eight lowest energy eigenmodes \[with light (dark) regions indicating a large (small) DOS\] for a corral consisting of 90 [*non-magnetic*]{} impurities with $U_i=4$ eV (the impurities are represented by filled yellow squares, separated by $\Delta
r=2$). In this unitary scattering limit, the eigenmodes (i.e., their spatial structure and ordering in energy) are well described by the eigenstates, $\phi_{lm}$, of an infinitely deep triangular potential well (TPW) [@Kri82] \[the corresponding quantum numbers $(m,l)$ are shown in the upper left corner of Figs. \[DOSTriangle\](a)-(h)\].
With decreasing $U_i$, the energy separation of the eigenmodes is reduced. In addition, new eigenmodes emerge, such as the ones shown in Figs. \[DOSTriangle\](i),(j) for $U_i=0.5$ eV, which are similar to those observed experimentally (cf. Fig.1(d) in Ref. [@Bra02]).
The spatial form of the corral’s eigenmodes opens the possibility to form multiple [*quantum mirages*]{}. To demonstrate the qualitative nature of this effect, we retain the above parameters and take the superconducting gap to be $\Delta_0=4$ meV, yielding a superconducting coherence length $\xi_c=k_F/(m\Delta_0)=135$. As the quantum candle whose image is formed we use the spectroscopic signature of a fermionic bound state induced by a single magnetic impurity ($J_i=1.0$ eV), located in the center of the corral at ${\bf r}_1=(0,0)$.
This signature consists of a particlelike and holelike peak in the DOS at energies $\Omega_b^{(1,2)} = \mp 2.4$ meV, as shown in Fig. \[DOSTriangle1imp\]a. In Figs. \[DOSTriangle1imp\](b),(c) we present spatial DOS plots at $\Omega_b^{(1,2)}$ (the location of the magnetic impurity is shown as a filled red circle). The formation of the impurity bound state is accompanied by the excitation of the $(2,4)$-eigenmode \[Fig. \[DOSTriangle\](c)\] and by the emergence of three images of the bound state peaks inside the corral. Note that only eigenmodes that possesses sufficiently large spectral weight at the impurity site and are close in energy to $\Omega_b^{(1,2)}$ are relevant for the formation of quantum images. Since the energy of the $(2,4)$-eigenmode, $E_{(2,4)}=-4$ meV, is closer to $\Omega_b^{(1)}$ than to $\Omega_b^{(2)}$, the spectral weight of the quantum images is larger at $\Omega_b^{(1)}$ than at $\Omega_b^{(2)}$. This result demonstrates that a triangular quantum corral acts as a [*quantum copying machine*]{} for distinct features in the DOS. Moreover, the corral’s imaging properties can be specifically designed since changing the corral’s size leads to a shift in the eigenmode energies [@Kri82]. For example, in a corral consisting of 117 impurities, the imaging properties in the SC state are determined by the $(1,5)$-mode \[$E_{(1,5)}=0$ meV\] leading to a different spatial pattern of the quantum images [@MorrFP]. Finally, we note as an important result that the formation of an impurity bound state can be completely suppressed inside the corral. To demonstrate this effect, we place the magnetic impurity at a node \[${\bf r}_1=(-5,-5)$\] of the $(2,4)$- and $(1,4)$-eigenmodes, as shown in Fig. \[DOSSuppression\](a) for $|\omega|>\Delta_0$. In this case, the DOS at ${\bf r}_1$ \[Fig. \[DOSSuppression\](b)\] does not possess any signature of an induced bound state and hence, no image is observed anywhere inside the corral. This complete suppression arises from the incompatibility of the bound state with the boundary conditions provided by the corral’s wall. In other words, an impurity bound state can only be formed if it can couple to one of the corral’s eigenmodes. The importance of this result lies in the fact that while non-magnetic impurities cannot induce a fermionic bound state in an $s$-wave superconductor, they can suppress its formation and thus reverse the pair-breaking effect of a [*magnetic defect*]{}.
Next, we consider the effects of two magnetic impurities located inside the corral at ${\bf r}_1=(-10,-10)$ and ${\bf
r}_2=(20,-10)$. The angle, $\alpha$, between the impurity spins is determined by the interaction between them. In what follows, we assume a ferromagnetic alignment of the spins ($\alpha=0$), however, results similar to those shown below are also obtained for $\alpha \not = 0$. Quantum interference of scattered electronic waves leads to the formation of even and odd impurity bound states (with respect to a vertical axis midway between the two impurities) and a splitting of the bound state energies [@com1]. As a result, the DOS exhibits four peaks at $\Omega_o^{(1,2)}=\mp 2.8$ meV and $\Omega_e^{(1,2)}=\mp 2.0$ meV \[see Fig. \[DOSTriangle2imp\](a)\]. The formation of quantum images requires that the even/odd bound states couple to corral eigenmodes of the same symmetry. The eigenstates of a TPW and thus the corral eigenmodes transform under reflection at the vertical axis as $\phi_{lm} \rightarrow -\exp\left[ i 2 \pi (m+l)/3
\right] \phi^*_{lm}$ [@Kri82]. While $\phi_{24}$ is imaginary and thus even under reflection, an even and odd wave-function is formed from $\phi_{14}$ via $\phi^{(e,o)}_{14}(x)=\phi_{14}(x) \pm
\phi_{14}(-x)$.
In Figs. \[DOSTriangle2imp\](b),(e) we plot the DOS at $\Omega_{o}^{(1,2)}$ whose spatial form agrees well with that of $|\phi^{(o)}_{14}|^2$ shown in the inset of Fig. \[DOSTriangle2imp\](a). We therefore conclude that $\Omega_{o}^{(1,2)}$ are the frequencies of the odd bound state, while $\Omega_{e}^{(1,2)}$ are the energies of the even bound state whose spatial DOS is shown in Figs. \[DOSTriangle2imp\](c),(d). The imaging properties of the corral are thus frequency dependent due to the interplay between the corral’s geometry and the location of the quantum candles.
This interplay can be further studied by placing three magnetic impurities with parallel spins at the corners of an equilateral triangle at ${\bf r}_1=(-10,-10)$, ${\bf r}_2=(20,-10)$, and ${\bf
r}_3=(-10,20)$. Since the degeneracy of the impurity bound states is again lifted via quantum interference, we expect to find six peaks in the DOS. Instead, the DOS exhibits four peaks, as shown in Fig. \[DOSTriangle3imp\](a), corresponding to the presence of only two non-degenerate impurity states.
This reduction to two impurity states arises from a new type of selection rule that is based on the interplay between the corral geometry and the location of the quantum candles. Under a rotation of $2\pi/3$ around the corral’s center, $\phi_{lm}$, and hence the corral eigenmodes, transform as $\phi_{lm} \rightarrow \exp [ i 2
\pi (m+l)/3 ] \phi_{lm}$. Due to their geometry the non-degenerate impurity bound states possess the same transformation properties, and their formation thus requires that they couple to eigenmodes with $n=(m+l){\rm mod} 3=1,2,3$. However, the eigenmodes with $n=1$ are at energies $|E_{(m,l)}|
\gg \Delta_0$, thus preventing the creation of the bound state with $n=1$. As a result, only the bound states with $n=0$ \[Figs. \[DOSTriangle3imp\](c) and (d)\] and $n=2$ \[Figs. \[DOSTriangle3imp\](b) and (e)\] are formed via their coupling to the $(2,4)$- \[Fig. \[DOSTriangle\](c)\] and $(1,4)$-eigenmodes \[Fig. \[DOSTriangle\](d)\], respectively.
More complex corral structures can be employed to project quantum images “around the corner". To demonstrate this effect, we insert a triangular corral with 42 [*non-magnetic*]{} impurities ($U_i=4$ eV) into the corral discussed above. In the normal state, two eigenmodes of this double corral \[Figs. \[DoubleTriangle\](a) and (b)\] possess energies that render them relevant for the formation of quantum images.
By placing a magnetic impurity ($J_i=2$ eV) between the apices of the triangles in the SC state \[Fig. \[DoubleTriangle\](c)\] a quantum image is formed “around the corner" between the triangles’ bases. Similarly, a quantum image is created between the apices by placing the magnetic impurity between the triangles’ bases \[Fig. \[DoubleTriangle\](d)\]. Note the significant shift in $\Omega_b$ when the position of the impurity is changed. This result opens the interesting possibility to custom design the imaging properties of quantum corrals to form mirages at arbitrary locations.
Finally, we note that the qualitative features of our results presented above are robust against changes in the band parameters or $\Delta_0$, as long as $\xi_c$ is larger than the size of the corral. Moreover, whether a Kondo-effect can occur inside a corral and how it is affected by the corral’s eigenmodes is an interesting but non-trivial question whose study we reserve for future work [@MorrFP].
In summary, we demonstrate that the eigenmode spectrum of a triangular quantum corral can be employed (i) to create multiple images of a quantum candle, and (ii) to suppress the formation of impurity bound states. We obtain new selection rules for quantum imaging that arise from the interplay of the corral’s geometry and the location of quantum candles. Finally, we show that more complex nanostructures allow the projection of quantum images “around the corner".
We would like to thank K.-F. Braun, J.C. Davis, O. Pietzsch, and R. Wiesendanger for stimulating discussions, and would especially like to thank K.-H. Rieder for a series of discussions that motivated this work. D.K.M. acknowledges financial support from the Alexander von Humboldt foundation.
[99]{}
H.C. Manoharan, C.P. Lutz, and D.M. Eigler, Nature (London) [**403**]{}, 512 (2000).
A.W. Holleitner [*et al.*]{}, Phys. Rev. Lett. [**87**]{}, 256802 (2001); D.J. Derro [*et al.*]{}, Phys. Rev. Lett. [**88**]{}, 097002 (2002); C. Chicanne [*et al.*]{}, Phys. Rev. Lett. [**88**]{}, 097402 (2002).
K.-F. Braun and K.-H. Rieder, Phys. Rev. Lett. [**88**]{}, 096801 (2002).
O. Pietzsch [*et al.*]{}, Phys. Rev. Lett. [**92**]{},057202 (2004).
J. Repp [*et al.*]{}, Phys. Rev. Lett. [**92**]{}, 036803 (2004).
S.A. Wolf [*et al.*]{}, Science [**294**]{}, 1488 (2001); B.E. Kane, Nature (London) [**393**]{}, 133 (1998).
G.A. Fiete [*et al.*]{}, Phys. Rev. Lett. [**86**]{}, 2392 (2001); A.A. Aligia, Phys. Rev. B [**64**]{}, 121102 (2001); K. Hallberg, A.A. Correa, and C.A. Balseiro, Phys. Rev. Lett. [**88**]{}, 066802 (2002); D. Porras [*et al.*]{}, Phys. Rev. B [**63**]{}, 155406 (2001);O. Agam and A. Schiller, Phys. Rev. Lett. [**86**]{}, 484 (2001); Y. Shimada [*et al.*]{}, Surf. Sci [**514**]{}, 89 (2002); M. Weissmann and H. Bonadeao, Physica E [**10**]{}, 544 (2001); M. Schmid and A.P. Kampf, Ann. Phys. [**12**]{}, 463 (2003).
For a general review see G.A. Fiete and E.J. Heller, Rev. Mod. Phys. [**75**]{} 933 (2003), and references therein.
D. K. Morr and N. A. Stavropoulos, Phys. Rev. Lett. [**92**]{}, 107006 (2004); [*ibid.*]{}, Phys. Rev. B [**67**]{}, 020502(R) (2003).
D.K. Morr and A.V. Balatsky, Phys. Rev. Lett. [**90**]{}, 067005 (2003).
Y. Lu, Acta Physics Sinica **21**,75 (1965); H. Shiba, Prog. Theoret. Phys. [**40**]{}, 435 (1968).
K. Satori [*et al.*]{}, J. Phys. Soc. Jpn. [**61**]{}, 3239 (1992).
A. Yazdani [*et al.*]{}, Science [**275**]{}, 1767 (1997).
H.R. Krishnamurthy [*et al.*]{}, J. Phys. A: Math. Gen. [**15**]{}, 2131 (1982).
D.K. Morr and N.A. Stavropoulos, in preparation.
No even/odd bound states are formed for $\alpha =
\pi$ [@Morr04].
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Majid Janzamin[^1]'
- 'Hanie Sedghi[^2]'
- 'Anima Anandkumar[^3]'
title: |
Beating the Perils of Non-Convexity:\
Guaranteed Training of Neural Networks using Tensor Methods
---
### Acknowledgements {#acknowledgements .unnumbered}
We thank Ben Recht for pointing out to us the Barron’s work on approximation bounds for neural networks [@Barron93], and thank Arthur Gretton for discussion about estimation of score function. We also acknowledge fruitful discussion with Roi Weiss about the presentation of proof of Theorem \[thm:approx-guarantees\] on combining estimation and approximation bounds, and his detailed editorial comments about the preliminary version of the draft. We thank Peter Bartlett for detailed discussions and pointing us to several classical results on neural networks. We are very thankful for Andrew Barron for detailed discussion and for encouraging us to explore alternate incremental training methods for estimating remaining parameters after the tensor decomposition step and we have added this discussion to the paper. We thank Daniel Hsu for discussion on random design analysis of ridge regression. We thank Percy Liang for discussion about score function.
M. Janzamin is supported by NSF BIGDATA award FG16455. H. Sedghi is supported by NSF Career award FG15890. A. Anandkumar is supported in part by Microsoft Faculty Fellowship, NSF Career award CCF-$1254106$, and ONR Award N00014-14-1-0665.
[52]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{}
A. Agarwal, A. Anandkumar, P. Jain, P. Netrapalli, and R. Tandon. . In *Conference on Learning Theory (COLT)*, June 2014.
Guillaume Alain and Yoshua Bengio. What regularized auto-encoders learn from the data generating distribution. *arXiv preprint arXiv:1211.4246*, 2012.
A. Anandkumar, R. Ge, D. Hsu, and S. M. Kakade. . In *Conference on Learning Theory (COLT)*, June 2013.
Anima Anandkumar, Rong Ge, and Majid Janzamin. . *arXiv preprint arXiv:1408.0553*, Aug. 2014.
Animashree Anandkumar, Rong Ge, Daniel Hsu, Sham M. Kakade, and Matus Telgarsky. Tensor decompositions for learning latent variable models. *Journal of Machine Learning Research*, 15:0 2773–2832, 2014.
Animashree Anandkumar, Rong Ge, and Majid Janzamin. . *arXiv preprint arXiv:1402.5180*, Feb. 2014.
Animashree Anandkumar, Rong Ge, and Majid Janzamin. . In *Proceedings of the Conference on Learning Theory (COLT)*, Paris, France, July 2015.
Alexandr Andoni, Rina Panigrahy, Gregory Valiant, and Li Zhang. Learning polynomials with neural networks. In *Proceedings of the 31st International Conference on Machine Learning (ICML-14)*, pages 1908–1916, 2014.
Martin Anthony and Peter L Bartlett. *Neural network learning: Theoretical foundations*. cambridge university press, 2009.
Sanjeev Arora, Aditya Bhaskara, Rong Ge, and Tengyu Ma. Provable bounds for learning some deep representations. *arXiv preprint arXiv:1310.6343*, 2013.
Antonio Auffinger, Gerard Ben Arous, et al. Complexity of random smooth functions on the high-dimensional sphere. *The Annals of Probability*, 410 (6):0 4214–4247, 2013.
Pierre Baldi and Kurt Hornik. Neural networks and principal component analysis: Learning from examples without local minima. *Neural networks*, 20 (1):0 53–58, 1989.
Andrew R. Barron. Universal approximation bounds for superpositions of a sigmoidal function. *IEEE Transactions on Information Theory*, 390 (3):0 930–945, May 1993.
Andrew R Barron. Approximation and estimation bounds for artificial neural networks. *Machine Learning*, 14:0 115–133, 1994.
Peter Bartlett and Shai Ben-David. Hardness results for neural network approximation problems. In *Computational Learning Theory*, pages 50–62. Springer, 1999.
Peter L Bartlett. The sample complexity of pattern classification with neural networks: the size of the weights is more important than the size of the network. *Information Theory, IEEE Transactions on*, 440 (2):0 525–536, 1998.
A. Bhaskara, M. Charikar, A. Moitra, and A. Vijayaraghavan. Smoothed analysis of tensor decompositions. *arXiv preprint arXiv:1311.3651*, 2013.
Avrim L Blum and Ronald L Rivest. Training a 3-node neural network is np-complete. In *Machine learning: From theory to applications*, pages 9–28. Springer, 1993.
Martin L Brady, Raghu Raghavan, and Joseph Slawny. Back propagation fails to separate where perceptrons succeed. *Circuits and Systems, IEEE Transactions on*, 360 (5):0 665–674, 1989.
Venkat Chandrasekaran, Pablo Parrilo, Alan S Willsky, et al. Latent variable graphical model selection via convex optimization. In *Communication, Control, and Computing (Allerton), 2010 48th Annual Allerton Conference on*, pages 1610–1613. IEEE, 2010.
Anna Choromanska, Mikael Henaff, Micha[ë]{}l Mathieu, G[é]{}rard Ben Arous, and Yann LeCun. The loss surface of multilayer networks. In *Proc. of 18th Intl. Conf. on Artificial Intelligence and Statistics (AISTATS)*, 2015.
G. Cybenko. approximation by superpositions of a sigmoidal function. *Mathematics of Control, Signals, and Systems*, 2:0 303–314, 1989.
George Cybenko. Approximation by superpositions of a sigmoidal function. *Mathematics of control, signals and systems*, 20 (4):0 303–314, 1989.
Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In *Advances in Neural Information Processing Systems*, pages 2933–2941, 2014.
P Frasconi, M Gori, and A Tesi. Successes and failures of backpropagation: A theoretical investigation. *Progress in Neural Networks: Architecture*, 5:0 205, 1997.
Marco Gori and Alberto Tesi. On the problem of local minima in backpropagation. *IEEE Transactions on Pattern Analysis and Machine Intelligence*, 140 (1):0 76–86, 1992.
Benjamin D. Haeffele and Ren[é]{} Vidal. Global optimality in tensor factorization, deep learning, and beyond. *CoRR*, abs/1506.07540, 2015.
Moritz Hardt, Benjamin Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. *arXiv preprint arXiv:1509.01240*, 2015.
Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. *arXiv preprint arXiv:1207.0580*, 2012.
K. Hornik, M. Stinchcombe, and H. White. multilayer feedforward networks are universal approximators. *Neural Networks*, 2:0 359–366, 1989.
Kurt Hornik. Approximation capabilities of multilayer feedforward networks. *Neural networks*, 40 (2):0 251–257, 1991.
Daniel Hsu, Sham M Kakade, and Tong Zhang. Random design analysis of ridge regression. *Foundations of Computational Mathematics*, 140 (3):0 569–600, 2014.
Aapo Hyv[ä]{}rinen. Estimation of non-normalized statistical models by score matching. In *Journal of Machine Learning Research*, pages 695–709, 2005.
Majid Janzamin, Hanie Sedghi, and Anima Anandkumar. . *arXiv preprint arXiv:1412.2863*, Dec. 2014.
Michael J Kearns and Umesh Virkumar Vazirani. *An introduction to computational learning theory*. MIT press, 1994.
Pravesh Kothari and Raghu Meka. Almost optimal pseudorandom generators for spherical caps. *arXiv preprint arXiv:1411.6299*, 2014.
Christian Kuhlmann. Hardness results for general two-layer neural networks. In *Proc. of COLT*, pages 275–285, 2000.
Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of training neural networks. In *Advances in Neural Information Processing Systems*, pages 855–863, 2014.
Robert J Marks II and Payman Arabshahi. Fourier analysis and filtering of a single hidden layer perceptron. In *International Conference on Artificial Neural Networks (IEEE/ENNS), Sorrento, Italy*, 1994.
Nelson Morgan and Herv[é]{} Bourlard. Generalization and parameter estimation in feedforward nets: Some experiments. In *NIPS*, pages 630–637, 1989.
Ra[ú]{}l Rojas. *Neural networks: a systematic introduction*. Springer Science & Business Media, 1996.
Mark Rudelson and Roman Vershynin. Smallest singular value of a random rectangular matrix. *Communications on Pure and Applied Mathematics*, 620 (12):0 1707–1739, 2009.
Hanie Sedghi and Anima Anandkumar. Provable methods for training neural networks with sparse connectivity. *NIPS workshop on Deep Learning and Representation Learning*, Dec. 2014.
Shai Shalev-Shwartz and Shai Ben-David. *Understanding Machine Learning: From Theory to Algorithms*. Cambridge University Press, 2014.
Ji[ř]{}[í]{} [Š]{}[í]{}ma. Training a single sigmoidal neuron is hard. *Neural Computation*, 140 (11):0 2709–2728, 2002.
L. Song, A. Anandkumar, B. Dai, and B. Xie. Nonparametric estimation of multi-view latent variable models. *arXiv preprint arXiv:1311.3287*, Nov. 2013.
Bharath Sriperumbudur, Kenji Fukumizu, Revant Kumar, Arthur Gretton, and Aapo Hyv[ä]{}rinen. Density estimation in infinite dimensional exponential families. *arXiv preprint arXiv:1312.3516*, 2013.
Kevin Swersky, David Buchman, Nando D Freitas, Benjamin M Marlin, et al. On autoencoders and score matching for energy based models. In *Proceedings of the 28th International Conference on Machine Learning (ICML-11)*, pages 1201–1208, 2011.
Martin J Wainwright and Michael I Jordan. Graphical models, exponential families, and variational inference. *Foundations and Trends in Machine Learning*, 10 (1-2):0 1–305, 2008.
Yining Wang, Hsiao-Yu Tung, Alexander Smola, and Animashree Anandkumar. Fast and guaranteed tensor decomposition via sketching. In *Proc. of NIPS*, 2015.
T. Zhang and G. Golub. Rank-one approximation to high order tensors. *SIAM Journal on Matrix Analysis and Applications*, 23:0 534–550, 2001.
Yuchen Zhang, Jason D. Lee, and Michael I. Jordan. $\ell_1$-regularized neural networks are improperly learnable in polynomial time. *CoRR*, abs/1510.03528, 2015.
[^1]: University of California, Irvine. Email: [email protected]
[^2]: University of California, Irvine. Email: [email protected]
[^3]: University of California, Irvine. Email: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the influence of spin polarization on the degree of coherence of electron transport through interacting quantum dots. To this end, we identify transport regimes in which the degree of coherence can be related to the visibility of the Aharonov-Bohm oscillations in the current through a quantum-dot Aharonov-Bohm interferometer with one normal and one ferromagnetic lead. For these regimes, we calculate the visibility and, thus, the degree of coherence, as a function of the degree of spin polarization of the ferromagnetic lead.'
author:
- 'Bastian Hiltscher$^1$, Michele Governale$^2$, and Jürgen König$^1$'
title: 'Spin-dependent transport through quantum-dot Aharonov-Bohm interferometers'
---
Introduction
============
The investigation of electric transport properties of nano-structured devices defines a field of increasing importance. The issue of quantum coherence and its limitation by Coulomb interaction can be conveniently studied in devices that contain quantum dots (QDs) in multiply-connected geometries. The interplay between interference and Coulomb interaction has been extensively studied in these so-called quantum-dot Aharonov-Bohm interferometers (QD-ABIs) both experimentally[@aikawa04; @ihn07; @yacoby94; @schuster97; @ji00; @wiel00; @holleitner01; @sigrist06; @gustavsson08] and theoretically. [@aleiner97; @khym06; @moldo07; @hofstetter01; @koenig01; @koenig02; @akera93; @yeyati95; @bruder96; @hackenbroich96; @wu98; @kang99; @silvestrov00; @gerland00; @kang00; @boese01; @silvestrov03; @lopez05; @urban08; @silva09] Observed oscillations of the current through quantum-dot ABIs as a function of the magnetic flux enclosed by the interferometer arms [@yacoby94; @schuster97; @ji00; @wiel00; @holleitner01; @aikawa04; @sigrist06; @gustavsson08; @ihn07] prove that transport through a quantum dot is at least partially coherent. The degree of coherence may be suppressed by interaction. This can, e.g., be studied in a controlled way by electrostatically coupling a quantum-point contact (QPC) to the quantum dot in the ABI. The current through the QPC serves as a which-path detector that diminishes the amplitude of the interference signal. [@aleiner97; @moldo07; @khym06; @gustavsson08]
But even in the absence of any coupling to the outside world the degree of coherence may be limited by Coulomb interaction among the electrons within the QD-ABI. This is the issue that we will concentrate on for the rest of the paper. Similarly, the effect of different interdot- and intradot-interactions in T-shaped quantum-dot interferometers on the amplitude of the Fano resonance has been studied.[@moldoveanu08] The two central questions that we will address are: (1) What fraction $c=I^{\rm coh}/I^{\rm total}$ of the total current through a single-level quantum dot weakly coupled to the electrodes is coherent? (2) How and under which circumstances can this fraction $c$ be extracted from a current measurement in an Aharonov-Bohm setup? For transport through a single-level quantum dot with strong Coulomb interaction weakly coupled to normal leads the answer was given in Refs. and . If any coupling of the quantum dot to some bath is negligibly small then the only source of decoherence is connected to the spin degree of freedom in the quantum dot. In general, transport through the quantum dot can be divided into spin-flip and non-spin-flip processes. Spin-flip processes due to spin-orbit coupling are neglected in the following considerations. Furthermore, we restrict our analysis to temperatures larger than the Kondo temperature.[@remark] When the dot is initially empty, transferred electrons keep their spin orientation, and the transport is fully coherent. In contrast, when the dot is occupied with a single electron, then the transferred electron may either keep or flip its spin, i.e., only half of the processes (the non-spin-flip ones) are coherent. As a result, the fraction of coherent to total linear conductance in the limit of weak tunnel coupling is $c=1/[1+f(\epsilon)]$, where $f(\epsilon)$ is the Fermi function and $\epsilon$ the quantum dot level, measured relative to the Fermi energy of the leads. It was theoretically predicted[@koenig01; @koenig02] and experimentally confirmed[@aikawa04; @ihn07] that this fraction $c$ of coherent transport can be extracted from measuring the Aharonov-Bohm oscillation amplitude as a function of level energy for a quantum dot embedded in an Aharonov-Bohm ring. The asymmetry of the oscillation amplitude for $\epsilon>0$ as compared to the one for $\epsilon<0$ was in agreement with the theoretical prediction. The restriction to a single level is justified as long as the level spacing on the dot is larger than temperature and bias voltage such that only one orbital participates in transport. The influence of many levels on coherence and the crossover from large to small level spacing is discussed in Ref. .
In order to substantiate the role played by the spin, we suggest in this paper to replace one of the electrodes by a lead with a finite degree of spin polarization $p$. The main idea behind this proposal is that a large degree of spin polarization should, in general, increase the fraction of coherent transport since spin-flip processes are less frequent. However, introducing a spin-polarized lead breaks the spin symmetry and, thus, changes the transport characteristics in a non-trivial way. This includes the possibility of spin accumulation on the dot,[@linde09; @braun04; @souza07; @barnas08], tunnel magneto resistance[@stefanski09; @hamaya07] or a negative differential conductance.[@braun04; @bulka00; @rudzinski05; @braig05] Therefore, both questions 1. and 2. have to be reanalyzed carefully. Since the physics of spin accumulation may introduce an asymmetry of the current between the cases $\epsilon>0$ and $\epsilon<0$, that is not related to decoherence, an asymmetry of the AB oscillation amplitude does not necessarily indicate decoherence. The measurable quantity to compare $c$ with is the visibility $v$. In case of weak tunneling, where only one Fourier component of the flux-dependent current needs to be considered, the visibility $v$ (with $v>0$) is defined via $$I^{\rm total}(\varphi) = I^{\rm av} \left[ 1 + v \cos(\varphi +\delta) \right]\, ,$$ where $I^{\rm av}$ is the flux-averaged current and $\varphi$ the AB phase. As we will argue below, a clear correspondence between $c$ and $v$ can be established in the regime of uni-directional cotunneling, with an extra condition for the polarity of the applied bias voltage in the case $\epsilon<0$. In the latter case, it is possible to extract a polarization-dependent coherence factor $c=(1+p^2)/2$ by measuring the visibility $v$.
This paper is structured as follows. In Section \[model\] we introduce the model under consideration. The theoretical method that we employ is described in Sec. \[method\]. Results are presented in Sec. \[results\], which is subdivided in four parts: expressions for the charge current in different orders are given in Sec. \[current\]; we discuss the fraction of coherent transport in Sec. \[rescoherence\] ; Sec. \[secfluxdep\] concerns the visibility of the current and Sec. \[seccoherence\] elucidates the relation between visibility and coherence. Conclusions are drawn in Sec. \[conclusions\].
Model
=====
![\[setup\]Setup of single-dot Aharonov-Bohm interferometer with one spin-polarized lead. ](fig1.eps){width="7cm"}
We consider a closed single-dot ABI, i.e., a two-terminal ABI with a single-level quantum dot embedded in one of the arms, see Fig \[setup\]. The total Hamiltonian of our system consists of four parts, $$\begin{aligned}
H=H_{\text{dot}}+H_{\text{leads}}+H_{\text{tunn}}+ H_{\text{ref}}\, .\end{aligned}$$
The QD is assumed to accommodate a single, spin-degenerate level. It is described by the Anderson-impurity model: $$\begin{aligned}
\label{dothamil}Ê
H_{\text{dot}}&=&{\varepsilon}\sum\limits_\sigma n_{\sigma} + Un_{\uparrow} n_{\downarrow \,} .\end{aligned}$$ Here, $n_{{\sigma}}=d_{\sigma}^\dagger d_{\sigma}$, with $d_{\sigma}^\dagger$ being the creation operator for an electron with spin $\sigma$ on the quantum dot. The dot-level position is denoted by $\epsilon$ and the onsite Coulomb-repulsion energy by $U$. In the results, we will concentrate on the two limits of non-interacting electrons, $U=0$, and infinite charging energy, $U=\infty$.
We consider a two terminal setup, with the index $r=\text{F}$ labeling the ferromagnet and $r={\mathrm{N}}$ labeling the normal conductor. Both leads are large, non-interacting reservoirs, whose Hamiltonian reads $$\begin{aligned}
H_{\text{leads}}&=&\sum\limits_{r k\sigma}{\varepsilon}_{rk\sigma}c_{rk\sigma }^\dagger c_{rk\sigma
}\, ,\end{aligned}$$ where $ c_{r k\sigma}^\dagger$ is the creation operator for an electron in lead $r$ in a state labeled by the quantum number $k$ and with spin $\sigma$.
The tunnel coupling between the dot and the two leads is modeled by the tunneling Hamiltonian: $$\begin{aligned}
\label{tunnelhamil}
H_{\rm{tunn}}&=&\sum\limits_{r k\sigma }t_{r} c_{r k\sigma}^\dagger d_{\sigma}
+\mathrm{H.c.} \, .\end{aligned}$$ We assume the tunnel matrix elements $t_{r}$ and the density of states of the leads $N_{r\sigma}$ to be energy independent in the energy window relevant for transport. In the ferromagnetic lead we also have to distinguish between the density of states of electrons with majority ($\sigma=+$) and minority spin ($\sigma=-$). For the normal lead this distinction is not necessary ($N_{\mathrm{N}}/2\equiv N_{{\mathrm{N}}+}= N_{{\mathrm{N}}-}$). The spin polarization $p=(N_{{\mathrm{F}}+}-N_{{\mathrm{F}}-})/(N_{{\mathrm{F}}+}+N_{{\mathrm{F}}-})$ characterizes the asymmetry of the density of states. Tunnel-coupling strengths are then defined as $\Gamma_{\mathrm{N}}=2\pi |t_{\mathrm{N}}|^2 N_{\mathrm{N}}$ and $\Gamma_{\mathrm{F \pm}}=2\pi |t_{\mathrm{F}}|^2 N_{\mathrm{F\pm}}= (1\pm p)\Gamma_{\mathrm{F}}$. The intrinsic line width of the quantum dot’s level is the sum of the tunnel couplings, $\Gamma=\Gamma_{\mathrm{N}}+\Gamma_{\mathrm{F}}$.
The second (“reference”) interferometer arm is modeled by a direct tunnel coupling between the leads. The Hamiltonian of the reference arm reads $$\begin{aligned}
H_{\text{ref}}=\sum\limits_{k \in {\mathrm{N}}, q \in {\mathrm{F}}, \sigma} (\tilde{t}c^\dagger_{\text{N} k\sigma}c_{\text{F} q\sigma }+{\mathrm{H.c.}})\, .\end{aligned}$$ with transmission amplitude $t^{\text{ref}}_\sigma=2\pi \tilde{t} \sqrt{N_{\text{F}\sigma}N_{\text{N}}}$. The magnetic flux $\Phi$ threading the interferometer is included in the phases of the tunneling amplitudes. We choose the gauge in which $t_{\text{F}},t_{\text{N}} \in \Re^+$ and $\arg \tilde{t}=\varphi=2\pi \Phi/\Phi_0$, where $\Phi$ is the magnetic flux and $\Phi_0$ the flux quantum. In analogy to the tunnel coupling to the dot, we define the total transmission probability $|t^{\text{ref}}|^2= |t^{\text{ref}}_+|^2 + |t^{\text{ref}}_-|^2$.
Method
======
The dynamics of the quantum dot’s degree of freedom, i.e., the probabilities $P_{\chi}$ to find the dot in state $\chi=0, \uparrow, \downarrow, d$, is governed by a generalized master equation. In the stationary limit, it reads $0=\sum_{\chi'}W_{\chi\chi'}P_{\chi'}$, where $W_{\chi\chi'}$ are the transition rates from state $\chi'$ to $\chi$. Having solved the master equation for the probabilities, the stationary current can be computed from $I={\mathrm{e}}\sum_{\chi'}W_{\chi\chi'}^IP_{\chi'}$, where the current transition rates $W_{\chi\chi'}^I$ are obtained from the transition rates $W_{\chi\chi'}$ by multiplying with the net number of electrons that are transfered from source to drain in the transition described by $W_{\chi\chi'}$.
Our method is applicable for arbitrary values of the Coulomb repulsion $U$. However, for simplicity we only consider the two limits $U=0$ and $U=\infty$ from now on. In the latter case, double occupancy of the dot is prohibited. We aim at a systematic perturbation expansion for weak coupling ($\Gamma\lesssim k_{\mathrm{B}}T$ and $|t^{\mathrm{ref}}|\ll1$) of the current $I=\sum_{m,n} I^{(m,n)}$, where $m$ indicates the power in the tunnel coupling $\Gamma$ between dot and leads and $n$ the power in the direct tunnel coupling $|t^{\mathrm{ref}}|$ between the two leads. A direct coupling between the leads can be made small in experiments with the help of a tunable barrier in the reference arm. We perform a corresponding expansion for the probabilities and the transition rates. We restrict ourselves to the lowest-order contributions. This means, we include the current through the reference arm in the absence of the quantum dot, $I^{\mathrm{(0,2)}}$, the interference term $I^{\mathrm{(1,1)}}(\varphi)$, which is the lowest-order contribution that depends on the Aharonov-Bohm phase $\varphi$, and the current through the quantum dot in the absence of the reference arm.
For the last contribution, it is important to distinguish two different transport regimes. If the dot level ${\varepsilon}$ lies inside the energy window for which occupied states in the source electrode and simultaneously empty states in the drain are available, i.e., $|{\varepsilon}| \lesssim \max\{k_{\mathrm{B}}T,|{\mathrm{e}}V/2|\}$, then transport is dominated by transition rates $W^{\mathrm{(1,0)}}$ (and $W^{I{\mathrm{(1,0)}}}$) that are first order in $\Gamma$. It is clear that in this case only first-order rates are required to evaluate the zeroth-order probability distribution $P_\chi^{(0,0)}$. We refer to this procedure as calculation scheme 1.
The situation is different in the cotunneling regime, $|{\varepsilon}| \gg \max\{k_{\mathrm{B}}T,|{\mathrm{e}}V/2|\}$, for which some of the rates $W^{\mathrm{(1,0)}}$ are exponentially suppressed and the lowest-order contribution is $W^{\mathrm{(2,0)}}$. Then, as discussed e.g. in Ref. , some second-order rates are required to evaluate the zeroth-order probability distribution $P_\chi^{(0,0)}$. This we call calculation scheme 2.
For scheme 1, we use a real-time diagrammatic technique to perform the perturbation expansion in the tunnel-coupling strengths.[@koenig96] The advantage of this technique is that it is systematic in the sense that all contributions of given order are properly taken into account. The downside is that including higher-order contributions becomes increasingly cumbersome. In the cotunneling regime, where scheme 2 needs to be used, the expressions for the rates obtained from the diagrammatic technique drastically simplify. In that case it is easier to directly identify all the cotunneling processes and evaluate the corresponding rates by second-order perturbation theory rather than employing the real-time diagrammatics.
To discuss the results obtained by scheme 1, we will only provide the final expressions for the current. As we will discuss below, for connecting the degree of coherence with the visibility of the Aharonov-Bohm oscillations, the cotunneling regime is more important. In this case, we use scheme 2 with the cotunneling rates $W^{(2,0)}_{\mathrm{\chi\chi'}}=\sum\limits_{r,r'} \gamma^{\chi\chi'(2,0)}_{rr'}$, where $\gamma^{\chi\chi'}_{rr'}$ is the rate of a transition where an electron is transfered from reservoir $r'$ to reservoir $r$, accompanied by a change of the dot state from $\chi'$ to $\chi$. An example for the calculation of such a cotunneling rate, as introduced in Refs. for metallic islands and applied for single-level quantum dots, e.g., in Ref. , is given in Appendix \[secordcalc\]. For $|{\varepsilon}| \gg \max\{k_{\mathrm{B}}T,|{\mathrm{e}}V/2|\}$ and $U=0$, the cotunneling rates simplify to $$\begin{aligned}
\gamma^{00(2,0)}_{rr'}&=&\gamma^{\sigma\sigma(2,0)}_{rr'}=\gamma^{dd(2,0)}_{rr'}=
\sum_{\sigma'} \frac{\Gamma^{\sigma'}_{r}\Gamma^{\sigma'}_{r'}}{2\pi{\varepsilon}^2}
F(\mu_r-\mu_{r'})
\\
\gamma^{\bar{\sigma}\sigma(2,0)}_{rr'}&=&0\end{aligned}$$ with $F(x)=x/[\exp (x/k_{\mathrm{B}}T)-1]$. For $U=\infty$, we obtain $$\begin{aligned}
\gamma^{00(2,0)}_{rr'}&=&\sum_\sigma \frac{\Gamma^\sigma_{r}\Gamma^\sigma_{r'}}{2\pi{\varepsilon}^2}
F(\mu_r-\mu_{r'})
\label{rate_empty}
\\
\gamma^{\sigma\sigma(2,0)}_{rr'}&=&\frac{\Gamma^\sigma_{r}\Gamma^\sigma_{r'}}{2\pi{\varepsilon}^2} F(\mu_r-\mu_{r'})
\label{rate_nonsf}
\\
\label{rate_sf}
\gamma^{\bar{\sigma}\sigma(2,0)}_{rr'}&=&\frac{\Gamma^{\sigma}_{r}\Gamma^{\bar{\sigma}}_{r'}}{2\pi{\varepsilon}^2} F(\mu_r-\mu_{r'})
\, .\end{aligned}$$ The rates in Eqs. (\[rate\_empty\]) and (\[rate\_nonsf\]) are associated with non-spin-flip processes while Eq. (\[rate\_sf\]) describes spin-flip processes.
Finally, we also need the rates to first order in $\Gamma$ and first order in $|t^{\mathrm{ref}}|$. Only those contributions with $r \neq r'$ exist. We obtain $$\begin{aligned}
\gamma^{00(1,1)}_{rr'}&=& \delta_{r,\bar r'}\sum_\sigma \frac{|t^{\mathrm{ref}}_\sigma| \sqrt{\Gamma^\sigma_{r}\Gamma^\sigma_{r'}}}{\pi{\varepsilon}}F(\mu_r-\mu_{r'}) \cos\varphi
\\
\gamma^{\sigma\sigma'(1,1)}_{rr'}&=& 2 \delta_{\sigma,\sigma'} \delta_{r,\bar r'} \frac{|t^{\mathrm{ref}}_\sigma| \sqrt{\Gamma^\sigma_{r}\Gamma^\sigma_{r'}}}{\pi{\varepsilon}}F(\mu_r-\mu_{r'}) \cos\varphi
\\
\gamma^{dd(1,1)}_{rr'}&=&- \gamma^{00(1,1)}_{rr'}\end{aligned}$$ for $U=0$ and $$\begin{aligned}
\gamma^{00(1,1)}_{rr'}&=& \delta_{r,\bar r'}\sum_\sigma \frac{|t^{\mathrm{ref}}_\sigma| \sqrt{\Gamma^\sigma_{r}\Gamma^\sigma_{r'}}}{\pi{\varepsilon}}F(\mu_r-\mu_{r'}) \cos\varphi
\\
\gamma^{\sigma\sigma'(1,1)}_{rr'}&=& \delta_{r,\bar r'} \frac{|t^{\mathrm{ref}}_\sigma| \sqrt{\Gamma^\sigma_{r}\Gamma^\sigma_{r'}}}{\pi{\varepsilon}}F(\mu_r-\mu_{r'}) \cos\varphi \end{aligned}$$ for $U=\infty$. Here, $\bar r'$ indicates the lead other than $r'$.
Results
=======
Charge current {#current}
--------------
The quantity that is directly measured in experiment is the charge current. As indicated above, the total current can be split into three contributions: the current through the reference arm in the absence of the quantum dot, $I^{\mathrm{(0,2)}}$, the current through the quantum dot in the absence of the reference arm, $I^{\mathrm{(1,0)}}$ or $I^{\mathrm{(2,0)}}$ (for scheme 1 and 2, depending on the level position $\epsilon$, respectively), and the interference term, $I^{\mathrm{(1,1)}}(\varphi)$. Only the last one depends on the Aharonov-Bohm phase $\varphi$.
Direct tunneling through the reference arm can be calculated with Fermi’s golden rule and contributes to the current with $$I^{\mathrm{(0,2)}}=\frac{e^2}{\pi}V|t^{\mathrm{ref}}|^2,$$ where $V$ is the bias voltage applied between the ferromagnet and the normal conductor.
We now consider the transport through the quantum dot in the absence of the direct interferometer arm. If the dot level lies in between the transport voltage defined by the Fermi energies of the electrodes then transport through the dot will be dominated by first-order tunneling, $I^{\mathrm{(1,0)}}$, and we use scheme 1. We find $$\begin{aligned}
I^{\mathrm{(1,0)}}=-2{\mathrm{e}}\frac{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}\left(\Gamma-p^2\Gamma_{\mathrm{F}} \right)} {\Gamma^2-p^2\Gamma^2_{\mathrm{F}}}\left[f_{\mathrm{F}}({\varepsilon})-f_{\mathrm{N}}({\varepsilon})\right]\, ,\end{aligned}$$ for noninteracting electrons, $U=0$, where $f_{\mathrm{F/N}}$ is the Fermi function of the normal/ferromagnetic lead. For an infinite interaction, $U=\infty$, we obtain $$\begin{aligned}
I^{\mathrm{(1,0)}}=-2{\mathrm{e}}A^{-1}\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}\left[f_{\mathrm{F}}({\varepsilon})-f_{\mathrm{N}}({\varepsilon})\right]\nonumber \\
\times\left[\Gamma_{\mathrm{F}}\left(1-p^2\right)(1-f_{\mathrm{F}}({\varepsilon}))+\Gamma_{\mathrm{N}}\left(1-f_{\mathrm{N}}({\varepsilon})\right)\right] \, ,\end{aligned}$$ with $$\begin{aligned}
A=\Gamma^2-p^2\Gamma_{\mathrm{F}}^2-[\Gamma_{\mathrm{F}}f_{\mathrm{F}}({\varepsilon})+ \Gamma_{\mathrm{N}}
f_{\mathrm{N}}({\varepsilon})]^2+p^2\Gamma_{\mathrm{F}}^2f_{\mathrm{F}}^2({\varepsilon}) \nonumber \, .\end{aligned}$$
If the dot level lies outside the energy window defined by the Fermi energies of the leads ($|{\varepsilon}|\gg\max\{k_{\mathrm{B}}T,|{\mathrm{e}}V/2|\}$) then $I^{\mathrm{(1,0)}}$ is exponentially suppressed and transport through the dot is dominated by cotunneling, $I^{\mathrm{(2,0)}}$. In this case, we employ scheme 2, see Sec. \[method\]. In this regime the current for noninteracting electrons ($U=0$) reads, $$I^{\mathrm{(2,0)}}={\mathrm{e}}^2\frac{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}{\pi{\varepsilon}^2}V\, .$$
In the case of an infinite Coulomb interaction on the dot ($U=\infty$) we have to distinguish different cases. For a dot-level position well above the Fermi energy of the leads, ${\varepsilon}>0$, the current through the quantum dot is the same as for noninteracting electrons. In the opposite case, ${\varepsilon}<0$, we get $$I^{\mathrm{(2,0)}}={\mathrm{e}}^2\frac{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}{\pi{\varepsilon}^2}V\left[1+\frac{pm}{1-{\mathrm{exp}}(-{\mathrm{e}}V/k_{\mathrm{B}}T)}\right]\, ,$$ where $m$ is the spin accumulation on the dot, which depends on the transport direction. In the regime of unidirectional cotunneling, $|{\varepsilon}|\gg|{\mathrm{e}}V/2|\gg k_{\mathrm{B}}T$, it simplifies to $m=p$ for transport from the ferromagnetic into the normal lead ($V<0$) and $m=-p$ for the opposite transport direction ($V>0$).
The flux-dependent part is given by $I^{(1,1)}(\varphi) = I^{(1,1)}_{\mathrm{even}}\cos \varphi + I^{(1,1)}_{\mathrm{odd}}\sin \varphi$. For noninteracting electrons, the coefficients are $$\begin{aligned}
I^{\mathrm{(1,1)}}_{\mathrm{odd}}=0 \end{aligned}$$ and $$\begin{aligned}
I^{\mathrm{(1,1)}}_{\mathrm{even}}=2{\mathrm{e}} |t^{\mathrm{ref}}|\sqrt{\Gamma_{\mathrm{N}}\Gamma_{\mathrm{F}}}\sigma({\varepsilon}) \,\end{aligned}$$ with $$\begin{aligned}
\sigma({\varepsilon})=\frac{1}{\pi}{\mathrm{Re}}\left[\int d\omega\frac{f_{\mathrm{F}}(\omega)-f_{\mathrm{N}}(\omega)}{{\varepsilon}-\omega+{\mathrm{i}}0^+}\right]\, ,\end{aligned}$$ independent of the polarization $p$.
For an infinitely strong charging energy, both the contributions even and odd in the flux are present. They read in the sequential tunneling regime (scheme 1) $$\begin{aligned}
I^{\mathrm{(1,1)}}_{\mathrm{odd}}&=&2{\mathrm{e}}A^{-2}|t^{\mathrm{ref}}|(\Gamma_{\mathrm{N}}\Gamma_{\mathrm{F}})^{3/2}\left(f_ {\mathrm{F}}({\varepsilon})-f_{\mathrm{N}}({\varepsilon})\right)^2 \nonumber \\
&\times& \left\{\left[\Gamma_{\mathrm{F}}(1-p^2)(1-f_{\mathrm{F}}({\varepsilon}))+\Gamma_{\mathrm{N}}(1-f_{\mathrm{N}}({\varepsilon}))\right]^2\right.\nonumber\\
&-&\left.\Gamma_{\mathrm{N}}^2\left[p^2\left(1-f_{\mathrm{N}}^2({\varepsilon})\right)\right]\right\} \,\end{aligned}$$ and $$\begin{aligned}
I^{\mathrm{(1,1)}}_{\mathrm{even}}&=&2{\mathrm{e}}A^{-1}|t^{\mathrm{ref}}|\sqrt{\Gamma_{\mathrm{N}}\Gamma_{\mathrm{F}}} \sigma({\varepsilon}) \nonumber \\
&\times& \left\{\Gamma_{\mathrm{F}}^2(1-f_{\mathrm{F}}({\varepsilon}))(1-p^2)+\Gamma_{\mathrm{N}}^2(1-f_{\mathrm{N}}({\varepsilon}))\right.\\
&+&\left.\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}\left(2-f_{\mathrm{F}}({\varepsilon})(1-p^2)-f_{\mathrm{N}}({\varepsilon})(1+p^2)\right)\right\} \nonumber\,\end{aligned}$$ respectively.
In the cotunneling regime (scheme 2) the odd contribution drops out, $I^{\mathrm{(1,1)}}_{\mathrm{odd}}=0$ while the even part is given by $$I^{\mathrm{(1,1)}}_{\mathrm{even}}=-\frac{{\mathrm{e}}^2V}{\pi{\varepsilon}}\sqrt{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}|t^{\mathrm{ref}}|
(1+pm)\; .$$
The odd and even parts of the first flux dependent correction differ in many respects. The odd part $I^{\mathrm{(1,1)}}_{\mathrm{odd}}$ describes transport processes where an electron cotunnels through a lead.[@urban08] It only occurs for a nonvanishing Coulomb interaction. Figure \[Iodd\] shows both transport directions of $I_{\mathrm{odd}}^{\mathrm{(1,1)}}$ as a function of the dot-level position ${\varepsilon}$ for an infinite Coulomb interaction. The current has its maximum value where the dot level lies between the chemical potential of the two leads. Beside this range the current decreases exponentially.
Figure \[strom\] shows $I_{\mathrm{even}}^{\mathrm{(1,1)}}$ as function of dot-level position for vanishing and infinite Coulomb interaction. For an infinite Coulomb interaction two different lines are shown. The red, dashed line is calculated by means of scheme 1 while the blue, dashed-dotted line is obtained by means of scheme 2, see Sec. \[method\]. Figure \[strom\]a) shows the current of electrons from the ferromagnetic into the normal lead. For noninteracting electrons the current $I^{\mathrm{(1,1)}}$ is an odd function of the dot level position ${\varepsilon}$. The sign change around ${\varepsilon}=0$ relies on a phase shift of the transmission amplitude of the quantum dot. An infinite Coulomb interaction excludes the double occupation of the dot. Hence it has a higher influence for negative than for positive values of ${\varepsilon}$. The transport from the normal conductor into the ferromagnet is for ${\varepsilon}<0$ strongly suppressed, see Fig. \[strom\]b). Transport through the quantum dot is blocked by an accumulation of the minority spin on the dot.
What can we conclude from this for the fraction $c$ of coherent transport through a quantum dot? Not much, as long as the dot’s level is inside the energy window of lowest-order transport. And even for the cotunneling regime, an interpretation is difficult for $|{\mathrm{e}}V/2|\lesssim k_{\mathrm{B}}T$, i.e., when transport processes from source to drain are partially compensated by processes from drain to source. For the further discussion, we will, therefore, turn to the regime of unidirectional cotunneling, $|{\varepsilon}|\gg|{\mathrm{e}}V/2|\gg k_{\mathrm{B}}T$. We emphasize that our method is applicable to arbitrary values of $U$. For $U=0$, no spin-flip processes occur since contributions with intermediate empty and double occupation of the dot cancel out each other. As long as $U\ll\min\{|eV|,k_{\mathrm{B}}T\}$ this also holds for a finite $U$. In the opposite limit, $U=\infty$, double occupancy is fully suppressed, and this cancellation does not occur anymore. This will remain true as long as $U\gg\max\{|eV|,k_{\mathrm{B}}T\}$ . Between these two limits there will be a smooth crossover. Therefore, we focus on the limits $U=0$ and $U=\infty$ only. In particular we will distinguish the four different cases summarized in the Table \[cases\].
case 1 $U=0$
--------- ---------------------------------------------------------------------------------------
case 2a $U=\infty$, ${\varepsilon}\gg|{\mathrm{e}}V/2|\gg k_{\mathrm{B}}T$
case 2b $U=\infty$, $-{\varepsilon}\gg|{\mathrm{e}}V/2|\gg k_{\mathrm{B}}T$, $F\rightarrow N$
case 2c $U=\infty$, $-{\varepsilon}\gg|{\mathrm{e}}V/2|\gg k_{\mathrm{B}}T$, $N\rightarrow F$
: \[cases\]The considered cases
For reference, we always compare to the non-interacting limit (case 1). For strong Coulomb interaction, the dot level may either lie well above the Fermi level of the leads (case 2a), or it may lie well below. In the latter case, the results will strongly depend on the polarity of the applied transport voltage. Case 2b refers to the limit when electrons are transported from the ferromagnet to the normal lead, and case 2c describes the opposite transport direction.
 Odd part of the first flux dependent order of the current $I^{\mathrm{(1,1)}}_{\mathrm{odd}}$ for polarization $p=0.7$ and Coulomb interaction $U=\infty$ as a function of ${\varepsilon}$. In the total current a negative bias voltage corresponds to a transport from the ferromagnet into the normal conductor. The value of the parameters used in the calculations are: $\varphi=\pi/2$, $|t^{\mathrm{ref}}|=0.1$, $k_{\mathrm{B}}T=\Gamma$, $\Gamma_{\mathrm{F}}=\Gamma_{\mathrm{N}}=\Gamma/2$.](fig2.eps){width="6cm"}
 Even part of the first flux dependent order of the current $I^{\mathrm{(1,1)}}_{\mathrm{even}}$ for polarization $p=0.7$ as a function of ${\varepsilon}$ for vanishing (solid line) and infinite Coulomb interaction for two different calculation schemes (see text). Scheme 1 (dashed line) is more accurate for ${\varepsilon}\gtrsim -\max\{k_{\mathrm{B}}T, |{\mathrm{e}}V/2|\}$ while scheme 2 (dashed-dotted line) is more accurate for ${\varepsilon}\ll -\max\{k_{\mathrm{B}}T, |{\mathrm{e}}V/2|\}$. **a)** Electrons are transported from ferromagnet into normal lead ($eV=-15\Gamma$). **b)** Electrons are transported from normal lead into ferromagnet ($eV=15\Gamma$). The value of the parameters used in the calculations are: $\varphi=0$, $|t^{\mathrm{ref}}|=0.1$, $k_{\mathrm{B}}T=\Gamma$, $\Gamma_{\mathrm{F}}=\Gamma_{\mathrm{N}}=\Gamma/2$. ](fig3.eps){width="6cm"}
Fraction of coherent transport {#rescoherence}
------------------------------
How can the fraction of coherent transport be measured in an experiment? Coherence can be tested by interferometry. We consider here an Aharonov-Bohm interferometer in which a single-level quantum dot is embedded in one of the arms. Electrons entering from the source electrode can either travel through the quantum dot or through the direct arm to the drain. If no spin flip occurs, there will be an interference of both paths, which gives rise to a flux-dependent current. The amplitude of the Aharonov-Bohm oscillations relative to the flux-averaged current contains information about the degree of coherence.
There is, however, a major problem in quantitatively connecting the degree of coherence $c$ and the visibility $v$. A quantum-dot Aharonov-Bohm interferometer probes many different transport channels, distinguished by the energy of the incoming electron, simultaneously. If for the participating electrons the transmission through the quantum dot is strongly energy dependent, then the expected visibilities for the individual channels will be very different from each other. This is the case, when the dot’s level position lies inside the energy window defined by the Fermi energies of the leads. For establishing a connection between visibility and fraction of coherent transport, we need to identify a situation in which for all participating electrons the transmission through the dot is the same. This is possible in the cotunneling regime, i.e., when the energy level of the quantum dot is outside this energy window, $|{\varepsilon}|\gg\max \{|{\mathrm{e}}V/2|,k_{\mathrm{B}}T\}$. Furthermore, for the case when the fraction of coherent transport depends on the transport direction, we need $|{\mathrm{e}}V| \gg k_{\mathrm{B}}T$, i.e., uni-directional cotunneling, as an extra condition to separate the two directions.
What do we expect for the fraction $c$ of coherent transport in this regime of uni-directional cotunneling from lead $r$ to lead $\bar r$? We assume that flipping the spin in the quantum dot provides the only source of decoherence. The coherence fraction $c$ is the ratio $$c=\frac{\sum_{\chi} \gamma^{\chi\chi(2,0)}_{\bar{r}r}P_{\chi}^{\mathrm{(0,0)}}}
{\sum_{\chi,\chi'} \gamma^{\chi\chi'(2,0)}_{\bar{r}r}P_{\chi'}^{\mathrm{(0,0)}}} \, ,$$ with $P_\chi^{\mathrm{(0,0)}}$ being the probability to find the dot in state $\chi$ and $\gamma^{\chi\chi'(2,0)}_{\bar{r}r}$ the transition rate from initial dot state $\chi'$ to final dot state $\chi$ where an electron is transferred from lead $r$ to lead $\bar{r}$. In the numerator, only rates are taken into account that do not change the dot state. In particular, no spin-flip processes are included. This contrasts with the expression in the denominator, in which spin-flip processes, i.e. $\chi=\sigma$ and $\chi'=\bar\sigma$ are taken into account.
In the limit of vanishing Coulomb interaction, $U=0$ (case 1), no spin-flip processes occur, which yields $c=1$. Now we consider the limit of strong Coulomb interaction, $U=\infty$. If the dot’s level lies well above the Fermi energies of the leads, ${\varepsilon}\gg |{\mathrm{e}}V/2|$ (case 2a), then the dot will be predominately empty. Electrons passing through the quantum dot cannot flip their spin, and therefore $c=1$. The situation becomes different for $-{\varepsilon}\gg |{\mathrm{e}}V/2|$. Then the dot is mostly singly occupied with either spin with probabilities $p_\uparrow$ and $p_\downarrow$ (such that $p_\uparrow^{\mathrm{(0,0)}}+p_\downarrow^{\mathrm{(0,0)}}=1$). If the electrons travel from the ferromagnet to the normal lead (case 2b) we get $c=(1+p^2)/2$. In the opposite case (case 2c), transport from the normal lead to the ferromagnet, we get always $c=1/2$ since an electron enters the dot from the normal lead and hence carries in one half of the cases the same spin as the electron initially occupying the dot.
These results are given in the last column of Table \[summary\]. The remaining question now is whether and how they are reflected in the visibility of the Aharonov-Bohm oscillations.
Visibility {#secfluxdep}
----------
In the regime of unidirectional cotunneling, the leading order of transport through the quantum dot is $I^{(2,0)}$. In this limit the transmission through the QD is for all energies in good approximation the same. Furthermore, the $\sin \varphi$ part of the current which describes cotunneling through the lead but not through the quantum dot [@urban08] vanishes. The total current then is $$I^{\mathrm{total}}=I^{\mathrm{(2,0)}}+I^{\mathrm{(0,2)}}+I^{\mathrm{(1,1)}}=I^{\mathrm{av}}\left(1+v \cos\varphi\right)\, ,$$ where $v$ is the visibility and $I^{\mathrm{av}}$ the flux averaged current.
For the average current, measured in units of $I_0={\mathrm{e}}^2V/\pi$ we find $$\frac{I^{\mathrm{av}}}{I_0}= \left\{
\begin{array}{ll}
\displaystyle |t^{\mathrm{ref}}|^2+\frac{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}{{\varepsilon}^2}
& \text{for cases 1,2a,2b}
\\
\displaystyle |t^{\mathrm{ref}}|^2+(1-p^2)\frac{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}{{\varepsilon}^2}
& \text{for case 2c}
\end{array} \right.
\, .$$
We express the visibility in terms of $$v_0 = \frac{2\frac{\sqrt{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}}{|{\varepsilon}|}|t^{\mathrm{ref}}|}{|t^{\mathrm{ref}}|^2+\frac{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}{{\varepsilon}^2}} \; .$$ We obtain $$\frac{v}{v_0} = \left\{
\begin{array}{ll}
\displaystyle 1
& \text{for cases 1,2a}
\\
\displaystyle \frac{1+p^2}{2}
& \text{for case 2b}
\\
\displaystyle \frac{1}{2} - \frac{p^2 |t^{\mathrm{ref}}|^2}{2|t^{\mathrm{ref}}|^2+
2(1-p^2)\frac{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}{{\varepsilon}^2}}
& \text{for case 2c}
\end{array} \right.
\, .$$
In the cases 1, 2a, and 2b, the visibility can be maximized by tuning $|t^{\mathrm{ref}}|$ to $\frac{\sqrt{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}}{|{\varepsilon}|}$. Then, $v_0=1$ and $v=v_{\mathrm{max}}$, independent of the degree of spin polarization $p$. However, for the case 2c, the maximal visibility $v_{\mathrm{max}}$ can only be obtained by tuning $|t^{\mathrm{ref}}|$ in a $p$-dependent way to $\frac{\sqrt{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}}{|{\varepsilon}|}\sqrt{1-p^2}$. In this case $v_{\mathrm{max}}$ turns out to be $v_{\mathrm{max}}=\frac{\sqrt{1-p^2}}{2}$.
The visibility of the total current is a quantity that can be measured in an experiment. To what extent the visibility provides information about coherence of transport is discussed in the next section.
visibility versus coherence {#seccoherence}
---------------------------
In order to investigate the measurability of coherence we compare the fraction $c$ with the maximal visibility $v_{\mathrm{max}}$. Because coherence is a essential assumption for flux dependence in general $c\ge v_{\mathrm{max}}$. In the case of a vanishing Coulomb interaction (case 1) or a very high dot’s level position (case 2a) $v_{\mathrm{max}}=1$ and, hence, $c=1$, see Fig.\[coherence\]a). If the dot level is very low and electrons are transferred from the ferromagnet into the normal conductor (case 2b) the coherent fraction $c$ is equal to the maximal visibility $v_{\mathrm{max}}$, see Fig. \[coherence\]b). For a vanishing polarization $1/2$ of the electrons leaving the source carry the same spin as the electron initially occupying the dot. Hence, in one half of the cases the spin on the dot is not flipped and transport is coherent. The higher the polarization the more electrons with majority spin take part in transport and, thus, less spin-flip processes take place.\
For reversed transport voltages (case 2c) independent of the polarization one half of the processes are coherent. The source is a normal lead and, hence, one half of the electrons which tunnel onto the dot carry the same spin as the electron which initially occupied the dot. On the other hand the visibility is low for a high polarization, see Fig. \[coherence\](c), due to spin blockade on the dot. While transport through the reference arm is spin independent transport through the quantum dot is not. This prevents the possibility to tune the transmission through the reference arm and the transmission through the quantum dot to the same value.\
In all cases the maximal visibility is obtained by tuning $|t^{\mathrm{ref}}|$ to a certain value. While in the cases 1, 2a and 2b this value is independent of the polarization $p$, in case 2c $|t^{\mathrm{ref}}|$ has to be tuned in a $p$-dependent way, see Sec. \[secfluxdep\].\
![\[coherence\] (Color online) Maximal visibility and coherent fraction of the total current for the two different transport directions. ](fig4.eps){width="6cm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$v/v_0$ $v_{\mathrm{max}}$ c
------------------------------------------------------------------------------------------------ --------------------------------------------------------------------------------- -------------------- -------------
$U=0$ (case 1) $1$ 1 1
$U=\infty$, ${\varepsilon}\gg|{\mathrm{e}}V/2|\gg k_{\mathrm{B}}T$ (case 2a) $1$ 1 1
$U=\infty$, $-{\varepsilon}\gg|{\mathrm{e}}V/2|\gg k_{\mathrm{B}}T$,$F\rightarrow N$ (case 2b) $(1+p^2)/2$ $(1+p^2)/2$ $(1+p^2)/2$
$U=\infty$, $-{\varepsilon}\gg|{\mathrm{e}}V/2|\gg k_{\mathrm{B}}T$,$N\rightarrow F$ (case 2c) $\frac{1}{2} - \frac{p^2 |t^{\mathrm{ref}}|^2}{2|t^{\mathrm{ref}}|^2+2 $\sqrt{1-p^2}/2$ 1/2
(1-p^2)\frac{\Gamma_{\mathrm{F}}\Gamma_{\mathrm{N}}}{{\varepsilon}^2}}$
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[summary\]Summary of results
Conclusions
===========
We have investigated the current through an AB-interferometer coupled to one normal and one ferromagnetic lead with a quantum dot embedded in one of the arms. In particular we elucidated the influence of polarization on the visibility of transport and studied the relation between visibility and coherence. We found that in the lowest flux-dependent order transport of noninteracting electrons is fully coherent and the maximal visibility is $1$. In the case of an infinite intra-dot Coulomb repulsion the coherence as well as the visibility of the current are strongly influenced by the polarization and the transport direction. As long as no spin blockade on the dot occurs the maximal visibility is equal to the coherent fraction of the current and can be obtained by tuning the transmission through the reference arm $|t^{\mathrm{ref}}|$ in a $p$-independent way.
We acknowledge stimulating discussions with Saskia Fischer and Sven Buchholz, as well as support from DFG via SPP 1285, SFB 491 and EU STREP GEOMDISS.
Calculation of a cotunneling rate {#secordcalc}
=================================
In this appendix we give an example of how to calculate the cotunneling rates. For this, we choose the rate $\gamma^{\bar{\sigma}\sigma(2,0)}_{rr'}$, where an electron is transfered from lead $r'$ to lead $r$ accompanied by a change of the dot state from $\sigma$ to $\bar{\sigma}$. We start with $$\begin{aligned}
\label{apprate1}
\gamma^{\bar{\sigma}\sigma(2,0)}_{rr'}&=\frac{1}{2\pi}\int\limits_{-\infty}^{\infty}d\omega f(\omega-\mu_{r'})\left(1-f(\omega-\mu_r)\right) \nonumber\\
&\times {\rm Re}\left[\frac{\sqrt{\Gamma_{r'}^{\bar{\sigma}}\Gamma_r^\sigma}}{{\varepsilon}-\omega+i0^+}+\frac{\sqrt{\Gamma_{r'}^{\bar{\sigma}}\Gamma_r^{\sigma}}}{{\varepsilon}+U-\omega+i0^+}\right]^2\, .\end{aligned}$$ The regularization $+i0^+$ in the resolvents is added here by hand; however, it appears naturally when employing the real-time diagrams. In the present case, there are two possible intermediate states, the dot being empty or doubly occupied. The corresponding energy differences to the final state appear in the denominator of the resolvents that have to be added coherently before performing the square. The Fermi functions guarantee that the lead state from the electron enters the dot is occupied and that the lead state to which the dot electron leaves is unoccupied. The integral sums over all possible energies of the incoming electron. For an infinite Coulomb repulsion $U=\infty$ and $|{\varepsilon}|\gg \max\{k_{\mathrm{B}}T,\mu_r\}$ the rate simplifies to $$\begin{aligned}
\gamma^{\bar{\sigma}\sigma(2,0)}_{rr'}&=\frac{1}{2\pi{\varepsilon}^2}\Gamma_{r'}^{\bar{\sigma}}\Gamma_r^\sigma\int\limits_{-\infty}^{\infty}d\omega f(\omega-\mu_{r'})\left(1-f(\omega-\mu_r)\right)\nonumber\\
&=\frac{1}{2\pi{\varepsilon}^2}\Gamma_{r'}^{\bar{\sigma}}\Gamma_r^\sigma\frac{\mu_r-\mu_{r'}}{e^{\beta(\mu_r-\mu_{r'})}-1}\, .\end{aligned}$$ The calculation of the current rate is similar. The only difference is that one has to multiply the rates with the charge that is being transfered through the dot during the cotunneling process.
[24]{}
H. Aikawa, K. Kobayashi, A. Sano, S. Katsumoto, and Y. Iye, Phys. Rev. Lett. [**92**]{}, 176802 (2004).
T. Ihn, M. Sigrist, K. Ensslin, W. Wegscheider, and M. Reinwald, New J. Phys. [**9**]{}, 111 (2007).
A. Yacoby, M. Heiblum, D. Mahalu, and H. Shtrikman, Phys. Rev. Lett. [**74**]{}, 4047 (1995).
R. Schuster, E. Buks, M. Heiblum, D. Mahalu, V. Umansky, and H. Shtrikman, Nature [**385**]{}, 417 (1997).
Y. Ji, M. Heiblum, D. Sprinzak, D. Mahalu, and H. Shtrikman, Science [**290**]{}, 779 (2000).
W. G. van der Wiel *et al.*, Science [**289**]{}, 2105 (2000).
A. W. Holleitner, C. R. Decker, H. Qin, K. Eberl, and R. H. Blick, Phys. Rev. Lett. [**87**]{}, 256802 (2001).
M. Sigrist, T. Ihn, K. Ensslin, D. Loss, M. Reinwald, and W. Wegscheider , Phys. Rev. Lett. [**96**]{}, 036804 (2006).
S. Gustavsson, R. Leturcq, M. Studer, T. Ihn, K. Ensslin, D. C. Driscoll, A. C. Gossard, Nano Lett. [**8**]{}, 2547 (2008).
I. L. Aleiner, N. S. Wingreen, and Y. Meir, Phys. Rev. Lett. [**79**]{}, 3740 (1997).
G. L. Khym and K. Kang, Phys. Rev. B [**74**]{}, 153309 (2006).
V. Moldoveanu, M. Tolea, and B. Tanatar, Phys. Rev. B [**75**]{}, 045309 (2007).
W. Hofstetter, J. König, and H. Schoeller Phys. Rev. Lett. [**87**]{}, 156803 (2001).
J. König and Y. Gefen, Phys. Rev. Lett. [**86**]{}, 3855 (2001).
J. König and Y. Gefen, Phys. Rev. B [**65**]{}, 045316 (2002).
H. Akera , Phys. Rev. B [**47**]{}, 6835 (1993).
A. Levy Yeyati and M. Büttiker, Phys. Rev. B [**52**]{}, R14360 (1995).
C. Bruder, R. Fazio, and H. Schoeller, Phys. Rev. Lett. [**76**]{}, 114 (1996).
G. Hackenbroich and H. A. Weidenmüller, Phys. Rev. Lett. [**76**]{}, 110 (1996); Phys. Rev. B [**53**]{}, 16379 (1996).
J. Wu, B.-L. Gu, H. Chen, W. Duan, and Y. Kawazoe, Phys. Rev. Lett. [**80**]{}, 1952 (1998).
K. Kang, Phys. Rev. B [**59**]{}, 4608 (1999).
P. G. Silvestrov and Y. Imry, Phys. Rev. Lett. [**85**]{}, 2565 (2000).
U. Gerland, J. von Delft, T. A. Costi, and Y. Oreg, Phys. Rev. Lett. [**84**]{}, 3710 (2000).
K. Kang and S. C. Shin, Phys. Rev. Lett. [**85**]{}, 5619 (2000).
D. Boese, W. Hofstetter, and H. Schoeller, Phys. Rev. B [**64**]{}, 125309 (2001).
P. G. Silvestrov and Y. Imry, Phys. Rev. Lett. [**90**]{}, 106602 (2003).
R. López, D. Sánchez, M. Lee, M.-S. Choi, P. Simon, and K. Le Hur, Phys. Rev. B [**71**]{}, 115312 (2005).
D. Urban, J. König, and R. Fazio, Phys. Rev. B [**78**]{}, 075318 (2008).
L. G. G. V. Dias da Silva, N. Sandler, P. Simon, K. Ingersent, and S. E. Ulloa, Phys. Rev. Lett. [**102**]{}, 166806 (2009).
V. Moldoveanu and B. Tanatar, Phys. Rev. B, [**77**]{}, 195302 (2008).
For temperatures below the Kondo temperature, the lead electrons screen the local spin on the quantum dot, and, as a consequence, no spin-flip scattering occurs anymore, see, e.g., D. C. Langreth, Phys. Rev. [**150**]{}, 516 (1966).
S. Lindebaum, D. Urban, and J. König, Phys. Rev. B [**79**]{}, 245303 (2009).
M. Braun, J. König, and J. Martinek, Phys. Rev. B [**70**]{}, 195345 (2004).
J. Barnas and I. Weymann, J. Phys.: Condens. Matter [**20**]{}, 423202 (2008).
F. M. Souza, J. C. Egues, and A. P. Jauho, Phys. Rev. B [**75**]{}, 165303 (2007).
P. Stefanski, Phys. Rev. B [**79**]{}, 085312 (2009).
K. Hamaya, S. Masubuchi, M. Kawamura, T. Machida, M. Jung, K. Shibata, K. Hirakawa, T. Taniyami, S. Ishida, and Y. Arakawa, Appl. Phys. Lett. [**90**]{}, 053108 (2007).
B. R. Bułka, Phys. Rev. B [**62**]{}, 1186 (2000).
W. Rudziński, J. Barnaś, R. Świrkowicz, and M. Wilczyński, Phys. Rev. B [**71**]{}, 205307 (2005).
S. Braig and P. W. Brouwer, Phys. Rev. B [**71**]{}, 195324 (2005).
I. Weymann, J. Barnaś, J. König, J. Martinek, and G. Schön, Phys. Rev. B [**72**]{}, 113301 (2005).
J. König, H. Schoeller, and G. Schön, Phys. Rev. Lett. [**76**]{}, 1715 (1996); J. König, J. Schmid, H. Schoeller, and G. Schön, Phys. Rev. B [**54**]{}, 16820 (1996).
D. V. Averin and A. A. Odintsov, Phys. Lett. A [**140**]{}, 251 (1989).
D. V. Averin and Y. V. Nazarov, Phys. Rev. Lett. [**65**]{}, 2446 (1990).
I. Weymann, J. König, J. Martinek, J. Barnaś, and G. Schön, Phys. Rev. B [**72**]{}, 115334 (2005).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We report laser Doppler ophthalmoscopic fundus imaging in the rat eye, with near infrared heterodyne holography. Sequential sampling of the beat of the reflected radiation against a frequency-shifted optical local oscillator is made onto an array detector. Wide-field maps of fluctuation spectra in the 10 Hz to 25 kHz band exhibit angiographic contrasts in the retinal vascular tree without requirement of exogenous marker.
OCIS : 170.4470; 170.0110; 170.3340
author:
- 'M. Simonutti'
- 'M. Paques'
- 'J. A. Sahel'
- 'M. Gross'
- 'B. Samson'
- 'C. Magnain'
- 'M. Atlan'
title: 'Holographic laser Doppler ophthalmoscopy.'
---
Retinal blood flow plays a pivotal role in several blinding diseases such as diabetic retinopathy and vascular occlusions. While currently available optical instrumentation is well adapted for imaging retinal vessels, there are still technological limitations of currently available methods for retinal blood flow measurement that impair their clinical applications. Self-beating intensity fluctuations from a laser spot focused in the eye fundus was demonstrated to enable retinal flow assessment [@Riva1979; @Aizu1992; @Fujii1994] non-invasively. It paved the way for the development of spatial scanning techniques such as confocal spot [@Michelson1996] or line [@Ferguson2004] scanning Doppler ophthalmoscopy. Yet, laser Doppler methods for retinal blood flow mapping are still limited in term of velocity resolution and mapping capabilities.
Optical coherence tomography, commonly used for structural retinal layers imaging, was translated to a depth-resolved functional Doppler-contrast technique [@ChenMilner1997b; @Izatt1997]. Optical micro-angiography further improved depth-resolved Doppler measurements for eye fundus imaging [@AnWang2008]. Such techniques can enable quantitative and directional flow assessment, but vessel segmentation is required [@WangBower2007; @MichaelyBachmann2007].
Our approach to functional Doppler imaging is based on the detection of light fluctuations with an array detector. Multiply-scattered light yields Doppler spectra from which the directional information of flow is lost because of wave vector randomization. Nevertheless, it provides a spatially-resolved hemodynamic contrast in low-light illumination conditions. In this letter, we demonstrate experimentally the feasibility of wide-field holographic laser Doppler ophthalmoscopy, in vivo.
The experimental ophthalmoscope realized for this study is based on the heterodyne imaging scheme described in [@AtlanGrossVitalis2006]. It consists of a Mach-Zehnder laser interferometer in off-axis and frequency-shifting configuration. The detection scheme is sketched in Fig.\[fig\_setup\]. A laser diode provides the main near-infrared radiation at wavelength $\lambda = 785 \, \rm nm$, polarized linearly. In the object arm, a polarizing beam splitter cube is used to illuminate the preparation and collect the cross-polarized backscattered light component, in order to increase the relative weight of multiply scattered Doppler-shifted photons with respect to photons scattered once [@Schmitt1992]. Three adult rats were used for the preparations. Anesthesia was induced by intraperitoneal injection of 100 mg/kg ketamine and 25 mg/kg xylazine (both from Sigma-Aldrich). Topical tropicamide (CibaVision) was administered for pupil dilation. Each rat was placed on its side under the illumination beam. The head was supported so that the iris was perpendicular to the illumination axis. After administration of topical oxybuprocaine (CibaVision), a coverslip was applied on a ring surrounding the globe. Methylcellulose (Goniosol) was applied as contact medium. The incident optical field ${E_\mathrm{I}}$ is expanded to form a plane wave. Illumination power within $5 \times 5 \, {\rm mm}^2$ at the object plane is $\sim$ 1 mW. In the reference arm (local oscillator : LO), an attenuator, a half wave plate, and a beam expander (not shown) are used to control the beam power, polarization angle, and to ensure a roughly flat LO illumination of the detector. The optical frequency detuning ${\Delta \omega}$ between the two optical channels is imposed by acousto-optic modulators. The backscattered field ${E}$ is combined with the LO field ${E_\mathrm{LO}}$ with a non-polarizing beam splitter cube. The detuning ${\Delta \omega}$ shifts a component of interest of the scattered field temporal fluctuation spectrum within the actual camera bandwidth (temporal heterodyning). Moreover, a small angular tilt $\theta_0, \phi_0$ of $\sim 1 ^\circ$ ensures off-axis mixing conditions that shift the spatial frequency spectrum of the recorded object field (spatial heterodyning). The interference pattern $I= \left| {E}+ {E_\mathrm{LO}}\right|^2$ is measured by a Sony ICX 285AL charge-coupled device (CCD) array sensor (gain: 3.8 e/count), from which the central $1024 \times 1024$ pixels region is readout at $10 \, \rm Hz$. The detector is set $\sim 30 \, \rm cm$ away from the object plane. The recorded intensity pattern $I_n$ at time $t_n$ in the detector plane ($z=z_0$) is : $$\begin{aligned}
\label{eq_I_CMOS}
\nonumber I_n &=& \left| {{\cal E}}\right|^2 + \left| {{\cal E}_\mathrm{LO}}\right|^2\\
\nonumber &+& {{\cal E}}{{\cal E}_\mathrm{LO}}^* {\rm e}^{- i ({\Delta \omega}t_n + \Delta k_{x} x + \Delta k_{y} y )} \\
&+& {{\cal E}}^* {{\cal E}_\mathrm{LO}}{\rm e}^{+ i ({\Delta \omega}t_n + \Delta k_{x} x + \Delta k_{y} y )}\end{aligned}$$ where ${{\cal E}}$ and ${{\cal E}_\mathrm{LO}}$ are the complex envelopes of the scattered and LO field, respectively. $\Delta k_{x} = 2 \pi {\sin} ( \theta_0 ) / \lambda$, $\Delta k_{y} = 2 \pi {\sin} ( \phi_0 ) / \lambda$ are the projections of the difference between the LO and signal wave vectors in the transverse ($x,y$) plane. $^*$ denotes the complex conjugate. The two first terms of the right member of eq. \[eq\_I\_CMOS\] are the self-beating (homodyne) contributions of ${E}$ and ${E_\mathrm{LO}}$. The heterodyne signal of interest lies is the third term. The fourth term is the twin-image (ghost) contribution.
A typical interference pattern $I_n$ of ${E}$ beating against ${E_\mathrm{LO}}$ is shown in fig. \[fig\_080421\_holographic\_processing\](a). The magnitude of the fast spatial Fourier transform (FFT) of one recorded frame $| {\rm FFT} (I_n) |^2 $ mainly carries the LO contribution in high heterodyne gain regime, when the optical power in the LO channel is larger than in the object channel \[fig. \[fig\_080421\_holographic\_processing\](b)\]. Since the LO beam exhibits flat-field fluctuations, its self-beating contribution is gathered in the low frequency region (1) in reciprocal space $(k_x, k_y)$. Making the difference of two consecutive frames before (or after) applying the spatial Fourier transform yields $| {\rm FFT} (I_{n+1} - I_n) |^2$; it decreases substantially the relative weight of the LO self-beating term, showing up the object-against-LO beating term and the ghost term \[regions (2) and (4), respectively, in fig. \[fig\_080421\_holographic\_processing\](c)\]. The object beam self-beating term also appears, in region 3. Detection noise is assessed in region (5). Because of off-axis geometry, the heterodyne signal contribution is shifted-away, by $\pm(\Delta k_{x},\Delta k_{y})$, from self-beating fields contributions. In on-axis geometry $(\Delta k_{x} = 0,\Delta k_{y} = 0)$, all interferometric terms would overlap spatially. The object field can hence be assessed with much better sensitivity than if measured in either on-axis-only or off-axis-only conditions and used for numerical reconstruction of the signal $S$ with a discrete Fresnel transform $S = | {\rm FFT} ( (I_{n+1} - I_n) {\rm e}^{i \frac{\pi}{\lambda \Delta z} (x^2 + y^2)} )|^2$, where the distance parameter used for free-space back-propagation of the optical field is $\Delta z = z_1 - z_0 = 26.6$ cm. Once accurate focus is found, the fundus image appears as reported in fig. \[fig\_080421\_holographic\_processing\](d), in region 6, magnified in fig. \[fig\_080421\_holographic\_processing\](e). These Doppler images are time-averaged $\left< S \right>$; angular brackets $\left< \, \right>$ account for averaging over 32 frames. The star-shaped vascular tree of the rat eye fundus is clearly visible and consistent with the white-light endoscopic image performed afterwards in the same preparation, displayed in fig. \[fig\_080421\_holographic\_processing\](f).
Detuning ${\Delta \omega}$ slows down selectively the drift rate of the set of fringes associated to a given Doppler component, and sets it within the actual camera bandwidth. Frequency-selective eye fundus images $\left< S \right>$ of a healthy rat at four LO detunings (25 Hz, 845 Hz, 4.6 kHz, and 22 kHz) are reported in Fig. \[fig\_4img\_080428\]. A contrast reversal is observed between vessels and surrounding retinal tissue (and most likely in the choroid) from low to high frequency detunings \[figs. \[fig\_4img\_080428\](a) and (c)\]. Fluctuation spectra lines $\left< S \right>$ throughout a vessel cross section are reported in Fig. \[fig\_080428\_spectres\], as a function of position (a), and as a function of ${\Delta \omega}$, at four locations within the vessel (b); typical spatial standard deviation of $\left< S \right>$ are shown in the inset. It can be estimated that choroidal flow significantly contributed to the signal along the retinal vessel section $[AA']$. Within this vessel, the Doppler spectrum is clearly broader from those of immediate surrounding areas, showing that the retinal flow predominantly contributed to the signal in this specific zone. The broadest Doppler signal, which is still visible at 22 kHz in fig. \[fig\_4img\_080428\](d) is found in the optic nerve head region (circled); it is interpreted as a consequence of vessels orientation and increased density in this area.
In conclusion, we have demonstrated the feasibility of holographic laser Doppler ophthalmoscopy with near infrared radiation. The illumination power over the whole eye fundus can be kept low ($\sim$ 1 mW). High detection sensitivity, in moderate to high heterodyne gain regime, is achieved by spatiotemporal heterodyning, which enables an efficient rejection of parasitic interferometric contributions. Fluctuation spectra discrepancies below 25 kHz between the superficially vascularized regions and the surrounding retinal tissue layers provide an optical contrast suitable for angiographic mapping. A high reproducibility of the signal acquired over up to several hours, over repeated trials, and between different animals is observed. Potential applications are essentially the investigation of retinal and possibly choroidal vascular diseases. At the current stage there are yet limitations due to suboptimal lateral, depth, and time resolution, which for instance do not allow heart-beat related flow variations detection. Technical improvements are expected to circumvent these problems.
We acknowledge financial support from Agence Nationale de la Recherche (ANR-09-JCJC-0113 grant), Fondation Pierre-Gilles de Gennes (FPGG014 grant), Fondation Voir & Entendre, Région Ile-de-France, ESPCI and CNRS.
[10]{}
C. E. Riva, G. T. [Feke]{}, B. [Eberli]{}, and V. [Benary]{}. Bidirectional ldv system for absolute measurement of blood speed in retinal vessels. , 18:2301–2306, July 1979.
Yoshihisa Aizu, Kouji Ogino, Toshiaki Sugita, Tetsuya Yamamoto, Nobukatsu Takai, and Toshimitsu Asakura. Evaluation of blood flow at ocular fundus using laser speckle. , 31(16):3020, 1992.
H. Fujii. Visualisation of retinal blood flow by laser speckle flowgraphy. , 32(3):302–304, 1994.
G Michelson, B Schmauss, MJ Langhans, J Harazny, and MJ Groh. Principle, validity, and reliability of scanning laser doppler flowmetry. , 5(2):99–105, 1996.
R. D. Ferguson, D. X. [Hammer]{}, A. E. [Elsner]{}, R. H. [Webb]{}, S. A. [Burns]{}, and J. J. [Weiter]{}. Wide-field retinal hemodynamic imaging with the tracking scanning laser ophthalmoscope. , 12:5198–+, August 2004.
Z. Chen, T.E. Milner, S. Srinivas, X. Wang, A.Malakzali, M.J.C van Gemert, and J.S. Nelson. Non invasive imaging of in-vivo blood flow velocity using optical doppler tomography. , 22:1119–1121, 1997.
J. A. Izatt, M. D. Kulkarni, S. Yazdanfar, J. K. Barton, and A. J. Welch. In vivo bidirectional color doppler flow imagingof picoliter blood volumes using optical coherence tomography. , 22:1439–1441, September 1997.
L. An and R.K. Wang. In vivo volumetric imaging of vascular perfusion within human retina and choroids with optical micro-angiography. , 16(15):11438–11452, 2008.
Yimin Wang, Bradley A. Bower, Joseph A. Izatt, Ou Tan, and David Huang. In vivo total retinal blood flow measurement by fourier domain doppler optical coherence tomography. , 12(4):041215, 2007.
Roland Michaely, Adrian H. Bachmann, Martin L. Villiger, Cedric Blatter, Theo Lasser, and Rainer A. Leitgeb. Vectorial reconstruction of retinal blood flow in three dimensions measured with high resolution resonant doppler fourier domain optical coherence tomography. , 12(4):041213, 2007.
M. Atlan, M. Gross, T. Vitalis, A. Rancillac, B. C. Forget, and A. K. Dunn. Frequency-domain, wide-field laser doppler in vivo imaging. , 31(18):2762–2764, 2006.
JM Schmitt, AH Gandjbakhche, and RF Bonner. Use of polarized light to discriminate short-path photons in a multiply scattering medium. , 31(30):6535, 1992.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Warm absorbers are present in many Active Galactic Nuclei (AGN), seen as mildly ionised gas outflowing with velocities of a few hundred to a few thousand kilometres per second. These slow velocities imply a large launch radius, pointing to the broad line region and/or torus as the origin of this material. Thermal driving was originally suggested as a plausible mechanism for launching this material but recent work has focused instead on magnetic winds, unifying these slow, mildly ionised winds with the more highly ionised ultra-fast outflows. Here we use the recently developed quantitative models for thermal winds in black hole binary systems to predict the column density, velocity and ionisation state from AGN. Thermal winds are sensitive to the spectral energy distribution (SED), so we use realistic models for SEDs which change as a function of mass and mass accretion rate, becoming X-ray weaker (and hence more disc dominated) at higher Eddington ratio. These models allow us to predict the launch radius, velocity, column density and ionisation state of thermal winds as well as the mass loss rate and energetics. While these match well to some of the observed properties of warm absorbers, the data point to the presence of additional wind material, most likely from dust driving.'
author:
- |
Misaki Mizumoto$^{1}$[^1], Chris Done$^{1,3}$, Ryota Tomaru$^{2,3}$, & Isaac Edwards$^{1}$\
$^{1}$Centre for Extragalactic Astronomy, Department of Physics, University of Durham, South Road, Durham DH1 3LE, UK\
$^{2}$Department of Physics, The University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-0033, Japan\
$^{3}$Kavli Institute for the Physics and Mathematics of the Universe (WPI), University of Tokyo, Kashiwa 277-8583, Japan\
bibliography:
- '00.bib'
date: 'Accepted XXX. Received YYY; in original form ZZZ'
title: Thermally driven wind as the origin of warm absorbers in AGN
---
\[firstpage\]
galaxies: nuclei – X-rays: galaxies – X-rays: ISM – quasars: general
Introduction
============
X-ray observations of active galactic nuclei (AGNs) often reveal the presence of mildly ionised material, with multiple absorption lines from partially ionised oxygen, neon, and iron in the 0.5–2 keV bandpass. These lines are blueshifted, indicating outflow velocities of a few hundred to a few thousand km s$^{-1}$. These ‘warm absorbers’ are best studied with high resolution X-ray data from gratings, the Reflection Grating Spectrometer (RGS) on board [*XMM-Newton*]{} (e.g. @sak01 and the compilations of @lah14) and the High/Low Energy Transmission Grating (HETG/LETG) on [*Chandra*]{} (e.g. @kas00b [@kaa00] and the compilation of @mck07). The inferred mass outflow rate is often comparable to or even larger than the mass accretion rate onto the supermassive black hole, so this must impact on the available material for accretion [@blu05; @lah14]. However, its kinetic energy is rather small, generally less than 1% of bolometric luminosity [@blu05; @lah14]. Thus the warm absorbers are probably not important in setting $M$-$\sigma$ relation as these typically require a wind with kinetic power of 0.5–5% of the bolometric luminosity [@hop10].
There are three main models for producing the warm absorber outflows: radiation pressure, magnetic force, and thermal pressure. The radiation force overcomes gravity when $L > L_{\rm Edd}$, where $L_{\rm Edd}$ is the Eddington limit, which is defined from Thompson scattering on free electrons. Most AGNs with warm absorbers are sub-Eddington, so continuum radiation pressure cannot be the main mechanism. However, there can be other processes which enhance the coupling of the gas to the radiation field. An additional cross-section, $\sigma_i$, leads to a decrease in the luminosity at which a wind can be driven, to $L_{\rm Edd}/ (1+ M)$ where $M=\sigma_i/\sigma_{\rm T}$ is the force multiplier. The force multiplier can be very large for low ionisation gas due to the enormous numbers of ultraviolet (UV) line (bound-bound) as well as edge (bound-free) transitions, allowing radiation pressure to drive strong winds in sub-Eddington AGNs (e.g. @pro04). However, this is most efficient where there is strong UV radiation, which is the inner disc in most bright AGN. Typically, winds have terminal velocity which is of order the escape velocity from their launch point so these winds are much faster than those observed in warm absorbers [@blu05].
Winds can also be driven by centrifugal acceleration along magnetic field lines anchored in the disc [@bla82; @kon94; @fuk10]. However, these winds depend on the (currently unknown) magnetic field configuration, so are impossible to calculate [*ab initio*]{}.
The third wind launch mechanism is thermal driving [@beg83; @woo96], which was first applied to the warm absorbers by @kro95 based on numerical studies by @bal93. X-rays from the AGN heats any illuminated material up to the Compton temperature, $T_{\rm IC}$. This is determined only by the spectrum of the radiation, as photons with $h\nu\ll kT_{\rm IC}$ will Compton cool the material, whereas photons with $h\nu\gg kT_{\rm IC}$ heat it. The heated skin expands due to the pressure gradient, producing a thermally driven wind at radii where the sound speed exceeds the local escape velocity.
In this paper we focus on the thermally driven wind model as these can be rather well predicted given the spectral energy distribution (SED) and luminosity of the source. We use the approach of @don18, who recast the analytic thermal wind solutions of @beg83 into a more tractable form, and used them to show that this is most likely the origin of the narrow, highly ionised, blueshifted absorption seen in black hole binaries (see also @hig18 [@tom19]). Here we apply this instead to AGN, predicting the thermal wind using the SED models of @kub18 to track how the Compton temperature varies with mass and $L/L_{\rm Edd}$. We compare these predictions with the data to assess the viability of a thermally driven wind model for the origin of warm absorbers in AGN.
Thermal winds in AGN
====================
![AGN SED model for $M_{\rm BH}=10^7 M_\odot$. The vertical axis has an arbitrary unit of $\nu F_\nu$. The colour lines are for Eddington ratios of $\log (L/L_{\rm Edd})=-1.5, -1.4, \dots, -0.1$, from bottom to top at 0.1 keV. The spectra become softer for larger Eddington ratios in this range. We use the same spectral shape for $\log (L/L_{\rm Edd})\leq-1.7$ (black line), which normalisation is changed for the different luminosities.[]{data-label="fig:SED"}](./m1e7.eps){width="0.8\columnwidth"}
![Compton temperature (upper), Compton radius (medium), and critical luminosity (low) for different Eddington ratios and black hole masses. In the low Eddington ratio regime ($\log(L/L_{\rm Edd})<-1.7$) no difference is seen among different black hole masses (black lines).[]{data-label="fig:param"}](./param.eps){width="1.1\columnwidth"}
Irradiation by X-rays from the inner region heats material to the Compton temperature, defined by $$kT_{\rm IC}=\frac{1}{4}\frac{\int E L(E) dE}{\int L(E) dE},$$ where $L(E)$ is luminosity. $T_{\rm
IC}$ depends on only the shape of SED; harder spectra produce higher $T_{\rm IC}$. The heated material expands with the sound velocity of $c_{\rm
IC}=\sqrt{kT_{\rm IC}/\mu}$, where $\mu=0.63\,m_p$ is the mean particle mass of ions and electrons for solar abundances. The Compton radius ($R_{\rm IC}$) is where this local sound speed exceeds the escape velocity, i.e., $R_{\rm IC}=GM_{\rm
BH}/c_{\rm IC}=6.4\times10^4 T_{{\rm IC},8}^{-1} R_g$, where $T_{{\rm IC},8}=10^{-8}T_{\rm IC}$ and $R_g=GM/c^2$ is the gravitational radius.
The spectral shape is critical to the thermal wind properties. Multiple papers have shown that the AGN SED changes systematically as a function of mass and mass accretion rate [@vas07; @vas09; @jin12; @don12]. We use the specific model [qsosed]{} of @kub18 to model them. This captures the main trends seen in the data by assuming that the accretion flow forms three different regions, an outer standard disc where the emission thermalises to the local blackbody temperature, an intermediate region where the accretion power is dissipated higher up in the photosphere, producing a ‘soft X-ray excess’ warm Comptonisation region, and an inner hot flow which is assumed to have constant hard X-ray luminosity, $L_{\rm X}=0.02L_{\rm
Edd}$. At low luminosity, almost all of the accretion power is taken by the hot flow, whereas for $L\sim L_{\rm Edd}$, this forms only a very small fraction of the bolometric luminosity, as required (see @kub18 for details). Fig. \[fig:SED\] shows examples of the assumed SED for a black hole of mass $10^7M_\odot$ and zero spin for $L/L_{\rm Edd}=0.03$ to $1$. For $L\le 0.03L_{\rm Edd}$ the spectrum is assumed to be purely a power law with photon spectral index of $\Gamma=1.7$, i.e., that the entire disc is replaced by an advection dominated accretion flow (ADAF).
Fig. \[fig:param\] shows the resulting Compton temperature, $T_{\rm IC}$, and corresponding Compton radius, $R_{\rm IC}$, at which material at this temperature is able to escape the black hole gravity. The softer SED at higher $L/L_{\rm Edd}$ gives a lower Compton temperature, so the radius at which the heated material can escape is larger. We show these predictions for a range of black hole masses, from $10^6M_\odot$ (red) to $10^{10}M_\odot$ (magenta), but this makes little difference as the Compton temperature is much more sensitive to higher energy photons (essentially it is a spectral average of $\nu^2 F_v$ rather than $\nu F_v$). Thus the expected decrease in $T_{\rm IC}$ due to the lower disc temperature at higher mass only becomes noticeable at the highest $L/L_{\rm Edd}$, where the disk component almost completely dominates the spectrum.
![Specific mass loss rate for $M_{\rm BH}=10^7\,M_\odot$. Peaks constantly appear at $R\sim0.2R_{\rm IC}$. []{data-label="fig:specific"}](./specificmassloss_m1e7.eps){width="0.8\columnwidth"}
The luminosity is of secondary importance to the spectral shape as long as it is high enough to heat the gas up to the Compton temperature before it escapes. However, if the luminosity is lower than this critical luminosity, $L_{\rm crit}$, then the gas is heated only to a characteristic temperature ($T_{\rm ch} < T_{\rm IC}$). The critical luminosity and the characteristic temperature are written as $$\begin{split}
L_{\rm crit}&=\frac{1}{8}\left(\frac{m_e}{\mu}\right)^{1/2}\left(\frac{m_ec^2}{kT_{\rm IC}}\right)^{1/2}L_{\rm Edd}\\
&\simeq 0.03T_{\rm IC,8}^{-1/2}L_{\rm Edd}.
\end{split}$$ and $$T_{\rm ch}=T_{\rm IC}\left(\frac{L}{L_{\rm crit}}\right)^{2/3}\left(\frac{R}{R_{\rm IC}}\right)^{-2/3}$$ [@beg83; @don18]. The lower panel of Fig. \[fig:param\] shows $L/L_{\rm crit}$ for each of the SEDs. The luminosity is around or above the critical luminosity for $\log (L/L_{\rm Edd})\geq-1.5$, predicting that all AGN with some UV emitting outer disc have $L>L_{\rm crit}$ in our models, so can efficiently produce thermal winds.
![Geometry of the torus and the self-shielding radius. $R_{\rm torus}$ is the radius of the torus, and $R_{\rm dust}$ is the distance from the central black hole to the inner radius of the torus. $R_{\rm shield}$ is the distance from the black hole to the contact point of the torus, beyond which the wind cannot be launched because the seed gas does not exist. $\theta$ is set to be $34^\circ$. []{data-label="fig:geometry"}](./geometry.eps){width="0.8\columnwidth"}
@woo96 provided an equation for the specific mass loss rate (per unit area) at radius $R$ from hydrodynamic simulations as $$\begin{aligned}
\begin{autobreak}
\dot{m}(R)=\frac{L/c}{4\pi R^2 \Xi_{\rm max} c_{\rm ch}}
\left\{\frac{1+[(0.125L/L_{\rm crit}+0.00382)/\zeta]^2}{1+[(L/L_{\rm crit})^4(1+262\zeta^2)]^{-2}}\right\}^{1/6}
\exp\left\{\frac{-[1-(1+0.25\zeta^{-2})^{-1/2}]^2}{2\zeta}\right\},
\end{autobreak}\label{eq:specific}\end{aligned}$$ Here the characteristic sound speed is $c_{\rm ch}=\sqrt{kT_{\rm
ch}/\mu}$, $\zeta=R/R_{\rm IC}$, and the pressure ionisation parameter marking the base of the X-ray heated atmosphere, $\Xi_{\rm max}$, is assumed constant at $\sim 40$ [@don18]. This equation matches well with the analytic expectations [@beg83; @don18] and more recent hydrodynamic results [@hig18]. Fig. \[fig:specific\] shows the results of equation (\[eq:specific\]), which are calculated with a step of $0.05R_{\rm
IC}$. These peak at $R \sim 0.2R_{IC}$ as long as the heating is rapid ($L>L_{\rm crit}$), as is the case here for all AGN with $L/L_{\rm Edd}\ge 0.03$. Hence we consider this to be the launch radius for thermal winds. @don18 showed that in the more general case the wind launch radius is $R_{\rm in}=0.2R_{\rm IC}$ for $L/L_{\rm crit}>1$ and $0.2R_{\rm
IC}/(L/L_{\rm crit})$ for $L/L_{\rm crit}\leq1$.
{width="0.85\columnwidth"}
Illuminated material
====================
Thermal winds are produced where the AGN spectrum illuminates cool material, with typical launch radii of $10^5\,R_g$ for most of the SEDs considered (Fig. \[fig:param\]). In black hole binaries it is common for the outer accretion disc to extend to these radii, but in AGN self gravity should truncate the disc at a radius [@lao89] of $$R_{\rm sg}=2150M_9^{-2/9}\dot{m}^{4/9}\alpha^{2/9},$$ where $M_9=M_{\rm BH}/10^9M_\odot$, $\dot{m}=\dot{M}_{\rm
acc}/\dot{M}_{\rm Edd}$, and $\alpha$ is the disc viscosity parameter of @sha73, assumed here as $\alpha=0.02$ [@sta04]. Thus thermal winds cannot be produced from AGN discs.
Nonetheless, there is gas at larger radii, in the self gravitating regime, from either the broad line region (BLR) and/or molecular torus [@kro01]. Both of these are radially extended structures, so we consider that there is a continuous distribution of gas connecting the accretion disc to the molecular torus. Nonetheless, there are characteristic radii which can be identified within this. Reverberation mapping shows that this gas produces the broad H$\beta$ line at $$\begin{split}
%\begin{autobreak}
&\log(R_{\rm BLR}/1\,{\rm lt \mathchar"712D day}) =\\
&\qquad\quad(1.527\pm0.031)
+0.533^{+0.035}_{-0.033} \log (\lambda L_\lambda
/10^{44}\,\mathrm{erg\,s}^{-1})
%\end{autobreak}
\end{split}$$ [@ben13], where $\lambda L_\lambda$ is taken at 5100Å from the AGN SED model (Fig. \[fig:SED\]).
The inner radius of the dust torus ($R_{\rm dust}$) is determined by the dust sublimation temperature, $T_{\rm sub}\simeq1500$ K [@bar87]. Dust evaporates when the irradiation flux reaches $L/(4\pi R^2) = \sigma_{\rm SB}T_{\rm sub}^4$, where $\sigma_{\rm SB}$ is the Stefan-Boltzmann constant. Therefore, $$R_{\rm dust}=\left(\frac{L}{4\pi\sigma_{\rm SB} T_{\rm sub}^4}\right)^{1/2}.$$
There should also be an outer radius for illumination of the torus, depending on the geometry. When we simply assume a circular cross-section torus, it has a self-shielding radius ($R_{\rm shield}$, see Fig. \[fig:geometry\]), as $R_{\rm shield}/R_{\rm
dust}=\cos\theta/(1-\sin\theta)$, where $\theta$ is the half opening angle subtended by the torus from the central black hole. We set $\theta=34^\circ$ to match the hydrodynamic simulation of @dor08a, which assume that the torus obstructs all lines of light of AGN with inclinations of $i>56^\circ$. Therefore $R_{\rm
shield}=1.88R_{\rm dust}$. The dust torus may not have a simple toroidal shape, but be conical/flared made by a dusty wind (e.g. @eli06 [@dor12a; @cha16]). In this case the dusty gas in the torus can supply more gas to the wind, which will be discussed in section \[sec7\].
The BLR and inner and outer torus radii all scale as $R\propto L^{1/2}\propto (M\dot{m})^{1/2}$. Hence $R/R_g \propto
(\dot{m}/M)^{1/2}$ so the typical radii of these gas structures increases with $L/L_{\rm Edd}$ but decreases with increasing black hole mass. Fig. \[fig:param2\] shows how the H$\beta$ broad line radius (green) and inner/outer (blue/black) radius of the dusty torus all decrease systematically in terms of $R_g$ for increasing black hole masses of $M=10^6, 10^7, 10^8$ and $10^9M_\odot$, while the predicted thermal wind launch radius (red crosses) is approximately constant in $R/R_g$. Thus while the lowest mass AGN can launch a thermal wind from radii close to the H$\beta$ emitting part of the BLR, the highest mass AGN cannot launch a thermal wind at all as the torus self shields before the material can escape in a wind with our assumed geometry. The range in wind producing radii is largest when $\log(L/L_{\rm Edd})\sim-1.5$, and at lowest black hole masses, so highlighting these system parameters as the ones where the thermal wind mass loss rates should be largest.
Wind mass, momentum and energy loss rates
=========================================
The total mass loss rate in the wind is the integration of the specific mass loss rate up to $R_{\rm shield}$, i.e., $$\dot{M}_{\rm wind}=\int_0^{R_{\rm shield}} \dot{m}(R)\times2\times2\pi R dR.$$ Fig. \[fig:total\_massloss\] shows the ratio of the total mass loss rate to the mass accretion rate ($\dot{M}_{\rm acc}=L/\eta c^2$, where $\eta=0.057$ for a non-spinning black hole). The mass-loss is indeed larger when the black hole mass is smaller and the Eddington ratio is closer to $10^{-1.5}$ as expected from the discussion above. It exceeds the mass accretion rate for the most efficient case.
![The total mass-loss rate. The vertical axis is normalised by the mass accretion rate.[]{data-label="fig:total_massloss"}](./total_massloss4.eps){width="0.85\columnwidth"}
We calculate the expected wind velocity following @don18 as the mass loss weighted average sound speed ($c_{\rm ch}$). The wind velocity becomes slower for larger Eddington ratios (Fig. \[fig:velocity\]), which due to their lower $T_{\rm IC}$ (Fig. \[fig:param\]). This enables us to calculate the momentum and kinetic energy carried by the wind, as shown in Figs. \[fig:momentum\] and \[fig:energy\]. Because the wind velocity is much slower than the light velocity ($\lesssim1500$ km s$^{-1}=0.05c$), the momenta are much less than $L_{\rm
AGN}/c$ and the energy loss rates do not exceed $0.5\%$ of the AGN luminosity. Thermal winds make only a small contribution to the AGN feedback and evolution of the host galaxy. This is consistent with results of full radiation hydrodynamic models of X-ray heated thermal winds from the torus [@dor08a; @dor08b].
![Wind velocity, calculated as the mass loss weighted average sound speed[]{data-label="fig:velocity"}](velocity.eps){width="0.8\columnwidth"}
![Momentum of the winds. The vertical axis is normalised by $L_{\rm AGN}/c$.[]{data-label="fig:momentum"}](./momentum2.eps){width="0.8\columnwidth"}
![The energy-loss rate. The vertical axis in the upper panel is normalised by the AGN luminosities.[]{data-label="fig:energy"}](./energy2.eps){width="0.8\columnwidth"}
Observational characteristics of thermal winds in AGN
=====================================================
While the thermal wind does not carry enough energy to determine AGN feedback, it may still produce the observed warm absorber features. @don18 show that the column density and ionisation along the line of sight could be analytically estimated for the black hole binary case, (i.e., when the wind is launched from the disc and there are no obstacles such as torus in the line of sight). They could match to the results of radiative hydrodynamic models [@woo96] by assuming radial streamlines and $n(R,\mu) =
n_0(R)(1-\mu)$, where $n_0$ is the number density for an edge-on inclination and $\mu=\cos i$. However, in the AGN case, the geometry is more complex, with the BLR clouds and torus having some scale height so the wind does not extend down to edge-on inclination angles. Our torus geometry has opening angle of $56^\circ$, so we assume that the same amount of thermal wind gas exists within $0^\circ<i<56^\circ$, i.e., $$\int_0^1\!n_0(R)(1-\mu) d\mu
= A \int_{\mu_0}^1\!n_0(R)(1-\mu) d\mu,$$ where $\mu_0=\cos 56^\circ$. This gives a flat disc to funnel correction coefficient of $A=(1-\mu_0)^{-2}$. The column density of the wind is then $$\begin{split}
N_{\rm H}(\mu)&=A\int_0^{R_{\rm shield}}\! n_0(R)(1-\mu) dR\\
&=\frac{(1-\mu)}{(1-\mu_0)^2}\int_0^{R_{\rm shield}}\! \frac{\dot{m}(R)}{c_{\rm ch}(R)m_{\rm I}} dR, \label{eq:NH}
\end{split}$$ where $m_{\rm I}=1.26m_{\rm p}$ is the mean ion mass per electron.
![Column densities of the wind when $i=30^\circ$[]{data-label="fig:column"}](./column2.eps){width="0.8\columnwidth"}
![Ionisation parameters of the wind. The cases of $N_{\rm H}<10^{19}$ cm$^{-2}$ are not shown.[]{data-label="fig:logxi"}](./logxi2.eps){width="0.8\columnwidth"}
Fig. \[fig:column\] shows the column density predicted for $i=30^\circ$ for a range of mass and mass accretion rates. We also calculate the ionisation parameter as $\xi(R)=L_{\rm
ion}/n(R)R^2$, where $L_{\rm ion}$ is the luminosity from 13.6 eV to 13.6 keV. Our radial streamlines mean that this is constant with radius, with the value shown Fig. \[fig:logxi\]. These predictions result in typical columns of $10^{20-21}~\mathrm{cm}^{-2}$ at $\log\xi\sim 2-3$ for black hole masses $<10^7\ M_\odot$, which are close to the observed values for warm absorbers, as are the typical velocities of few hundred to a few thousand km s$^{-1}$ (see Fig. \[fig:velocity\]).
Comparison to observations
==========================
The predictions above for thermal winds appear to be quite well matched to the observed properties of warm absorbers. However, the individual AGNs all have different mass, luminosity, inclination, and presumably spin. This makes it difficult to compare the change in wind properties using tracks of constant mass. Therefore, we instead make predictions for each individual system from its own estimated mass and luminosity. We use the radio quiet AGN sample of @lah14, which tabulates mass and luminosity along with the observed warm absorber properties, keeping our assumption of zero spin.
![Eddington ratio dependence of warm absorber parameters shown in @lah14, with our models. Each model point corresponds to each black hole mass and Eddington ratio. The same colour show the same black hole mass, from red ($M_{\rm BH}=10^6\,M_\odot$) to blue ($M_{\rm BH}=10^9\,M_\odot$). Some targets have more observation points than the model ones because they have multiple warm absorber components and/or because the model points are outside of the shown region. In the middle and lower panels, the start points in the model arrows (with short black horizontal lines) show when $i=56^\circ$, and the end points show when $i=30^\circ$. []{data-label="fig:warmabs"}](./vel_Edd.eps "fig:"){width="0.75\columnwidth"} ![Eddington ratio dependence of warm absorber parameters shown in @lah14, with our models. Each model point corresponds to each black hole mass and Eddington ratio. The same colour show the same black hole mass, from red ($M_{\rm BH}=10^6\,M_\odot$) to blue ($M_{\rm BH}=10^9\,M_\odot$). Some targets have more observation points than the model ones because they have multiple warm absorber components and/or because the model points are outside of the shown region. In the middle and lower panels, the start points in the model arrows (with short black horizontal lines) show when $i=56^\circ$, and the end points show when $i=30^\circ$. []{data-label="fig:warmabs"}](./Nh_Edd.eps "fig:"){width="0.75\columnwidth"} ![Eddington ratio dependence of warm absorber parameters shown in @lah14, with our models. Each model point corresponds to each black hole mass and Eddington ratio. The same colour show the same black hole mass, from red ($M_{\rm BH}=10^6\,M_\odot$) to blue ($M_{\rm BH}=10^9\,M_\odot$). Some targets have more observation points than the model ones because they have multiple warm absorber components and/or because the model points are outside of the shown region. In the middle and lower panels, the start points in the model arrows (with short black horizontal lines) show when $i=56^\circ$, and the end points show when $i=30^\circ$. []{data-label="fig:warmabs"}](./logxi_Edd.eps "fig:"){width="0.75\columnwidth"}
The upper panel of Fig. \[fig:warmabs\] shows a comparison between the observed wind velocities (circle bins with error bars) and the velocities we have calculated (crosses). The thermal wind model velocities are clearly consistent with the observed velocities of $v\sim1000$ km s$^{-1}$.
Next, we compare the column density in the middle panel in Fig. \[fig:warmabs\]. The calculated $N_{\rm H}$ values depend on the inclination angle from $0^\circ\leq i \leq 56^\circ$ (see equation \[eq:NH\]), and thus we plot the upper limit (when $i=56^\circ$) as a vertical arrow. The end point of the arrow corresponds to $i=30^\circ$. Again the agreement with the observed columns is generally good, with some exceptions.
Finally, we show our results for the ionisation parameter and the corresponding observed values are shown in the bottom panel in Fig. \[fig:warmabs\], again with a vertical arrow marking the lower limit of ionisation expected from the possible range of inclination angles. Here the predicted values form an upper bound of the observed ionisation states, but there are many points which lie 2–4 orders of magnitude below. Most of these are from the same objects as are also detected with absorption columns at the much higher ionisation stage, as most of the warm absorbers detected in AGN are multi-phase.
We first consider whether the thermal wind can naturally produce such a multi-phase absorber. Studies of individual objects show that many of these phases appear to be in pressure equilibrium (e.g. @kro03 [@kro07; @kro09; @net03]). The red line in Fig. \[fig:Scurve\] shows the thermal equilibrium curves for gas in pressure balance for our SED for $\log(L/L_{\rm Edd})=-1.3$ and $M_{\rm BH}=10^7M_\odot$. Regions to the lower right of the line have heating larger than cooling, while the upper left has cooling larger than heating. Where the thermal stability curve has an S-bend, this indicates that material on the middle branch is thermally unstable as it will rapidly heat up to join the upper branch, or cool down to join the lower branch. Two phase gas can stably co-exist in pressure equilibrium for $\Xi$ between the minimum value on the upper branch and the maximum value on the lower [@kro81]. There are two regions of S curve for this SED, one at $\log\Xi\sim
1.3$ where the gas can be at $10^7$ K on the upper branch, or $10^6$ K on the lower, and another separate region at $\log\Xi\sim 1.2$ where the gas can be at $\sim 6\times 10^5$ K or $1.5\times 10^5$ K. Since pressure $\propto nT$ then these separate phases have densities which are higher in proportion to their lower temperature, so they are characterised by a standard ionisation parameter ($\xi=L/nR^2$) which is lower [@roz06; @gon06].
@hol07 introduced the absorption measure distribution (AMD) which defines the distribution of column density as a function of the standard ionisation parameter. They interpret a distinct dip in this distribution for the warm absorber gas in several objects as evidence for the thermal instability (see @beh09 [@hol12]).
The thermal instabilities are always triggered in material with large enough column density (radiation pressure confinement: @ste14 [@goo16; @adh19]). Fig.\[fig:cooling\] shows the temperature versus column density into an irradiated slab in pressure balance for this SED and black hole mass. Given that Fig. \[fig:param\]b shows that the luminosity is around the critical luminosity, this means that the Compton heating time is only just short enough to get material onto the upper branch, i.e. that its maximum ionisation is quite close to the critical ionisation parameter $\Xi_{\rm max}\sim 40$. The material then starts out close to the instability point so a multi-phase outflow may quite easily develop even with a total column of only $10^{21}$ cm$^{-2}$.
However, the more material drops down onto the lower branch, the less there is on the upper branch. The predicted column densities with all the material in the upper branch were already a little low, so putting more material on the lower branch to decrease the tension with the ionisation parameters will increase the tension with the column density. This is especially an issue with the high mass accretion rates, where the column density is predicted to be lower, and there is no true instability (black line in Fig. \[fig:Scurve\]). Instead, it seems more likely that there are additional processes at work which enhance the column density of wind material.
![Thermal equilibrium curves for gas in pressure balance for our SED model. The surface density and ionisation parameter are assumed to be $n=10^8$ cm$^{-3}$ and $\log\xi=5$, respectively.[]{data-label="fig:Scurve"}](./Scurve.eps){width="0.8\columnwidth"}
![Temperature of the slab gas irradiated by our SED model for $M_{\rm BH}=10^7M_\odot$ and $\log(L/L_{\rm Edd})=-1.3$. The setting is same as Fig. \[fig:Scurve\].[]{data-label="fig:cooling"}](./NH_T.eps){width="0.8\columnwidth"}
The thermal instability is also required to explain the UV absorption lines. In some AGNs (e.g. NGC 4151) the lower ionisation phase of the warm absorber gas can be traced through both soft X-ray absorption and UV absorption. The low ionisation parameter means that it cannot be on the upper branch of the thermal heating/cooling curve, so it has much lower temperature than required to escape as a thermal wind [@cre07]. This material could be accelerated via thermal driving to its current velocity, and then cool (and probably clump) via the thermal instability.
The requirement for additional wind material is seen even more clearly in radio loud AGN. These can show warm absorbers, though the columns are generally towards the lower range of those seen in radio quiet AGN [@meh19]. However, the black holes are very massive ($M_{\rm BH}\sim10^{8-10}M_\odot$) where our model gives a thermal wind launching radius which exceeds the self-shielding radius of the torus (see Fig. \[fig:param2\]). The thermal wind column density is then negligible.
Discussion {#sec7}
==========
Our thermal wind model successfully predicts the velocity of the warm absorbers (not the ultrafast outflows) seen in AGN, but the column density of material is often a little lower than observed especially if this material has to be shared between multiple phases of gas with different ionisation state. Instead, it seems more likely that there are additional sources of material and/or launch mechanisms for these outflows. This is especially the case for higher mass AGN, where the launch radius for thermal winds is much larger than the inner edge of the dust torus. Our assumed circular torus geometry means that there is self-shielding at these radii, so very little wind is predicted. A different assumed torus geometry would change this, e.g. a torus with conical/flared rather than circular cross-section would allow material further out to be illuminated, increasing the total column density of the thermal wind predictions as these go as $N_{\rm H}\propto \log (R_{\rm out}/R_{\rm in})$. Similarly, a clumpy rather than smooth torus (e.g. @nen08) will allow more irradiation at larger radii as well as make a more inhomogeneous thermal wind.
The torus shape should be determined self consistently, as it is sculpted by the mass loss via the wind. Hydrodynamic simulations of this are shown in @dor08a [@dor08b] and [@dor09], where they start with a smooth torus of finite optical depth, and follow the evolution as the X-ray heating produces a thermal wind. However, in these simulations, the inner edge of the torus is arbitrary rather than being tied to the dust sublimation radius. They assume a torus central radius of $R_0=0.5-1.5$ pc, with shape such that the inner edge is at $0.23R_0$ (for their $d=2.5$), which is much larger than the inner edge of the dust sublimation region at 0.02–0.04 pc for their black hole mass of $10^6M_\odot$ and assumed $L/L_{\rm Edd}=0.1-0.5$ (see Fig. \[fig:param2\]). They predict large column densities of gas at high inclinations, but much of this material is part of the torus rather than being in a wind as their torus density distribution means that it is optically thin around its edge. This is more realistic than the infinitely optically thick torus used here, but cannot explain the observed [*outflow*]{} columns unless the torus (and/or BLR) is itself part of an additional wind rather than just rotating as assumed here. Nor can it explain the observed statistics of the incidence of warm absorbers in more than 50% of AGN as these large columns only cover a small fraction of the torus opening [@dor09]. The question then becomes what drives the wind which provides the material for the torus and BLR.
Dust is an obvious component of the torus, and may well also be responsible for the underlying driving mechanism for the BLR [@cze11]. Dust opacity is much larger than the electron scattering cross-section, particularly for UV wavelengths, so there is again a ‘force multiplier’ effect, reducing the effective Eddington limit. @fab08 showed that this is $\ge 100$ for realistic AGN SEDs for columns $\lesssim 10^{21}$ cm$^{-2}$. Thus all of the bright AGN considered here are effectively super-Eddington for low columns of dusty gas, so should power a wind from the surface of the torus. Hydrodynamic simulations including dust have focused on the role of trapped infrared radiation in producing the scale height of the torus [@dor12a; @dor12b], but models including dynamics show a strong dust driven wind from the inner edge of the torus [@wad12; @wad15; @cha16; @dor16].
Hence it seems most likely that the warm absorbers are thermal-radiative winds, rather than solely thermal, and where the radiation pressure is mainly on the dust grains present in the torus.
Conclusion
==========
@kro01 showed that the warm absorbers seen in AGN result could plausibly result from a thermal wind from the torus, and highlighted the role of the thermal instability in producing the observed multi-phase absorbing gas. Thermal winds are sensitive to the SED, and there is now more observational data on how the SED changes as a function of mass and mass accretion rate in AGN. We use the specific models of @kub18 which can reproduce these general trends ([qsosed]{}). These predict that the Compton temperature decreases with Eddington ratio, which increases the radius from which the wind is launched in terms of $R/R_g\propto \dot{m}^{-1}$. The irradiated material in the BLR and torus has radius $R/R_g\propto (\dot{m}/M)^{1/2}$ so thermal winds are launched from the BLR for low mass, low mass accretion rate AGN, and from the outer edge of the torus for high mass, high mass accretion rates. We use the analytic models of thermal winds developed by @don18 to make quantitative predictions of the amount of column produced by thermal winds in AGN. We show generally that these predictions are in tension with the observations, especially if the thermal instability is invoked in order to produce the observed low ionisation gas components. Instead, dust driven winds from the inner edge of the torus is more likely to be the origin of much of the observed outflow material, with the X-ray heating contributing more to the observed multi-phase ionisation structure. Hydrodynamic simulations of this complex, dynamic environment show that this produces outflows and obscuration from the torus in line with the observations [@wad15; @cha16; @dor16].
Magnetic winds are not required in this picture. There are two main arguments used as evidence for such winds. In NGC 4151 (and other well studies AGN) the lower ionisation phase traced by UV absorption must have much lower temperature than required to escape as a thermal wind [@cre07]. This is true, but this material could be accelerated via either thermal or dust driving and then cool (and probably clump) via the thermal instability. The second piece of evidence is that there is an anti-correlation of warm absorber column density with radio loudness in a sample of radio loud AGN [@meh19]. This could link the winds causally to the jet, which is clearly a magnetic structure and similar anti-correlation of wind and jet is seen directly in the stellar mass black hole binaries [@nei09; @pon12]. Nonetheless, in binaries the link to the jet is more likely to be via the changing SED with changing spectral state rather than causally via the jet magnetic field [@tom19]. The origin of the correlation in the small sample of radio loud AGN is yet to be clarified, but dust and thermal driving together clearly have the potential to explain most of the warm absorbers seen.
Acknowledgements {#acknowledgements .unnumbered}
================
MM is financially supported by Japan Society for the Promotion of Science (JSPS) Overseas Research Fellowship. This work is partly supported by the Science and Technology Facilities Council (STFC) grant ST/P000541/1 (CD), Kavli Institute for the Physics and Mathematics of the Universe (IPMU) funding from the National Science Foundation No. NSF PHY17-48958 (CD), and JSPS KAKENHI Grant Number 19J13373 (RT).
\[lastpage\]
[^1]: E-mail: [email protected], [email protected] (MM)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This study performs parameter inference in a partial differential equations system of pulmonary circulation. We use a fluid dynamics network model that takes selected parameter values and mimics the behaviour of the pulmonary haemodynamics under normal physiological and pathological conditions. This is of medical interest as it enables tracking the progression of pulmonary hypertension. We show how we make the fluids model tractable by reducing the parameter dimension from a 55D to a 5D problem. The Delayed Rejection Adaptive Metropolis (DRAM) algorithm, coupled with constraint nonlinear optimization is successfully used to learn the parameter values and quantify the uncertainty in the parameter estimates. To accommodate for different magnitudes of the parameter values, we introduce an improved parameter scaling technique in the DRAM algorithm. Formal convergence diagnostics are employed to check for convergence of the Markov chains. Additionally, we perform model selection using different information criteria, including Watanabe Akaike Information Criteria.'
author:
- 'L. Mihaela Paun'
- 'M. Umar Qureshi'
- Mitchel Colebank
- 'Nicholas A. Hill'
- 'Mette S. Olufsen'
- 'Mansoor A. Haider'
- Dirk Husmeier
bibliography:
- 'references.bib'
title: MCMC methods for inference in a mathematical model of pulmonary circulation
---
Introduction
============
The cardiovascular circulation is composed of the systemic circulation and the pulmonary circulation. Extensive work has been done to model the systemic circulation [@ma1], however pulmonary hypertension is one of the leading causes of right heart failure [@ma2]. Our work focuses on predicting the observed haemodynamic behaviour in the pulmonary circulation under normal and pathological conditions (hypoxia). Hypoxia is a pathological condition in which the body tissues are not sufficiently well oxygenated, leading to pulmonary hypertension.
Pulmonary hypertension (PH) is characterised by high mean blood pressure in the lungs (above $25 \textrm{ mmHg}$). PH leads to vascular remodelling, including stiffening, thickening, constriction of the small and large pulmonary arteries and microvascular rarefaction [@PulmCirc]. Microvascular rarefaction occurs in patients suffering of pulmonary hypertension, and is a pathological condition in which there are fewer capillaries per unit volume of body tissue. Developing a reliable predictive model for the pulmonary haemodynamics allows us to assist clinicians in diagnosing and treating pulmonary hypertension in a systematic manner. In addition, it helps reduce the number of invasive procedures for patients as, currently, pulmonary pressure is measured invasively via right heart catheterization [@Tabima]. In our problem we test the model in the context of data from mice, however, the work can be extended to human data.
One of the most important aspects in cardiovascular modelling is parameter inference and uncertainty quantification [@Eck]. This is because the number of parameters increases significantly when modelling several interconnected components at different scales. What is more, the limited amount of data available makes parameter estimation challenging in terms of finding physiologically sensible parameters.
Inferring key parameters for disease diagnosis and treatment planning is therefore an essential and challenging step in predicting the observed haemodynamics. One such parameter is the arterial stiffness, which is significantly higher for patients having pulmonary hypertension. Often, the parameters cannot be measured in-vivo, which creates the need for them to be learnt indirectly from the blood flow and pressure measurements.
In this paper, we present statistical techniques for inference of parameters of a mathematical model in order to reliably simulate and predict blood pressure and flow in the lungs. We first employ a nonlinear constraint optimization scheme to find the maximum likelihood estimates. We then quantify the uncertainty around these estimates by approximately sampling parameters from their posterior distribution using the Delayed Rejection Adaptive Metropolis (DRAM) algorithm and the Adaptive Metropolis (AM) algorithm.
Simulating blood flow and pressure in the 1D fluid-structure model described in the next section, requires solving a system of nonlinear partial differential equations (PDEs), which gets computationally expensive as the number of vessels increases in the model or a component model is made more complex. Therefore, starting the MCMC algorithm from optimised values allows us to save computational time. The goal of this paper is to show how optimization, combined with uncertainty quantification, can be used to learn the parameters relevant for pulmonary disease detection and, thus, to reliably predict the measured blood flow and pressure in the lungs. We show how by making a few restrictive assumptions about the parameters characterising the vessels’ geometry or the boundary conditions in the PDEs, we manage to reduce the parameter dimension from a 55D problem to 5D. Furthermore, we illustrate via pseudocode how to deal with different parameter magnitudes by incorporating a parameter scaling technique in the MCMC Matlab toolbox [@DRAMtoolbox]. We test this code on real data coming from a healthy and a hypoxic mouse. Furthermore, we show the importance of parameter scaling on a concrete example, and the consequences of not including the scaling in the analysis.
The remainder of the paper is structured as follows: Section 2 introduces the mathematical model, Section 3 describes the statistical model used in this study, Section 4 provides an overview of the methodology used, including the DRAM algorithm allowing for parameter scaling, Section 5 gives a few details of the computer simulations, Section 6 presents results of the methods applied on real data, and we finalise the paper by a discussion of the main findings of the study, as well as limitations and further work.
Mathematical model
==================
[\[sssec:mathsmodel\]]{} This study predicts pulmonary arterial flow and pressure using a 1D fluid dynamics network model, described in detail in our recent studies [@QureshiMice] and [@NewPaper]. The model is derived from the incompressible axisymmetric Navier–Stokes equations for a Newtonian fluid, and coupled with a constitutive wall model predicting elasticity (i.e. stiffness) of the blood vessels. In addition, the model assumes that the vessels are cylindrical and tapering, and that the wavelength is significantly longer than vessels’ radii. We use a tapering factor for the large arteries, as there is evidence of the vessels radii decreasing along their length and this was not quantified during the segmentation process. Under these assumptions, conservation of mass and momentum give $${\label{eq:1}}
\frac{\partial{A}}{\partial{t}} + \frac{\partial{q}}{\partial{x}} = 0, \hspace{0.5in} \frac{\partial{q}}{\partial{t}} + \frac{\partial}{\partial{x}} \frac{q^2}{A} + \frac{A}{\rho}\frac{\partial{p}}{\partial{x}} = -\frac{2\pi\mu r}{\delta}\frac{q}{A},$$ where $x \textrm{ (cm)}$ and $t \textrm{ (s)}$ are the axial and temporal coordinates, $p \textrm{ (mmHg)}$ is the blood pressure, $q \textrm{ (ml/s)}$ is the blood flow rate, $A \textrm{ (cm}^2)$ is the cross-sectional area, $\rho = 1.055 \textrm{ g/ml}$ is the blood density, $\mu = 0.049\, \textrm{g/(cm\,s)}$ is the viscosity and $\delta = \sqrt{\mu T/2\pi\rho} \textrm{ } \textrm{cm}$ is the boundary-layer thickness of velocity profile.
The arterial walls are assumed to be homogeneous, isotropic and thin. We use a linear wall model, where the relationship between pressure, $p$ (stress) and cross-sectional area $A$ deformation (strain) can be expressed as follows: $${\label{eq:2}}
p = p_0 + \frac{4}{3}f \left(1-\sqrt \frac{A_0}{A}\right) $$ where $A_0$ is the unstressed vessel cross-sectional area, and $f \textrm{ (mmHg)}$ is the arterial network stiffness.
![3D smoothed network from a micro-CT image of a healthy mouse lung (left) and the connectivity graph of of the same network (right). A similar network was extracted for the hypoxic mouse (not shown here) with the same connectivity graph. Note, the connectivity graph only shows the network structure and does not indicate length and radii of individual vessels. A three-element Windkessel model with two resistors and a capacitor is attached at the end of each terminal vessel. Detailed network dimensions for healthy and hypoxic mouse can be can be found in [@NewPaper].[]{data-label="fig:1"}](LungTree.png "fig:") ![3D smoothed network from a micro-CT image of a healthy mouse lung (left) and the connectivity graph of of the same network (right). A similar network was extracted for the hypoxic mouse (not shown here) with the same connectivity graph. Note, the connectivity graph only shows the network structure and does not indicate length and radii of individual vessels. A three-element Windkessel model with two resistors and a capacitor is attached at the end of each terminal vessel. Detailed network dimensions for healthy and hypoxic mouse can be can be found in [@NewPaper].[]{data-label="fig:1"}](Network.pdf "fig:")
The flow in the network is driven by prescribing a flow waveform measured in the main pulmonary artery (MPA) of a healthy and hypoxic mouse. Conservation of blood flow ($q_p = q_{d_1} + q_{d_2}$) and continuity of pressure ($p_p = p_{d_1} = p_{d_2}$) are enforced at the bifurcations ($p$ - parent, $d_1$ and $d_2$ - daughters), and an outflow condition predicting vascular bed impedance is prescribed using 3-element Windkessel model, with two resistors $R_1, R_2$ and a capacitor $C$, of the form $${\label{eq:3}}
Z(\omega) = R_1 + \frac{R_2}{1+i\omega C R_2} \implies p(L,t) = \frac{1}{T} \int_{0}^{T} q(L, t-\tau)Z(\tau)d\tau,$$ The system of equations (\[eq:1\]) and (\[eq:2\]) were solved numerically using the two step Lax-Wendroff scheme [@LaxWendroff]. Simulations were set up to fit pressure waveforms measured in the MPA in a healthy and hypoxic mouse.
The coupled model contains a total of 100 parameters, used to determine pressure and flow in the network, associated with the vessel fluid dynamics (24 parameters), the geometry (43 parameters), and the outflow boundary conditions (33 parameters). Of these, 45 parameters describing physical properties are fixed. The remaining 55 parameters characterize vessel stiffness (21 parameters - one per vessel), outflow boundary conditions (33 parameters, 3 parameters per terminal vessel), and a tapering factor.
To reduce the dimension of the parameter space, we assume that the vessel stiffness $f$ is constant throughout the network [@Krenz] (reducing the dimension by 20). Second, as described in [@NewPaper], we predict the nominal outflow boundary condition for each vessel from estimates of total resistance and compliance and use a constant scaling factor to optimize each of the resistances and capacitances in the Windkessel models (reducing the dimension by 30). Briefly, total system resistance $R_T \equiv R_1 + R_2 = \overline{p}/\overline{q}$, where $R_1 = 0.2 R_T$ and total capacitance $C_T$ is estimated from diastolic pressure decay. These quantities are distributed to each terminal vessel, indexed $j$, as described in [@NewPaper], giving a nominal set of parameters ${\theta}_{0\text{wk}}^j = \{R^j_{01},R^j_{02},C^j_{0}\}$. Introducing the scaling factors $r_1,r_2$ and $c$, final Windkessel parameters are estimated by optimizing these scaling factors as $$\label{eq:13}
R^j_1 = (1- 0.5 r_1)R^j_{01}, \quad R^j_2 = (1- 0.5 r_2)R^j_{02}\quad\text{and}\quad C^j = (1- 0.5c)C^j_0,$$ To account for the physiologically observed tapering in large arteries, a tapering factor $\xi$ is introduced, such that $$\label{eq:14}
r_{\textrm{B}} = r_{\textrm{T}} (1-0.5\xi),$$ where $r_{\textrm{T}}$ and $r_{\textrm{B}}$ are the inlet and outlet radii of the vessels.
Above assumptions and criteria has allowed us to reduce the dimension of free parameters from 55 to 5. The parameters to be estimated include: vessel stiffness $f$, scaling factors $r_1,r_2,c$ (equation (\[eq:13\])) and the tapering factor $\xi$.\
Two network configurations are analysed: $$\begin{aligned}
{\label{eq:12}}
\begin{split}
&\textrm{Network with straight vessels (4D model):}\qquad \boldsymbol{\theta} = \{f,r_1,r_2,c\}\\
&\textrm{Network with tapered vessels (5D model):}\quad \boldsymbol{\theta} = \{f,r_1,r_2,c,\xi\} \\
\end{split}\end{aligned}$$
Statistical Model
=================
We start the analysis by specifying the statistical model: $$\begin{aligned}
{\label{eq:4}}
\begin{split}
y^q_i = m_q(x_i; \boldsymbol{\theta}) + \epsilon_{1i}, \\
y^p_i = m_p(x_i; \boldsymbol{\theta}) + \epsilon_{2i}, \\
\end{split}\end{aligned}$$ where
- $\mathbf{y}^q = [y^q_1, y^q_2, ..., y^q_n]$ are the noisy measured flows, and $\mathbf{y}^p = [y^p_1, y^p_2, ... y^p_n]$ are the noisy measured pressures,
- $m(.)$ describes the system behaviour that comes from numerically solving the fluids model equations ([\[eq:1\]]{}) - ([\[eq:2\]]{}), i.e. $m_q(.)$ is the output (simulated) flow and $m_p(.)$ is the output (simulated) pressure,
- $\boldsymbol{\theta}$ are the parameters to be inferred from the observed flow and pressure. These relate to the Windkessel model parameters $R_1, R_2, C$ defined in (\[eq:3\]), to the vessel stiffness introduced in (\[eq:2\]) and to the lung network geometry. For a more detailed description of the parameters see Section \[sssec:mathsmodel\],
- $x_i\in\mathbf{x}$ denote other input variables (e.g. inflow into the MPA),
- $\boldsymbol{\epsilon}$ are the errors, which we assume are i.i.d following a Gaussian distribution and are different for flow and pressure, due to different measurement techniques.
Our objective function is chosen to be the residual sum of squares (S), which measures the deviation of our simulated signal from the measured signal and which, under the Gaussian assumption of the errors, is the negative log likelihood of a Normal distribution. In this study, S is calculated for pressure in the main pulmonary artery, and we aim to find the parameters that minimise S. We can write our objective function for pressure as $$\begin{aligned}
{\label{eq:5}}
\begin{split}
S = (\mathbf{y}^p - m_p(\mathbf{x}; \boldsymbol{\theta}))^2 = \sum_{i=1}^{n} (y^p_i - m_p(x_i; \boldsymbol{\theta}))^2, \\
\end{split}\end{aligned}$$ and the log likelihood of the pressure is $$\log(L) = -n\log \sqrt{2\pi\sigma^2} - \frac{S}{2\sigma^2}.$$
For simplicity, in the remainder of the paper, we will no longer use the $p$ index or the $x_i$ term, but it should be noted that all the calculations are done for pressure, and $\mathbf{x}$ is needed in solving the PDEs.
Statistical analysis methods
============================
In this paper, we apply Markov Chain Monte Carlo (see [@MonteCarlo] for a comprehensive review) to the problem of pulmonary circulation to approximately sample the parameters from the posterior distribution. Parameters are inferred using the Delayed Rejection Adaptive Metropolis (DRAM) algorithm [@DRAM] and we compare the performance of the DRAM algorithm with that of the Adaptive Metropolis (AM) algorithm [@AM]. Since MCMC is too slow to converge when started from random points in parameter space, we first obtain point estimates using the maximum likelihood approach, and these estimates are used as initial values in the DRAM or AM algorithm. The concepts of optimization in a maximum likelihood context and uncertainty assessment of these estimates using MCMC are reviewed in subsections \[sssec:mle\] and subsubsection \[sssec:mcmc\]. Our work builds on and further improves on existing literature by allowing for a novel parameter scaling technique in the DRAM algorithm with informative priors for the parameters (see subsubsection \[sssec:dramscaled\]). Following work in existing studies, we employ formal convergence diagnostics for the DRAM algorithm (Multivariate Potential Scale Reduction Factor, Geweke test, Integrated Autocorrelation Time and Effective Sample Size), as described in \[sssec:convdiag\]. The last step is to use existing scores for selecting the better model for explaining the data (AIC, AICc, BIC, DIC, WAIC), reviewed in subsection \[sssec:modelselection\].
Maximum likelihood with nonlinear optimization {#sssec:mle}
----------------------------------------------
We first apply nonlinear constraint optimization to find the global or local optimum of the objective function, as this exhibits unimodality in low dimensions (see Results section) and additionally, it is nonlinear in the parameters, which are bounded (see the Simulations section for a justification of the parameter ranges). We apply the Sequential Quadratic Programming (SQP) algorithm, which iteratively generates estimates that converge to a solution of (\[eq:6\]) by constructing and solving quadratic sub-problems that approximate the nonlinear programming problem defined by (\[eq:6\]). $${\label{eq:6}}
\begin{split}
\text{minimize \hspace{0.5in}} e(\boldsymbol{\theta}),\boldsymbol{\theta} \in \mathbb{R}^d, \text{where } e(\boldsymbol{\theta}) \text{ is the residual-sum-of-squares, } S \text{ (see (\ref{eq:5}))} \\
&\text{\hspace{-4.15in} subject to \hspace{0.45in}} g(\boldsymbol{\theta}) = 0, g:\mathbb{R}^d \rightarrow \mathbb{R}^t \text{ equality constraints defining relations} \\
&\text{\hspace{-2.1in} between parameters,} \\
& \text{\hspace{-3.2in}} h(\boldsymbol{\theta}) \leq 0, h:\mathbb{R}^d \rightarrow \mathbb{R}^u \text{ inequality constraints defining a} \\
& \text{\hspace{-2.15in} restricted parameter ranges.}
\end{split}$$
The gradient-based method used here involves Taylor series expansion and, so, it requires the function to be twice continuously differentiable in the neighbourhood of the point tested and the Hessian matrix of e to be positive definite This method has been shown to be a very good tool for solving highly nonlinear, fairly low dimensional problems, where we have equality and inequality constraints and where the function displays unimodality. In this algorithm, information about the function to be minimised, $e$, and about the equality and inequality functions $g$ and $h$ is stored in a Lagrangian function: $L(\theta, \Lambda, \Gamma) = e(\theta) + \sum_d \Lambda_d g_d(\theta) + \sum_d \Gamma_d h_d(\theta)$, where $\Lambda$ and $\Gamma$ are the Lagrange multipliers. Next, the Lagrangian function is minimised using Newton’s method, due to the nonlinearity of the problem, which calls for numerically iterative methods (finding the critical points that make $\nabla L = 0$ is challenging since some of the derivatives may be highly nonlinear). We can optimise the Lagrangian function instead of the objective because the critical points of the Lagrangian are the same as those of the objective function. The problem is then turned into a quadratic subproblem. At iteration k, the direction, v, in which the algorithm searches the minimum is a solution to the equation: $$\min_v e(\theta_k) + \nabla e(\theta_k)^T v + \frac{1}{2}v^T \nabla_{\theta \theta}^2 L(\theta_k, \Lambda_k, \Gamma_k)v,$$ $$\hspace{0.7cm} \text{s.t. } g(\theta_k) + \nabla g(\theta_k)^T v \geq 0 \text{ and } h(\theta_k) + \nabla h(\theta_k)^T v = 0.$$ For more details of this algorithm, see [@sqpPaper1] and [@sqpPaper2].
Uncertainty quantification with MCMC
------------------------------------
### Overview of MCMC (DRAM) {#sssec:mcmc}
The Metropolis–Hastings (MH) algorithm [@mh1953] is one of the most widely used MCMC methods. In this algorithm, new samples are generated from the proposal distribution and are either accepted or rejected depending on how likely it is that these samples come from the target distribution, i.e. accept with probability: $$\alpha_1(\boldsymbol{\theta}^{k-1}, \boldsymbol{\theta}^*) = \min \left(1, \frac{\pi(\boldsymbol{\theta}^*) q_1(\boldsymbol{\theta}^*,\boldsymbol{\theta}^{k-1})}{\pi(\boldsymbol{\theta}^{k-1}) q_1(\boldsymbol{\theta}^{k-1}, \boldsymbol{\theta}^*)} \right),$$ where $\boldsymbol{\theta}_{k-1}$ is the current point, $\boldsymbol{\theta}^*$ the proposed point, $q_1(\boldsymbol{\theta}^*,\boldsymbol{\theta}^{k-1})$ the proposal distribution from $\boldsymbol{\theta}^{k-1}$ to $\boldsymbol{\theta}^*$ and $\pi(.)$ the stationary distribution. If a symmetric proposal distribution $q_1$ (i.e. $q_1(\boldsymbol{\theta}^*,\boldsymbol{\theta}^{k-1}) = MVN(\boldsymbol{\theta}^{k-1}, V_k)$), is used, then the ratio of the proposal distributions will be 1. A poor proposal distribution will result in a high rejection rate. To overcome this, upon rejection of $\boldsymbol{\theta}^*$, instead of retaining $\boldsymbol{\theta}^{k-1}$, one can make a second attempt to move [@Tierney]; in the $2^{nd}$ proposal, a different distribution is used (i.e. $q_2(\boldsymbol{\theta}^{*2}, \boldsymbol{\theta}^*, \boldsymbol{\theta}^{k-1}) = MVN(\boldsymbol{\theta}^{k-1}, \beta^2 V_k)$ for a symmetric proposal distribution) and this is allowed to depend on the rejected sample, which can be proved to not destroy the Markovian property of the chain [@DR]. Having started in $\boldsymbol{\theta}^{k-1}$ and after rejecting $\boldsymbol{\theta}^*$, the second proposed point, $\boldsymbol{\theta}^{*2}$ is accepted with probability: $${\label{eq:7}}
\alpha_2(\boldsymbol{\theta}^{k-1}, \boldsymbol{\theta}^*, \boldsymbol{\theta}^{*2}) = \min \left(1,\frac{\pi(\boldsymbol{\theta}^{*2}) q_1(\boldsymbol{\theta}^{*2}, \boldsymbol{\theta}^*) q_2(\boldsymbol{\theta}^{*2}, \boldsymbol{\theta}^*, \boldsymbol{\theta}^{k-1})[1-\alpha_1(\boldsymbol{\theta}^{*2}, \boldsymbol{\theta}^*)]}{\pi(\boldsymbol{\theta}^{k-1}) q_1(\boldsymbol{\theta}^{k-1}, \boldsymbol{\theta}^*) q_2(\boldsymbol{\theta}^{k-1}, \boldsymbol{\theta}^*, \boldsymbol{\theta}^{*2})[1-\alpha_1(\boldsymbol{\theta}^{k-1}, \boldsymbol{\theta}^*)]} \right),$$
The acceptance probability in the $2^{nd}$ stage is calculated in such a way that detailed balance is preserved. The ratio of the $q_2$ proposal distributions will become 1 if the proposal distribution is symmetric. The Delayed Rejection (DR) algorithm may have a smaller rejection rate, which implies a smaller asymptotic variance of the estimates compared to the MH algorithm [@Mira2001a].
In cases where the parameters to be learnt are strongly correlated, the parameter dimensionality is large or the problem is highly nonlinear in the parameters, then running the classical MH algorithm can be inefficient, since the chain will move slowly in the parameter space due to the random walk behaviour of the algorithm, thus taking a long time to approach the target density. A solution to this is to adapt the Gaussian proposal covariance depending on the shape (or spatial orientation) and size of the target distribution, i.e. we adapt the proposal covariance matrix, $V_k$ based on the past chain [@AM]. It can be shown that this Adaptive Metropolis (AM) algorithm still produces an ergodic Markov chain [@AM].
The proposal distribution in the AM algorithm is a Gaussian distribution centred at the current point $\boldsymbol{\theta}^{k-1}$ and covariance matrix equal to $V_k = V_k(\boldsymbol{\theta}^0, ... \boldsymbol{\theta}^{k-1})$. The proposal covariance matrix, $V_k = R_k^TR_k$ is adapted at given intervals, after possibly some non-adaptation time $t_{\textrm{ad}}$, according to the following formula: $$\label{eq:8}
V_k =
\begin{cases}
V_0, & \text{if}\ k \leq t_{\textrm{ad}} \\
s_d cov(\boldsymbol{\theta}^0, ...,\boldsymbol{\theta}^{k-1})+\epsilon I_d, & \text{if}\ k > t_{\textrm{ad}},
\end{cases}$$ where $V_0$ is the initial proposal covariance matrix, $s_d$ is a parameter that depends on the dimension d of the target ($s_d = 2.4^2/d$ has been shown to be optimal for Gaussian targets in terms of mixing) [@AM], $\epsilon > 0$ is a constant that ensures that $V_k$ does not become singular, and $(\boldsymbol{\theta}^0, ... \boldsymbol{\theta}^{k-1})$ are the past chain samples.
DRAM [@DRAM] increases the convergence through the use of the rejected values in the DR step (local adaptation), and of the past chain samples to learn about the posterior distribution through the adapted covariance matrix, in the AM step (global adaptation).
### Delayed Rejection Adaptive Metropolis (DRAM) with parameter scaling {#sssec:dramscaled}
A pseudocode of the DRAM algorithm can be found in [@UncertaintySmith] (pages 175–176). In practice, we implement the algorithm by making use of the Matlab DRAM toolbox [@DRAMtoolbox]. However, the toolbox does not account for the case where the parameters differ by orders of magnitude, and so the code had to be modified to allow scaling the parameters. The pseudocode for DRAM with an informative prior, together with the parameter scaling adjustment can be found in Algorithms \[alg:algorithm1\]-\[alg:algorithm4\]. Here we improve the Algorithms 8.8 and 8.10 in [@UncertaintySmith]. We introduce a novel extension to DRAM by adding an informative prior and improving the scaling of the parameters. This improvement concerns the way that the acceptance probabilities are calculated. More precisely, the jump densities should be calculated for the transformed parameters (see step \[dramScaled1\] in Algorithm \[alg:algorithm4\]). Using the original parameters can cause the ratio of the jump densities to become very large in exceptional cases, making us falsely accept the proposals. For example, if the starting value, which is also the optimum, is close to the boundaries, the algorithm will often jump outside boundaries. This will make some parameters get very high values (e.g. in the second DR try), and so the density in the numerator will become much higher than the density in the denominator in the jump ratio (see equation (\[eq:7\])). This, however, does not represent a problem for the first stage acceptance probabilities when a symmetric proposal distribution is used, as the proposals will cancel out in the ratio. Hence, the correction only affects DRAM and DR, while no correction is needed for AM.
Initialise design parameters: M: number of chain iterations, $\beta$: scaling factor for proposal covariance matrix in DR, $n_{\textrm{dr}}$: number of DR tries, $t_{\textrm{ad}}$: adaptation interval, $b_{\textrm{s}}$: burn-in scale, $b_\textrm{t}$: burn-in time, $\gamma_\textrm{s}^2$: prior for $\sigma^2$ (error variance), $n_\textrm{s}$: prior accuracy for $\gamma_\textrm{s}^2$.
Compute $\boldsymbol{\theta}^0 =\operatorname*{arg\,min}_{\boldsymbol{\theta}} \sum_{i=1}^{n} [y_i - m_i(\boldsymbol{\theta})]^2$, where $a_j\leq \theta_j \leq b_j, j = 1 ... d$, $d$: number of parameters, using nonlinear constraint optimization (SQP algorithm) with 20 overdispersed starting values generated from a Sobol sequence. \[dram2\]
Initialise $\sigma_0^2 = \frac{S_0}{n-d}$, where $S_0$: optimised residual-of-squares value, $n$: number of observations.
Assume that the prior distribution for $\boldsymbol{\theta}$ is a Multivariate Normal distribution (independent $\theta$s), $\boldsymbol{\theta} \sim \mathcal{MVN}(\boldsymbol{\mu}, T) $
Set initial $S_{\textrm{prior}}^0 = \sum_{j=1}^d \frac{(\theta^0_j-\mu_j)^2}{t_j^2} $.
Either set the initial covariance matrix $V_0$ to be the Hessian matrix from optimization or set $V_0$ via monitoring of acceptance rate.\
Set $R_0$ = $(V_0)$, where “chol” stands for the Choleski factorization, as an efficient numerical solution of $V_0 = R_0^TR_0$.
Sample $\mathbf{u}_k \sim \mathcal{MVN}(\mathbf{0},\mathcal{I})$.
Construct candidate from Gaussian proposal: $\boldsymbol{\theta}^* = \boldsymbol{\theta}^{k-1} + R_{k-1} \boldsymbol{u}_k.$ \[dram9\]
Jump to DR Algorithm \[alg:algorithm3\] for iteration $k$ Calculate $S_{\boldsymbol{\theta}^*} = \sum_{i=1}^n [y_i - m_i(\boldsymbol{\theta}^*)]^2,
S_{\textrm{pri}}^* = \sum_{j=1}^d \frac{(\theta^*_j-\mu_j)^2}{t_j^2}$ \[dram13\]
Calculate the acceptance probability,
$\alpha_1(\boldsymbol{\theta}^*, \boldsymbol{\theta}^{k-1}) = \min\left(1, e^{-0.5 \left[\frac{S_{\boldsymbol{\theta}^*} - S_{\boldsymbol{\theta}^{k-1}}}{\sigma^2_{k-1}} + S_{\textrm{pri}}^* - S_{\textrm{pri}}^{k-1}\right]}\right).$ \[dram15\]
Sample $v_{\alpha_1} \sim \mathcal{N}(0,1).$
$\boldsymbol{\theta}^k = \boldsymbol{\theta}^*, S_{\boldsymbol{\theta}^k} = S_{\boldsymbol{\theta}^*}, S_{\textrm{pri}}^k = S_{\textrm{pri}}^*$ \[dram18\] Enter DR Algorithm \[alg:algorithm3\] for iteration $k$
Update $\sigma_k^2 \sim \textrm{Inv-Gamma} \left(\frac{n_\textrm{s}+n}{2}, \frac{n_\textrm{s} \gamma_\textrm{s}^2+S_{\boldsymbol{\theta}^k}}{2} \right)$
Enter AM Algorithm \[alg:algorithm2\] for iteration $k$
$R_k \leftarrow R_{k-1}/b_\textrm{s}$, $R_k \leftarrow b_\textrm{s} R_{k-1}$, Set $V_k = s_d cov(\boldsymbol{\theta}^0, \boldsymbol{\theta}^1, ... \boldsymbol{\theta}^k) + \epsilon \mathcal{I}_d,
R_k = chol(V_k)$ $V_k = V_{k-1},
R_k =R_{k-1}$ Output $R_k$
Sample $\mathbf{u}_k \sim \mathcal{MVN}(\mathbf{0},\mathcal{I})$.
Construct $2^{nd}$ stage candidate from Gaussian proposal: $\boldsymbol{\theta}^{*2} = \boldsymbol{\theta}^{k-1} + \beta R_{k-1} \boldsymbol{u}_k.$ \[dr2\]
$\boldsymbol{\theta}^k = \boldsymbol{\theta}^{k-1}, S_{\boldsymbol{\theta}^k} = S_{\boldsymbol{\theta}^{k-1}},
S_{\textrm{pri}}^k = S_{\textrm{pri}}^{k-1}$
and jump to AM Algorithm \[alg:algorithm2\].
$S_{\boldsymbol{\theta}^{*2}} = \sum_{i=1}^n [y_i - m_i(\boldsymbol{\theta}^{*2})]^2,
S_{\textrm{pri}}^{*2} = \sum_{j=1}^d \frac{(\theta^{*2}_j-\mu_j)^2}{t_j^2}$ \[dr4\]
Calculate the acceptance probability, $ \alpha_2(\boldsymbol{\theta}^{*2}, \boldsymbol{\theta}^*, \boldsymbol{\theta}^{k-1}) $ according to equation (\[eq:7\]). \[dr6\]
Sample $v_{\alpha_2} \sim \mathcal{N}(0,1).$
,
$\boldsymbol{\theta}^k = \boldsymbol{\theta}^{*2}, S_{\boldsymbol{\theta}^k} = S_{\boldsymbol{\theta}^{*2}}, S_{\textrm{pri}}^k = S_{\textrm{pri}}^{*2}$ \[dr8\] $\boldsymbol{\theta}^k = \boldsymbol{\theta}^{k-1}, S_{\boldsymbol{\theta}^k} = S_{\boldsymbol{\theta}^{k-1}}, S_{\textrm{pri}}^k = S_{\textrm{pri}}^{k-1}$ \[dr9\]
Denote $\boldsymbol{\theta}_s = \boldsymbol{\theta}./\mathbf{s}$, where $\mathbf{s}$ are the scaling factors.
In Algorithm \[alg:algorithm1\]: DRAM \[dram2\]: Compute $\boldsymbol{\theta}^0 =\operatorname*{arg\,min}_{\boldsymbol{\theta}} \sum_{i=1}^{n} [y_i - m_i(\boldsymbol{\theta}.\times \mathbf{s})]^2$
DRAM \[dram9\]: Construct candidate from Gaussian proposal: $$\boldsymbol{\theta}_s^* = \boldsymbol{\theta}_s^{k-1} + R_{k-1} \boldsymbol{u}_k.$$
and set $$\boldsymbol{\theta}^* = \boldsymbol{\theta}_s^* .\times \mathbf{s}$$
DRAM \[dram13\]: Set $$S_{\textrm{pri}}^* = \sum_{j=1}^d \frac{(\theta^*_j-\mu_j)^2}{t_j^2} + \log(det(J))$$ where $J$ is the Jacobian of the parameter transformation. If the parameters are assumed independent and $\mathbf{s}$ is constant, then $\log(det(J))=\sum_{j=1}^d s_j$.
DRAM \[dram15\]: Calculate the acceptance probability, $$\alpha_1(\boldsymbol{\theta}_s^*, \boldsymbol{\theta}_s^{k-1})$$
**Note**: if a noninformative prior is used and $\boldsymbol{s}$ is a constant, then we are left with the likelihood contribution in the acceptance probability, i.e. $\alpha = \min\left(1, e^{-0.5 \left[\frac{S_{\boldsymbol{\theta}^*} - S_{\boldsymbol{\theta}^{k-1}}}{\sigma^2_{k-1}}\right]}\right)$. In addition, in the likelihood calculations, we always use the unscaled parameters, whereas the priors and proposals are done in the transformed parameter space.
DRAM \[dram18\]: In addition, set $\boldsymbol{\theta}_s^k = \boldsymbol{\theta}_s^*.$
In Algorithm \[alg:algorithm3\]:
DR \[dr2\]: Construct $2^{nd}$ stage candidate from Gaussian proposal: $$\boldsymbol{\theta}_s^{*2} = \boldsymbol{\theta}_s^{k-1} + \beta R_{k-1} \boldsymbol{u}_k,$$ and set $$\boldsymbol{\theta}^{*2} = \boldsymbol{\theta}_s^{*2} .\times \mathbf{s}$$
DR \[dr4\]: Set $$S_{\textrm{pri}}^{*2} = \sum_{j=1}^d \frac{(\theta^{*2}_j-\mu_j)^2}{t_j^2} + \log(det(J))$$
DR \[dr6\]: Calculate the acceptance probability in the scaled parameter space, $ \alpha_2(\boldsymbol{\theta}_s^{*2}, \boldsymbol{\theta}_s^*, \boldsymbol{\theta}_s^{k-1}) $. \[dramScaled1\]
DR \[dr8\] or DR \[dr9\]: In addition, set $\boldsymbol{\theta}_s^k = \boldsymbol{\theta}_s^{*2}$ or $\boldsymbol{\theta}_s^k = \boldsymbol{\theta}_s^{k-1}$.
**Note**: If a noninformative prior is used, then $S_{\textrm{pri}} = 0$ in Algorithms \[alg:algorithm1\], \[alg:algorithm3\] and \[alg:algorithm4\].
Convergence Diagnostics {#sssec:convdiag}
-----------------------
Once the Markov chains are produced using the algorithms discussed above, we test for convergence, i.e. test whether our Markov chains have converged in distribution to the posterior distribution of interest.
### Integrated Autocorrelation Time (IACT)
One way to assess convergence is to assess the autocorrelations between the draws of the Markov chain. The lag $l$ autocorrelation $\rho_l$ is the correlation between every draw and its $l^{\textrm{th}}$ lag: $$\rho_l = \frac{\sum_{k=1}^{M-l}(\theta_k - \bar{\theta})(\theta_{k+l}-\bar{\theta})}{\sum_{k=1}^{M}(\theta_k-\bar{\theta})^2}$$
The Integrated Autocorrelation Time [@IACT] is given by: $\tau = 1 + 2\sum_{l=1}^{\infty}\rho_l$.
It would be expected that the $l^{\textrm{th}}$ lag autocorrelation is smaller as $l$ increases ($2^{\textrm{nd}}$ and $50^{\textrm{th}}$ draws should be less correlated than $2^{\textrm{nd}}$ and $4^{\textrm{th}}$ draws). Relatively high autocorrelation for higher values of $l$ suggests high degree of correlation between draws and thus, slow mixing.
Having IACT, we can also derive the Effective Sample Size (ESS) [@ESS], which gives an estimate of the equivalent number of independent iterations that the chain represents. Low ESS can indicate that the sampler used is inefficient.
The formula for ESS is given by: $$ESS = \frac{M}{1 + 2\sum_{l=1}^{\infty}\rho_l} = \frac{M}{\tau}.$$
### Geweke test
The Geweke test [@Geweke] formally tests for equality of the mean of two sub-chains (typically the first $10\%$ and last $50\%$ of the iterations). A low p-value for the Z-statistic will indicate non-convergence.
### Brooks Gelman Rubin test
The Brooks Multivariate Potential Scale Reduction Factor (MPSRF) [@BrooksGelman] can be calculated for multiple chains from starting values overdispersed with respect to the stationary distribution. MPSRF is a multivariate extension to the Gelman-Rubin test [@GelmanRubin], assessing convergence of the parameters simultaneously. Not only does this take into account the variances of individual parameters, but also the interactions between the parameters (the covariances). The MPSRF is calculated in such a way that it is an approximate estimate of the maximum upper bound of the univariate SRF for every parameter. Suppose we have $m$ chains in parallel, and $\boldsymbol{\theta}^k_i$ is the set of parameters drawn at the $k^{\textrm{th}}$ iteration, $k = 1, ...M$ from the $i^{\textrm{th}}, i = 1, ...m$ chain. Then we can calculate the posterior variance-covariance matrix, as follows: $$\hat{V} = \frac{M-1}{M} W + \left( 1+\frac{1}{m}\right)\frac{B}{M},$$ where $$W = \frac{1}{m(M-1)}\sum_{i=1}^m \sum_{k=1}^M (\boldsymbol{\theta}_i^k - \bar{\boldsymbol{\theta}}_{i.}) (\boldsymbol{\theta}_i^k - \bar{\boldsymbol{\theta}}_{i.})^{'}$$ and $$B = \frac{M}{m-1}\sum_{i=1}^m (\bar{\boldsymbol{\theta}}_{i.} - \bar{\boldsymbol{\theta}}_{..}) (\bar{\boldsymbol{\theta}}_{i.} - \bar{\boldsymbol{\theta}}_{..})^{'}$$ $W$ is the within-chain covariance matrix and $B$ the between-chain covariance matrix. If $\lambda_1$ is the largest eigenvalue of the symmetric and positive definite matrix $W^{-1}B$ (see proof in [@BrooksGelman]), then the MPSRF, $$\hat{R}^p = \frac{M-1}{M} + \frac{m+1}{m} \lambda_1$$ approaches 1 if $M \rightarrow \infty$ and the chains have converged to the stationary distribution, i.e. $W \approx B$ and, hence $\lambda_1 \rightarrow 0$.
In addition, a MPSRF plot is useful in indicating whether the lack of convergence is due to the correlation between parameters.
Model selection {#sssec:modelselection}
---------------
Like in any other Bayesian application, where we have two or more competing models, we may be interested in comparing them, in order to choose the better model. A more complex model is not necessarily favoured, despite the fact that it provides a better fit. Penalising for complexity (too many parameters) is a necessary adjustment to prevent from choosing an overparameterised model that overfits the data and is not generalisable to future data. In this paper, we use the corrected Akaike Information Criterion (AICc) [@AICc], the Bayesian Information Criteria (BIC) [@BIC], the Deviance Information Criteria (DIC) [@DIC] and the Watanabe AIC (WAIC) [@WAIC2010] score for model selection, and the model with the lowest score will be preferred.
### AICc and BIC
AICc [@AICc] and BIC [@BIC] can be used for model selection in the maximum likelihood framework, where we obtain a maximum likelihood estimate. We use the maximum log likelihood (i.e. the minimum residual sum-of-squares, S) to represent the goodness of fit, and we apply a penalty to avoid overfitting. AICc and BIC penalise for the model complexity differently: AICc applies a lower penalty, while BIC tends to penalise more and choose simpler models. AICc is a modification of AIC [@AIC] by adding a correction for finite sample sizes, hence applying a higher penalty to avoid overfitting. The formula for AICc is given for a univariate Gaussian errors model: $$\begin{aligned}
\begin{split}
AICc &= -2\log(\mathbf{y}|\hat{\boldsymbol{\theta}}_{\textrm{mle}}) + 2d + \frac{2d(d+1)}{n-d-1}, \\
BIC &= -2\log(\mathbf{y}|\hat{\boldsymbol{\theta}}_{\textrm{mle}}) + d\log n,
\end{split}\end{aligned}$$ where $\log(\mathbf{y}|\hat{\boldsymbol{\theta}}_{\textrm{mle}})$ is the maximum log-likelihood, $d$ is the number of parameters in the model and $n$ is the total number of observations. We can derive these formulas to depend on the residual sum-of-squares and the formula involving $S$ is given in equation (\[eq:9\]): $$\begin{aligned}
\label{eq:9}
\begin{split}
AICc &= \frac{S}{\sigma^2} + n\log \sigma^2 + 2d + \frac{2d(d+1)}{n-d-1} + C,\\
BIC &= \frac{S}{\sigma^2} + n\log \sigma^2 + d\log n + C,
\end{split}\end{aligned}$$ where $C = n\log 2\pi$ is a constant.
### DIC
DIC [@DIC] is a partially Bayesian version of AIC, where $\hat{\boldsymbol{\theta}}_{\textrm{mle}}$ is replaced by the posterior mean $\hat{\boldsymbol{\theta}}_{\textrm{Bayes}} = E(\boldsymbol{\theta}|\mathbf{y})$ and the penalisation is data driven. Then, $$DIC = -2\log p(\mathbf{y}|\hat{\boldsymbol{\theta}}_{\textrm{Bayes}}) + 2p_{\textrm{DIC}},$$ where $$\label {eq:10}
p_{\textrm{DIC}} = 2 \left(\log p(\mathbf{y}|\hat{\boldsymbol{\theta}}_{\textrm{Bayes}}) - E_{\boldsymbol{\theta}|\mathbf{y}}(\log p(\mathbf{y}|\boldsymbol{\theta}))\right),$$ where $E_{\boldsymbol{\theta}|\mathbf{y}}(\log(\mathbf{y}|\boldsymbol{\theta}))$ is the expectation of the log posterior predictive density, which in practice is calculated by replacing the expectation with the average over the posterior draws, as follows [@gelman2013]: $$\text{computed } p_{\textrm{DIC}} = 2 \left(\log p(\mathbf{y}|\hat{\boldsymbol{\theta}}_{\textrm{Bayes}}) - \frac{1}{S}\sum_{s=1}^S \log p(\mathbf{y}|\boldsymbol{\theta}^s)\right),$$ where the sum is over the set of parameters $\boldsymbol{\theta}^s, s = 1, ...S$, that have approximately been drawn from the posterior distribution with MCMC (i.e. $\boldsymbol{\theta}^s$ is the MCMC sample).
### WAIC
WAIC [@WAIC2010] is fully Bayesian in the sense that it is computed using the whole posterior distribution. It takes the computed log posterior predictive density in equation (\[eq:11\]) and adjusts for overfitting by adding a correction for the effective number of parameters [@gelman2013]. $$WAIC = -2 lppd + 2 p_{\textrm{WAIC}},$$ where $$\label{eq:50}
lppd = log \prod_{i=1}^n p_{\boldsymbol{\theta}|\mathbf{y}}(y_i) = \sum_{i=1}^n \log \int p(y_i|\boldsymbol{\theta})p(\boldsymbol{\theta}|\mathbf{y}) d\boldsymbol{\theta}$$ The left hand side term of equation (\[eq:50\]), $p_{\boldsymbol{\theta}|\mathbf{y}}(y_i)$ equals the integral on the right hand side because $\boldsymbol{\theta}$ is unknown, and so we take an expectation of $p_{\boldsymbol{\theta}|\mathbf{y}}(y_i)$ over $\boldsymbol{\theta}$ given $\mathbf{y}$, i.e. we integrate out $\boldsymbol{\theta}$. In practice, this is computed as follows: $$\label{eq:11}
\text{computed } lppd = \sum_{i=1}^n \log \left (\frac{1}{S} \sum_{s=1}^S p(y_i|\boldsymbol{\theta}^s)\right)$$ where $\boldsymbol{\theta}^s, s = 1, ...S$, are posterior samples from $p(\boldsymbol{\theta}|\mathbf{y})$, and $$p_{\textrm{WAIC}} = \sum_{i=1}^n var_{\boldsymbol{\theta}|\mathbf{y}} (\log p(y_i|\boldsymbol{\theta}))$$ In practice, to compute the posterior variance for every term in the log predictive density $V_{s=1}^S \log p(y_i|\boldsymbol{\theta}^s)$, where $V_{s=1}^S$ is the sample variance, we use: $V_{s=1}^S a_s = \frac{1}{S-1}\sum_{s=1}^S (a_s - \bar{a})^2.$ The effective number of parameters is then given by the sum of $V_{s=1}^S$ over all the data points: $$\text{computed } p_{\textrm{WAIC}} = \sum_{i=1}^n V_{s=1}^S (\log p(y_i|\boldsymbol{\theta}^s)).$$
AIC, AICc, DIC, WAIC can be used for model selection through predictive accuracy, whereas BIC is used to approximate the marginal probability density of the data under the model $p(\mathbf{y})$, useful for discrete model selection when estimating the relative posterior probabilities [@gelman2013].
Out of AIC, AICc, DIC, WAIC, the latter one is preferred as it has better asymptotic properties. This is because in their derivation using the Taylor series expansion of the information distance between the model and the real system around the optimal parameter that minimises this distance, terms above the third order are neglected for WAIC, whereas for AIC for example, terms above second order are neglected, so the error associated with WAIC is lower, i.e. the predictive accuracy is higher for WAIC [@WAIC2015]. In addition, WAIC is asymptotically equivalent to Bayesian leave-one-out cross-validation (LOO-CV) [@WAIC2010]. In Bayesian Loo- CV, one point is removed at a time from the data, the model is trained on the $n-1$ observations and the model fit to the data assessed by computing the log predictive density of the holdout data. A model that has a higher Bayesian LOO- CV estimate of the out-of-sample predictive fit (or lower -2 times this estimate, to be on the deviance scale) is preferred. The Baysian LOO-CV estimate of the out-of-sample predictive fit is: $$lppd_{LOO-CV} = \sum_{i=1}^n \log p(y_i|\mathbf{y}_{-i}) = \sum_{i=1}^n \log \int p(y_i|\boldsymbol{\theta})p(\boldsymbol{\theta}|\mathbf{y}_{-i})d\boldsymbol{\theta},$$ which in practice is calculated as: $$\text{computed } lppd_{LOO-CV} = \sum_{i=1}^n \log \left(\frac{1}{S}\sum_{s=1}^S p(y_i|\boldsymbol{\theta}^s)\right),$$ where ${\boldsymbol{\theta}^s}$ is a sample from $p(\boldsymbol{\theta}|\mathbf{y}_{-i})$.
AIC, AICc, DIC and WAIC are scores for the adjusted within-sample predictive accuracy, i.e. all points in $\mathbf{y}$ are used to calculate the scores and prediction is done for the same points; the bias associated with doing this goes to zero as the number of observations increases, i.e. asymptotically. The scores use a penalty term for the model complexity. In contrast, LOO-CV measures the out-of-sample prediction accuracy, as its calculation relies on the points in $\mathbf{y}_{-i}$, i.e. on all points, except the $i^{\textrm{th}}$, and we calculate the predictive accuracy for the $i^{\textrm{th}}$ point. What is more, when comparing WAIC and DIC, WAIC can produce a reliable result even when dealing with multimodal posterior distributions, as it uses the entire posterior distribution, while DIC would be misleading as it uses the mean posterior estimate (e.g. in a bimodal posterior distribution, the mean would be between the two modes). It is known that DIC cannot deal with singular models [@WAIC2010].
Simulations
===========
The parameters that we wish to infer, described in detail in Section 2 are taken as arguments to the PDEs (\[eq:1\]) - (\[eq:2\]), and predicted flow and pressure are obtained along several locations in the 21 vessels. Since measured data are only available from one location in the main pulmonary artery, we only use the corresponding prediction from the MPA. We drive the system by the measured flow data, and compare our predictions to pressure measurements in the MPA. This is done using the residual sum-of-squares, S (equation (\[eq:5\])). The set of parameters that give the closest prediction to the measured data is our estimate (maximum likelihood estimate) [@PaunIWSM]. In practice, this is performed using the nonlinear constrained optimization scheme (SQP algorithm). The uncertainty around these estimates is obtained by approximately sampling from the posterior distribution using MCMC (DRAM, AM). We implement these algorithms by making use of the DRAM toolbox implemented in Matlab [@DRAMtoolbox].
The parameters are on different scales: $f \in [10^4, 10^6], r_1, r_2, c \in [-3, 2.5], \xi \in [0, 0.5]$, and these parameter bounds are chosen to be biologically meaningful. To avoid having an ill-conditioned problem induced by a high condition number in the Hessian matrix [@Illcond], we rescale the parameters to have the same order of magnitude and use these scaled parameters in the optimization routine, as well as in the DRAM algorithm. There are certain parameter configurations outside the domain of convergence of the numerical scheme. This prompted us to assign a very high value to the RSS $(10^{10})$ that marks the unsuccessful simulations (data are unlikely given those parameter values).
The 20 different initial values for the parameters passed in the optimization algorithm are uniformly drawn from a Sobol sequence to ensure a good coverage of the multidimensional parameter space [@Sobol]. The SQP algorithm is iterated until it satisfies the convergence criterion, i.e. $|\boldsymbol{\theta_{i}} - \boldsymbol{\theta_{i+1}}| < 1e-11$. Once convergence has been reached, we take the optimum parameter values as starting values for the DRAM algorithm in order to speed up the simulations.
Let us now define the likelihood, priors, proposal distribution and posterior distribution used in the DRAM algorithm.
- Data likelihood: $y_i|\boldsymbol{\theta} \sim N(m_i(\boldsymbol{\theta}), \sigma^2)$, i.e. $$p(\mathbf{y}|\boldsymbol{\theta}, \sigma^2) = \left(\frac{1}{\sqrt{2\pi\sigma^2}}\right)^n \exp \left(-\frac{\sum_{i=1}^n(y_i-m_i(\boldsymbol{\theta}))^2}{2\sigma^2} \right)$$
- Priors: $\boldsymbol{\theta} \sim \text{Truncated } MVN(\boldsymbol{\mu}, T), T: \text{positive definite matrix}$, and due to lack of prior knowledge, we assign high variances to the parameters, to make the prior noninformative.\
$\sigma^2 \sim \textrm{Inv-Gamma}(a,b)$, where following Algorithm \[alg:algorithm1\], we denote $a = \frac{n_s}{2}>0, b = \frac{n_s\gamma_s^2}{2}>0$ to be the shape and scale parameters, with $\gamma_s^2$ being the prior for $\sigma^2$ and $n_s$ the prior accuracy for $\gamma_s^2$ – a large value of $n_s$ means we are certain of our prior, i.e. the prior dominates, and so the posterior samples will be close to $\sigma_s^2$. The reason for choosing this prior is that the Inv-Gamma family is conditionally conjugate (i.e. if the prior distribution for $\sigma^2$ is an Inv-Gamma, then the conditional posterior distribution for $\sigma^2$ is also Inv-Gamma), allowing us to easily draw posterior samples for $\sigma^2$ from its full conditional distribution in a Gibbs step. We assume a noninformative prior for $\sigma^2$, so we set $n_s = 1$ and let $\gamma_s^2$ be the initial noise variance, $\sigma_2$ chosen by the user (the defaults in the toolbox). $$\begin{aligned}
\begin{split}
p(\boldsymbol{\theta}) = \det(2\pi T)^{-\frac{1}{2}} \exp \left(-\frac{1}{2}(\boldsymbol{\theta}-\boldsymbol{\mu})^{'}T^{-1}(\boldsymbol{\theta}-\boldsymbol{\mu})\right), \\
p(\sigma^2) = \frac{b^a}{\Gamma(a)} (\sigma^2)^{-a-1} \exp \left(-\frac{b}{\sigma^2}\right)
\end{split}
\end{aligned}$$
- Proposal distribution: $q(\boldsymbol{\theta}^*, \boldsymbol{\theta}^{k-1}) = \text{Truncated } MVN(\boldsymbol{\theta}^{k-1}, V)$.
For $\boldsymbol{\theta}^*$ inside boundaries: $$q(\boldsymbol{\theta}^*,\boldsymbol{\theta}^{k-1}) = \det(2\pi V)^{-\frac{1}{2}} \exp \left(-\frac{1}{2}(\boldsymbol{\theta}^*-\boldsymbol{\theta}^{k-1})^{'}V^{-1}(\boldsymbol{\theta}^*-\boldsymbol{\theta}^{k-1})\right), \\$$
- $\text{Posterior distribution} \propto \text{Likelihood} \times \text{Prior}$, i.e. $$p(\boldsymbol{\theta}, \sigma^2|\mathbf{y}) \propto p(\mathbf{y}|\boldsymbol{\theta}, \sigma^2) p(\boldsymbol{\theta}) p(\sigma^2)$$ which yields: $$\begin{aligned}
\begin{split}
p(\boldsymbol{\theta}, \sigma^2|\mathbf{y}) \propto (\sigma^2)^{-\frac{n}{2}-a-1} \exp \left(-\frac{0.5\sum_{i=1}^n (y_i-m_i(\boldsymbol{\theta}))^2 + b}{\sigma^2}\right) \times \\
\exp \left(-0.5 (\boldsymbol{\theta}-\boldsymbol{\mu})^{'}T^{-1}(\boldsymbol{\theta}-\boldsymbol{\mu})\right).
\end{split}
\end{aligned}$$ In this expression, $\mathbf{m}(\boldsymbol{\theta})$ is the prediction from the PDEs and $\mathbf{y}$ is the measured data.
DRAM is used to sample the parameters $\boldsymbol{\theta}$ in a Metropolis-Hastings step. We then fix $\boldsymbol{\theta}$ to the sample drawn, and we update $\sigma^2$, the error variance, in a Gibbs step, by drawing samples from an Inverse-Gamma distribution, $\textrm{Inv-Gamma}\left(\frac{n}{2}+a, 0.5\sum_{i=1}^n (y_i-m_i(\boldsymbol{\theta}))^2 + b\right) = \textrm{Inv-Gamma}\left(\frac{n_s+n}{2}, \frac{n_s\gamma_s^2 + \sum_{i=1}^n (y_i-m_i(\boldsymbol{\theta}))^2}{2}\right)$. We notice that when going from the prior to the posterior, the distribution is modified to include the contribution from the data, $\frac{\sum_{i=1}^n (y_i-m_i(\boldsymbol{\theta}))^2}{2}$, and the accuracy in the $\sigma^2$ sample increases when data become available (we add $\frac{n}{2}$ to the prior shape parameter - the more data points, the higher the accuracy). We use the $\boldsymbol{\theta}$ samples as an input in the mathematical equations (\[eq:1\]) - (\[eq:2\]), and thus, for every DRAM sample we solve the PDEs. Since every forward simulation takes around 20s of CPU time, in the interest of time, DRAM is started from optimised values. In our analysis, we compare the pressure prediction obtained using the DRAM samples to the measured pressure.
To check for convergence of the Markov chains, we have employed the Geweke test, the Integrated Autocorrelation Time, the Effective Sample Size, and the Brooks Gelman Rubin test for sub-chains of the same chain started from optimised values. We acknowledge the fact that we do not satisfy a fundamental condition of this test (using parallel chains started from overdispersed values), but to make the method computationally viable, adopting these heuristics is indispensable.
Regarding model selection, we have compared the model with tapering (5D) to the model without the tapering factor (4D).
Results
=======
This analysis is done on two mice, a control and a hypoxic mouse. For each mouse, we aim to compare two models: a 5D model which includes all the parameters considered (see (\[eq:12\])), and a 4D model (nested within the 5D model) which excludes the tapering factor $\xi$, as it is uncertain whether it is needed in the model.
Optimization Results
--------------------
We start our analysis with an exploration of the objective function S ( equation (\[eq:5\])), calculated for pressure in the MPA. We can visualize the landscape of S in 1D and 2D space (in 2D we vary two parameters at the time, and in 1D we vary one parameter at the time, while keeping all the others fixed to some value). We choose a representative mouse (the control mouse) and illustrate some of the 2D plots of S in Figure \[fig:2\] (4D model: no tapering) and Figure \[fig:3\] (5D model: tapering).
S exhibits unimodality and it is highly skewed in 2D, and this is the case for both the control and the hypoxic mouse tested, for both models. In some of the figures, there is a sudden jump from low S values to high S values ($10^{10}$). These very high S value region marks the parameter domain that is invalid under the mathematical model. This shows that the biological parameter bounds are valid in an univariate sense, but this domain gets more restricted in higher dimensions, with some parameter combinations that produce no solutions to the PDEs (\[eq:1\]) - (\[eq:2\]), due to being outside the domain of convergence of the numerical scheme. The unimodality in S motivates applying the nonlinear constraint optimization to our problem to find the set of parameters that minimise S. We do so for all of the parameters in the model (see equation (\[eq:12\])), for both mice, and to ensure good coverage of the multidimensional space, we generate 20 initial values from a Sobol sequence. Our results reveal that regardless of the starting value, we manage to recover the same parameter values. control mouse: $f = 85529; r_1 = 1.96; r_2 = 0.25; c = -2.03; \xi = 0.26$, and $S_{\textrm{final}} = 104.5$; hypoxic mouse: $f = 228383; r_1 = 0.59; r_2 = 0.14; c = -0.22; \xi = 0.09$, and $S_{\textrm{final}} = 80$, and we consider these to be the maximum likelihood estimates. We can use these estimates as input into the PDEs to predict the pressure signal, and compare this against the measured pressure (Figure \[fig:4\]), and we do so for both mice.
Figure \[fig:4\] shows the optimised pressure waveform for the 5D problem plotted alongside the measured pressure data for both mice. Panel (a) shows that in the case of the healthy mouse, the simulated pressure closely follows the measured pressure except near the peak, where an offset is registered. The overall model prediction appears to be better for the hypoxic mouse (panel (b)), which is also reinforced by the lower S score.
Next, we set $\xi = 0$ and repeat the analysis for a parameter set that excludes $\xi$. Our findings indicate that for the Control mouse, the fit is poorer if $\xi$ is not included, but the fit is very similar for the Hypoxic mouse. Control mouse: $f = 100398; r_1 = 1.69; r_2 = 0.17; c = -1.30$, and $S_{\textrm{final}} = 126.5$; Hypoxic mouse: $f = 243645; r_1 = 0.52; r_2 = 0.13; c = -0.24$, and $S_{\textrm{final}} = 80.3$. See Figure \[fig:5\] for a visual inspection of the fit.
However, we cannot decide on whether to include $\xi$ in our model without a formal model selection analysis, which will follow later in the paper.
MCMC Results
------------
Next, we aim to quantify the uncertainty around our parameter estimates, and to do so, we implement the DRAM algorithm, the AM algorithm and the DR algorithm to our problem. This is done for each mouse for the two models (with and without the taper factor, $\xi$).
### AM Results
In the AM algorithm there are a couple of parameters that need to be set prior to the commencement of the algorithm. One such parameter is $s_d$ (see (\[eq:8\])), which we take to be the recommended value in literature, $2.4^2/d$. Another such parameter is the length of the adaptation interval, $t_{\textrm{ad}}$ (see (\[eq:8\])), which in practice is chosen to ensure mixing while having enough points that we use for updating the chain covariance matrix. The recommended $t_{\textrm{ad}}$ is 100 in literature, however we have found that for our problem, a much larger $t_{\textrm{ad}}$ was needed. We experimented with values of 100, 500 and 1000. Our findings indicate that $t_{\textrm{ad}} = 1000$ gives higher acceptance rates by up to $5\%$ than $t_{\textrm{ad}} = 500$. The acceptance rate stays between $19\% - 23\%$, and the mixing appears to be better than for $t_{\textrm{ad}} = 100$, so we have chosen $t_{\textrm{ad}} = 1000$. In Figures \[fig:6\] to \[fig:11\] (in the text), and \[fig:10\] and \[fig:12\] (in the Appendix) we show traceplots, marginal posterior densities and scatterplots of parameters sampled during the AM simulation for the two mice, for both models and Table \[table:1\] provides posterior summaries for the parameters, and we notice that the posterior means are very similar to the optimised values from the Optimization Results section.
For the control mouse, for the two models (Figures \[fig:6\] - \[fig:9\] in the text and \[fig:10\] in the Appendix), we can see that the chains fluctuate around the parameter values obtained from optimisation, suggesting that the chain has reached the highest posterior density region, and that the optimization might have found the global optimum.
There are no signs of multimodality in the objective function, which was reinforced by the convergence results we obtained through optimization. In addition, the S values are steady around the initial, lowest S value and the $\sigma^2$ chain does not indicate non-convergence (see Figure \[fig:7\]). Also, the marginal posterior densities for some parameters have long tails (Figure \[fig:8\]), which reinforces the skewness evident in the 2D S plots (Figures \[fig:2\] and \[fig:3\]). The scatterplots (Figures \[fig:9\]) reveal strong correlations between some of the parameters, which is a reason why the algorithm needs longer runs to converge. For example, the stiffness $f$ is highly correlated to the tapering factor, $\xi$, as the wall stiffness increases, the vessels will become less tapered (top and bottom vessel radii get closer together), which is what we would expect (see (\[eq:14\])). Another example is the high correlation between stiffness, $f$ and compliance adjustment parameter, $c$, i.e. as the stiffness of the vessel wall increases, the compliance adjustment increases, and so the nominal compliance or elasticity decreases, which is expected (see (\[eq:13\])).
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the Adaptive Metropolis algorithm for the 5D model corresponding to a control mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $22\%$.[]{data-label="fig:6"}](AM_param_C06taper_grey.png)
![Markov chains traceplots for the error variance $\sigma^2$, and the residual-sum-of-squares $S$ obtained using the Adaptive Metropolis algorithm for the 5D model corresponding to a control mouse.[]{data-label="fig:7"}](AM_s2S_C06taper_grey.png)
![Marginal posterior densities of parameters (defined in the Simulations section) inferred using the Adaptive Metropolis algorithm for the 5D model corresponding to a control mouse. Plots were obtained using a kernel density estimator with customised bandwidth.[]{data-label="fig:8"}](AM_margDens_C06taper_grey.png)
![Scatterplots of parameter (defined in the Simulations section) chains obtained using the Adaptive Metropolis algorithm for the 5D model corresponding to a control mouse.[]{data-label="fig:9"}](AM_scatterplots_C06taper_grey.png)
Regarding the hypoxic mouse, for the 5D model, we note a more pronounced skewness in the distribution for two of the parameters: $f$ and $\xi$ (Figure \[fig:11\]), which are strongly correlated. We also notice that the biological range for $\xi$ is overly conservative, and a wider range should be used. In addition, the posterior mean appears to vary slightly from the maximum likelihood estimates for some of the parameters, but the two are more similar in the 4D model for the hypoxic mouse (Figure \[fig:12\] in the Appendix).
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the Adaptive Metropolis algorithm for the 5D model corresponding to a hypoxic mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $23\%$.[]{data-label="fig:11"}](AM_param_H01taper_grey.png)
### DR Results
Figure \[fig:13\] shows 10000 iterations of a chain produced using the DR algorithm for the control mouse, 5D model. Despite the acceptance rate which is close to the optimal target acceptance rate given in literature ($23\%$), the plots reveal non-convergence of the chains, and the parameter space is explored slowly, leading to bad mixing. DR seems to be affected by the strong correlations between the parameters. Therefore, the chains would have to be run for a very long time, implying that the AM algorithm is essential in reaching convergence in a reasonable time frame.
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the Delayed Rejection algorithm for the 5D model corresponding to a control mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $22\%$.[]{data-label="fig:13"}](DR_param_C06taper_grey.pdf)
### DRAM Results
Lastly, we implement the DRAM algorithm starting from optimised values for every mouse and model type. We find that the estimates are very similar to the estimates obtained using the AM algorithm (see Table \[table:1\]), which is expected. The algorithm used to sample draws is irrelevant if the samples are approximately drawn from the posterior distribution; the algorithm only influences the convergence of the chains. This is supported by the traceplots showing these samples (Figures \[fig:14\] and \[fig:15\] in the text and \[fig:16\], \[fig:17\], \[fig:18\] in the Appendix), which do not indicate non-convergence, as the chains appear to be exploring the space around the optimum values well and are steady around these values. We find that the DRAM algorithm has increased the acceptance rate compared to the AM algorithm (from $20\%$ up to $40\%$). This is not surprising, considering the extra DR step in the DRAM algorithm, which proposes a second point at every iteration of the algorithm. This clearly increases the acceptance rate, and thus, fewer iterations are needed to reach convergence when compared to AM. This, however, does not imply less computational time, since in every DRAM iteration, when a second point is proposed, the PDEs have to be solved twice. Therefore, AM and DRAM are computationally equivalent in our case, since double the number of iterations were needed for AM to decrease the correlation between chain samples. The inspection of the convergence diagnostics (discussed in the MCMC Convergence Diagnostics section below), as well as the improvement in the acceptance rate has made us favour DRAM over AM.
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the Delayed Rejection Adaptive Metropolis algorithm for the 5D model corresponding to a control mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $39\%$.[]{data-label="fig:14"}](DRAM_param_C06taper_grey.png)
![Markov chains traceplots for the error variance $\sigma^2$, and the residual-sum-of-squares $S$ obtained using the Delayed Rejection Adaptive Metropolis algorithm for the 5D model corresponding to a control mouse.[]{data-label="fig:15"}](DRAM_s2S_C06taper_grey.png)
Let us now illustrate (Figures \[fig:19\] and \[fig:20\]) what happens if we use the DRAM algorithm in [@UncertaintySmith], that does not work in the transformed parameter space when calculating the acceptance probabilities, in particular the $2^{\textrm{nd}}$ stage acceptance probability.
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the original Delayed Rejection Adaptive Metropolis algorithm in [@UncertaintySmith] without the correction term discussed in this paper, for the 5D model corresponding to a control mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $30\%$.[]{data-label="fig:19"}](DRAMuncorrected_param_C06taper_grey.png)
![Markov chains traceplots for the error variance $\sigma^2$, and the residual-sum-of-squares $S$ obtained obtained using the original Delayed Rejection Adaptive Metropolis algorithm in [@UncertaintySmith] without the correction term discussed in this paper, for the 5D model corresponding to a control mouse.[]{data-label="fig:20"}](DRAMuncorrected_s2S_C06taper_grey.png)
When looking at Figure \[fig:20\], it is readily apparent that several proposals with high S are being falsely accepted. Figure \[fig:15\] for the correct DRAM algorithm does not display such spikes, and the traceplots in Figure \[fig:14\] are well-behaved. We falsely accept points because the ratio of the proposal distributions in equation (\[eq:7\]) becomes very large, since it is calculated in the unscaled parameter space, and the starting values (optimum values) are very close to the boundaries or to the invalid paramater domain (see Figure \[fig:3\]). This implies that many of the proposed points will either lie outside boundaries or inside the invalid domain, making $f$, the parameter with dominating magnitude, and S become high. When this ratio becomes large, the proposed point will be falsely accepted, as the acceptance rate is the minimum between 1 and the $\log$ of a high number, i.e. the acceptance probability will be 1. This would however not happen if the optimum value was somewhere far from the boundaries or the invalid domain, because the algorithm would explore a region of fairly similar values. This is tested for the case of the control mouse, model 4D, and the findings (not shown here) back up our hypothesis.
MCMC Convergence Diagnostics
----------------------------
It is now of interest to perform convergence diagnostics on the DRAM chains for every mouse, for the two models. To do so, we employ the Geweke test, the Brooks Gelman Rubin test, the Integrated Autocorrelated Time (IACT), and the Effective Sample Size (ESS) tests and the results are summarised in Table \[table:2\].
The Geweke test reveals a high p-value (0.99), well above 0.05 for all 4 scenarios, suggesting that there is no evidence of a difference between the mean of the first $10\%$ samples and last $50\%$ samples in the chain. The MPSRF is below 1.1 for all 4 cases, the highest one being 1.03 and the lowest 1.0004. The 4D model for both mice provides a lower MPSRF $(< 1.01)$ than the 5D model. Nevertheless, the within and between chain covariance matrices are similar for both models. The estimates for IACT and ESS suggest that there is some correlation still left between the samples, in particular for the 5D model, where we have seen that there is a very strong correlation between $f$ and $\xi$. This strong correlation delays the algorithm from sampling independent draws. For the 4D model, for both mice, we notice that on average, one independent sample is drawn every 13 samples. Therefore, the ESS is somewhere in the range \[1200, 1800\] out of a total number of 20000 iterations. For the 5D model, there is a more pronounced correlation effect, with an average IACT of 27 and ESS ranging between \[700, 1500\]. While the Geweke test and the MPSRF do not indicate non-convergence of the chains, the IACT values (slightly too high) and ESS values (slightly too low) indicate that the DRAM algorithm should be run for longer in order to decrease the correlation between samples.
When analysing Table \[table:3\] which displays the MCMC convergence diagnostics for the AM algorithm, we notice that the AM algorithm produces even more correlated samples, the IACT being consistently higher than that for the DRAM algorithm.
Model Selection {#model-selection}
---------------
We further consider model selection, since we have two competing models for the two mice: a 4D model (without tapering factor, $\xi$ in), and a 5D model (with $\xi$ in). For the purpose of model selection, we show the performance of 4 criteria: AICc, BIC, DIC and WAIC. We have seen in the previous section that there is still some correlation left between the chain samples. We therefore only consider a subset of 1000 samples of the total number of samples when computing WAIC. Table \[table:4\] summarises our model selection scores.
The results indicate that the scores are similar and fairly consistent across different criteria. For the control mouse, all the scores are lower for the 5D model, compared to the 4D model, which implies that in the case of the control mouse, adding the tapering factor, $\xi$ makes the model more appropriate for the data, and so, it is worth including $\xi$ in the model. On the other hand, for the hypoxic mouse, the scores are very similar across the two models, WAIC favours the 4D model, while AICc, BIC and DIC favour the 5D model. Considering the very low difference in S between the 4D and 5D models (0.56), and the fact that there is low difference between the scores, then we choose the 4D model to be used in the case of the hypoxic mouse. Hence, there is evidence that the tapering factor is not significant for the hypoxic mouse, but the opposite is true for the control mouse. A possible explanation for this could be that in hypoxia, the vessel walls become stiff, and so there will not be a pronounced decrease in the radii along the vessels’ length.
Discussion
==========
In this paper we have applied the Delayed Rejection Adaptive Metropolis algorithm and its variations (Adaptive Metropolis and Delayed Rejection) to measure uncertainty of the parameter estimates in a partial differential equations system of the pulmonary circulation. The method, which builds on and further improves existing literature by allowing for novel parameter scaling due to different magnitudes of the parameters has been applied on real data, coming from a healthy and a diseased mouse. Making a few restrictive assumptions about the parameters has allowed us to decrease the parameter dimension considerably (from a 55D problem – $21 f, 11 r_1$, $11 r_2, 11 c, 1 \xi$ parameters, down to a 5D problem – $1f, 1 r_1, 1 r_2, 1c, 1 \xi$). Further studies should be done to check these assumptions. Some preliminary exploration of non-constant Windkessel adjustments, $r_1, r_2, c$ and stiffness $f$ on a smaller network has indicated that fitting different sets of parameters for every vessel would result in data overfitting. Even if the goodness of fit has improved, the model complexity has increased significantly. Model selection criteria have revealed that a very low error variance would be needed to favour use of the more complex model. Since our error variance is believed to be higher, we have rejected the complex model. The parameters had different orders of magnitude, which motivated us to improve existing DRAM code from the literature in order to assess uncertainty in a system of partial differential equations, which is the pulmonary circulation. To make the computations feasible, we have started the MCMC algorithm from optimised parameter values (a nonlinear constraint optimisation scheme was used). Our findings indicate that DRAM is preferred over DR and AM. The DR algorithm alone is very slow in converging, and the chains appear far from converging even after 10000 iterations. The AM algorithm retains an acceptance rate of about $20\%$, which appears to be too low, and there is correlation between the samples even after 36000 iterations. The DRAM algorithm appears to be the best, with an acceptance rate of about $40\%$, and fairly good mixing. While the Geweke test and the Multivariate Scale Reduction Factor test did not indicate non-convergence, the Integrated Autocorrelation Time appeared to be slightly too high after 20000 iterations. The strong dependence between some of the parameters makes the algorithm slow in generating more independent samples, and should therefore be run for longer. Another focus of this paper was to perform model selection to identify whether one of the parameters should be included in the model. The AICc, BIC, DIC and WAIC criterion consistently chose the more complex model for the control mouse. For the hypoxic mouse, the scores were inconsistent, but since the difference between them was low, and the most reliable score, WAIC, favoured the simpler model, we preferred this model. All these results are based on a mathematical model that looks more appropriate for the hypoxic mouse than for the control mouse. Possible causes for the slight model mismatch are: (i) the simplicity of the model specifying the elastic behaviour of the blood vessels and/or the boundary conditions, (ii) uncertainty of the geometry measurements which are not specific to a given mouse, (iii) a combination of (i) and (ii). This slight model mis-specification has the undesired effect of breaking the i.i.d. error Normality assumption. Therefore, to account for the correlated structure in the errors, we aim to use a Markov model that will also model the correlation structure, and not only the parameters and the error variance. In addition, future work will include implementing faster methods to infer the relevant parameters. This includes the Hamiltonian Monte Carlo algorithm [@HMC], which suppresses the random behaviour of the algorithm by using a momentum variable that leads the direction in which the proposals are made. To speed up the simulations, we will run this algorithm on an objective function emulated using Gaussian Processes [@GPs]. Another alternative to learn the unobserved parameters is by employing the Ensemble Kalman filter, following the idea in [@EnKF].
acknowledgements {#acknowledgements .unnumbered}
================
This work is part of the research programme of the Centre for multiscale soft tissue mechanics with application to heart & cancer (SofTMech), funded by the Engineering and Physical Sciences Research Council (EPSRC) of the UK, grant reference number EP/N014642/1. Olufsen, Haider and Qureshi were supported by National Science Foundation (NSF-DMS \# 1615820). Data were made available by N. Chesler, Department of Biomedical Engineering, University of Wisconsin, Madison.
------------ --------------- -------------- -------------- -------------- ------------- -------------- -------
Control 4D 100374 (391) 1.69 (0.01) 0.17 (0.005) -1.30 (0.04) - 0.12 (0.005) 127
Control 5D 85471 (658) 1.96 (0.006) 0.25 (0.01) -2.03 (0.07) 0.26 (0.01) 0.10 (0.004) 105
Hypoxic 4D 243653 (1180) 0.52 (0.01) 0.13 (0.003) -0.24 (0.01) - 0.08 (0.003) 80.56
Hypoxic 5D 231627 (3720) 0.57 (0.02) 0.14 (0.003) -0.23 (0.01) 0.07 (0.02) 0.08 (0.003) 80
------------ --------------- -------------- -------------- -------------- ------------- -------------- -------
: Parameter ($f, r_1, r_2, c, \xi$) estimates and $\sigma^2$ estimate with their standard deviations inside brackets and S (residual-sum-of squares) for every type of mouse (control and hypoxic) and model (4D and 5D) obtained from running the Adaptive Metropolis algorithm. These parameters are defined in the Simulations section.[]{data-label="table:1"}
------------ -------------- ------- ------------------ -------------------------
Control 4D all $>$ 0.05 1.004 (11,12,11,14) (1443,1362,1424,1180)
Control 5D all $>$ 0.05 1.01 (22,31,28,21,22) (912,639,702,930,927)
Hypoxic 4D all $>$ 0.05 1.006 (14,15,14,11) (1412,1333,1465,1807)
Hypoxic 5D all $>$ 0.05 1.03 (23,22,13,17,27) (862,897,1530,1144,732)
------------ -------------- ------- ------------------ -------------------------
: MCMC Convergence Diagnostics: Geweke test, Multivariate Potential Scale Reduction Factor (MPSRF), Integrated Autocorrelation Time (IACT) for each of the parameters, Effective Sample Size (ESS) for each of the parameters, corresponding to the results obtained using the Delayed Rejection Adaptive Metropolis algorithm for the two mice (control, hypoxic) and the two models (4D containing 4 parameters and 5D containing 5 parameters).[]{data-label="table:2"}
------------ -------------- ------- ------------------ -------------------------
Control 4D all $>$ 0.05 1.01 (22,22,20,22) (1630,1671,1811,1625)
Control 5D all $>$ 0.05 1.04 (39,47,49,27,43) (918,767,731,1317,837)
Hypoxic 4D all $>$ 0.05 1.004 (17,18,18,18) (2123,2030,2056,1990)
Hypoxic 5D all $>$ 0.05 1.06 (49,37,19,26,58) (730,965,1899,1375,622)
------------ -------------- ------- ------------------ -------------------------
: MCMC Convergence Diagnostics: Geweke test, Multivariate Potential Scale Reduction Factor (MPSRF), Integrated Autocorrelation Time (IACT) for each of the parameters, Effective Sample Size (ESS) for each of the parameters, corresponding to the results obtained using the Adaptive Metropolis algorithm for the two mice (control, hypoxic) and the two models (4D containing 4 parameters and 5D containing 5 parameters).[]{data-label="table:3"}
------------ ----- ----- ----- -----
Control 4D 772 792 772 774
Control 5D 579 603 578 580
Hypoxic 4D 307 326 306 304
Hypoxic 5D 301 325 300 310
------------ ----- ----- ----- -----
: Model selection scores (AICc, BIC, DIC, WAIC) for the 4D and 5D models for the control and the hypoxic mouse. AICc is very similar to AIC (up to 4 dp), since the sample size is large (1024 data points). For the calculation of AICc and BIC, we have estimated $\hat{\sigma}^2 = \frac{S}{n-d}$, for DIC, $\hat{\sigma}^2$ is the posterior mean of the $\sigma^2$ samples, and for WAIC, every $\sigma^2$ sample in the chain is used.[]{data-label="table:4"}
![Markov chains traceplots for the parameters obtained using the Adaptive Metropolis algorithm for the 4D model corresponding to a control mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $20\%$.[]{data-label="fig:10"}](AM_param_C06notaper_grey.png)
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the Adaptive Metropolis algorithm for the 4D model corresponding to a hypoxic mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $19\%$.[]{data-label="fig:12"}](AM_param_H01notaper_grey.png)
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the Delayed Rejection Adaptive Metropolis algorithm for the 4D model corresponding to a control mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $48\%$.[]{data-label="fig:16"}](DRAM_param_C06notaper_grey.png)
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the Delayed Rejection Adaptive Metropolis algorithm for the 5D model corresponding to a hypoxic mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $45\%$.[]{data-label="fig:17"}](DRAM_param_H01taper_grey.png)
![Markov chains traceplots for the parameters (defined in the Simulations section) obtained using the Delayed Rejection Adaptive Metropolis algorithm for the 4D model corresponding to a hypoxic mouse. Starting values for the algorithm are the optimised values and are superimposed in black horizontal lines. Acceptance rate is $46\%$.[]{data-label="fig:18"}](DRAM_param_H01notaper_grey.png)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We generalize to the case of spinning black holes a recently introduced “effective one-body” approach to the general relativistic dynamics of binary systems. We show how to approximately map the conservative part of the third post-Newtonian (3 PN) dynamics of two spinning black holes of masses $m_1$, $m_2$ and spins $\mbox{\boldmath$S$}_1$, $\mbox{\boldmath$S$}_2$ onto the dynamics of a non-spinning particle of mass $\mu
\equiv m_1 \, m_2 / (m_1 + m_2)$ in a certain effective metric $g_{\mu \nu}^{\rm eff}
(x^{\lambda} ; M , \nu , \mbox{\boldmath$a$})$ which can be viewed either as a spin-deformation (with deformation parameter $\mbox{\boldmath$a$} \equiv
\mbox{\boldmath$S$}_{\rm eff} / M$) of the recently constructed 3 PN effective metric $g_{\mu \nu}^{\rm eff} (x^{\lambda} ; M , \nu)$, or as a $\nu$-deformation (with comparable-mass deformation parameter $\nu \equiv m_1 \, m_2 / (m_1 + m_2)^2$) of a Kerr metric of mass $M \equiv m_1 + m_2$ and (effective) spin $\mbox{\boldmath$S$}_{\rm eff}
\equiv (1+3 \, m_2 / (4 \, m_1)) \, \mbox{\boldmath$S$}_1 + (1+3 \, m_1 / (4 \, m_2)) \,
\mbox{\boldmath$S$}_2$. The combination of the effective one-body approach, and of a Padé definition of the crucial effective radial functions, is shown to define a dynamics with much improved post-Newtonian convergence properties, even for black hole separations of the order of $6~GM / c^2$. The complete (conservative) phase-space evolution equations of binary spinning black hole systems are written down and their exact and approximate first integrals are discussed. This leads to the approximate existence of a two-parameter family of “spherical orbits” (with constant radius), and of a corresponding one-parameter family of “last stable spherical orbits” (LSSO). These orbits are of special interest for forthcoming LIGO/VIRGO/GEO gravitational wave observations. The binding energy and total angular momentum of LSSO’s are studied in some detail. It is argued that for most (but not all) of the parameter space of two spinning holes the approximate (leading-order) effective one-body approach introduced here gives a reliable analytical tool for describing the dynamics of the last orbits before coalescence. This tool predicts, in a quantitative way, how certain spin orientations increase the binding energy of the LSSO. This leads to a detection bias, in LIGO/VIRGO/GEO observations, favouring spinning black hole systems, and makes it urgent to complete the conservative effective one-body dynamics given here by adding (resummed) radiation reaction effects, and by constructing gravitational waveform templates that include spin effects. Finally, our approach predicts that the spin of the final hole formed by the coalescence of two arbitrarily spinning holes never approaches extremality.
address: '[*Institut des Hautes Etudes Scientifiques, 91440 Bures-sur-Yvette, France*]{}'
author:
- Thibault Damour
date: 'Sept 18, 2001'
title: |
Coalescence of Two Spinning Black Holes:\
An Effective One-Body Approach
---
Introduction {#sec1}
============
The most promising candidate sources for the LIGO/VIRGO/GEO/$\ldots$ network of ground based gravitational wave (GW) interferometric detectors are coalescing binary systems made of massive (stellar) black holes [@postnov; @FH98; @BCT98; @PZ99; @DIS00]. Signal to noise ratio (SNR) estimates [@DIS00] suggest that the first detections will concern black hole binaries of total mass $\gaq \, 25 M_{\odot}$. Modelling the GW signal emitted by such systems poses a difficult theoretical problem because the observationally most “useful” part of the gravitational waveform is emitted in the last $\sim 5$ orbits of the inspiral, and during the “plunge” taking place after crossing the last stable circular orbit. The transition between (adiabatic) inspiral and plunge takes place in a regime where the two bodies are moving at relativistic speeds ($v/c \sim 1 / \sqrt 6 \sim 0.4$) and where their gravitational interaction becomes (nearly by definition) highly non-linear ($GM / c^2 r \sim 1/6$).
Several authors (notably [@cutler93; @BCT98]) have taken the view that the modelling of this crucial transition between inspiral and plunge is (in the general case of comparable-mass systems) beyond the reach of analytical tools and can only be tackled by (possibly special-purpose [@BCT98]) numerical simulations. By contrast, other authors [@DIS98; @BD99; @BD00; @DIS00; @DJS2; @BD01] have introduced new “resummation methods” to improve the analytical description of the last few GW cycles near this transition and have argued that these resummed analytical results gave a reliable description of the gravitational physics near the transition. The purpose of the present paper is to further the latter approach by generalizing the (resummed) “effective one-body” (EOB) methods introduced in [@BD99; @BD00; @DJS2] to the case of binary systems of [*spinning*]{} black holes. Before doing this, we wish to clarify what is the rationale for arguing that the “resummed” analytical approach can describe the last stages of inspiral and the transition between inspiral and plunge.
Let us first recall that a lot of effort has been devoted in recent years to the analytical computing, by means of post-Newtonian (PN) expansions in powers of $v^2 /
c^2 \sim GM / c^2 r$, of the equations of motion, and the GW emission, of comparable-mass binary systems. The equations of motions have been computed to $v^6 / c^6$ ($3$ PN) accuracy by two separate groups, [@JS98; @JS99; @DJS1; @DJS3] and [@BF1; @BF2; @BF3], and the two results have been shown to agree [@DJS4; @BF4]. Until recently, there remained (in both approaches) an ambiguous parameter, $\omega_s$, linked to the problem of regularizing some badly divergent integrals arising at the 3 PN level. In a recent work, using an improved regularization method (dimensional continuation), the first group [@DJSd] has succeeded in determining without ambiguity the value of $\omega_s$, namely $\omega_s = 0$. This unique determination of the 3PN equations of motion is consistent with an old argument of [@D83] showing that it should be possible to model black holes by point particles without ambiguity up to the 5 PN level (excluded). The emission of GW is unambiguously known to $v^5 / c^5$ ($2.5$ PN) accuracy [@2.5PN], and has recently been formally computed to $v^7 / c^7$ ($3.5$ PN) accuracy [@3.5PN], modulo the appearance of several ambiguous parameters ($\xi, \kappa, \zeta$) linked to the problem of regularizing some divergences arising at the 3PN level. Dimensional regularization is expected to determine without ambiguity the values of $\xi, \kappa$, and $ \zeta$, but has not (yet) been applied to the radiation problem.
We wish to emphasize that such high-order PN results are a [*necessary*]{}, but [ *not*]{} by themselves [*sufficient*]{}, ingredient for computing with adequate accuracy the gravitational waveform of coalescing binaries. Indeed, it was emphasized long ago [@cutler93] that the PN series (written as straightforward Taylor series in powers of some parameter $\varepsilon \sim v/c$) become slowly convergent in the late stages of binary inspiral. A first attempt was made in [@KWW] to improve the convergence of the PN-expanded equations of motion so as to determine the (crucial) location of the last stable (circular) orbit (LSO) for comparable-mass systems. However, further work [@WS93; @SW93; @DIS98] has shown the unreliability (and coordinate-dependence) of this attempt. There is, however, no reason of principle preventing the existence of gauge-invariant “resummation methods” able to give reliable results near the LSO. Indeed, as emphasized in [@DIS98] and [@DJS2] most coordinate-invariant functions (of some invariant quantity $x \sim v^2 / c^2
\sim GM / c^2 r$) that one wishes to consider when discussing the dynamics and GW emission of circular orbits are expected to have a singularity only at the “light ring” (LR) value of $x$ \[last possible [*unstable*]{} circular orbit\]. If we trust (for orders of magnitude considerations) the small mass-ratio limit ($\nu \equiv \mu /
M \equiv m_1 \, m_2 / (m_1 + m_2)^2 \ll 1$), we know that $x_{\rm LR} \simeq 1/3$ is smaller by a factor 2 than $x_{\rm LSO} \simeq 1/6$. If the functions $f(x)$ we are dealing with are meromorphic functions of $x$, the location of the expected closest singularity $(x_{\rm LR})$ determines their radius of convergence. Therefore, we expect that, for $x < x_{\rm LR}$, the Taylor expansion of $f(x)$ will converge and will behave essentially like ${\sum_n \ (x/x_{\rm LR})^n}$. In particular, one expects ${ f (x_{\rm LSO}) \sim \sum_n \ (x_{\rm LSO} / x_{\rm LR})^n
\sim \sum \ 2^{-n}}$. This heuristic argument suggests a rather slow convergence, but the crucial point is to have some convergence, so that the application of suitable [*resummation methods*]{} can be expected to accelerate the convergence and to lead to numerically accurate results from the knowledge of only a few terms in the Taylor expansion. There exist many types of resummation methods and none of them are of truly universal applicability. As a rule, one must know something about the structure of the functions $f(x) = f_0 + f_1 \, x + f_2 \, x^2 + \cdots$ one is trying to resum to be able to devise an efficient resummation method. Refs. [@DIS98], [@BD99], [@BD00] and [@DJS2] have studied in detail the various functions that might be used to discuss the GW flux and the dynamics of binary systems. This work has led to selecting some specific resummation methods, acting on some specific functions. For what concerns the GW flux we refer to Fig. 3 of [@DIS98] for evidence of the acceleration of convergence (near the LSO) provided by a specific resummation method combining a redefinition of the [*GW flux*]{} function with Padé approximants. We wish here, for the benefit of the skeptics, to exhibit some of the evidence for the acceleration of convergence (near the LSO) in the description of the 2-body [ *dynamics*]{} provided, at the 3 PN level, by a resummation method defined by combining [@BD99] and [@DJS2]. Specifically, we mean the combination of the effective-one-body (EOB) approach (further discussed below) and of a suitable Padé resummation of the effective radial potential at the $n$ PN level: $A_{P_n} (u) =
P_n^1 \, [T_{n+1} (A(u))]$ (see below).
Let us consider a sequence of [*circular*]{} orbits, near the LSO and for two non-spinning black holes. In the EOB approach the circular orbits are obtained by minimizing a certain effective radial potential, defined by fixing the total orbital angular momentum $L$ in the Hamiltonian. The most natural variable defining the one-parameter sequence of circular orbits is then simply the angular momentum $L$. It is therefore natural, for the purpose of this work, to measure the separation between the two holes (in a gauge-invariant and approximation-independent) way by [*conventionally*]{} defining a $\ell$-radius $R_{\ell} \equiv GM \, r_{\ell}$, such that the (invariantly defined) total orbital angular momentum $L \equiv G \mu M
\ell$ is given by $\ell^2 \equiv r_{\ell}^2 / (r_{\ell} - 3)$, i.e. by the relation holding for a test particle in a Schwarzschild spacetime. \[Here, and in the following, we shall often set $c=1$ and/or $G=1$, except in some (final) formulas where it might be illuminating to reestablish the dependence on $c$ and/or $G$.\] As the problem is to know whether the resummation method of the PN-expanded two-body dynamics is efficient, we compare in Table \[Tab1\] the total energies ${\cal E}_{\rm real}$ of the binary system, computed using 1 PN, 2 PN and 3 PN information, for circular orbits at $\ell$-radii $r_{\ell} = 12, 11, 10, 9,
8, 7$, and 6. We give the values of the binding energy per unit (total) mass, $e
\equiv ({\cal E} / M) - 1$, for the equal-mass case ($m_1 = m_2$; $\nu = 1/4$).
$r_{\ell}$ $12$ $11$ $10$ $9$ $8$ $7$ $6$
--------------------------------------------- ----------- ---------- ---------- ---------- ---------- ---------- ----------
$\ell$ $4.$ $3.889$ $3.780$ $3.674$ $3.578$ $3.5$ $3.464$
$100 \, e_{1 \, {\rm PN}}$ $-0.9482$ $-1.020$ $-1.101$ $-1.193$ $-1.291$ $-1.387$ $-1.440$
$100 \, e_{2 \, {\rm PN}}$ $-0.9441$ $-1.015$ $-1.094$ $-1.183$ $-1.277$ $-1.366$ $-1.412$
$100 \, e_{3 \, {\rm PN}}^{(\omega_s = 0)}$ $-0.9412$ $-1.011$ $-1.088$ $-1.174$ $-1.264$ $-1.346$ $-1.388$
: Binding energies $e \equiv ({\cal E}_{\rm real} / M) - 1$ for circular orbits of equal-mass (non-spinning) binary systems near the LSO. The invariant dimensionless $\ell$-radius $r_{\ell} \equiv R_{\ell} / GM$ is defined in the text. The binding energy is computed from the Padé-resummed Effective-one-body Hamiltonian at three successive post-Newtonian approximations: 1 PN, 2 PN and 3 PN (with $\omega_s = 0$ ).[]{data-label="Tab1"}
The numbers displayed in Table \[Tab1\] illustrate the efficiency of the resummation method advocated in [@BD99; @DJS2]. For $r_{\ell} = 12$ the fractional difference in binding energy between the 1 PN approximation and the 3 PN one is 0.74%, while even for $r_{\ell} = 6$ this difference is only 3.6%. These numbers indicate that, even near the LSO, the Padé-improved effective-one-body approach is a rationally sound way of computing the 2-body dynamics. There are no signs of numerical unreliability, as there were in the calculations based on the straightforward coordinate-dependent, PN-expanded versions of the equations of motion [@KWW] or of the Hamiltonian [@WS93; @SW93] (which gave results differing by ${\cal O}$ (100%) among themselves, and as one changed the PN order). We shall see below that the robustness of the PN-resummation exhibited in Table \[Tab1\] extends to a large domain of the parameter space of spinning black holes.
As we do not know the exact result, and as current numerical simulations do not give reliable information about the late stages of the quasi-circular orbital dynamics of two black holes (see below), the kind of internal consistency check exhibited in Table \[Tab1\] is about the only evidence we can set forth at present. \[Note that, from a logical point of view, the situation here is the same as for numerical simulations: in absence of an exact solution (and of experimental data) one can only do internal convergence tests.\] Ideally, it would be important to extend the checks of Table \[Tab1\] to the 4 PN level (to confirm the trend and see a real sign of convergence to a limit) but this seems to be an hopelessly difficult task with present analytical means. Finally, let us note that the fact that one can concoct many “bad” ways of using the PN-expanded information near the LSO (exhibiting as badly divergent results as wished) is not a valid argument against the reliability of the specific resummation technique used in [@DIS98; @BD99; @DJS2] and here. An ambiguity problem would arise only if one could construct two different resummation methods, both exhibiting an internal “convergence” (as the PN order increases) as good as that illustrated in Table \[Tab1\], but yielding very different predictions for physical observables near the LSO. This is not the case at present because the comparative study of [@DJS2] (see Table I there) has shown that the EOB approach exhibited (when $\omega_s \ne -9$) significantly better PN convergence than a panel of other invariant resummation methods.
In the following we take for granted the soundness of the effective-one-body resummation approach and we show how to generalize it to the case of two (moderately) spinning black holes. Let us first recall that the basic idea of the EOB approach was first developed in the context of the electromagnetically interacting quantum two-body problem [@BIZ], [@T1] (see also [@B00]). A first attempt to deal with the gravitationally interacting two-body problem (at the 1 PN level) was made in [@T2] (see also [@FT01]). A renewed EOB approach (which significantly differs from the general framework set up by Todorov and coworkers [@T1; @T2]) was introduced in [@BD99]. The latter reference showed how to apply this method at the 2 PN level. It was then used to study the transition between the inspiral and the plunge for comparable masses, and, in particular, to construct a complete waveform covering the inspiral, the plunge and the final merger [@BD00] (see [@DIS3] for the physical consequences of this waveform). More recently, Ref. [@DJS2] showed how to extend the EOB approach to the 3 PN level (this required a non trivial generalization of the basic idea).
Before entering the details of our way of introducing the spin degrees of freedom in the EOB approach, let us state our general view of the usefulness of the EOB method in this context. As we shall discuss below, the present work (which incorporates spin effects at leading PN order) can only be expected to give physically reliable results in the case of moderate spins ( $\widehat{a}_p \alt 0.3$, see below). However, the EOB approach, far from being a rigid structure, is extremely flexible. One can modify the basic functions (such as $A(u)$) determining the EOB dynamics by introducing new parameters corresponding to (yet) uncalculated higher PN effects. \[ These terms become important only for orbits closer than $ 6 \, G M$, and/or for fast-spinning holes.\] Therefore, when either higher-accuracy analytical calculations are performed or numerical relativity becomes able to give physically relevant data about the interaction of (fast-spinning) black holes, we expect that it will be possible to complete the current EOB Hamiltonian so as to incorporate this information. As the parameter space of two spinning black holes (with arbitrarily oriented spins) is very large, numerical relativity will never be able, by itself, to cover it densely. We think, however, that a suitable “numerically fitted” (and, if possible, “analytically extended”) EOB Hamiltonian should be able to fit the needs of upcoming GW detectors. The present work should be viewed as a first step in this direction.
The present paper is organized as follows. In Section \[sec2\] we show how to incorporate (in some approximation) the spin degrees of freedom of each black hole within a 3 PN-level, resummed effective one-body approach. In Section \[sec3\] we study some of the predictions of our resummed dynamics, notably for what concerns the location of the transition between the inspiral and the plunge. Section \[sec4\] contains our concluding remarks.
Effective one-body approach, effective spin and a deformed Kerr metric {#sec2}
======================================================================
Effective one-body approach {#ssec2.1}
---------------------------
Let us recall the basic set up of the effective one-body (EOB) approach. One starts from the (PN-expanded) two-body equations of motion, which depend on the dynamical variables of two particles. One separates the equations of motion in a “conservative part”, and a “radiation reaction part”. Though this separation is not well-defined at the exact (general relativistic) level it is not ambiguous at the 3 PN level (in the conservative part) which we shall consider here [^1]. We shall henceforth consider only the conservative part of the dynamics. \[We leave to future work the generalization to spinning black holes of the definition of [ *resummed*]{} radiation reaction effects which was achieved in [@BD00] for non-spinning black holes.\] It has been explicitly shown that the 3 PN[^2] dynamics is Poincaré invariant [@DJS3], [@BF2]. The ten first integrals associated to the ten generators of the Poincaré group were constructed in [@DJS3] (see also [@BF4]). In particular, we have the “center of mass” vectorial constant $\mbox{\boldmath$K$} =
\mbox{\boldmath$G$} - t \, \mbox{\boldmath$P$}$. This constant allows one to define the center of mass frame, in which $\mbox{\boldmath$K$} = 0$, which implies $\mbox{\boldmath$P$} = 0$ and $\mbox{\boldmath$G$} = 0$. We can then reduce the PN-expanded two-body dynamics to a PN-expanded one-body dynamics by considering the relative motion in the center of mass frame. This reduction leads to a great simplification of the dynamics.
Indeed, the full 3 PN Hamiltonian in an arbitrary reference frame [@DJS3] contains ${\cal O} (100)$ terms, while its center-of-mass-reduced version contains only 24 terms. However, this simplification is, by itself, insufficient for helping in any way the crucial problem of the slow convergence of the PN expansion. One should also mention that the use of an Hamiltonian framework (like the ADM formalism used in [@JS98; @JS99; @DJS2; @DJS3; @DJSd]) is extremely convenient (much more so than an approach based on the harmonic-coordinates equations of motion, as in [@BF1; @BF2; @BF3; @BF4]). Indeed, on the one hand it simplifies very much the reduction to the center-of-mass relative dynamics (which is trivially obtained by setting $\mbox{\boldmath$p$}_{\rm
rel} = \mbox{\boldmath$p$}_1 = - \mbox{\boldmath$p$}_2$), and on the other hand it yields directly (without guesswork) an action principle for the dynamics[^3]. We shall find also below that an Hamiltonian approach is very convenient for dealing with the spin degrees of freedom.
Up to now, we only mentionned the dynamics of the orbital degrees of freedom, i.e. (in the order-reduced Hamiltonian formalism) the (ADM-coordinate) positions and momenta $\mbox{\boldmath$x$}_1 , \mbox{\boldmath$x$}_2 , \mbox{\boldmath$p$}_1 ,
\mbox{\boldmath$p$}_2$ of the two black holes[^4]. After reduction to the center-of-mass frame ($\mbox{\boldmath$P$} = \mbox{\boldmath$p$}_1 +
\mbox{\boldmath$p$}_2 = 0$), and to the relative dynamics ($\mbox{\boldmath$x$} \equiv
\mbox{\boldmath$x$}_1 - \mbox{\boldmath$x$}_2$, $\mbox{\boldmath$p$} \equiv
\mbox{\boldmath$p$}_1 = - \mbox{\boldmath$p$}_2$), one ends up with a canonical pair $\mbox{\boldmath$x$}$, $\mbox{\boldmath$p$}$ of phase-space variables.
The addition of spin degrees of freedom on each black hole is, a priori, a rather complicated matter. If one wished to have a [*relativistically covariant*]{} description of the dynamics of two spinning objects, one would need not only to add, in Einstein’s equations, extra (covariant) source terms proportional to suitable derivatives of delta-functions (spin dipoles), but also to enlarge the two-body action principle to incorporate the spin variables. The first task is doable, and has been performed (to the lowest order) in several works, such as Refs. [@tulczyjew; @infeld-plebanski; @D75; @D82; @martin]. For other works on the relativistic equations of motion of black holes or extended bodies (endowed with spin and higher multipole moments) see [@D'Eath75; @thorne-hartle; @DSX]. Recently, Ref. [@Tagoshi] has tackled the next-to-leading order contribution to spin-orbit effects. \[We consider here only the case of interacting, comparable massive objects. The problem of a spinning [ *test*]{} particle in an external field is simpler and has been dealt with by many authors, such as Mathisson [@mathisson], Papapetrou [@papapetrou], etc$\ldots$\] On the other hand, the second task (incorporating the spin degrees of freedom in an action principle) is quite intricate. First, it has been found that, within a relativistically covariant set up for a spinning particle, the Lagrangian describing the [*orbital*]{} motion could not, even at lowest order in the spin, be taken as an ordinary Lagrangian $L (x,\dot x)$, but needed to be a higher-order one $L (x,\dot x , \ddot x , \ldots)$ [@D82]. Second, a relativistically covariant treatment of the [*spin*]{} degrees of freedom is an intricate matter, involving all the subtleties of constrained dynamical systems, even in the simplest case of a free relativistic top [@regge]. Contrary to the case of the spin-independent EOB where it was easy to use the test-mass results to constrain the EOB Hamiltonian, the Mathisson-Papapetrou dynamics of spinning test masses in external gravitational fields is rather complicated and cannot easily be used to constrain the spin-dependent EOB Hamiltonian. (It might, however, be interesting to try to do so.) Fortunately, there is a technically much lighter approach which bypasses these problems and simplifies [*both*]{} the description of orbital degrees of freedom and that of spin degrees of freedom. This approach is [*not*]{} manifestly relativistically covariant. This lack of [*manifest*]{} Poincaré covariance is not (in principle) a problem at all for two reasons: (i) it does not prevent the expected global Poincaré covariance of the two-body dynamics to be realized as a phase-space symmetry (as was explicitly proven, at 3 PN, for the orbital degrees of motion in Ref. [@DJS3]), and (ii) as we are, at this stage, mainly interested in the description of the relative motion in a specific (center-of-mass) Lorentz frame, there is no physical need to enforce any boost invariance. This non covariant approach to the gravitational interaction of spinning objects stems from the classic work of Breit on the electromagnetic interaction of (quantum) spinning electrons (see, e.g. [@LL]) and has been developed in a sequence of papers [@BGH66; @BOC70; @BOC75; @BOC79]. \[The change of variables needed to pass from the covariant, higher-order Lagrangian description to the non-covariant, ordinary Lagrangian has been discussed in several papers, e.g. [@regge; @BOC74GRG; @DS88].\] In the present paper, we shall combine the non manifestely covariant, ADM-Hamiltonian treatment of the orbital degrees of freedom of [@JS98; @DJS3], with the similarly non covariant, but Hamiltonian, treatment of the spin degrees of freedom of [@BGH66; @BOC75]. Moreover, as is explained below, we shall improve upon [@BOC70; @BOC75; @BOC79] in using a direct Poisson-bracket treatment of the dynamical spin variables.
Finally, our starting point (for the effective one-body approach) is a [ *PN-expanded*]{} Hamiltonian for the [*relative*]{} motion of two spinning objects of the form $$\begin{aligned}
\label{eq2.1}
H_{\rm real}^{\rm PN} (\mbox{\boldmath$x$} , \mbox{\boldmath$p$} ,
\mbox{\boldmath$S$}_1 , \mbox{\boldmath$S$}_2) &= &H_{\rm orb}^{\rm PN}
(\mbox{\boldmath$x$} , \mbox{\boldmath$p$}) + H_S^{\rm PN} (\mbox{\boldmath$x$} ,
\mbox{\boldmath$p$} , \mbox{\boldmath$S$}_1 , \mbox{\boldmath$S$}_2) \nonumber \\
&+ &H_{SS}^{\rm PN} (\mbox{\boldmath$x$} , \mbox{\boldmath$p$} , \mbox{\boldmath$S$}_1
, \mbox{\boldmath$S$}_2) + H_{SSS}^{\rm PN} (\mbox{\boldmath$x$} , \mbox{\boldmath$p$}
, \mbox{\boldmath$S$}_1 , \mbox{\boldmath$S$}_2) + \cdots \end{aligned}$$
Here, $H_{\rm orb}^{\rm PN}$ denotes the PN-expanded [*orbital*]{} Hamiltonian, which is the sum of the free Hamiltonian $H_0 = \sqrt{m_1^2 \, c^4 + \mbox{\boldmath$p$}_1^2
\, c^2} + \sqrt{m_2^2 \, c^4 + \mbox{\boldmath$p$}_2^2 \, c^2}$ and of the monopolar interaction Hamiltonian $H_{\rm int}^m$ generated by the source terms proportional to the masses. Before the reduction to the center-of-mass frame $H_{\rm
int}^m$ has the symbolic structure: $H_{\rm int}^m \sim m_1 \, m_2 + m_1^2 \, m_2 +
m_1 \, m_2^2 + m_1^3 \, m_2 + m_1^2 \, m_2^2 + m_1 \, m_2^3 + \cdots \sim m_1 \, m_2
(1 + m_1 + m_2 + (m_1 + m_2)^2 + (m_1 + m_2)^3 + \cdots)$. It is explicitly known up to the 3 PN level (i.e. up to velocity-independent terms $\propto \, m_1 \, m_2 (m_1 +
m_2)^3$). After reduction to the center-of-mass frame the PN expansion of $H_{\rm
orb}^{\rm PN}$ reads (with $M \equiv m_1 + m_2$, $\mu \equiv m_1 \, m_2 / M$, $\nu
\equiv \mu / M \equiv m_1 \, m_2 / (m_1 + m_2)^2$, $r \equiv \vert \mbox{\boldmath$x$}
\vert$, $\widehat{\mbox{\boldmath$p$}} \equiv \mbox{\boldmath$p$} / \mu$, $\widehat{\mbox{\boldmath$x$}} \equiv \mbox{\boldmath$x$} / GM$) \[eq2.2\] H\_[orb]{}\^[PN]{} ( , ) = M c\^2 + H\_N ( , ) + H\_[1 [PN]{}]{} ( , ) + H\_[2 [PN]{}]{} ( , ) + H\_[3 [PN]{}]{} ( , ) , \[eq2.3\] H\_N ( , ) = - = , \[eq2.4\] H\_[1 [PN]{}]{} ( , ) = , \[eq2.5\] H\_[2 [PN]{}]{} ( , ) = , $$\begin{aligned}
\label{eq2.6}
H_{3 \, {\rm PN}} (\mbox{\boldmath$x$} , \mbox{\boldmath$p$}) &= &\frac{\mu}{c^6}
\Biggl[ \frac{1}{128} \, (-5 + 35 \nu - 70 \nu^2 + 35 \nu^3) \,
\widehat{\mbox{\boldmath$p$}}^8 + \cdots \nonumber \\
&+ &\left[ \frac{1}{8} + \left( \frac{109}{12} - \frac{21}{32} \, \pi^2 +
\omega_s \right) \nu \right] \, \frac{1}{\widehat{r}^4} \Biggl]
\, .\end{aligned}$$
We have exhibited (for illustration) in Eqs. (\[eq2.5\]) and (\[eq2.6\]) only the first and the last terms. We refer to [@DS88] for the full (center-of-mass) 2 PN Hamiltonian (7 terms in all), and to [@JS98; @DJS1; @DJS3] for the full (center-of-mass) 3 PN Hamiltonian (11 terms in all). Our effective one-body treatment will take into account the [*full*]{} 2 PN and 3 PN structures, but in a very streamlined way which will be explicitly displayed below.
The other terms in Eq. (\[eq2.1\]) denote the various spin-dependent contributions to the Hamiltonian: respectively the terms linear $(H_S^{\rm PN})$, quadratic $(H_{SS}^{\rm PN})$, cubic $(H_{SSS}^{\rm PN})$, etc$\ldots$ in the spins $\mbox{\boldmath$S$}_1$, $\mbox{\boldmath$S$}_2$. Before reduction to the center-of-mass frame they have the symbolic structure: $$\begin{aligned}
\label{eq2.7}
H_S^{\rm PN} &\sim &S_1 \, m_2 (1 + m_1 + m_2 + ( m_1 + m_2)^2 + \cdots) \nonumber \\
&+ &S_2 \, m_1 (1 + m_1 + m_2 + ( m_1 + m_2)^2 + \cdots) \, ,\end{aligned}$$ $$\begin{aligned}
\label{eq2.8}
H_{SS}^{\rm PN} &\sim &S_1^2 \, m_2 \left( \frac{1}{m_1} + 1 + m_1 + m_2 + \cdots
\right) + S_1 \, S_2 (1 + m_1 + m_2 + \cdots) \nonumber \\
&+ &S_2^2 \, m_1 \left(\frac{1}{m_2} + 1 + m_1 + m_2 + \cdots \right) \, ,\end{aligned}$$ etc$\ldots$ We shall explain below the occurence of the terms quadratic in the spins and inversely proportional to a mass. In contradistinction with the case of the orbital Hamiltonian which has been worked out with a high PN accuracy, only the simplest spin-dependent terms have been explicitly derived, namely the lowest PN-order term in $H_S^{\rm PN}$, whose center-of-mass reduction reads [@BOC75] \[eq2.9\] H\_S\^[PN]{} (, , \_1 , \_2) = ( ) + [O]{} ( ) , and the lowest PN-order one-graviton exchange contribution to the bilinear term $(\propto \, S_1 \, S_2)$ in $H_{SS}^{\rm PN}$. Ref. [@Tagoshi] contains some information about the ${\cal O} ( \frac{1}{c^4})$ corrections in Eq. (\[eq2.9\]), but, because of the use of different gauge and spin conditions, not in a form which can be directly used to derive these corrections. We shall discuss the spin-bilinear contribution $(\propto \, S_1 \, S_2)$ below, together with the leading spin-quadratic contributions $\propto \, S_1^2 \, m_2 / m_1 + S_2^2
\, m_1 / m_2$.
Before going further, let us make clear that, before and after any type of resummation, the dynamics entailed by the Hamiltonians we shall consider $H (\mbox{\boldmath$x$},
\mbox{\boldmath$p$}, \mbox{\boldmath$S$}_1 , \mbox{\boldmath$S$}_2)$ follow, for all degrees of freedom, from the basic Poisson brackets \[eq2.10\] { x\^i , p\_j } = \_j\^i , \[eq2.11\] { S\_1\^i , S\_1\^j } = \^[ijk]{} S\_1\^k , \[eq2.12\] { S\_2\^i , S\_2\^j } = \^[ijk]{} S\_2\^k , \[eq2.13\] 0 = { x\^i , x\^j } = { p\_i , p\_j } = { S\_1\^i , S\_2\^j } = { x\^i , S\_1\^j } = { x\^i , S\_2\^j } = { p\_i , S\_1\^j } = {p\_i , S\_2\^j } . The (real) time evolution of any dynamical quantity $f (\mbox{\boldmath$x$},
\mbox{\boldmath$p$}, \mbox{\boldmath$S$}_1 , \mbox{\boldmath$S$}_2)$ is given by \[eq2.14\] f (, , \_1 , \_2) = { f , H\_[real]{} } , where the Poisson bracket (PB) $\{ f , H_{\rm real} \}$ is computed from the basic PB’s (\[eq2.10\])–(\[eq2.13\]) by using the standard PB properties (skew symmetry: $\{ f,g \} = - \{ g,f \}$, Leibniz rule: $\{ f,gh \} = \{ f,g \} \, h + g \,
\{ f,h \}$, and the Jacobi identity: $\{ f , \{ g,h \}\} + \{ g , \{ h,f \}\} + \{ h ,
\{ f,g \}\} = 0$). The simplest way to prove the statements (\[eq2.10\])–(\[eq2.14\]) is to consider our dynamics as the classical limit of the quantum dynamics of a system of gravitationally interacting spinning particles. Surprisingly, though Refs. [@BGH66; @BOC70] derived (à la Breit) the spin-dependent contributions to the Hamiltonian by a quantum route, they never noticed that they could very simply derive the spin evolution equations by using the PB’s (\[eq2.11\]), (\[eq2.12\]). They had to go back to a Lagrangian formalism and add some explicit spin kinetic energy terms $\left( \frac{1}{2} \, I_1 \,
\mbox{\boldmath$\omega$}_1^2 + \frac{1}{2} \, I_2 \, \mbox{\boldmath$\omega$}_2^2
\right)$ to derive the spin evolution equations. Note also that we have kept the label “real” on the Hamiltonian in Eq. (\[eq2.14\]) to distinguish the evolution with respect to the real time (associated with the original two-body system) from the evolution generated by the effective Hamiltonian to be introduced below (which is associated with an auxiliary, effective time).
Before generalizing it, by including the spin degrees of freedom, let us recall the results of [@BD99] (2 PN level) and [@DJS2] (3 PN level) concerning the effective one-body “upgrading” of the PN-expanded orbital Hamiltonian $H_{\rm
orb}^{\rm PN} (\mbox{\boldmath$x$} , \mbox{\boldmath$p$})$. Again, the simplest way to motivate it is to think of our dynamics as the classical limit of a quantum dynamics defined by some hermitian Hamiltonian operator $H_{\rm orb} (\mbox{\boldmath$x$} , - i
\hbar \mbox{\boldmath$\nabla$})$. We are mainly interested in the bound states of $H_{\rm orb}$. It is crucial to note that the orbital Hamiltonian (\[eq2.2\])–(\[eq2.6\]) is symmetric under an $O(3)$ group (corresponding to arbitrary rotations of the relative position $\mbox{\boldmath$x$} =
\mbox{\boldmath$x$}_1 - \mbox{\boldmath$x$}_2$, in the center-of-mass frame). Therefore the quantum (and classical) bound states will be labelled (besides parity) by only two quantum numbers: (i) the total orbital angular momentum $\mbox{\boldmath$L$}^2 = L (L + \hbar)$, and (ii) some “principal quantum number” $N$ (such that $(N-L) / \hbar$ counts the number of nodes in the radial relative wave function). Both $L$ and $N$ are quantized in units of $\hbar$. The full list of two-body bound states is thereby encoded in the formula giving the bound state energy as a function of the two quantum numbers $L$ and $N : E_{\rm real} = E_{\rm real}
(L,N) = M c^2 - \frac{1}{2} \, \mu (G \,m_1 \, m_2)^2 / N^2 + E_{1 \, {\rm PN}} (L,N)
+ E_{2 \, {\rm PN}} (L,N) + E_{3 \, {\rm PN}} (L,N)$. The basic idea of the effective one-body method is to map (in a one-to-one manner) the discrete set of real two-body bound states $E_{\rm real} (L,N)$ onto the discrete set of bound states of an auxiliary (“effective”) one-body Hamiltonian $H_{\rm eff} (\mbox{\boldmath$x$}_{\rm
eff} , \mbox{\boldmath$p$}_{\rm eff}$). Because of the special labelling by (only) two integer quantum numbers $L/\hbar$, $N/\hbar$ one is naturally led to imposing that: (i) the EOB Hamiltonian be spherically symmetric[^5], and (ii) the integer valued quantum numbers be identified in the two problems, i.e. $L/\hbar = L_{\rm eff}
/ \hbar$ and $N/\hbar = N_{\rm eff} / \hbar$. On the other hand, one can (and one a priori should) leave free a (one-to-one) continuous function $f$ mapping the real energies onto the effective ones: $E_{\rm eff} (L,N) = f (E_{\rm real} (L,N))$. Evidently, for this method to buy us anything we wish the effective dynamics to be significantly simpler than the original $H_{\rm real} (\mbox{\boldmath$x$} ,
\mbox{\boldmath$p$})$, and, in particular, to reduce, in some approximation, to the paradigm of the simplest gravitational one-body problem, namely the dynamics of an (effective) test particle in some (to be determined) effective metric $g_{\mu
\nu}^{\rm eff} (x_{\rm eff}^{\lambda})$. Remarkably enough, it was found in [@BD99] that such a mapping between the very complicated real two-body orbital 2 PN Hamiltonian (\[eq2.2\])–(\[eq2.5\]) and the usual (“geodesic”) dynamics of a test particle of mass $\mu = m_1 \, m_2 / (m_1 + m_2)$ in a very simple (spherically symmetric) effective metric \[eq2.15\] ds\_[eff]{}\^2 = - A (r\_[eff]{}) c\^2 dt\_[eff]{}\^2 + d r\_[eff]{}\^2 + r\_[eff]{}\^2 (d \^2 + \^2 d \^2) , is possible, if and only if the energy mapping $E_{\rm eff} = f (E_{\rm real})$ is given by \[eq2.16\] = .
Remarkably, the simple energy map (\[eq2.16\]) (which is here determined by our requirements) coincides with the energy map introduced in several other investigations [@BIZ], [@DIS98] (and is simply related to the one defined a priori in [@T1; @T2; @FT01]).
Recently, the problem of mapping the extremely complicated real two-body 3 PN Hamiltonian (\[eq2.6\]) onto an effective one-body dynamics has been solved [@DJS2]. Again the result is remarkably simple, though less simple than at the 2 PN level. Indeed, it was found[^6] that the effective one-body dynamics was given by an Hamilton-Jacobi equation of the form (with $p_{\alpha}^{\rm eff} = \partial S /
\partial \, x_{\rm eff}^{\alpha}$) \[eq2.17\] 0 = \^2 + g\_[eff]{}\^ p\_\^[eff]{} p\_\^[eff]{} + Q\_4 (p\^[eff]{}) , with a simple (spherically symmetric) effective metric of the form (\[eq2.15\]) and some additional quartic-in-momenta contribution $Q_4 (p)$. Remarkably, it was found that, at the 3 PN level, the energy mapping is again uniquely determined to be the simple relation (\[eq2.16\]). As for the metric coefficients of the (covariant) effective metric $g_{\alpha \beta}^{\rm eff}$, and the quartic terms $Q_4 (p)$, they were found to be \[eq2.18\] A (r) = 1 - 2 u + 2 [u]{}\^3 + a\_4 () [u]{}\^4 , \[eq2.19\] D (r) = 1 - 6 [u]{}\^2 + 2 (3 - 26) [u]{}\^3 , \[eq2.20\] = 2 (4 - 3 ) [u]{}\^2 ( )\^4 , where $\widehat u \equiv GM / r$, $\widehat{\mbox{\boldmath$p$}} =
\mbox{\boldmath$p$} / \mu$ and \[eq2.21\] a\_4 () = ( - \^2 + 2 \_s ) . As we said above, the correct value of $\omega_s$ has been recently found [@DJSd] to be simply $\omega_s = 0$. We shall find, however, convenient to keep $\omega_s$ as a free parameter in order to assess the quantitative importance of 3PN effects.
Let us emphasize again the streamlined nature of the effective one-body description of the orbital dynamics. Successively, as the PN order increases, one can say that: (i) the 6 terms of the Newtonian plus first post-Newtonian relative Hamiltonian (\[eq2.3\]), (\[eq2.4\]) can be mapped (via (\[eq2.16\]) and a canonical transformation of $(\mbox{\boldmath$x$} , \mbox{\boldmath$p$})$) onto geodesic motion in a Schwarzschild spacetime of mass $M$ (i.e. $A_{1 \, {\rm PN}} = 1 - 2 \, {\widehat
u}$, $D_{1 \, {\rm PN}} = 1$), (ii) to take into account the 7 additional terms entering $H_{2 \, {\rm PN}}$, Eq. (\[eq2.5\]), it is enough to add $+ \, 2 \, \nu \,
{\widehat u}^3$ to $A (r)$ and $- \, 6 \, \nu \, {\widehat u}^2$ to $D (r)$, and (iii) to take into account the 11 additional terms entering $H_{3 \, {\rm PN}}$, it is enough to further add $+ \, a_4 (\nu) \, {\widehat u}^4$ to $A (r)$ and $+ \, 2 (3 \nu - 26) \,
\nu
\, {\widehat u}^3$ to $D(r)$, and to add the simple quartic term (\[eq2.20\]) to the mass-shell condition (\[eq2.17\]). Note that the effective one-body dynamics is a “deformation” of a geodesic dynamics for a particle of mass $\mu$ in a Schwarzschild spacetime of mass $M$, with deformation parameter the symmetric mass ratio $\nu = \mu
/ M = m_1 \, m_2 / (m_1 + m_2)^2$ ($0 < \nu \leq \frac{1}{4}$, with $\nu$ being small if $m_1 \ll m_2$ or $m_2 \ll m_1$, and reaching its maximal value of $\frac{1}{4}$ when $m_1 = m_2$). Note also that, at this stage, we have not yet introduced any particular resummation technique. The effective quantities (\[eq2.18\])–(\[eq2.20\]) are still given as straightforward $PN$ expansions in powers of $\widehat u = GM / c^2 \, r$. However, this already means an appreciable gain over the original $PN$ expansions, Eqs. (\[eq2.3\])–(\[eq2.6\]). Indeed, there are far less terms in the “effective” PN expansions and they generically have significantly smaller coefficients (which are now all multiplied by $\nu < 1/4$). For instance, the “radial” potential determining the circular orbits is now fully encoded in the simple function $g_{00}^{\rm eff} = - A (r_{\rm eff})$ which differs from the test-mass (Schwarzschild) result, $A_{(\nu \, = \, 0)} = 1 - 2
\, \widehat u$, only by two numerically smallish terms when one is above the last stable orbit. Indeed, when $r_{\rm eff} > 6 \, GM / c^2$, $2 \, \nu \, {\widehat
u}^3 < 0.23\%$ and $a_4 (\nu) \, {\widehat u}^4 < 0.36\%$ (for $\omega_s = 0$). When working at the 2 PN level it was, in fact, found unnecessary [@BD99] to further “resum” the effective PN expansions of $A(r)$ and $D(r)$. As they stand, they led to small deviations from the test-mass ($\nu \rightarrow 0$) results. However, it was found in [@DJS2] that the situation is not quite as rosy at the 3 PN level. Because of the largish coefficient in Eq. (\[eq2.21\]), $\frac{94}{3} - \frac{41}{32} \, \pi^2 \simeq 18.688$, the additional term $+ \, a_4 (\nu)
\, \widehat{u}^4$ in Eq. (\[eq2.18\]) significantly modifies the qualitative behaviour of the metric coefficient $A(r)$ for $r_{\rm eff} < 6~GM$. In particular, when $\omega_s = 0$, the straightforward PN-expanded function $A(r)$ no longer features a zero near $r_{\rm eff} = 2~GM$ for all possible values of the deformation parameter $0 < \nu \leq \frac{1}{4}$. As this zero (which corresponds to the Schwarzschild horizon) is a crucial qualitative feature of the $\nu = 0$ limit, it was argued in [@DJS2] that one should Padé-resum the PN-expansion of $A(r)$ so as to ensure the stable presence of a similar “horizon” when $0 < \nu \leq \frac{1}{4}$. We shall also do so here, but only after having introduced the spin effects, which modify the radial function which is the analog of $A(r)$.
Effective spin {#ssec2.2}
--------------
Let us now tackle the central task of this work: to introduce spin effects in the effective one-body approach. Let us first emphasize that the ambition of the present work is somewhat limited. Our main goal is to derive a spin-dependent EOB Hamiltonian which is physically reliable for small and moderate spins. We shall proceed toward this goal in successive steps. First, we consider only (at the lowest PN order where they enter) the interactions which are [ *linear*]{} in the spins. Then, we shall incorporate (to some approximation) the interactions which are quadratic in the spins. Contrary to the case of the spin-independent interactions where many years of work have yielded high-PN-accuracy information which allowed one to refine and test the EOB approach, one does not have in hand enough information for gauging the reliability of the spin-dependent interactions in the case of fast spins. As a consequence, we will be able to trust the presently introduced EOB Hamiltonian only when spin effects are not too large. As we said in the Introduction, one will need new information (either from numerical relativity, or from improved analytical methods) to find a reliable form of the EOB Hamiltonian for large spins.
We have written down in Eq. (\[eq2.9\]) above the contribution to the real, PN-expanded, two-body Hamiltonian which is linear in the two spins. Our first proposal is to map this contribution to the spin-orbit coupling of a (spinless) effective particle moving in a suitably “spinning” effective metric, i.e. some type of generalized Kerr metric. If we formally consider $\mbox{\boldmath$S$}_1$ and $\mbox{\boldmath$S$}_2$ as deformation parameters (on top of the basic [*orbital*]{} deformation parameter $\nu$), the effective dynamics we are looking for should be a “spin deformation” of the currently most accurate orbital dynamics, as described by the 3 PN effective dynamics Eqs. (\[eq2.15\])–(\[eq2.21\]). In particular, we should keep the non-geodesic Hamilton-Jacobi equation (\[eq2.17\]). Note that the spin effects (notably Eq. (\[eq2.9\])) break the $O(3)$ symmetry of the orbital interaction. At the quantum level, this means that spin interactions lift the degeneracy ($=$ lack of dependence on the “magnetic” quantum number $L_z / \hbar$) of the orbital energy states. This shows that the effective (co)metric $g_{\rm eff}^{\alpha
\beta}$ entering Eq. (\[eq2.17\]) should no longer be spherically symmetric, but should contain special directions linked to the spins. For generality, let us first consider an arbitrary (time-independent) effective cometric \[eq2.22\] g\_[eff]{}\^ p\_ p\_ = g\_[eff]{}\^[00]{} p\_0\^2 + 2 g\_[eff]{}\^[0i]{} p\_0 p\_i + g\_[eff]{}\^[ij]{} p\_i p\_j . Let us define \[eq2.23\] (-g\_[eff]{}\^[00]{})\^[-1/2]{} , \^i , \^[ij]{} g\_[eff]{}\^[ij]{} - , i.e. \[eq2.24\] g\_[eff]{}\^[00]{} = - , g\_[eff]{}\^[0i]{} = - , g\_[eff]{}\^[ij]{} = \^[ij]{} - .
The effective energy $E_{\rm eff} \equiv -p_0^{\rm eff}$ is conserved (because of the assumed stationarity of $g_{\rm eff}^{\alpha\beta}$). Using the parametrization (\[eq2.23\]), Eq. (\[eq2.17\]) reads \[eq2.25\] (E\_[eff]{} - \^i p\_i)\^2 = \^2 \[\^2 + \^[ij]{} p\_i p\_j + Q\_4 (p)\] .
Solving Eq. (\[eq2.25\]) for $E_{\rm eff}$ (using the fact that $Q_4 (p)$, Eq. (\[eq2.20\]), depends only on the $p_i$’s) yields the effective Hamiltonian \[eq2.26\] E\_[eff]{} = H\_[eff]{} ( , , …) = \^i p\_i + , where we have suppressed, for readability, the labels “eff” on the orbital phase space variables $\mbox{\boldmath$x$}^{\rm eff}$, $\mbox{\boldmath$p$}^{\rm eff}$. The ellipsis in the arguments of $H_{\rm eff}$ are added to remind us that $H_{\rm eff}$ will ultimately also depend on the spin variables $\mbox{\boldmath$S$}_1$, $\mbox{\boldmath$S$}_2$, which enter the metric coefficients $\alpha$, $\beta^i$, $\gamma^{ij}$ as parameters.
We also assume that Eq. (\[eq2.16\]) (which was found to hold at 1 PN, 2 PN and 3 PN) still holds. Solving it for the real energy $E_{\rm real}$ in terms of the effective one finally yields the real Hamiltonian \[eq2.27\] E\_[real]{} = H\_[real]{} ( , , …) = M .
We recall that, at the linearized level and at the lowest PN order, the addition of a spin $\mbox{\boldmath$S$}_{\rm eff}$ onto an initially spherical symmetric metric leads to an off-diagonal term in the metric: \[eq2.28\] \^i - g\^[0i]{} - g\_[0i]{} + \^[ijk]{} S\_[ eff]{}\^j x\^k .
Inserting this term in (\[eq2.26\]), and expanding (\[eq2.27\]) in a PN series yields, as leading spin-orbit coupling (linearized in $S_{\rm eff}$ and taken to formal order ${\cal O} (1/c^2)$) in $H_{\rm real}$, the term \[eq2.29\] \_[S\_[eff]{}]{} H\_[real]{} \^i p\_i \^[ijk]{} p\_i S\_[eff]{}\^j x\^k . This term can exactly reproduce the leading[^7] two-body spin-orbit coupling (\[eq2.9\]) if we define \[eq2.30\] \^[eff]{} \_1 \_1 + \_2 \_2 , with \[eq2.31\] \_1 1 + , \_2 1 + .
A deformed Kerr metric {#ssec2.3}
----------------------
Remembering that the main message of the effective one-body method is that the orbital dynamics of two comparable-mass black holes can be described in terms of a slightly deformed (with deformation parameter $\nu$) Schwarzschild metric, we expect that the orbital-plus-spin dynamics of two black holes can be described in terms of some deformation of the Kerr metric. In other words, we are expecting that not only the effects linear in the spins, such as Eq. (\[eq2.9\]), but also the spin-dependent non-linear effects, can be described in terms of some deformed, effective Kerr metric. At this stage it is therefore very natural to construct a suitable “deformed Kerr metric” which combines the orbital deformations (\[eq2.18\]), (\[eq2.19\]) with the full spin effects linked to the “effective spin” (\[eq2.30\]). After constructing this deformed Kerr metric, we shall a posteriori check that it approximately incorporates the expected two-body interactions which are [*quadratic*]{} in the spins.
Let us start from the simplest form of the Kerr cometric, underlying its separability properties [@carter] $$\begin{aligned}
\label{eq2.32}
g_{\rm Kerr}^{\alpha\beta} \, p_{\alpha} \, p_{\beta} && \, = \frac{1}{r^2 + a^2 \cos^2
\theta} \nonumber \\
\times&&\left[ \Delta_K (r) \, p_r^2 + p_{\theta}^2 + \frac{1}{\sin^2 \theta} \,
(p_{\varphi}
+ a \sin^2 \theta \, p_t)^2 - \frac{1}{\Delta_K (r)} \, ((r^2 + a^2) \, p_t + a \,
p_{\varphi})^2 \right] \, ,\end{aligned}$$ with $\Delta_K (r) = r^2 - 2Mr + a^2$. In the non-spinning limit $(a \rightarrow 0)$ the coefficients of $p_r^2$ and $p_t^2$ become, respectively, $\Delta_K (r) / r^2$ and $-
r^2 / \Delta_K (r)$. However, we know that in this limit we should get (from (\[eq2.15\])) $A(r) / D(r)$ and $- 1 / A(r)$, respectively. It is therefore very natural to generalize the Kerr metric (\[eq2.32\]) (while still keeping its separability properties) by assuming that the coefficients of the first and last terms in the square brackets of (\[eq2.32\]) involve two different functions of $r$, say $\Delta_r (r)$ and $-1 / \Delta_t (r)$, whose product reduces to $-1/D(r)$ when $a
\rightarrow 0$. This reasoning leads us, as simplest[^8] possibility for combining spin effects with orbital effects, to postulating that the effective metric entering (\[eq2.17\]) has the form $$\begin{aligned}
\label{eq2.33}
g_{\rm eff}^{\alpha\beta} \, p_{\alpha} \, p_{\beta} && \, = \frac{1}{r^2 + a^2 \cos^2
\theta} \nonumber \\
\times&&\left[ \Delta_r (r) \, p_r^2 + p_{\theta}^2 + \frac{1}{\sin^2 \theta} \,
(p_{\varphi}
+ a \sin^2 \theta \, p_t)^2 - \frac{1}{\Delta_t (r)} \, ((r^2 + a^2) \, p_t + a \,
p_{\varphi})^2 \right] \, ,\end{aligned}$$ with \[eq2.34\] \_r\^[PN]{} (r) = , \_t\^[PN]{} (r) = r\^2 A\^[PN]{} (r) + a\^2 . Here, the superscripts “PN” indicate that, at this stage, we are only comparing PN expansions. We already know from the 3 PN study of [@DJS2] that this is unsatisfactory because it tends to change the qualitative behaviour of the radial functions, and, in particular, the presence of a horizon in the metric (\[eq2.33\]). To get a regular horizon in Eq. (\[eq2.33\]) we need the two functions $\Delta_t (r)$ and $\Delta_r(r)$ to have a zero at the same value of $r$. The simplest (and most robust) way of ensuring this is (as discussed in [@DJS2]) to define them as \[eq2.35\] \_t (r) r\^2 P\_3\^1 , \_r (r) \_t (r) ( )\^[PN]{} . Here, $P_m^n [f^{\rm PN} (u)]$, with $u \equiv 1/r$, denotes the $N_n (u) / D_m (u)$ Padé of a certain PN-expanded function $f^{\rm PN} (u) = c_0 + c_1 \, u + c_2 \, u^2
+ \cdots + c_{n+m} \, u^{n+m}$ ($N_n (u)$ and $D_m (u)$ being polynomials in $u$ of degrees $n$ and $m$, respectively). We do not write down the (uniquely defined) explicit expression of $$\begin{aligned}
\bar A (u) \equiv P_3^1 \, [A^{\rm PN} (u) + a^2 \, u^2] &= &P_3^1 \, [1 - 2 \widehat{u} +
\widehat{a}^2 \, \widehat{u}^2 + 2 \nu \, \widehat{u}^3 + a_4 (\nu) \, \widehat{u}^4]
\nonumber \\
&= &\frac{1 + n_1 \, \widehat{u}}{1 + d_1 \, \widehat{u} + d_2 \, \widehat{u}^2 + d_3 \,
\widehat{u}^3} \nonumber\end{aligned}$$ (where $\widehat{u} = GM / r$, $\widehat{a} = a / GM$) because: (i) it is rather complicated and not illuminating, and (ii) modern algebraic manipulators compute it directly from its Padé definition.
In the definition of $\Delta_r (r)$ (which is less important than that of $\Delta_t (r)$) we have factorized the Padéed $\Delta_t (r)$ and assumed that it was enough to work with the non-resummed PN-expansion of the inverse of the $D$-function, i.e. (from (\[eq2.19\])) \[eq2.36\] (D\^[-1]{} (r))\^[PN]{} 1 + 6 \^2 + 2 (26 - 3 ) \^3 . If the need arises, it would be easy to define improved (resummed) versions of $D^{-1}
(r)$. Because of the positive coefficients in (\[eq2.36\]) the present definition does not interfere (as would the consideration of $(D^{\rm PN} (r))^{-1}$) with the desired feature of having a simple zero in $\Delta_r (r)$ located at the same value as that in $\Delta_t (r)$.
Finally, we shall see later that there are some advantages in defining the quartic-in-momenta contribution $Q_4 (p)$ in the following (deformed) way \[eq2.36new\] Q\_4 (p) = 2 (4 - 3 ) ( )\^4 .
Eq. (\[eq2.33\]) defines $g_{\rm eff}^{\alpha\beta}$ only with respect to some (instantaneous) polar coordinate system with $z$-axis aligned with the effective spin (\[eq2.30\]). Such a coordinate system cannot be used for describing the evolution of two gravitationally interacting spinning black holes. Indeed, we expect (and shall check below) that the total real Hamiltonian imposes some type of precession motion for $\mbox{\boldmath$S$}_1$, $\mbox{\boldmath$S$}_2$ and therefore $\mbox{\boldmath$S$}_{\rm eff}$. To get the full dynamics of the system we need to rewrite $g_{\rm eff}^{\alpha\beta}$ in a general, cartesian-like coordinate system. This is achieved by explicitly introducing, besides $n^i \equiv x^i / r$, the quantities $(S_{\rm eff} \equiv (\delta_{ab} \, S_{\rm eff}^a \, S_{\rm eff}^b)^{1/2})$ \[eq2.37\] s\^i , a , n\^i s\^i , \^2 r\^2 + a\^2 \^2 . This leads to the following, cartesian-like, effective metric \[eq2.38\] - \^2 g\_[eff]{}\^[00]{} = , \[eq2.39\] \^2 g\_[eff]{}\^[0i]{} = - ( )\^i , \[eq2.40\] \^2 g\_[eff]{}\^[ij]{} = \_r n\^i n\^j + r\^2 (\^[ij]{} - n\^i n\^j) - ( )\^i ( )\^j , from which follows \[eq2.41\] = , \[eq2.42\] \^i = , \[eq2.43\] \^[ij]{} = g\_[eff]{}\^[ij]{} + . Note that near the “horizon”, i.e. as $\Delta_t \rightarrow 0$, the quantity $\alpha$ tends to zero like $\sqrt{\Delta_t}$, while $\beta^i$ and $\gamma^{ij}$ have finite limits. \[The singular last term on the right-hand side of Eq. (\[eq2.40\]) is cancelled near the horizon by the contribution $+ \beta^i \, \beta^j / \alpha^2$ to $\gamma^{ij}$, Eq. (\[eq2.43\]).\]
Finally, the spin-dependent, real two-body Hamiltonian $H_{\rm real}
(\mbox{\boldmath$x$} , \mbox{\boldmath$p$} , \mbox{\boldmath$S$}_1 ,
\mbox{\boldmath$S$}_2)$ is defined by ($\widehat{p}_i \equiv p_i / \mu$, $\widehat
u_{\rho}
\equiv GM / \sqrt{\rho^2}$, $n^i \equiv x^i / r$) $$\begin{aligned}
\label{eq2.44}
H_{\rm real} (\mbox{\boldmath$x$} , \mbox{\boldmath$p$} , \mbox{\boldmath$S$}_1 ,
\mbox{\boldmath$S$}_2) \nonumber \\
\equiv &M \sqrt{1 + 2\nu \left( \beta^i \, \widehat{p}_i + \alpha \sqrt{1 + \gamma^{ij}
\, \widehat{p}_i \, \widehat{p}_j + 2 \, (4-3\nu) \, \nu \, \widehat{u}_{\rho}^2 (n^i \,
\widehat{p}_i)^4} - 1 \right)} \, ,\end{aligned}$$ where we recall that the basic effective Kerr spin vector is defined by \[eq2.45\] M a s\^i S\_[eff]{}\^i \_1 S\_1\^i + \_2 S\_2\^i , with $\sigma_1$ and $\sigma_2$ defined in Eq. (\[eq2.31\]). The phase-space coordinates appearing in this Hamiltonian are the effective ones $(x_{\rm eff}^i ,
p_i^{\rm eff})$. They differ [@BD99; @DJS2] by ${\cal O} (1/c^2)$ terms from the coordinates used in usual PN calculations, such as ADM ones. The evolution equations defined by the Hamiltonian (\[eq2.44\]) are obtained by the Poisson bracket equations (\[eq2.10\])–(\[eq2.13\]). Before discussing them let us show how the Hamiltonian (\[eq2.44\]) contains spin-quadratic effects of the good sign and magnitude.
Effects quadratic in the spins {#ssec2.4}
------------------------------
Note first that if we introduce the “non relativistic” effective Hamiltonian $H_{\rm
eff}^{\rm NR} \equiv H_{\rm eff} - \mu \, c^2$, and similarly $H_{\rm real}^{\rm NR}
\equiv H_{\rm real} - M \, c^2$, one has \[eq2.46\] H\_[eff]{}\^[NR]{} = H\_[real]{}\^[NR]{} ( 1 + ) . Therefore, if we are interested in the leading PN approximation to any additional term in $H_{\rm real}^{\rm PN}$, one can neglect the (${\cal O} (1/c^2)$ smaller) difference between $H_{\rm eff}$ and $H_{\rm real}$. By this argument, the leading PN approximation to the term linear in the spins is \[eq2.47\] H\_[[real]{} S]{} H\_[[eff]{} S]{} \^i p\_i = ( )\^i p\_i . We write it explicitly in the form in which it appears in our Hamiltonian for the reader to see how the term (\[eq2.29\]) is generated. \[The important feature here being that $r^2 + a^2 - \Delta_t \simeq 2 \, GM \, r$ at the leading PN approximation.\]
Let us now consider the interaction terms in $H_{\rm real}$ or $H_{\rm eff}$ which are quadratic in the spins, and therefore quadratic in the Kerr-like parameter $a$, Eq. (\[eq2.45\]). First, one should remember that most terms of order $a^2$, as they appear in the effective metric (\[eq2.38\])–(\[eq2.40\]), do not directly correspond to physical effects proportional to $\mbox{\boldmath$S$}_{\rm eff}^2$. Indeed, we are using here Boyer-Lindquist-type coordinates which differ, even in the flat space limit $GM \rightarrow 0$, from usual (flat-space, cartesian-like) coordinates by terms of order ${\cal O} (a^2)$. As we are interested in the leading PN effects quadratic in the spins, we can view the Kerr-like metric (\[eq2.38\])–(\[eq2.40\]) as a deformation, by the $a$-dependent terms, of the Schwarzschild metric (which is the leading PN version of the orbital effective metric). We then expect that the leading physical effects quadratic $a$ will be those linked to the $a$-dependent quadrupole moment deformation of the Schwarzschild metric. The quadrupole moment of the Kerr metric (which coincides with our metric when we neglect additional 2 PN fractional corrections) has been determined to be [@thorne] \[eq2.48\] Q\_[ij]{} = -M a\^2 s\_i s\_j + M a\^2 \_[ij]{} . This corresponds to an additional term in the interaction Hamiltonian equal to \[eq2.49\] H\_Q\^[real]{} H\_Q\^[eff]{} - Q\_[ij]{} \_[ij]{} = + Ma\^2 s\^i s\^j \_[ij]{} . In terms of the effective spin this reads (in standard units) \[eq2.50\] H\_Q\^[real]{} + S\_[eff]{}\^i S\_[eff]{}\^j \_[ij]{} . Such is the prediction from our Hamiltonian. Let us now compare it to the expected real two-body, spin-quadratic effects. As sketched in Eq. (\[eq2.8\]) there are several sources of spin-quadratic effects. At leading PN order, it is enough to consider: (i) the term $\propto \, m_2 \, S_1^2 / m_1$ which arises because of the interaction of the monopole $m_2$ with the spin-induced quadrupole moment of the spinning black hole of mass $m_1$, (ii) the term $\propto \, m_1 \, S_2^2 / m_2$ obtained by exchanging $1
\rightarrow 2$, and (iii) the term $\propto \, S_1 \, S_2$ coming from the direct, one-graviton interaction between the two spin-dipoles. The first term is obtained by relabelling the result (\[eq2.50\]) by $\mu \rightarrow m_2$, $M \rightarrow m_1$, $S_{\rm eff}^i \rightarrow S_1^i$. Therefore the sum of (i) and (ii) reads \[eq2.51\] H\_[S\_1 S\_1]{} + H\_[S\_2 S\_2]{} + ( S\_1\^i S\_1\^j + S\_2\^i S\_2\^j ) \_[ij]{} . The term (iii) has been computed in [@BGH66; @BOC70] and reads \[eq2.52\] H\_[S\_1 S\_2]{} + S\_1\^i S\_2\^j (3 n\^i n\^j - \^[ij]{}) = + S\_1\^i S\_2\^j \_[ij]{} . It is easily checked that the sum of (\[eq2.51\]) and (\[eq2.52\]), say $H_{SS}
\equiv H_{S_1 S_1} + H_{S_2 S_2} + H_{S_1 S_2}$, can be written as \[eq2.53\] H\_[SS]{} S\_0\^i S\_0\^j \_[ij]{} , with $S_0^i/M \equiv a_0^i \equiv a_1^i + a_2^i \equiv S_1^i/m_1 + S_2^i/m_2$, i.e. explicitly, \[eq2.54\] S\_0\^i (1 + ) S\_1\^i + (1 + ) S\_2\^i . The result (\[eq2.53\]), (\[eq2.54\]) is remarkably similar to the prediction (\[eq2.50\]) (with (\[eq2.30\]), (\[eq2.31\])). The only discrepancy is a 25% difference in the coefficient of the mass ratios in the definition of the effective spin. Though there might be physical situations where this smallish difference might play a significant role, we think that in most cases where one will be entitled to trust the approximate spin-dependent EOB Hamiltonian introduced here this difference will not matter. Indeed, because of the partially ad hoc way in which we constructed our deformed Kerr metric, we cannot trust our predictions beyond the domain where spin effects are moderate corrections to orbital effects. However, it is useful to incorporate in a qualitatively correct manner the non-linear spin effects. This is what our prescription achieves. For instance: (i) in the limit where, say, $m_2 \ll m_1$ (and $\vert \mbox{\boldmath$S$}_2 \vert \leq
m_2^2$) (\[eq2.50\]) and (\[eq2.53\]) become equivalent, or (ii) in the case where $\mbox{\boldmath$S$}_1$ and $\mbox{\boldmath$S$}_2$ are parallel (in the same direction), (\[eq2.50\]) and (\[eq2.53\]) differ only by a numerical factor which is near one for all mass ratios.
It might, however, be useful to define another Hamiltonian, say $H'_{\rm real}$, which: (a) reduces (like $H_{\rm real}$) to the Kerr one in the test-mass (and test-spin) limit, (b) contains (like $H_{\rm real}$) the spin-orbit terms (\[eq2.9\]), and (c) contains the exact spin-spin terms (\[eq2.53\]) (instead of their “25%” approximation (\[eq2.50\]) contained in $H_{\rm real}$). A simple way to do that is to define $H'_{\rm real} \equiv M \, \sqrt{1 + 2 \, \nu (H'_{\rm eff} - \mu) /
\mu}$ with a modified effective Hamiltonian defined as the sum of (\[eq2.26\]), written with the replacement $\mbox{\boldmath$S$}_{\rm eff} \rightarrow \mbox{\boldmath$S$}_0$, and of an additional spin-orbit interaction term, $\Delta \beta^i \, p^i$, with $\Delta
\beta^i$ proportional to the difference $\sigma^i \equiv S_{\rm eff}^i - S_0^i$: \[eq2.56\] H’\_[eff]{} ( , , \_1 , \_2) = H\_[eff]{} ( , , \_0) + H\_[SO]{} ( , , \_0 , ) , where (denoting $\mbox{\boldmath$a$}_0 \equiv \mbox{\boldmath$S$}_0 / M$, $\cos \theta_0
\equiv n^i \, S_0^i / \vert \mbox{\boldmath$S$}_0 \vert$) \[eq2.57\] H\_[SO]{} ( , , \_0 , ) , with \[eq2.58\] \^i S\_[eff]{}\^i - S\_0\^i - ( S\_1\^i + S\_2\^i ) .
The consideration of the new Hamiltonian $H'_{\rm real}$ would considerably complicate (even at the qualitative level) the discussion of the following section. As we are not sure that this complication really entails a better [*quantitative*]{} description of spin effects, when these become important, we shall, in the following, content ourselves with studying the consequences of the simpler (though slightly less “accurate”) Hamiltonian $H_{\rm real}$, Eq. (\[eq2.44\]) with (\[eq2.45\]). However, we mention that it might be useful to consider simultaneously $H_{\rm real}$ and $H'_{\rm real}$, and to trust their predictions only in the cases where they differ only by a slight amount. This gives a useful measure of the domain of validity of the present spin-dependent effective-one-body approach.
Dynamics of two spinning black holes {#sec3}
====================================
Equations of motion and exact or approximate first integrals {#ssec3.1}
------------------------------------------------------------
In the previous section we have explicitly constructed an Hamiltonian $H_{\rm real}
(\mbox{\boldmath$x$} , \mbox{\boldmath$p$} , \mbox{\boldmath$S$}_1 ,
\mbox{\boldmath$S$}_2)$ describing (to some approximation) the (conservative part of the) gravitational interaction of two spinning black holes in the center-of-mass frame of the binary system. In the present section, we shall describe some consequences of this Hamiltonian. Let us start by writing down explicitly the evolution equations for all the dynamical variables. From the basic PB’s (\[eq2.10\])–(\[eq2.14\]) we get \[eq3.1\] = { x\^i , H\_[real]{} } = + , \[eq3.2\] = { p\_i , H\_[real]{} } = - , \[eq3.3\] = { S\_1\^i , H\_[real]{} } = \^[ijk]{} S\_1\^k , \[eq3.4\] = { S\_2\^i , H\_[real]{} } = \^[ijk]{} S\_2\^k . In vectorial notation, the spin evolution equations read (e.g. for the first spin) \[eq3.5\] = \_1 \_1 , \_1 . A first consequence of these results is that the magnitudes of the two spins are exactly conserved: \[eq3.6\] \_1\^2 = [const]{}. , \_2\^2 = [const]{}. Another general consequence is the exact conservation of the total angular momentum \[eq3.7\] + \_1 + \_2 , where $L^i \equiv \varepsilon^{ijk} \, x^j \, p_k$. Indeed, it is easily checked that $J^i$ generates, by Poisson brackets, global rotations of all the vectorial dynamical quantities: $\{ J^i , V^j \} = \varepsilon^{ijk} \, V^k$ for $\mbox{\boldmath$V$} =
\mbox{\boldmath$x$}$, $\mbox{\boldmath$p$}$, $\mbox{\boldmath$S$}_1$ or $\mbox{\boldmath$S$}_2$. As the Hamiltonian is a scalar constructed out of $\mbox{\boldmath$x$}$, $\mbox{\boldmath$p$}$, $\mbox{\boldmath$S$}_1$ and $\mbox{\boldmath$S$}_2$, we have \[eq3.8\] J\^i = { J\^i , H\_[real]{} } = 0 .
Therefore \[eq3.9\] J\^i = [const]{}, \^2 = [const]{}. Evidently, we have also the conservation of the total energy: \[eq3.10\] = { H\_[real]{} , H\_[real]{} } = 0 H\_[ real]{} = [const]{}. This closes the list of generic first integrals of the evolution. It should be noted that, in general, quantities such as $\mbox{\boldmath$L$}^2$ or $\mbox{\boldmath$S$}_1
\cdot \mbox{\boldmath$S$}_2$ are not conserved in time. This means, in particular, that the magnitude of the effective spin, $a^2 = M^{-2} \, \mbox{\boldmath$S$}_{\rm eff}^2$, will not stay constant during the evolution.
Evidently, in particular situations, more quantities might be approximately conserved. An interesting case is that in which the spins are small enough for one to retain only the terms linear in them. In this approximation \[eq3.11\] H\_[eff]{} (, , \_1 , \_2) H\_0 (, ) + \_[eff]{} , where $H_0 (\mbox{\boldmath$x$}, \mbox{\boldmath$p$})$ is spherically symmetric.
Let us, more generally[^9], assume that $H_{\rm eff}$, as well as $H_{\rm real}$, are spherically symmetric, functions of $\mbox{\boldmath$x$}$ and $\mbox{\boldmath$p$}$ except for a dependence on the combination $\mbox{\boldmath$L$} \cdot \mbox{\boldmath$S$}_{\rm eff}$: \[eq3.12\] H\_[real]{} = H\_[real]{} (r , p\_r , \^2 , \_[eff]{}) , where $p_r \equiv n^i \, p_i$ is canonically conjugated to $r$ ($\{ r , p_r \} = 1$). Under the assumption (\[eq3.12\]) the angular momenta evolution equations become (with $({\rm LS}) \equiv \mbox{\boldmath$L$} \cdot \mbox{\boldmath$S$}_{\rm eff}$) \[eq3.13\] = \_1 \_1 , \[eq3.14\] = \_2 \_2 , \[eq3.15\] = \_[eff]{} .
These evolution equations imply not only (as in the general case) the conservation of $\mbox{\boldmath$J$} = \mbox{\boldmath$L$} + \mbox{\boldmath$S$}_1 +
\mbox{\boldmath$S$}_2$, and of $\mbox{\boldmath$S$}_1^2$ and $\mbox{\boldmath$S$}_2^2$, but also that of: \[eq3.16\] \^2 = [const]{}, \_[eff]{} = [const]{}. Note, however, that $\mbox{\boldmath$S$}_{\rm eff}^2$ is not conserved. Moreover, the radial motion is governed by the equations \[eq3.17\] r = , \[eq3.18\] \_r = - . In view of the constancy of $\mbox{\boldmath$L$}^2 = C_0$ and $\mbox{\boldmath$L$} \cdot
\mbox{\boldmath$S$}_{\rm eff} = C_1$, we see from these equations that the function of $r$ and $p_r$, $H_{\rm rad} (r , p_r) = H_{\rm real} (r , p_r , C_0 , C_1)$, defines a reduced Hamiltonian describing the radial motion, separately from the angular degrees of freedom. In particular, we see (using the fact that $p_r$ enters at least quadratically in $H_{\rm real} $) that, under our current (approximate) assumption (\[eq3.12\]), there exists a class of [*spherical orbits*]{}, i.e. of orbits satisfying \[eq3.19\] r = [const]{} , p\_r = 0 , = 0 . Because of the (possibly non-linear) spin-orbit coupling, i.e. the dependence of $H_{\rm
real}$ on $\mbox{\boldmath$L$} \cdot \mbox{\boldmath$S$}_{\rm eff}$, the orbital plane of these “spherical” orbits is not fixed in space. But the radial coordinate $r$ being constant, these orbits trace a complicated path on a sphere (hence the name). These orbits are the analogs, in our two-body problem, and in the approximation (\[eq3.12\]), of similar exact “spherical” orbits for the geodesic motion of test particles in a Kerr spacetime [@wilkins]. Their existence (under some approximation) in the two-body problem is interesting for the following reason. One expects most black hole binary sources of interest for the LIGO/VIRGO/GEO network to have had the time to relax, under radiation reaction, to circular orbits. When the two black holes will get closer, these circular orbits will adiabatically shrink until they come close enough for feeling the effect of the spin-orbit coupling (which varies $\propto \, r^{-3}$). In some intermediate domain where the spin-orbit coupling is significant, but couplings quadratic in the spins are still small, the initially circular orbit will evolve into an adiabatic sequence of “spherical” orbits of the type just discussed. \[We are here adding by hand the effect of radiation reaction, treated as an adiabatic perturbation of the conservative dynamics discussed in this paper.\] These considerations indicate that, in first approximation, the total amount of gravitational radiation emitted by coalescing spinning black holes will be determined by the binding energy of the Last Stable Spherical Orbit (LSSO), i.e. the last stable solution of Eqs. (\[eq3.19\]), which will satisfy \[eq3.20\] (r , p\_r = 0 ,\^2 , \_[eff]{}) = 0 , (r , p\_r = 0 ,\^2 , \_[eff]{}) = 0 .
Before studying the energetics of these LSSO’s let us mention the existence of other approximate first integrals in the dynamics of binary spinning black holes. Let us keep all the terms non-linear in $\mbox{\boldmath$S$}_{\rm eff}$, i.e. the full expression of $H_{\rm real} (\mbox{\boldmath$x$} , \mbox{\boldmath$p$} , \mbox{\boldmath$S$}_1 ,
\mbox{\boldmath$S$}_2)$, but let us try to approximately decouple the orbital motion from the spin degrees of freedom by considering that the two spin vectors evolve adiabatically (i.e. slowly on the orbital time scale), through Eqs. (\[eq3.3\]), (\[eq3.4\]). In this adiabatic-spin approximation, the orbital motion is described by the Hamilton-Jacobi equation (\[eq2.17\]), with an adiabatically fixed effective metric $g_{\rm eff}^{\alpha\beta}$. With the definition (\[eq2.36new\]) of the quadratic-in-momenta term $Q_4 (p)$, one can check that, in this approximation, there will exist a two-body analog of the Carter constant for geodesic motion in Kerr [@carter]. Indeed, we have constructed our deformed Kerr metric (\[eq2.33\]) so as to respect its separability properties. Let us work in an (adiabatic) Boyer-Lindquist-type coordinate system $(t,r,\theta,\varphi)$, as in Eq. (\[eq2.33\]). We find that the separability of the effective Hamilton-Jacobi equation yields the following first integrals (of the effective Hamiltonian) \[eq3.21\] p\_t = - E\_[eff]{} , p\_ = L\_z , \[eq3.22\] p\_\^2 + + \^2 a\^2 \^2 = [K]{} + (L\_z - a E\_[eff]{})\^2 , \[eq3.23\] p\_\^2 + \^2 = [Q]{} - (L\_z - a E\_[eff]{})\^2 . The last two equations are equivalent to each other, but, depending on the context, one can be more convenient than the other. Let us note the connection of the first integrals (\[eq3.21\])–(\[eq3.23\]) with the above analysis of the first integrals of the Hamiltonian depending only on the combination $\mbox{\boldmath$L$} \cdot
\mbox{\boldmath$S$}_{\rm eff}$. The conservation of $L_z$, Eq. (\[eq3.21\]), corresponds to the conservation of $\mbox{\boldmath$L$} \cdot \mbox{\boldmath$S$}_{\rm
eff}$, Eq. (\[eq3.16\]), while the conservation of ${\cal K}$ or ${\cal Q}$ corresponds to the conservation of $\mbox{\boldmath$L$}^2$. Indeed, if we neglect the terms $\propto \, a^2$ in (\[eq3.23\]) we get \[eq3.24\] [Q]{} p\_\^2 + L\_z\^2 = \^2 - L\_z\^2 . This suggests that, even beyond the adiabatic-spin approximation, the quantities, now defined in an arbitrary frame as \[eq3.25\] L\_z , [Q]{} \^2 - ( )\^2 + a\^2 ( )\^2 (\^2 - E\_[eff]{}\^2) , will be (as well as $\mbox{\boldmath$S$}_{\rm eff}^2$ and ${\cal K} \equiv {\cal Q} +
(L_z - a \, E_{\rm eff})^2$) conserved to a good approximation. We are mentionning here these approximate conservation laws because they could be helpful in qualitatively understanding the full two-body dynamics.
Spherical orbits and last stable spherical orbits {#ssec3.2}
-------------------------------------------------
We discussed above the existence of spherical orbits under the assumption (or the approximation) that $H_{\rm real}$ depend only on the “spin-orbit” combination $\mbox{\boldmath$L$} \cdot \mbox{\boldmath$S$}_{\rm eff}$ (as it does at the linear-in-spin level). More generally, we have seen that if we treat the evolution of the spins as being adiabatic, we have the (approximate) first integrals (\[eq3.25\]). If we use (as a heuristic mean of studying the main features of the orbital dynamics) this adiabatic approximation, we can define a family of spherical orbits by drawing on the conservation of the quantities (\[eq3.25\]). Indeed, inserting the definitions (\[eq3.25\]) into Eq. (\[eq2.17\]) we get an equation controlling the radial motion: $$\begin{aligned}
\label{eq3.26}
&\Delta_r \, p_r^2 + 2 \, (4 - 3 \nu) \, \nu \, (GM)^2 \, p_r^4 / \mu^2 = \nonumber
\\
&\displaystyle \frac{1}{\Delta_t \, (r)} \ [(r^2 + a^2) \, E_{\rm eff} - a \, L_z]^2 -
(\mu^2 \, r^2 +
{\cal Q} + (L_z - a \, E_{\rm eff})^2 ) \, .\end{aligned}$$ The right-hand-side of Eq. (\[eq3.26\]) defines (when, $a^2 = \mbox{\boldmath$S$}_{\rm
eff}^2 / M^2$, $L_z$ and ${\cal Q}$ are considered as adiabatic constants) a radial potential whose local minima, in $r$, determine (adiabatic) spherical orbits. The last stable spherical orbit is obtained when this radial potential has an inflection point. More precisely let us define \[eq3.27\] R (r , E\_[eff]{} , L\_z , [Q]{}) ((r\^2 + a\^2) E\_[eff]{} - a L\_z)\^2 - \_t (r) (\^2 r\^2 + [Q]{} + (L\_z - a E)\^2 ) . The spherical orbits are the solutions of \[eq3.28\] R (r , E\_[eff]{} , L\_z , [Q]{}) = R (r , E\_[ eff]{} , L\_z , [Q]{}) = 0 . The solutions of Eq. (\[eq3.28\]) yield a two-parameter family of solutions, along which, for instance, $r$ and $E_{\rm eff}$ are functions of $L_z$ and ${\cal Q}$. The last stable spherical orbit (LSSO) along such a family of solutions must satisfy the three equations \[eq3.29\] R (r , E\_[eff]{} , L\_z , [Q]{}) = R (r , E\_[ eff]{} , L\_z , [Q]{}) = R (r , E\_[eff]{} , L\_z , [Q]{}) = 0 . There is a one-parameter family of LSSO’s. For instance, one can take as free parameter the dimensionless ratio ${\cal Q} / L_z^2$ which is a measure of the maximum angle between the orbital plane and the equatorial plane defined by $\mbox{\boldmath$S$}_{\rm
eff}$. \[Note that ${\cal Q} = 0$ for an orbit in the equatorial plane.\] For each value of this angle, and for each value of the effective spin parameter $a$, there will be some LSSO, with particular values of $r$, $E_{\rm eff}$ and $L_z$.
To study the values of the (effective and real) binding energy, and of the orbital angular momentum along this one-parameter family of LSSO’s, it is convenient to work with slightly different variables. Let us introduce \[eq3.30\] |[L]{}\_z L\_z - a E\_[eff]{} , [K]{} + |[L]{}\_z\^2 . Let us also work with the radial variable $u \equiv 1/r$ and denote \[eq3.31\] |A (u) P\_3\^1 \[A\_[orb]{}\^[PN]{} (u) + a\^2 u\^2\] . We have \[eq3.32\] r\^[-4]{} R(r) U (u) (E\_[eff]{} - a |[L]{}\_z u\^2)\^2 - |A (u) (\^2 + [K]{} u\^2) . The equation $U (u) = 0$ (i.e. $R(r) = 0$) is now solved as \[eq3.33\] E\_[eff]{} = W\_a (u , |[L]{}\_z , [K]{}) a |[L]{}\_z u\^2 + . The two-parameter family of spherical orbits is now obtained (as functions of the parameters $\bar{L}_z$ and ${\cal K}$) by solving $\partial \, W / \partial \, u = 0$, while the one-parameter family of LSSO’s is obtained by solving $\partial \, W /
\partial \, u = \partial^2 \, W / \partial \, u^2 = 0$. The advantage of this formulation is that it exhibits in the simplest way the analogy with the effective radial potential discussed in [@BD99; @DJS2] for the pure (3 PN) orbital motion (without spin), namely \[eq3.34\] W\_0 (u,L) , with $A(u) \equiv P_3^1 \, [A_{\rm orb}^{\rm PN} (u)]$. Apart from the replacement $L^2
\rightarrow {\cal K}$, the only two differences between the spinning case ((\[eq3.33\])) and the spinless one ((\[eq3.34\])) is the addition of the spin-orbit energy term $+ \, a \, \bar{L}_z \, u^2$, and the additional $a^2 \, u^2$ term in the PN expansion of $\bar A (u)$. \[Note that $\bar A (u) \ne A(u) + a^2 \, u^2$ because the Padéeing is done after the addition of $a^2 \, u^2$.\] We have chosen to parametrize $W_a (u)$ in terms of $\bar{L}_z$ and ${\cal K}$ because it simplifies very much its expression and thereby renders more transparent the new physics incorporated in our effective one-body approach. The fact that $\bar{L}_z$ depends both on $L_z$ and $E_{\rm
eff}$ is not a problem for solving Eq. (\[eq3.32\]) for $E_{\rm eff}$. Indeed, we are discussing a continuous family of solutions and it is essentially indifferent to parametrize them in terms of $L_z$ or $\bar{L}_z$. We could have introduced another effective potential $W'_a (r , L_z , {\cal Q})$ by solving $R (r , E_{\rm eff} , L_z ,
{\cal Q})= 0$, with Eq. (\[eq3.27\]), which would be more complicated, but which would describe the same physics. \[Note that $W'_a (r)$ would directly exhibit the correct fact that the spin-orbit energy, for given $L_z$, decreases like $r^{-3}$, while this fact is hidden in $W_a (u)$ which assumes that $\bar{L}_z = L_z - a \, E_{\rm eff}$ is given.\]
Binding energy of last stable spherical orbits {#ssec3.3}
----------------------------------------------
To get a first idea of the physical consequences of our effective one-body description of coalescing spinning black holes we have numerically investigated the properties of the one-parameter family of LSSO’s. The most important quantity we are interested in is the binding energy at the last stable spherical orbit because it is the prime quantity determining the detectability of the GW emitted during the inspiral. We recall that the real, two-body energy is related to the effective energy entering the equations of the previous subsection through \[eq3.35\] E\_[real]{} = M . We are mostly interested in the (dimensionless) binding energy per unit total mass, say \[eq3.36\] e = - 1 . The value of $e$ at the LSSO depends on three dimensionless parameters \[eq3.37\] \_4 4 4 , a , \_[[LS]{}]{} . Here, the parameter $\nu_4$ (renormalized so that $0 < \nu_4 \leq 1$) determines the effect of having comparable masses ($\nu_4 \simeq 1$) rather than a large mass hierarchy ($\nu_4 \ll 1$). The dependence of $e^{\rm LSO}$ on $\nu_4$ in absence of spins was studied in [@BD99; @DJS2]. It was found that the ratio $e^{\rm LSO} / \nu_4$ was essentially linear in $\nu_4$ (even for $\nu_4$ as large as 1, corresponding to the equal-mass case) \[eq3.38\] e\_[S\_[eff]{} = 0]{}\^[LSO]{} - 0.014298 \_4 (1 + c\_1 \_4) . Here, the numerical value $- 1.4298\% = \frac{1}{4} \left(\sqrt{\frac{8}{9}} - 1
\right)$ is one fourth the specific binding LSO energy of a test particle in Schwarzschild. The numerical coefficient $c_1$ which condenses the effect of resummed PN interactions was found to have a value $c_1^{2 \, {\rm PN}} \simeq 0.048$ at 2 PN and $c_1^{3 \, {\rm PN}} (\omega_s = 0) \simeq 0.168$ at 3 PN, and for $\omega_s = 0$. \[The dependence of $c_1^{3 \, {\rm PN}}$ on $\omega_s$ is also roughly linear: $c_1^{3 \,
{\rm PN}} (\omega_s) \simeq 0.168 + 0.0126 \, \omega_s$, at least when $- \, 10 \, \laq \,
\omega_s \, \laq \, 0$.\]
We expect that the dependence on $\nu_4$ of the spin-dependent effects will also be roughly linear (after factorization of an overall factor $\nu_4$ which comes from expanding the square root in Eq. (\[eq3.36\])). In the following we shall generally consider (in our numerical investigations) the case $\nu_4 = 1$, and concentrate on the dependence on the other parameters.
Let us clarify the meaning of the parameters $\widehat a$ and $\cos \theta_{\rm LS}$ introduced in Eq. (\[eq3.37\]). The quantities $\mbox{\boldmath$S$}_{\rm eff}$, $\bar{L}_z$ and ${\cal K}$ entering these definitions are all supposed to be computed at the last stable spherical orbit of an adiabatic sequence of spherical orbits (in the sense discussed above). Physically, we have in mind the sequence of inspiralling orbits driven by radiation reaction. Technically, we define $e (\nu_4 , \widehat a , \cos \theta_{\rm
LS})$ by solving the effective radial potential problem defined in the previous subsection. We are aware of the fact that we cannot really attach to $\cos \theta_{\rm
LS}$ the meaning of being the cosinus between $\mbox{\boldmath$L$}$ and $\mbox{\boldmath$S$}_{\rm eff}$ (as the name would suggest), but this is not important. What is important is that there is indeed a physical degree of freedom related to the misalignment between $\mbox{\boldmath$L$}$ and $\mbox{\boldmath$S$}_{\rm eff}$ at the LSSO and that we measure it by a parameter normalized so that $\cos \theta_{\rm LS} = 1$ (or $-1$) when all angular momenta are aligned (in this limit the concept of last stable circular equatorial orbit is meaningful and coincides with the $\cos \theta_{\rm LS} =
1$ (or $-1$) limit of our formal definitions).
Note that, with the definitions (\[eq3.37\]) and the additional definition $\bar{\ell}
\equiv ({\cal K})^{1/2} / GM \, \mu$, the effective radial potential (\[eq3.33\]) yields (in dimensionless form, $\widehat u \equiv GM / r$) \[eq3.39\] = \_[a]{} (u , \_[LS]{} , |) = a \_[LS]{} | \^2 + . This form makes it clear that $E_{\rm eff}^{\rm LSSO} / \mu$, and therefore $e^{\rm
LSSO}$, Eq. (\[eq3.36\]), will depend primarily on the combination (“projected value of $\widehat a$”) \[eq3.40\] \_p a \_[LS]{} where $\mbox{\boldmath$k$} = \mbox{\boldmath$L$} / \vert \mbox{\boldmath$L$} \vert$ (at the LSSO). For smallish spins, the combination $\widehat{a}_p$ is the only one entering the problem [*linearly*]{}. As recalled by the notation in (\[eq3.39\]) the non-projected value of $\widehat a$ enters only quadratically in $\bar A (\widehat u ,
\widehat{a}^2)$.
Let us consider more closely the crucial quantity $$\begin{aligned}
\label{eq3.41}
\widehat{a}_p &= &\frac{1}{(m_1 + m_2)^2} \left[ \left( 1 + \frac{3}{4} \,
\frac{m_2}{m_1} \right) \, \mbox{\boldmath$k$} \cdot \mbox{\boldmath$S$}_1 + \left( 1 +
\frac{3}{4} \, \frac{m_1}{m_2} \right) \, \mbox{\boldmath$k$} \cdot
\mbox{\boldmath$S$}_2 \right] \nonumber \\
&= &\left( X_1^2 + \frac{3}{4} \, X_1 \, X_2 \right) \, \mbox{\boldmath$k$} \cdot
\widehat{\mbox{\boldmath$a$}}_1 + \left( X_2^2 + \frac{3}{4} \, X_1 \, X_2 \right) \,
\mbox{\boldmath$k$} \cdot \widehat{\mbox{\boldmath$a$}}_2 \, . \end{aligned}$$ In the second form, we have defined $X_1 \equiv m_1 / M$, $X_2 \equiv m_2 / M$ ($X_1 +
X_2 = 1$, $X_1 \, X_2 = \nu$) and $\widehat{\mbox{\boldmath$a$}}_1 \equiv
\mbox{\boldmath$S$}_1 / m_1^2$, $\widehat{\mbox{\boldmath$a$}}_2 \equiv
\mbox{\boldmath$S$}_2 / m_2^2$. \[We recall that a maximally spinning hole would have $\vert \widehat{\mbox{\boldmath$a$}}_1 \vert = 1$.\]
An important question (for the relevance of the present work) is: what are the plausible values of $\widehat{a}_p$ in the sources that will be detected by LIGO/VIRGO/GEO? Present astrophysical ideas about the formation of binary black holes [@postnov; @PZ99] suggest neither that the holes be typically maximally spinning, nor that there be any correlation between the spin and angular momenta, i.e. between the directions of $\mbox{\boldmath$k$}$, $\mbox{\boldmath$S$}_1$ and $\mbox{\boldmath$S$}_2$. Not much is known either about the probable value of the mass ratio. To have an idea of the plausible values of $\widehat{a}_p$ (which is an algebraic quantity which can take positive or negative values) let us consider the random mean square (rms) value of $\widehat{a}_p$ under the assumption of random, uncorrelated directions $\mbox{\boldmath$k$}$, $\mbox{\boldmath$S$}_1$ and $\mbox{\boldmath$S$}_2$ (so that $\langle \widehat{a}_p \rangle = 0$). Let us assume (for simplicity) that $m_1
= m_2$, i.e. $\nu_4 = 1$, which is the most favourable case because $e^{\rm LSSO}
\propto \nu_4$. We assume also that $\langle \widehat{\mbox{\boldmath$a$}}_1^2 \rangle =
\langle \widehat{\mbox{\boldmath$a$}}_2^2 \rangle \equiv (a_1^{\rm rms})^2$ is some given quantity (to be determined by astrophysical models). This yields for $\widehat{a}_p^{\rm rms} \equiv \sqrt{\langle \widehat{a}_p^2 \rangle}$ \[eq3.42\] \_p\^[rms]{} = a\_1\^[rms]{} = 0.357 a\_1\^[rms]{} . Even if $\widehat{a}_1^{\rm rms} = 1$ (which would mean that all black holes are maximally spinning) we get $\widehat{a}_p^{\rm rms} = 0.357$. However, we find it highly plausible that $\widehat{a}_1^{\rm rms}$ will be significantly smaller than 1. For instance, if we optimistically assume a uniform distribution of spin kinetic energy between 0 and the maximal value we would get $\widehat{a}_1^{\rm rms} = \frac{1}{\sqrt 2}$ and therefore $\widehat{a}_p^{\rm rms} = 7 / (16 \, \sqrt 3) = 0.253$. In view of these arguments, we find plausible that most LIGO/VIRGO binary black hole sources will have $\vert
\widehat{a}_p \vert \, \laq \, 0.3$. This consideration is important because we shall see later that for such smallish values of $\widehat{a}_p$ the simple analytical approach advocated here seems to be quite reliable. However, one should also be able to compute physically reliable (or, at least, sufficiently flexible) templates for fast spinning binary black holes. As we said in the Introduction, we think that the EOB approach can be an essential tool for this purpose, in conjunction with numerical data, by feeding the (necessarily sparse) numerical data into some multi-parameter version of the EOB Hamiltonian.
This statistical estimate of the plausible value of $a_p$ suggests that a typical value of $\cos \theta_{\rm LS} \simeq \mbox{\boldmath$k$} \cdot \mbox{\boldmath$S$}_{\rm eff}
/ \vert \mbox{\boldmath$S$}_{\rm eff} \vert$ is around $\pm \, 1 / \sqrt 3$. In our numerical estimates of $e^{\rm LSSO}$ we have used this value, as well as the (implausible) value $\cos \theta_{\rm LS} = \pm \, 1$ corresponding to perfect alignment. As a first step towards exploring the “parametric flexibility” of the EOB approach, we have studied the dependence of $e^{\rm LSSO}$ on the value of the parameter $\omega_s$, as it appears in Eq. (\[eq2.21\]). We have done numerical simulations for three fiducial values: $\omega_s = 0
\equiv \omega^{\rm DJS}$ [@DJSd], $\omega_s = - 1987 / 840 \equiv \omega_s^{\rm BF}$ [@BF1; @BF3][^10], and also for $\omega_s = - \frac{1}{2} \left( \frac{94}{3} - \frac{41 \, \pi^2}{32} \right) \equiv
\omega_s^*$. Note the numerical values: $\omega_s^{\rm BF} \simeq - 2.3655$, $\omega_s^* \simeq - 9.3439$. The original motivation (when writing this paper, before the completion of the work [@DJSd] which determined the correct value of $\omega_s$) was to study the sensitivity of our results to the “3PN ambiguity”. We kept it here as an interesting case study of the sensitivity of EOB results to modifications of the various coefficients entering the EOB Hamiltonian. The change from $\omega_s = 0$ to $\omega_s = \omega_s^{\rm BF}$ corresponds to a change of the coefficient $a_4 \equiv a_4(\nu) / \nu$ from $18.688$ to $13.957$, i.e. a fractional change of $ - 25.32 \%$. The value $\omega_s = \omega_s^*$ has the effect of completely cancelling the 3 PN contribution to the radial functions $A(u)$ and $\bar A (u)$. Therefore, choosing $\omega_s
= \omega_s^*$ gives for the LSSO quantities the same results as the 2 PN effective-one-body Hamiltonian [@BD99]. Its consideration is useful for exhibiting the difference between the 2 PN-based results and the 3 PN-based ones. Our results are displayed in Table \[Tab2\] and Fig. 1.
[cc]{} $\cos \theta_{\rm LS} = \pm \, 1 / \sqrt 3$ &$\cos \theta_{\rm LS} = \pm \, 1$\
----------------- -- -------------- -- ---------------- -- --------------------------
$\widehat{a}_p$ $\widehat a$ $e^{\rm LSSO}$ $\widehat{r}^{\rm LSSO}$
$-0.6$ $-1.039$ $-0.01319$ $6.298$
$-0.5$ $-0.8660$ $-0.01337$ $6.220$
$-0.4$ $-0.6928$ $-0.01364$ $6.109$
$-0.3$ $-0.5196$ $-0.01405$ $5.940$
$-0.2$ $-0.3464$ $-0.01463$ $5.700$
$-0.1$ $-0.1732$ $-0.01547$ $5.377$
$0.$ $0.$ $-0.01670$ $4.954$
$+0.1$ $+0.1732$ $-0.01859$ $4.391$
$+0.2$ $+0.3464$ $-0.02203$ $3.580$
$+0.3$ $+0.5196$ $-0.05146$ $1.344$
$+0.4$ $+0.6928$ $-0.1790$ $0.7752$
$ \ $ $ \ $ $ \ $ $ \ $
$ \ $ $ \ $ $ \ $ $ \ $
----------------- -- -------------- -- ---------------- -- --------------------------
: Binding energies $e \equiv ({\cal E}_{\rm real} / M) - 1$ and (effective Boyer-Lindquist) radii of last stable spherical orbits (LSSO) for equal-mass spinning binary systems. The LSSO’s depend on two independent parameters: $\vert \widehat{\mbox{\boldmath$a$}} \vert \equiv \vert \mbox{\boldmath$S$}_{\rm eff}
\vert / M^2$ and $\cos
\theta_{\rm LS} \equiv \bar{L}_z / \sqrt{{\cal K}}$ (which is, morally, the cosine of the angle between the orbital angular momentum and the effective spin). The combined parameter $\widehat{a}_p \equiv \vert \widehat{\mbox{\boldmath$a$}} \vert \, \cos \theta_{\rm LS}$ (projected spin) plays a primary role for moderate spins. The algebraic quantity $\widehat{a}$ is defined as $+ \, \vert
\widehat{\mbox{\boldmath$a$}} \vert$ if $\cos \theta_{\rm LS} > 0$ and $- \, \vert
\widehat{\mbox{\boldmath$a$}} \vert$ if $\cos \theta_{\rm LS} < 0$. All quantities are computed from the 3 PN-level Padé-resummed effective-one-body Hamiltonian (\[eq2.44\]) with $\omega_s = 0$.[]{data-label="Tab2"}
&
----------------- -- -------------- -- ---------------- -- --------------------------
$\widehat{a}_p$ $\widehat a$ $e^{\rm LSSO}$ $\widehat{r}^{\rm LSSO}$
$-0.6$ $-0.6$ $-0.01150$ $7.344$
$-0.5$ $-0.5$ $-0.01207$ $6.989$
$-0.4$ $-0.4$ $-0.01271$ $6.623$
$-0.3$ $-0.3$ $-0.01345$ $6.242$
$-0.2$ $-0.2$ $-0.01433$ $5.841$
$-0.1$ $-0.1$ $-0.01538$ $5.415$
$0.$ $0.$ $-0.01670$ $4.954$
$+0.1$ $+0.1$ $-0.01842$ $4.439$
$+0.2$ $+0.2$ $-0.02091$ $3.833$
$+0.3$ $+0.3$ $-0.02529$ $3.005$
$+0.4$ $+0.4$ $-0.04930$ $1.538$
$+0.5$ $+0.5$ $-0.1048$ $1.194$
$+0.6$ $+0.6$ $-0.1474$ $0.9792$
----------------- -- -------------- -- ---------------- -- --------------------------
: Binding energies $e \equiv ({\cal E}_{\rm real} / M) - 1$ and (effective Boyer-Lindquist) radii of last stable spherical orbits (LSSO) for equal-mass spinning binary systems. The LSSO’s depend on two independent parameters: $\vert \widehat{\mbox{\boldmath$a$}} \vert \equiv \vert \mbox{\boldmath$S$}_{\rm eff}
\vert / M^2$ and $\cos
\theta_{\rm LS} \equiv \bar{L}_z / \sqrt{{\cal K}}$ (which is, morally, the cosine of the angle between the orbital angular momentum and the effective spin). The combined parameter $\widehat{a}_p \equiv \vert \widehat{\mbox{\boldmath$a$}} \vert \, \cos \theta_{\rm LS}$ (projected spin) plays a primary role for moderate spins. The algebraic quantity $\widehat{a}$ is defined as $+ \, \vert
\widehat{\mbox{\boldmath$a$}} \vert$ if $\cos \theta_{\rm LS} > 0$ and $- \, \vert
\widehat{\mbox{\boldmath$a$}} \vert$ if $\cos \theta_{\rm LS} < 0$. All quantities are computed from the 3 PN-level Padé-resummed effective-one-body Hamiltonian (\[eq2.44\]) with $\omega_s = 0$.[]{data-label="Tab2"}
The most important conclusion we wish to draw from these results is that, when $\widehat a
\, \laq \, 0.3$ (which, as we argued above, covers a large domain of the physically relevant cases), the binding energy at the LSSO seems to be reliably describable by the simple analytical EOB Hamiltonian defined above. Indeed, the differences between: (i) the non-spinning case and the spinning ones, and (ii) the 2 PN orbital approximation and the 3 PN one, are all quite moderate (which indicates that the effective one-body approach is effective in resumming PN interactions near the LSO). Furthermore, the difference between: (iii) the spinning 3 PN case with $\omega_s = 0 = \omega_s^{\rm DJS}$ [@DJS2], and the same case with $\omega_s = \omega_s^{\rm BF}$ [@BF3] is rather small. This is a testimony of the robustness of the EOB approach. A change of its 3PN coefficient by $25 \%$ does not affect much the physical predictions. This robustness at the 3PN level is indicative of some robustness against the addition of higher PN effects. Note also (from table \[Tab2\]) the confirmation that when $\vert \widehat{a}_p \vert \, \laq \,
0.2$, the binding energy at the LSSO depends nearly only on the projected effective spin parameter $\widehat{a}_p = \widehat{a} \cos \theta_{\rm LS}$, with a very weak dependence on the value of $\cos \theta_{\rm LS}$.
On the other hand, it must be admitted that when, say, $\widehat{a} \, \gaq \, 0.4$ the differences between the three cases (i), (ii), (iii) become so large, and the radius of the LSSO becomes so small, that the present spin-dependent EOB predictions cannot be quantitatively trusted. \[However, as discussed in more detail below, we think that they remain qualitatively correct.\] If the orbital dynamics were well described by the 2 PN-level orbital EOB metric (i.e. if $\omega_s$ had been near $-9$; see upper curve in Fig. \[Fig1\]), the binding energy, even in such extreme cases, would differ only moderately from the non-spinning case, and we could trust the EOB-plus-spin predictions. However, as $\omega_s$ is zero the 3 PN EOB $+$ spin predictions become, for $\cos
\theta_{\rm LS} \simeq 1$ and $a \, \gaq \, 0.4$, very different from the 2 PN ones and quite sensitive to the numerical values of the expansion coefficients entering the EOB potentials. Let us note, however, that in all cases (even the most extremely spinning ones) if the spin parameter $\widehat{a}_p$ is [*negative*]{} (i.e. if the effective spin vector, whatever be its magnitude, has a negative projection on the orbital angular momentum) the EOB predictions become extremely reliable because all the differences between the cases (i), (ii), (iii) become quite small.
All these results are easy to interpret physically. This can be seen from the basic equations of the EOB approach which simplify so much the description of the physical interactions by representing them as slightly deformed versions of the well-known gravitational physics of test particles in Schwarzschild or Kerr geometries. Indeed, the basic equation of the EOB approach determining the binding of the LSO is Eq. (\[eq3.39\]) which differs from its well-known[^11] Kerr limit (i.e. $\nu \rightarrow 0$) only by the change \[eq3.43\] A\_K (u , \^2) = 1 - 2 u + \^2 \^2 |A (u , \^2) = P\_3\^1 \[1 - 2 u + \^2 \^2 + 2 \^3 + a\_4 () \^4 \] . The crucial point (which is, finally, the most important new information obtained by the 2 PN and 3 PN orbital calculations) is that the 2 PN and 3 PN additional terms to the radial function $A_{a \, = \, 0}^{\rm PN} (\widehat{u})$ have both [*positive*]{} coefficients. This means that, even before the addition of the effect of spin (which leads to a $+ \,
\widehat{a}^2 \, \widehat{u}^2$ additional term in $A_K (\widehat{u})$, corresponding to the famous $+ \, a^2$ term in $\Delta_K (r) = r^2 - 2 Mr + a^2$) the main effect of non-linear orbital interactions for comparable masses is [*“repulsive”*]{}, i.e. correspond to a partial screening of the basic Schwarzschild attractive term $1 - 2 \, \widehat{u}
= 1 - 2 \, GM / c^2 \, r$ by the addition of repulsive terms $\propto + \, \nu /
r^3$ and $+ \, \nu / r^4$. Now, paradoxically, the addition of a repulsive term leads to a more tightly bound LSSO because the [*less attractive*]{}, but still attractive[^12], radial function $A (\widehat{u} , \widehat{a}^2 = 0)$ will be able to “hold” a particle in spherical orbit down to a [*lower*]{} orbit. In other words, when a radial potential becomes less attractive, its LSSO gets closer to the horizon, and the binding energy of the LSSO becomes more negative. This being said, one understands immediately the additional effects due to the spin interaction. There are basically two such effects: (a) a linear “spin-orbit” effect linked to the $+ \, \widehat{a}_p \,
\bar \ell \, \widehat{u}^2$ term in (\[eq3.39\]) (with $\widehat{a}_p \equiv \widehat{a}
\cos
\theta_{\rm LS}$), and (b) a non-linear spin-quadratic modification of the metric coefficient, i.e. the additional $+ \, \widehat{a}^2 \, \widehat{u}^2$ term in $\bar A
(\widehat{u} , \widehat{a}^2)$ (or in $A_K (\widehat{u}) = 1 - 2 \, \widehat{u} +
\widehat{a}^2 \, \widehat{u}^2$). The crucial points are that: (1) when $\widehat{a}_p <
0$, i.e. $\cos \theta_{\rm LS} < 0$ (coarse antialignment of angular momenta) the dominant linear spin-orbit coupling is [*attractive*]{} and therefore pushes the LSSO [*upwards*]{}, towards a less bound orbit, while, (2) when $\widehat{a}_p > 0$, i.e. $\cos \theta_{\rm LS} > 0$ (coarse alignment of angular momenta) [*both*]{} the linear spin-orbit coupling $+ \, \widehat{a}_p \, \bar \ell \, \widehat{u}^2$ and the spin-quadratic additional term $+ \, \widehat{a}^2 \, \widehat{u}^2$ are [*repulsive*]{} and tend to draw the LSSO [*downwards*]{}, i.e. closer to the horizon, in a more bound orbit. Therefore we see that, when $\widehat{a}_p >0$, all the new effects (the $\nu$-dependent non-linear orbital interactions and the spin effects) tend in the same direction: towards a closer, more bound orbit. As the existence of a LSSO is due to a delicate balance between the attractive gravitational effects and the usual repulsive (“centrifugal”) effect of the orbital angular momentum (i.e. the term $+ \, \bar{\ell}^2 \, \widehat{u}^2 \propto + \,
L^2 / r^2$ in (\[eq3.39\])), when several attractive effects combine their action, they start having a large effect on the binding of the LSSO. This is well-known to be the case for circular, equatorial, corotating $(\widehat{a}_p = + \, \widehat{a})$ orbits of a test particle in Kerr, which feature, in the case of an extreme Kerr $(\widehat{a} = 1)$ an LSO at $\widehat r = 1$, with $\mu$-fractional binding $(E_{\rm eff} - \mu) / \mu = 1
/ \sqrt 3 - 1 = - 0.42265$ (corresponding to $e \simeq \nu \, (E_{\rm eff} - \mu) / \mu
\simeq - 0.10566 \, \nu_4$). It is also well-known that, again for extreme Kerr, a counterrotating $(\widehat{a}_p = - \, \widehat{a})$ circular, equatorial orbit in extreme Kerr has an LSO at $\widehat r = 9$, with $\mu$-fractional binding $(E_{\rm eff} - \mu)
/ \mu = 5 / (3 \, \sqrt 3 ) - 1 = - 0.037750$ (corresponding to $e \simeq - 0.0094374 \,
\nu_4$). What is less well-known is that the extreme binding of the circular, equatorial, corotating LSO around an extreme Kerr is not representative of the binding of typical LSSO’s around typical (or even extreme) Kerr holes. Indeed, when $\cos \theta_{\rm LS} \ne \pm \, 1$ (i.e. in more invariant language, when ${\cal Q} \ne
0$) and when $\widehat{a} < 1$, the LSSO is, in general, moderately perturbed away from the Schwarzschild value $\widehat{r}_{\rm LSO} = 6$ and its binding is correspondingly moderately different from its Schwarzschild limit $e_{\rm Schw}^{\rm LSO} \simeq -
0.014298 \, \nu_4$. The present work has shown that the location of the LSSO’s for binary spinning holes can be rather simply obtained, in the EOB approach, by balancing in the specific way of Eq. (\[eq3.39\]) the centrifugal effect of the orbital angular momentum against the overall attractive effect of gravity, but with the critical addition of the 2 PN and 3 PN repulsive terms, of the spin-quadratic repulsive term, and of the indefinite-spin effect of the spin-orbit interaction.
Expected spin of the hole formed by the coalescence of two spinning holes {#ssec3.4}
-------------------------------------------------------------------------
The last topic we wish to discuss concerns the expected result of the coalescence of two holes. In particular, we are interested in estimating the maximal spin that the final hole, resulting from the coalescence of two spinning holes, might have. It was estimated in [@BD00; @BD01] (by using the EOB approach) that the coalescence of two non-spinning holes of the same mass $m_1 = m_2 = M/2$ leads (after taking into account the effect of gravitational radiation on the orbital evolution and on the loss of energy and angular momentum) to the formation of a rotating black hole of mass $M_{\rm BH}
\simeq (1 - \varepsilon_{\rm rd}) \, 0.976 \, M$ and spin parameter $\widehat{a}_{\rm
BH} \simeq 0.80$ \[we have included a factor $1 - \varepsilon_{\rm rd}$ in $M_{\rm BH}$ to take into account the energy loss during the ring-down. Ref. [@BD01] found $\varepsilon_{\rm rd} \simeq 0.7\%$\]. The fractional energy $0.976 - 1 = - 0.024$ roughly corresponds to the (adiabatically estimated) LSO binding energy ($- 0.015$ in the 2 PN-based estimate of [@BD99]) minus the energy per unit mass radiated during the plunge ($\sim - 0.007$ [@BD01]). We shall leave to future work a similar estimate, for the 3 PN-plus-spin case, of the amount of energy emitted in GW. We wish here to focus on the issue of the spin of the final hole. The above value $\widehat{a}_{\rm BH} \simeq 0.80$ is rather close to the maximal value $\widehat{a}_{\rm max} = 1$ and there arises the question of whether an EOB treatment of the coalescence of two spinning holes might not formally predict a final value of $\widehat{a}_{\rm BH}$ larger than one! By “EOB treatment” we mean here a completed version of the EOB approach (as in [@BD00] at the 2 PN, non-spinning level) obtained by: (i) adding a resummed radiation force to the “conservative” EOB dynamics, and (ii) pushing the calculation of the EOB evolution down to its point of unreliability (near the last [*unstable*]{} orbit) where it is matched to a perturbed-single-black-hole description. A zeroth approximation to this completed EOB approach is the one we study in this paper: an adiabatic sequence of solutions of the conservative dynamics, terminated at the LSO. In this approach one entirely neglects the losses of energy and angular momentum during the plunge phase following the crossing of the LSO. The numbers recalled above show that the energy loss during the plunge (and the ring-down) is not negligible compared to the binding energy at the LSO. However, for the present question this is not a problem. What is important is that the angular momentum loss during plunge is a very small fraction (a percent or so) of the angular momentum at the LSO, and that the final mass of the black hole is nearly equal to $M = m_1 + m_2$. This leads us to the following zeroth order estimate of the spin parameter of the final hole: \[eq3.44\] \_[BH]{} . In view of the exact conservation of $\mbox{\boldmath$J$}$ in our conservative EOB (real) dynamics, it is clear that it is $\vert \mbox{\boldmath$J$} \vert^{\rm LSSO}$ which is a good measure of the total angular momentum of the final spacetime, i.e. of the final black hole.
We are facing here a potential consistency problem of this simple-minded EOB treatment: when computing (\[eq3.44\]) for spinning configurations does one always get $\widehat{a}_{\rm BH} < 1$? One might worry that, starting with a value of $\widehat{a}_{\rm BH} \simeq 0.80$ for non-spinning holes, the addition of large spins on the holes might quickly exceed the extremal limit. It is plausible that the most dangerous situation is the “aligned case”, where all the angular momenta, $\mbox{\boldmath$L$}$, $\mbox{\boldmath$S$}_1$ and $\mbox{\boldmath$S$}_2$ are parallel (or antiparallel). In this case the numerator of Eq. (\[eq3.44\]) reads \[eq3.45\] J\^[LSSO]{} = L\_z\^[LSSO]{} + S\_1 + S\_2 , while the spin parameter of the effective metric reads \[eq3.46\] a = \_p = = ( X\_1\^2 + ) \_1 + ( X\_2\^2 + ) \_2 . Here, we consider $S_1$, $S_2$ and $\widehat{a}_1 \equiv S_1 / m_1^2$, $\widehat{a}_2
\equiv S_2 / m_2^2$ as algebraic numbers (positive or negative). This allows us to investigate also the case where the spins might be antiparallel to $\mbox{\boldmath$k$}$. For simplicity, we shall only study the symmetric case where $m_1 = m_2$ and $S_1 =
S_2$. For this case \[eq3.47\] a = \_p = \_1 , and \[eq3.48\] \_[BH]{} = = \_z\^[ LSSO]{} + \_1 = \_z\^[LSSO]{} + \_p where the dimensionless orbital angular momentum $\widehat{L}_z \equiv L_z / \mu \, M$ is related to the dimensionless quantity (when $\cos \theta_{\rm LS} = 1$) $\bar{\ell}
\equiv \sqrt{\cal K} / \mu \, M = \bar{L}_z / \mu \, M$ appearing in (\[eq3.39\]) through \[eq3.49\] \_z = | + . It is interesting to note that, even in the case where both holes are extreme ($\widehat{a}_1 = \widehat{a}_2 = 1)$ the maximum value of the effective spin parameter is $\widehat{a}_{\rm max} = \frac{7}{8} < 1$. We have numerically investigated the quantity $\widehat{a}_{\rm BH}$, Eq. (\[eq3.48\]), as a function of the effective $\widehat{a} = \widehat{a}_p$. The result is plotted in Fig. \[Fig2\] for different values of the 3 PN parameter $\omega_s$.
We see that the final spin parameter reaches a maximum for a positive value of $\widehat{a}_p$, i.e. for parallel (rather than antiparallel) spins. For the correct 3PN value $\omega_s = 0$ the maximum value of $\widehat{a}_{\rm BH}$ is comfortably below 1: namely, $\widehat{a}_{\rm BH}^{\rm max} \simeq 0.87$, reached for $\widehat{a}_p
\simeq + \, 0.3$. This is not much larger than the value $\widehat{a}_{\rm BH} \simeq
0.82$ obtained for $\widehat{a}_p =
0$. We find that this is a nice sign of the consistency of the EOB approach. This consistency was not a priori evident. In fact for $\omega_s \leq -9$ one gets a maximum value of $\widehat{a}_{\rm BH}$ slightly larger than 1. In particular, note that the 2 PN treatment of the orbital dynamics (obtained for $\omega_s = \omega_s^* \simeq -
9.3439$; upper curve in Fig. \[Fig2\]) formally leads to problematic over-extreme values of $\widehat{a}_{\rm BH}$. This may be interpreted as a confirmation of the need of “repulsive” 3 PN effects (i.e. $\omega_s + 9 \gg 1$). It is (a posteriori!) easy to understand physically why $\widehat{a}_{\rm BH}$, after reaching a maximum, then decreases when one adds more spin on the two black holes. Indeed, there is here a competition between two effects: adding spin on the holes (i.e. increasing $\widehat{a}_p$), on the one hand directly contributes to augmenting $\widehat{a}_{\rm BH}$ through the second term of the RHS of Eq. (\[eq3.48\]), but, on the other hand, indirectly contributes to reducing the total $J^{\rm LSSO}$ by reducing $\widehat{L}_z^{\rm LSSO}$ (indeed, as we explained above, positive spin leads to an LSO orbit closer to the horizon, and therefore with less orbital angular momentum). The first effect wins for smallish spins, while the second (more non-linear) effect wins for larger spins.
Conclusions {#sec4}
===========
We started by recalling the need of techniques for accelerating the convergence of the post-Newtonian (PN) expansions in the last stages of the inspiral of binary systems. We summarized the evidence (Table \[Tab1\]) showing the remarkable convergence properties of the best current resummation technique: the effective one-body (EOB) approach of Refs. [@BD99; @DJS2]. We showed how to generalize the EOB approach to the case of two spinning black holes with comparable masses $(\nu = \mu / M \sim 1/4)$. As a first step towards computing the spin-dependent EOB Hamiltonian we constructed an effective metric, which can be viewed either as a $\nu$-deformation of the Kerr metric or as a spin-deformation of the $\nu$-deformed effective metric. The effective spin entering this deformed Kerr metric is $M \,
\mbox{\boldmath$a$} \equiv \mbox{\boldmath$S$}_{\rm eff} \equiv \left( 1 + \frac{3}{4} \,
\frac{m_2}{m_1} \right) \, \mbox{\boldmath$S$}_1 + \left( 1 + \frac{3}{4} \,
\frac{m_1}{m_2} \right) \, \mbox{\boldmath$S$}_2$. The introduction of this effective $\mbox{\boldmath$a$}$ allows one to combine in a simple manner all (PN leading) spin-orbit coupling effects, and most of the spin-spin ones, with the rather complex but important 3 PN effects, which have been incorporated only recently in the EOB approach [@DJS2].
We have also constructed a more complicated modified effective Hamiltonian, Eq. (\[eq2.56\]), which separately depends on two (effective) spin vectors, $M
\mbox{\boldmath$a$}_0 \equiv \mbox{\boldmath$S$}_0 \equiv \left(1 + \frac{m_2}{m_1}
\right) \,
\mbox{\boldmath$S$}_1 + \left(1 + \frac{m_1}{m_2} \right) \mbox{\boldmath$S$}_2$, and $\mbox{\boldmath$\sigma$} \equiv \mbox{\boldmath$S$}_{\rm eff} - \mbox{\boldmath$S$}_0
\equiv - \frac{1}{4} \left( \frac{m_2}{m_1} \ \mbox{\boldmath$S$}_1 + \frac{m_1}{m_2} \
\mbox{\boldmath$S$}_2 \right)$, and which allows a (hopefully) more accurate representation of spin-spin effects. We recommend the simultaneous consideration of $H_{\rm eff}$ and $H'_{\rm eff}$ to determine the domain of trustability of the presently constructed spin-dependent EOB Hamiltonian. Namely, when $H_{\rm real} = \bar f (H_{\rm eff})$ and $H'_{\rm real} = \bar f (H'_{\rm
eff})$ lead to numerically very similar evolutions, one is entitled to trusting them both; while a significant difference in their predictions signals a breakdown of the trustability of the simple EOB Hamiltonian proposed here.
The present paper has only investigated a few aspects of the physics predicted by our spin-generalized EOB approach. In particular, as a first cut toward understanding the relevance of our construction for gravitational wave (GW) observations we have discussed the approximate existence of “spherical orbits” (orbits with fixed radial coordinate, as in the Kerr metric) and we studied the binding energy of the last stable spherical orbits (LSSO). A message of this study is that, for most physically relevant cases (in the parameter space where one randomly varies all angles and all spin values), the results are only weakly dependent on the exact numerical values of the 3PN coefficients. Moreover, they exhibit moderate deviations from the non-spinning case (see Fig. 1 and Table II). To give a numerical flavor of the effects of spin we note that, when the projected spin parameter $\widehat{a}_p = \mbox{\boldmath$k$} \cdot \mbox{\boldmath$a$} / M$, Eq. (\[eq3.41\]), is smaller than about $+ \, 0.2$, its effect on the fractional binding energy ($e \equiv (E_{\rm real} - M) / M$) of the LSSO is, approximately, \[eq4.1\] 100 e\^[LSSO]{} - 1.43 \_4 (1 + 0.168 \_4) - 0.806 \_4 (1 + 0.888 \_4) \_p , where $\nu_4 \equiv 4 \, \nu \equiv 4 \, m_1 \, m_2 / (m_1 + m_2)^2 \leq 1$. As in most cases (random angles, random spin-kinetic energies) it is plausible that $\vert
\widehat{a}_p \vert \, \laq \, a_p^{\rm rms} \sim 0.25$, we expect that spin effects will only modify the energy emitted as gravitational waves up to the LSSO by less than about $0.6\%
\, \nu_4 \,
M$.
Such an increase, though modest, is still a significant fractional modification of the corresponding energy loss predicted for non-spinning systems ($e_0 = - 1.67\% \, M$ for $\nu_4 = 1$). In fact, this effect might cause an important [*bias*]{} in the first observations. If the intrinsic spins of the holes can (at all) take large values, the highest signal-to-noise-ratio events in the first years of LIGO observations might select binary systems with rather large and rather aligned spins. It is therefore important to include spin effects in the data analysis of coalescing black holes. We have argued that, in most cases, the simple-minded generalized EOB approach presented here should be a reliable analytical tool for describing the dynamics of two spinning holes and for computing a catalogue of gravitational waveforms, to be used as matched filters in the detection of GW’s. However, it must be admitted that, in the cases where the effective spin vector is coarsely (positively) aligned with the orbital angular momentum, and where the spins are so large that $\widehat a \,
\gaq \, 0.4$, the predictions from the above-introduced EOB Hamiltonian start predicting LSSO radii so near the “effective horizon” where $\Delta_t (r) = 0$, that they cannot be quantitatively relied upon. \[Though I would still argue that they can be qualitatively trusted, in view of the simple physics they use; see subsection \[ssec3.3\].\] We give some examples of that in Table \[Tab2\]. In such cases the EOB approach does predict much larger energy losses, possibly larger than $10\% \, M$. In these cases, the uncertainty in the waveform may be so large that one may need the type of non-linear filtering search algorithm advocated in Ref. [@FH98]. We wish, however, to emphasize the differences between our treatment and conclusions and those of Flanagan and Hughes. These authors defined the “merger” phase as (essentially) what comes after the binary system crosses the non-spinning LSO (around 6 $GM$), and they assumed that the signal from the “merger” phase can only be obtained from numerical relativity. Moreover, they optimistically assumed that (in all cases) $10\% \, M$ are emitted in GW energy during the merger phase, and 3% during the subsequent ring-down phase. By contrast, our treatment is based on the idea that a suitable resummed version of the PN-expanded dynamics, namely the EOB-plus-Padé approach, can, in most cases, give an analytical handle on the computation of the inspiral signal down to the spin-modified LSSO (and even during the subsequent plunge, as discussed for the spinless case in [@BD00]). We have argued in several ways that the simple EOB Hamiltonian (\[eq2.44\]) gives reliable answers in most cases, and allows one to analytically control the possible amplification (or deamplification, when $\widehat{a}_p < 0$) in GW energy loss due to spin effects. Moreover, it is only in rather extreme cases that we could agree with [@FH98] in predicting $\gaq \, 10\% \, M$ energy losses. In most other cases, we think that the EOB method provides a reliable basis for computing families of wave forms that will be useful templates for the detection of GW’s. Another difference with [@FH98] is that we have argued, on the basis of definite computations, that the spin of the final hole will never become nearly extremal (even if the initial spins are extremal). This is important for the data analysis of the ring-down signal, because the decay time of the least damped quasi-normal-mode starts becoming large only for near extremal holes.
Let us emphasize that the present work is only a first step toward an improved analytical understanding of the last stages of inspiral motion of two spinning compact object[^13]. The explicit spin-dependent Hamiltonian (\[eq2.44\]) (or, better, $H'_{\rm real}$ defined at the end of Section II) has only taken into account the leading effects (in a PN expansion) of the spin-dependent interactions. More work is needed to analytically determine more accurate versions of the EOB Hamiltonian. In particular, it would be interesting to explicitly derive the next-to-leading ( 1 PN) corrections to the EOB spin-orbit (and spin-spin) interactions. Furthermore, we have only provided a resummation of the conservative part of the dynamics. There remains the important complementary task of resumming the radiation reaction part. This was done in [@BD00], using previous results of [@DIS98], only for the spinless case.
Once this is done, we expect, as in our previous study [@BD00], that the presence of a LSSO along the sequence of adiabatic orbits will be blurred and will be replaced by a continuous transition between inspiral and plunge. There remains also the task of studying the effects of spin-dependent interactions on the gravitational waveform emitted during the last stages of inspiral and during the plunge (that we have not considered here). In other words, one needs to redo, by combining the EOB approach with resummed versions of radiation reaction, the studies, valid far from the LSO, which were based on straightforward PN-expanded results [@kidder], [@apostolatos]. Note that our result above about the primary importance of the single parameter $\widehat{a}_p$, combined with the understanding [@DIS00] that the number of “useful” cycles in the GW signal for massive binaries is rather small, suggests that a rather small number of “spinning templates” will be really needed in a matched filter data analysis. On the other hand, we recall that it was found in [@DIS00] that the plunge signal (but not the ring-down one, for stellar mass holes) plays a significant role in the data analysis.
It will also be interesting to see, within the EOB approach, the extent to which the non-linear spin-dependent interactions might, as has been recently suggested [@levine], lead to a chaotic dynamical evolution. We a priori suspect that two factors will diminish the significance of such chaotic evolutions: (1) they occur only in an improbably small region of phase space (involving, in particular, large spins), and (2) their effect on the crucial GW phasing is rather small.
It would be very useful to have independent means of testing the accuracy of the EOB approach. At this stage we see only three ways of doing that (beyond the performance of more internal checks of the robustness of the approach): (i) an analytical calculation of the 4 PN interaction Hamiltonian, (ii) a comparison between numerical computations and the EOB results, and/or (iii) a comparison between the EOB predictions and the forthcoming GW observations. (i) would be important for assessing the convergence of the PN-resummed EOB Hamiltonian. In view of the extreme difficulties involved in the 3 PN calculations [@JS98; @JS99; @DJS1; @DJS3; @BF1; @BF2; @BF3] it would seem hopeless to even mention the 4 PN level. But in fact, the EOB approach itself suggests that the current methods used in PN calculations are highly inefficient, and unnecessarily complicated. Indeed, as emphasized in [@DJS2] the final, gauge-invariant content of the 3 PN result is contained in only three quantities $a_4$, $b_3$ and $z_3$, and only one of them, $a_4 (\nu)$, is really important for determining the dynamics of inspiralling quasi-circular binaries. If one could invent a new approximation scheme which computes directly $a_4$ (at 3 PN), it might be possible to compute its 4 PN counterpart, $a_5
(\nu)$. (ii) is not yet possible (at least as a test of the EOB Hamiltonian) because numerical computations use as initial data geometrical configurations that do not take into account most of the crucial physics incorporated in PN calculations. Current numerical computations use somewhat ad hoc “binary-black-hole-like” data, often of the restricted spatially conformally flat type, without trying to match their initial data to the near LSO configurations predicted by (resummed) analytical approaches. On the other hand, let us stress that the value of the radial PN-expanded potential $A(\widehat{u}) = 1 - 2 \, \widehat{u} + a_3 (\nu) \,
\widehat{u}^3 + a_4 (\nu) \, \widehat{u}^4 +
\cdots$ crucially depends on the non-linear gravitational interactions linked to the $h_{ij}^{TT}$ part of the spatial metric, i.e. to its non conformally flat part, and also to the non-linear interactions linked to the ${\pi}^{ij}_{TT}$ part of the gravitational field momenta. For instance already at the 2 PN level, the truncation of the Einsteinian prediction for the two-body problem (driven into a close orbit by a long past interaction involving retarded GW interactions) corresponding to artificially assuming a conformally flat spatial metric changes the physically correct value $a_3^{2 \, {\rm PN}} (\nu) = 2 \nu$ into $a^{\rm
conf. \, flat}_3 (\nu) = \frac{1}{4} \, (18 - 5 \nu) \, \nu$ [@DJS2]. \[It also slightly changes the energy map $f$.\] For equal-mass systems, this corresponds to multiplying the positive $a_3 (\nu)$ by a factor $+
2.09375$. As we discussed above, this (artificial) increase of the “repulsive” character of the non-linear gravitational interactions tends to artificially increase the binding of the LSO. As the 2PN coefficient is anyway too small to have a large impact on the LSO characteristics, this 2PN change does not, by itself, change much the LSO energy [@DJS2]. However, if this tendency to increase the “repulsive” character of the PN expansion (caused by the neglect of the $h_{ij}^{TT}$-, and ${\pi}^{ij}_{TT}$-dependent interactions) persists at the (numerically more important) 3PN level, this might explain the current discrepancy between analytical and numerical estimates of LSO characteristics. In fact, we note that the initial data taken by a recent attempt [@baker] at fulfilling the proposal of [@BD99] to start a full numerical calculation only at the moment where it is really needed, i.e. after crossing the LSO, uses LSO initial data [@baumgarte] with a binding energy $e_{\rm LSO} = E / (m_1 +
m_2) - 1 \simeq -
2.3\%$ which is 38% larger than the value $e_{\rm LSO}^{3 \, {\rm PN}} \simeq -
1.67\%$ obtained at 3 PN (with $\omega_s = 0$) by analytical estimates. Similarly, the LSO orbital period of the initial data of [@baker] is $T_{\rm LSO} \simeq 35 \, (m_1 +
m_2)$ [@baumgarte], which is twice smaller than the 3 PN estimate $T_{\rm LSO}^{3 \, {\rm PN}} \simeq 71.2 \, (m_1 + m_2)$ [@DJS2]! These discrepancies between state-of-the-art numerical LSO initial data and state-of-the-art analytical estimates of LSO data are significantly larger than the natural “theoretical error bar” on the (resummed) analytical estimates (derived, say, by comparing 2PN estimates to 3PN ones). \[See, however, the new numerical approach of [@GGB] whose LSO data agree well with the 3 PN EOB estimates [@DGG].\] In our opinion, this makes it urgent for the numerical relativity community to develop ways of constructing initial data that correctly incorporates the crucial non-linear physics (linked to the $h_{ij}^{TT}$ and ${\pi}^{ij}_{TT}$ parts of the metric) which is taken into account in PN calculations. If a significant discrepancy remains after this is done, one will be entitled to blame the lack of convergence of the EOB-resummed PN calculations. If one finds agreement, this will be a confirmation of the claim made here that the Padé-improved EOB is a reliable description of the last orbits before coalescence. Once one succeeds in matching analytical and numerical results for non-spinning black holes, it will be very interesting to use numerical data on fast-spinning black holes to refine the EOB Hamiltonian by fitting the values of the extra parameters which can be introduced in the EOB Hamiltonian to represent higher PN effects. In the long term we expect that such a complementarity between numerical and analytical tools will be needed for defining a sufficiently dense set of GW templates. \[In view of the large dimensionality of the parameter space of the two spinning hole system, it seems hopeless to use only numerical techniques to define a dense network of templates.\]
Finally, even if no decisive progress is made on (i) or (ii) before the first sources are detected, there remains the possibility that the first observations might confirm the soundness of (or suggest specific modifications of) the EOB-based waveforms, and thereby facilitate further detections by narrowing the bank of templates. For instance, one might include a 4PN contribution $ + a_5 (\nu) u^5$ to $A(u)$, as a free parameter in constructing a bank of templates, and wait until LIGO/VIRGO/GEO get high signal-to-noise-ratio observations of massive coalescing binaries to determine its numerical value.
It is a pleasure to thank David Blair for a question which prompted this work, and the University of Western Australia for hospitality during its inception.
V.M. Lipunov, K.A. Postnov and M.E. Prokhorov, New Astron. [**2**]{}, 43 (1997). E.E. Flanagan and S.A. Hughes, Phys. Rev. [**D57**]{}, 4535 (1998). P.R. Brady, J.D.E. Creighton and K.S. Thorne, Phys. Rev. [**D58**]{}, 061501 (1998). S.F. Portegies Zwart and S.L. McMillan, Astrophys. J. Letters, in press, astro-ph/9910061. T. Damour, B.R. Iyer and B.S. Sathyaprakash, Phys. Rev. [**D62**]{}, 084036 (2000); and Phys. Rev. [**D63**]{}, 044023 (2001). C. Cutler et al., Phys. Rev. Lett. [**70**]{}, 2984 (1993). T. Damour, B.R. Iyer and B.S. Sathyaprakash, Phys. Rev. [**D57**]{}, 885 (1998). A. Buonanno and T. Damour, Phys. Rev. [**D59**]{}, 084006 (1999). A. Buonanno and T. Damour, Phys. Rev. [**D62**]{}, 064015 (2000). T. Damour, P. Jaranowski and G. Schäfer, Phys. Rev. [**D62**]{}, 084011 (2000). A. Buonanno and T. Damour, contributed paper to the IX$^{\rm th}$ Marcel Grossmann Meeting (Rome, July 2000); gr-qc/0011052. P. Jaranowski and G. Schäfer, Phys. Rev. [**D57**]{}, 7274 (1998). P. Jaranowski and G. Schäfer, Phys. Rev. [**D60**]{}, 124003 (1999). T. Damour, P. Jaranowski and G. Schäfer, Phys. Rev. [**D62**]{}, 044024 (2000). T. Damour, P. Jaranowski and G. Schäfer, Phys. Rev. [**D62**]{}, 021501 (R) (2000). L. Blanchet and G. Faye, Phys. Lett. [**A271**]{}, 58 (2000). L. Blanchet and G. Faye, gr-qc/0006100. L. Blanchet and G. Faye, gr-qc/0007051. T. Damour, P. Jaranowski and G. Schäfer, Phys. Rev. [**D63**]{}, 044021 (2001). V.C. de Andrade, L. Blanchet and G. Faye, Class. Quant. Grav. [**18**]{}, 753 (2001);gr-qc/0011063. T. Damour, P. Jaranowski and G. Schäfer, Phys. Lett. B [**513**]{}, 147 (2001). T. Damour, in [*Gravitational Radiation*]{}, edited by N. Deruelle et T. Piran (North-Holland, Amsterdam, 1983), pp. 59-144. L. Blanchet, T. Damour, B.R. Iyer, C.M. Will and A.G. Wiseman, Phys. Rev. Lett. [**74**]{}, 3515 (1995); L. Blanchet, T. Damour and B.R. Iyer, Phys. Rev. [**D51**]{}, 5360 (1995); C.M. Will and A.G. Wiseman, Phys. Rev. [**D54**]{}, 4813 (1996); L. Blanchet, Phys. Rev. [**D54**]{}, 1417 (1996). L. Blanchet, Class. Quantum Grav. [**15**]{}, 113 (1998); L. Blanchet, B.R. Iyer and B. Joguet, gr-qc/0105098. L.E. Kidder, C.M. Will and A.G. Wiseman, Class. Quantum Grav. [**9**]{}, L127 (1992); Phys. Rev. [**D47**]{}, 3281 (1993). N. Wex and G. Schäfer, Class. Quantum Grav. [**10**]{}, 2729 (1993). G. Schäfer and N. Wex, in [*XIIIth Moriond Workshop: Perspectives in neutrinos, atomic physics and gravitation*]{} (ed. by J. Trân Thanh Vân, T. Damour, E. Hinds and J. Wilkerson, Editions Frontières, Gif-sur-Yvette, 1993), pp. 513. E. Brézin, C. Itzykson and J. Zinn-Justin, Phys. Rev. [**D1**]{}, 2349 (1970). I.T. Todorov, Phys. Rev. [**D3**]{}, 2351 (1971); V.A. Rizov, I.T. Todorov and B.L. Aneva, Nucl. Phys. B [**98**]{}, 447 (1975). A. Buonanno, Phys. Rev. [**D62**]{}, 104022 (2000). A. Maheshwari, E.R. Nissimov and I.T. Todorov, Lett. Math. Phys. [**5**]{}, 359 (1981). P.P. Fiziev and I.T. Todorov, Phys. Rev. [**D63**]{}, 104007 (2001); gr-qc/0010104. T. Damour, B.R. Iyer and B.S. Sathyaprakash, second reference in [@DIS00]. W. Tulczyjew, Acta Phys. Pol. [**18**]{}, 37 (1959), and Errata in the same volume. L. Infeld and J. Plebanski, [*Motion and Relativity*]{} (Pergamon, Oxford, 1960). T. Damour in [*Physics and Astrophysics of Neutron Stars and Black Holes*]{}, Proceedings of the International School of Physics “Enrico Fermi”, Varenna, Italy, 1975, edited by R. Giacconi and R. Ruffini, Enrico Fermi Course LXV (North-Holland, Amsterdam, 1978), pp. 547-549. T. Damour, C. R. Acad. Sci. Paris [**294**]{}, série II, 1355 (1982). J. Ibañez and J. Martin, Gen. Rel. Grav. [**14**]{}, 439 (1982). P.D. D’Eath, Phys. Rev. [**D12**]{}, 2183 (1975). K.S. Thorne and J.B. Hartle, Phys. Rev. [**D31**]{}, 1815 (1985). T. Damour, M. Soffel and C. Xu, Phys. Rev. [**D43**]{}, 3273 (1991); and Phys. Rev. [**D45**]{}, 1017 (1992). H. Tagoshi, A. Ohashi and B. J. Owen, Phys. Rev. D [**63**]{}, 044006 (2001); gr-qc/0010014. M. Mathisson, Acta Phys. Polon. [**6**]{}, 163 (1937). A. Papapetrou, Proc. R. Soc. London [**A209**]{}, 248 (1951). A.J. Hanson and T. Regge, Annals Phys. [**87**]{}, 498 (1974). V. Berestetski, E. Lifchitz and L. Pitayevski, [*Théorie Quantique Relativiste, Première Partie*]{} (Mir, Moscow, 1972). B.M. Barker, S.N. Gupta and R.D. Haracz, Phys. Rev. [**149**]{}, 1027 (1966). B.M. Barker and R.F. O’Connell, Phys. Rev. [**D2**]{}, 1428 (1970). B.M. Barker and R.F. O’Connell, Phys. Rev. [**D12**]{}, 329 (1975). B.M. Barker and R.F. O’Connell, Gen. Rel. Grav. [**11**]{}, 149 (1979). B.M. Barker and R.F. O’Connell, Gen. Rel. Grav. [**5**]{}, 539 (1974). T. Damour and G. Schäfer, Nuov. Cim. [**101B**]{}, 127 (1988). B. Carter, Phys. Rev. [**174**]{}, 1559 (1968); and in [*Black Holes*]{}, ed. by C. and B.S. De Witt (Gordon and Breach, New York, 1973), pp. 58-214. K.S. Thorne, Rev. Mod. Phys. [**52**]{}, 299 (1980); see Eq. (11.28a) there. D.C. Wilkins, Phys. Rev. [**D5**]{}, 814 (1972). L.E. Kidder, C.M. Will and A.G. Wiseman, Phys. Rev. [**D47**]{}, R 4183 (1993); L.E. Kidder, Phys. Rev. [**D52**]{}, 821 (1995). T.A. Apostolatos, C. Cutler, G.J. Sussman and K.S. Thorne, Phys. Rev. [**D15**]{}, 6274 (1994); T.A. Apostolatos, Phys. Rev. [**D54**]{}, 2438 (1996). J. Levin, Phys. Rev. Lett. [**84**]{}, 3515 (2000); and gr-qc/0010100. J. Baker et al., gr-qc/0102037. T.W. Baumgarte, Phys. Rev. [**D62**]{}, 024018 (2000). E. Gourgoulhon, P. Grandclément, and S. Bonazzola, gr-qc/0106015; P. Grandclément, E. Gourgoulhon and S. Bonazzola, gr-qc/0106016. T. Damour, E. Gourgoulhon and P. Grandclément, in preparation.
[^1]: We expect real ambiguities to arise only at the $v^{10} / c^{10} \sim$ 5 PN level, because this corresponds to the [*square*]{} of the leading, $v^5 / c^5 \sim$ 2.5 PN, radiation reaction terms.
[^2]: Henceforth, ‘3 PN’ will mean the conservative 3 PN dynamics, i.e. $N + 1 \, {\rm PN} + 2 \, {\rm PN} + 3 \, {\rm PN}$.
[^3]: Note that a subtlety arises at 3 PN [@JS98] in that the Hamilton action principle involves [*derivatives*]{} of the phase space variables. However, it was shown in [@DJS1] how to reduce the problem to an ordinary Hamiltonian dynamics by means of a suitable ${\cal O} (v^6 / c^6)$ shift of phase-space variables. We henceforth assume that we work with the shifted variables defined in [@DJS1].
[^4]: We recall that the high-order perturbative, PN-expanded, calculations of the dynamics of two non-spinning compact objects model these objects by delta-function (monopole) sources. The supports of these delta-functions define the coordinate “positions” of the compact objects. As explained in [@D83] these “positions” physically correspond to some “centers of the gravitational field” generated by the objects.
[^5]: We are making this very explicit because some people, when they hear about the EOB approach, think that the effective metric describing the one-body dynamics should, at some level of approximation, include some Kerr-like features to model the velocity-dependent two-body interactions. This is not true for the orbital dynamics, whatever be the PN accuracy level. On the other hand, we shall see that we need Kerr-like features to accomodate the intrinsic spin effects.
[^6]: In fact, [@DJS2] found that it was possible to map the real dynamics onto the geodesic dynamics of a test particle. However, both the effective metric and the modified energy map needed for this representation are rather complicated. It was felt that it is more convincing to keep a simple effective metric, and a simple energy map, but to relax the constraint of [*geodesic*]{} motion.
[^7]: We use here the fact that the real phase-space coordinates $\mbox{\boldmath$x$}^{\rm real}$, $\mbox{\boldmath$p$}^{\rm
real}$ differ only by ${\cal O} (1/c^2)$ from the effective ones entering $H_{\rm eff}$ [@BD99].
[^8]: We leave untouched the dependence on $a$ to ensure that, when $GM \rightarrow 0$ with $a$ being fixed, the metric $g_{\rm eff}^{\alpha\beta}$ be Minkowski in disguise.
[^9]: For instance, we can assume Eq. (\[eq3.11\]) for $H_{\rm eff}$, but make no further approximation in computing $H_{\rm real} = f (H_{\rm
eff})$.
[^10]: This value corresponds to taking $\lambda
= 0$ in $\omega_s = - \, 11 \lambda / 3 - 1987/840$. Here, $\lambda$ denotes the natural ambiguity parameter entering the Blanchet-Faye framework. Note that the authors of Refs. [@BF1; @BF2; @BF3] do not claim that $\lambda = 0$ is a preferred value. However, as $\lambda$ is expected to be of order unity we use $\lambda = 0$, i.e. $\omega_s = - \, 1987/840$ as a fiducial deviation from $\omega_s = 0$.
[^11]: Actually, as far as we know, the Kerr limit of Eq. (\[eq3.39\]) has never been written down before. Usual treatments [@wilkins] use the more complicated effective radial potential $W'_a (r ,
L_z , {\cal Q})$. Anyway, the physics is the same, but it is more cleanly presented in (\[eq3.39\]).
[^12]: Remember that we Padé resum $A (\widehat u)$ and $\bar A (\widehat
u)$ to ensure that these functions qualitatively behave like $1 - 2 \, \widehat{u}$ or $1 - 2 \, \widehat{u} + \widehat{a}^2 \, \widehat{u}^2$ (for $\widehat{a}^2 < 1$), i.e. (generically) have a simple zero near $\widehat{r} = \widehat{u}^{-1} = 1 + \sqrt{1 -
\widehat{a}^2}$, which means that the effective metric becomes “infinitely attractive” at some deformed horizon.
[^13]: Though, in most of the paper we only spoke of binary black holes, it should be clarified that our EOB Hamiltonian also applies to binary spinning neutron stars or to spinning neutron-star-black-hole systems, at least down to the stage where the quadrupole deformation of the neutron star becomes significant.
|
{
"pile_set_name": "ArXiv"
}
|
---
address: |
Max-Plank-Institut für Physik, D-80805 Munich, Germany\
E-mail: [email protected]
author:
- 'S. KLUTH'
title: 'REVIEW OF [$\alpha_{\mathrm{S}}$]{} MEASUREMENTS'
---
Introduction {#sec_intro}
============
QCD is a highly predictive theory, because all processes involving strong interactions must be described by QCD with a universal parameter, the strong coupling constant [$\alpha_{\mathrm{S}}$]{}. Vice versa, all determinations of [$\alpha_{\mathrm{S}}$]{} using different processes must yield the same results once different momentum scales or renormalisation schemes are taken into account. A systematic comparison of measurements of [$\alpha_{\mathrm{S}}$]{} thus is a strong test of the theory.
Averages of different measurements of [$\alpha_{\mathrm{S}}$]{} are calculated with proper treatment of correlations from common uncertainties [@kluth06]. Uncertainties are classified as statistical, experimental, soft QCD or hard QCD. Statistical and experimental uncertainties stem from limited data samples and experimental systematics. Soft QCD uncertainties stem from hadronisation correction systematics, higher twist effects, influence of parton density functions (pdfs) and other non-perturbative effects. Hard QCD uncertainties arise from unknown or incomplete higher order corrections.
All results, intermediate averages and the final average are shown in table \[tab\_as\]. The intermediate averages of [$\alpha_{\mathrm{S}}$]{}from related analyses are found assuming uncorrelated statistical errors, partially correlated experimental and soft QCD errors and fully correlated hard QCD errors [@kluth06]. Results with total error on [${\ensuremath{\alpha_{\mathrm{S}}}}({\ensuremath{m_{{\ensuremath{\mathrm{Z^0}}}}}})$]{}significantly larger than 0.01 have not been considered.
Lattice QCD
===========
New implementations of unquenched lattice QCD (LQCD) with dynamical staggered light quarks (u, d and s) improve significantly the description of some low energy observables [@davies03] after tuning the simulation with precisely known hadron masses and mass differences. Due to quark staggering quark vacuum polarisation loops contribute 4-fold and the procedure has to be modified by hand to compensate this effect. The tuned LQCD is used to predict 28 selected short distance observables which in turn are compared with NNLO QCD calculations to extract [$\alpha_{\mathrm{S}}$]{} [@mason05]. The uncertainties of the result are dominated by limited simulation statistics and systematic uncertainties of the analysis. With this measurement a 1% accuracy for a determination of [${\ensuremath{\alpha_{\mathrm{S}}}}({\ensuremath{m_{{\ensuremath{\mathrm{Z^0}}}}}})$]{} is reached.
DIS Processes
=============
The analyses of scaling violation of structure functions (SFs) in deep inelastic scattering (DIS) of leptons on nucleons result in precision measurements of [$\alpha_{\mathrm{S}}$]{}. The scaling violation of the SF [$F_2^{\mathrm{ep}}$]{} in e-p DIS was studied in moment space in NNLO QCD but lacks a full analysis of the theoretical error [@santiago01]. Following [@bethke02] the theory error is doubled. The SF [$F_3^{\nu}$]{} for neutrino-nucleon DIS was analysed using Mellin moments and Jacobi polynomials in NNLO QCD [@kataev01]. The SF [$g_1^{\mathrm{N}}$]{} for DIS with polarised nucleons N was analysed in NLO QCD [@bluemlein02].
QCD predictions for some sum rules (SRs) in DIS are available in NNLO. The study of the Bjorken SR for polarised DIS [@ellis95] yields a precise value of [$\alpha_{\mathrm{S}}$]{} but the analysis of the experimental error has been criticised [@altarelli96; @knauf02]; we double this error for our averages. The GLS SR for $\nu$-N DIS was studied using CCFR data [@kim98].
The more recent determinations of [$\alpha_{\mathrm{S}}$]{} from jet production in e-p DIS (J. Terron, these proceedings) are NLO QCD analyses covering a wide range of $Q^2$ values. The comprehensive combination of HERA results for this process is used [@glasman05].
Y Decays
========
The Y resonances are [$\mathrm{b\bar{b}}$]{} bound systems with mass dominated by the large b quark mass. Properties of these systems are predicted with non-relativistic QCD (NRQCD) which takes the low velocities of the heavy quarks as an additional expansion parameter. Moments of $R_b(s={\ensuremath{m_{\mathrm{Y}}}}^2)=\sigma({\ensuremath{\mathrm{e^+e^-}}}\rightarrow{\ensuremath{\mathrm{b\bar{b}}}})/
\sigma({\ensuremath{\mathrm{e^+e^-}}}\rightarrow{\ensuremath{\mu^+\mu^-}})$ [@penin98] and the branching ratio ${\ensuremath{R_{\mathrm{Y\rightarrow\gamma gg}}}}=\Gamma(\mathrm{Y}\rightarrow\gamma+\mathrm{hadrons})/
\Gamma(\mathrm{Y}\rightarrow{\ensuremath{\mu^+\mu^-}})$ [@pdg04] have been predicted in NRQCD and used to extract [$\alpha_{\mathrm{S}}$]{}.
[$R_{\tau}$]{}
==============
For hadronic decays of $\tau$ leptons the invariant mass $s$ of the hadronic final state sets the energy scale for QCD processes. The hadronic branching ratio ${\ensuremath{R_{\tau}}}(s)=\Gamma(\tau\rightarrow\nu_{\tau}\mathrm{hadrons})
/\Gamma(\tau\rightarrow\nu_{\tau}\nu_{\ell}\ell)$ is predicted in NNLO QCD while non-perturbative effects are treated with the operator product expansion (OPE). Using $\tau$ decay data from [@pdg04] the determination of [$\alpha_{\mathrm{S}}$]{} is updated [@kluth06]. A recent analysis using the partially calculated NNNLO term is consistent but claims smaller uncertainties [@davier05].
Z Lineshape
===========
The precise data collected by the LEP experiments and SLD around the [$\mathrm{Z^0}$]{} resonance yield an accurate determination of [${\ensuremath{\alpha_{\mathrm{S}}}}({\ensuremath{m_{{\ensuremath{\mathrm{Z^0}}}}}})$]{} via QCD corrections to electroweak processes. The analysis [@kluth06] with data from [@zedometry05] uses as observables the hadronic width $\Gamma_h$ and the hadronic branching ratio ${\ensuremath{R_{{\ensuremath{\mathrm{Z^0}}}}}}=\Gamma({\ensuremath{\mathrm{Z^0}}}\rightarrow\mathrm{hadrons})
/\Gamma({\ensuremath{\mathrm{Z^0}}}\rightarrow\ell\bar{\ell})$ of the [$\mathrm{Z^0}$]{} and the on-peak hadronic and leptonic cross sections. The result is consistent with [@zedometry05] and has a more complete error analysis.
[$F_2^{\gamma}$]{}
==================
The scaling violations of the SF [$F_2^{\gamma}$]{} for hadron production in two-photon interactions at [$\mathrm{e^+e^-}$]{} colliders has been studied with NLO QCD. With recent data from LEP at high and low $Q^2$ a stable result with small errors is obtained [@albino02].
[$R_{\mathrm{had}}$]{}
======================
The analysis of hadron production in [$\mathrm{e^+e^-}$]{} annihilation at low ${\ensuremath{\sqrt{s}}}<2$ GeV uses the observable ${\ensuremath{R_{\mathrm{had}}}}(s)=\sigma({\ensuremath{\mathrm{e^+e^-}}}\rightarrow\mathrm{hadrons})/
\sigma({\ensuremath{\mathrm{e^+e^-}}}\rightarrow{\ensuremath{\mu^+\mu^-}})$. The NNLO analysis [@menke01] yields a fairly accurate result with errors dominated by experimental uncertainties.
[$\mathrm{e^+e^-}$]{} Jets and Event Shapes
===========================================
The determination of [$\alpha_{\mathrm{S}}$]{} from jet rates and event shape distributions is reviewed by J. Schieck in these proceedings. The LEP experiments have coordinated their final event shape analyses via a working group (LEPQCDWG) yielding directly comparable consistent results [@kluth06]. The re-analysis of JADE data uses the methods developed at LEP and thus the results can also be compared directly [@kluth06]. All analyses including the older TOPAZ study [@topaznlla] are based on NLO QCD calculations combined with resummed NLLA calculations leading to more stable results. The availability of NLO QCD predictions combined with resummed NLLA calculations for the 4-jet fraction with the Durham or Cambridge jet algorithms lead to precise measurements of [$\alpha_{\mathrm{S}}$]{} [@jader4; @kluth06] from the LEP experiments and from JADE data.
[$\mathrm{e^+e^-}$]{} Fragmentation
===================================
In [@kluth06] data from LEP experiments were used to update the measurement of [$\alpha_{\mathrm{S}}$]{} from the cross section [$\sigma_{\mathrm{L}}$]{} for hadron production via a longitudinally polarised virtual [$\mathrm{Z^0}$]{} or $\gamma$. A combination of results for [$\alpha_{\mathrm{S}}$]{} from analyses of the scaling violation of charged hadron momentum spectra was done in [@kluth06]. The data are from LEP and lower energy experiments. Both measurements are based on NLO QCD.
pp/[$\mathrm{p\bar{p}}$]{} Scattering Processes
===============================================
Both results stem from analysis of [$\mathrm{p\bar{p}}$]{} collisions by the CERN S[$\mathrm{p\bar{p}}$]{}S collider experiments UA1 and UA6. The cross section for production of final states with b-jets is determined with a cut on the angle between the two b-jets [@ua1asbbbar95]. Due to this cut the cross section measurement becomes sensitive to [$\alpha_{\mathrm{S}}$]{}; a NLO QCD prediction is used to extract [$\alpha_{\mathrm{S}}$]{}.
The cross section difference $\sigma({\ensuremath{\mathrm{p\bar{p}}}}\rightarrow\gamma\mathrm{X})-
\sigma(\mathrm{pp\rightarrow\gamma X})$ is sensitive to the parton process ${\ensuremath{\mathrm{q\bar{q}}}}\rightarrow\gamma\mathrm{g}$. Together with DIS data to constrain the valence quark pdfs [$\alpha_{\mathrm{S}}$]{} can be determined [@werlen99].
ICHEP 2006 World Average
========================
The final ICHEP 2006 world average is calculated from the values for [${\ensuremath{\alpha_{\mathrm{S}}}}({\ensuremath{m_{{\ensuremath{\mathrm{Z^0}}}}}})$]{} shown in table \[tab\_as\] for each class of analyses. In case of several analyses in a class the intermediate average as shown in table \[tab\_as\] is used. The statistical, experimental and soft QCD errors are assumed to be uncorrelated and the hard QCD errors are assumed to be partially correlated. The intermediate and final averages are shown in figure \[fig\_as\]. The final average is dominated by the LQCD result with ${\ensuremath{\chi^2/\mathrm{d.o.f.}}}=17/9$ and $P(\chi^2)=0.05$. Without the LQCD result the average is ${\ensuremath{{\ensuremath{\alpha_{\mathrm{S}}}}({\ensuremath{m_{{\ensuremath{\mathrm{Z^0}}}}}})}}=0.1200\pm0.0019$ with ${\ensuremath{\chi^2/\mathrm{d.o.f.}}}=14/8$ and $P(\chi^2)=0.07$. Our average is consistent with other recent results [@pdg04; @bethke06]. The small values for the $\chi^2$ probabilities might indicate that the systematic errors in some of the analyses are estimated aggressively.
![ Intermediate averages of [${\ensuremath{\alpha_{\mathrm{S}}}}({\ensuremath{m_{{\ensuremath{\mathrm{Z^0}}}}}})$]{} from each class of analyses as shown in table \[tab\_as\]. The dashed vertical line and grey band indicate the final average with total errors. []{data-label="fig_as"}](asfinal.eps){width="\columnwidth"}
[10]{}
S. Kluth: Rept. Prog. Phys. [**69**]{} (2006) 1771
C.T.H. Davies et al.: Phys. Rev. Lett. [**92**]{} (2004) 022001
Q. Mason et al.: Phys. Rev. Lett. [**95**]{} (2005) 052002
J. Santiago, F.J. Yndurain: Nucl. Phys. B [**611**]{} (2001) 447
S. Bethke: Nucl. Phys. Proc. Suppl. [**121**]{} (2003) 74
A.L. Kataev, G. Parente, A.V. Sidorov: Phys. Part. Nucl. [**34**]{} (2003) 20
J. Bl[ü]{}mlein, H. B[ö]{}ttcher: Nucl. Phys. B [**636**]{} (2002) 225
J.R. Ellis, M. Karliner: Phys. Lett. B [**341**]{} (1995) 397
G. Altarelli, R.D. Ball, S. Forte, G. Ridolfi: Nucl. Phys. B [**496**]{} (1997) 337
A. Knauf, M. Meyer-Hermann, G. Soff: Phys. Lett. B [**549**]{} (2002) 109
J.H. Kim et al.: Phys. Rev. Lett. [**81**]{} (1998) 3595
C. Glasman: AIP Conf. Proc. [**792**]{} (2005) 689
A.A. Penin, A.A. Pivovarov: Phys. Lett. B [**435**]{} (1998) 413
Particle Data Group Coll., S. Eidelman et al.: Phys. Lett. B [**592**]{} (2004) 1
M. Davier, A. H[ö]{}cker, Z. Zhang: LAL-05-37 (2005), hep-ph/0507078
Coll., S. Schael et al.: Phys. Rep. [**427**]{} (2006) 257
S. Albino, M. Klasen, S. S[ö]{}ldner-Rembold: Phys. Rev. Lett. [**89**]{} (2002) 122004
S. Menke: In: Stanford 2001, [$\mathrm{e^+e^-}$]{} physics at intermediate energies, D. Bettoni (ed.), 22. SLAC, SLAC-R-573, 2001
TOPAZ Coll., Y. Ohnishi et al.: Phys. Lett. B [**313**]{} (1993) 475
J. Schieck et al.: Acc. by Eur. Phys. J.C
UA1 Coll., C. Albajar et al.: Phys. Lett. B [**369**]{} (1996) 46
UA6 Coll., M. Werlen et al.: Phys. Lett. B [**452**]{} (1999) 201
S. Bethke: MPP-2006-54 (2006), hep-ex/0606035
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $X$ be a nonempty real variety that is invariant under the action of a reflection group $G$. We conjecture that if $X$ is defined in terms of the first $k$ basic invariants of $G$ (ordered by degree), then $X$ meets a $k$-dimensional flat of the associated reflection arrangement. We prove this conjecture for the infinite types, reflection groups of rank at most $3$, and $F_4$ and we give computational evidence for $H_4$. This is a generalization of Timofte’s degree principle to reflection groups. For general reflection groups, we compute nontrivial upper bounds on the minimal dimension of flats of the reflection arrangement meeting $X$ from the combinatorics of parabolic subgroups. We also give generalizations to real varieties invariant under Lie groups.'
address:
- 'Fachbereich Mathematik und Informatik, Freie Universität Berlin, Germany'
- 'Aalto Science Institute, PO Box 11000, FI-00076 Aalto'
- 'Institut für Mathematik, Goethe-Universität Frankfurt, Germany'
author:
- Tobias Friedl
- Cordian Riener
- Raman Sanyal
bibliography:
- 'DegreePrinciples.bib'
title: 'Reflection groups, reflection arrangements, and invariant real varieties'
---
Introduction {#sec:intro}
============
A real variety $X \subseteq {\mathbb{R}}^n$ is the set of real points simultaneously satisfying a system of polynomial equations with real coefficients, that is, $$X \ = \ {\mathrm{V}_{{\mathbb{R}}}}(f_1,\dots,f_m) \ := \ \{ {\mathbf{p}}\in {\mathbb{R}}^n : f_1({\mathbf{p}}) = f_2({\mathbf{p}}) =
\cdots = f_m({\mathbf{p}}) = 0 \},$$ for some $f_1,\dots,f_m \in {\mathbb{R}}[{\mathbf{x}}] := {\mathbb{R}}[x_1,\dots,x_n]$. In contrast to working over an algebraically closed field, the question if $X \neq {\varnothing}$ is considerably more difficult to answer, both theoretically and in practice; see [@roy]. Timofte [@tim] studied real varieties invariant under the action of the symmetric group ${\mathfrak{S}}_n$, and proved an interesting structural result. A ${\mathfrak{S}}_n$-invariant variety can be defined in terms of [**symmetric**]{} polynomials, that is, polynomials $f \in {\mathbb{R}}[{\mathbf{x}}]$ such that $f(x_{\tau(1)}, \dots, x_{\tau(n)}) = f(x_1,\dots,x_n)$ for all permutations $\tau \in {\mathfrak{S}}_n$. Recall that the fundamental theorem of symmetric polynomials states that a polynomial $f$ is symmetric if and only if $f$ is a polynomial in the elementary symmetric polynomials $e_1,\dots,e_n$. Let us call a ${\mathfrak{S}}_n$-invariant variety $X$ [**$\boldsymbol k$-sparse**]{} if $X = {\mathrm{V}_{{\mathbb{R}}}}(f_1,\dots,f_m)$ for some symmetric polynomials $f_1,\dots,f_m \in {\mathbb{R}}[e_1,\dots,e_k]$.
\[thm:Sn-degree\] Let $X \subseteq {\mathbb{R}}^n$ be a nonempty ${\mathfrak{S}}_n$-invariant real variety. If $X$ is $k$-sparse, then there is a point ${\mathbf{p}}\in X$ with at most $k$ distinct coordinates.
Viewing the symmetric group ${\mathfrak{S}}_n$ as a reflection group in ${\mathbb{R}}^n$ yields a sound geometric perspective on this result: As a group of linear transformations, ${\mathfrak{S}}_n$ is generated by reflections in the hyperplanes $H_{ij} = \{{\mathbf{p}}: p_i = p_j\}$ for $1 \le i < j \le n$. The ambient space ${\mathbb{R}}^n$ is stratified by the arrangement of reflection hyperplanes ${\mathcal{H}}= \{
H_{ij} : i < j \}$. The closed strata ${\mathcal{H}}_k \subseteq {\mathbb{R}}^n $ are the intersections of $n-k$ linearly independent reflection hyperplanes. Timofte’s result then states that a $k$-sparse variety $X$ is nonempty if and only if $X
\cap {\mathcal{H}}_k \neq {\varnothing}$. Such a point of view can be taken for general real reflection groups and the aim of this paper is a generalization of Theorem \[thm:Sn-degree\].
A (real) [**reflection group**]{} $G$ acting on $V \cong {\mathbb{R}}^n$ is a finite group of orthogonal transformations generated by reflections. The reflection group $G$ is [**irreducible**]{} if $G$ is not the product of two nontrivial reflection groups. Associated to $G$ is its [**reflection arrangement**]{} $${\mathcal{H}}\ = \ {\mathcal{H}}(G) \ := \ \{ H = \ker g : g \in G \text{ reflection} \}.$$ The [**flats**]{} of ${\mathcal{H}}$ are the linear subspaces arising from intersections of hyperplanes in ${\mathcal{H}}$. The arrangement of linear hyperplanes stratifies $V$ with strata given by $${\mathcal{H}}_i \ = \ {\mathcal{H}}_i(G) \ := \ \{ {\mathbf{p}}\in V : {\mathbf{p}}\text{ is contained in a flat
of dimension $i$} \}.$$ In particular, ${\mathcal{H}}_n = V$. We call $G$ [**essential**]{} if $G$ does not fix a nontrivial linear subspace or, equivalently, if ${\mathcal{H}}_0 = \{0\}$. If $G$ is essential, then the [**rank**]{} of $G$ is $ \operatorname{rank}(G) := \dim V$. Reflection groups naturally occur in connection with Lie groups/algebras and are well-studied from the perspective of geometry, algebra, and combinatorics [@hum; @FH91; @BB05]. A complete classification of reflection groups can be given in terms of Dynkin diagrams (see [@hum]). There are four infinite families of irreducible reflection groups ${\mathfrak{S}}_n \cong A_{n-1},B_n,D_n,I_2(m)$ and six exceptional reflection groups $H_3, H_4, F_4, E_6, E_7,$ and $E_8$.
The linear action of $G$ on $V$ induces an action on the symmetric algebra ${\mathbb{R}}[V]\cong{\mathbb{R}}[x_1,\ldots,x_n]$ by $g\cdot f({\mathbf{x}}):=f(g^{-1}\cdot {\mathbf{x}})$. Chevalley’s Theorem [@hum Ch. 3.5] states that the ring ${\mathbb{R}}[V]^G$ of polynomials invariant under $G$ is generated by algebraically independent homogeneous polynomials $\pi_1,\pi_2,\dots,\pi_n \in {\mathbb{R}}[V]$. The collection $\pi_1,\dots,\pi_n$ is called a set of [**basic invariants**]{} for $G$. The basic invariants are not unique, but their degrees $d_i(G) := \deg \pi_i$ are. Throughout, we will assume that the basic invariants are labelled such that $d_1 \le d_2 \le \cdots \le d_n$. In accord with ${\mathfrak{S}}_n$-invariant varieties, we call a $G$-invariant variety $X = {\mathrm{V}_{{\mathbb{R}}}}(f_1,\dots,f_m)$ [**$\boldsymbol k$-sparse**]{} if $f_1,\dots,f_m$ can be chosen in ${\mathbb{R}}[\pi_1,\dots,\pi_k]$ for some choice of basic invariants $\pi_1,\dots,\pi_n$. The following is the main result of the paper.
\[thm:main\] Let $G$ be a reflection group of type $I_2(m), A_{n-1},B_n,D_n, H_3,$ or $F_4$ and $X$ a nonempty $G$-invariant real variety. If $X$ is $k$-sparse, then $X \cap {\mathcal{H}}_k(G) \neq {\varnothing}$.
Since the first basic invariant of an essential reflection group is a scalar multiple of $p_2({\mathbf{x}}) = \|{\mathbf{x}}\|^2$, Theorem \[thm:main\] is trivially true for reflection groups of rank $\le 2$. The infinite families $A_{n-1}$, $B_n$, and $D_n$ are treated in Section \[sec:infinite\]. Timofte’s original proof and its simplification given by the second author in [@rie] use properties of the symmetric group that are not shared by all reflection groups (such as $D_n$) and we highlight this difference in Example \[ex:babacon\] and Remark \[rem:babacon\]. In Section \[sec:codim1\], we prove the following general result for invariant real varieties that strictly implies Theorem \[thm:main\] for $k=n-1$.
\[thm:all-but-one\] Let $G$ be an essential reflection group of rank $n$ and $X =
{\mathrm{V}_{{\mathbb{R}}}}(f_1,\dots,f_m)$ nonempty. If there is $j \in \{1,\dots,n\}$ such $f_1,\dots,f_m \in
{\mathbb{R}}[\pi_i : i \neq j]$, then $X \cap {\mathcal{H}}_{n-1}(G) \neq {\varnothing}$.
In particular, this result yields Theorem \[thm:main\] for all reflection groups of rank $\le 3$. The group $F_4$ is treated in Section \[sec:codim1\] and we provide computational evidence that Theorem \[thm:main\] also holds for $H_4$. That supports the following conjecture.
\[conj:main\] Let $G$ be an irreducible and essential reflection group. Then any nonempty and $k$-sparse $G$-invariant real variety $X$ intersects ${\mathcal{H}}_k(G)$.
In Section \[sec:higher\_codim\], we prove a weaker form of Conjecture \[conj:main\] under an extra assumption on the defining polynomials of $X$. In Section \[sec:StrNum\], we obtain upper bounds on the dimension of the stratum that meets $X$ in terms of the combinatorics of parabolic subgroups of $G$. Our results generalize to varieties invariant under the adjoint action of Lie groups and we explore this connection in Section \[sec:lie\].
Acevedo and Velasco [@velasco] independently considered the related problem of certifying nonnegativity of $G$-invariant homogeneous polynomials. They show that low-degree forms (where the exact degree depends on the group) are nonnegative if and only if they are nonnegative on ${\mathcal{H}}_{n-1}(G)$. Questions of nonnegativity of polynomials $f \in {\mathbb{R}}[V]^G$ are subsumed by our results. Let us call a $G$-invariant semialgebraic set $S \subseteq V$ $k$-sparse if $S$ is defined in terms of equations and inequalities with polynomials in ${\mathbb{R}}[\pi_1,\dots,\pi_k]$.
\[prop:nonneg\] Let $G$ be a reflection group for which Conjecture \[conj:main\] holds. Let $S \subseteq V$ be a $k$-sparse semialgebraic set and let $f \in {\mathbb{R}}[\pi_1,\dots,\pi_k]$. Then $f$ is nonnegative/positive on $S$ if and only if $f$ is nonnegative/positive on ${\mathcal{H}}_k(G) \cap S$.
If $S$ is $k$-sparse, then the $G$-invariant variety $$\label{eqn:X_k}
X_k({\mathbf{q}}) \ := \ \{ {\mathbf{p}}\in V : \pi_i({\mathbf{p}}) = \pi_i({\mathbf{q}}) \text{ for } i =
1,\dots,k \}$$ is contained in $S$ for any ${\mathbf{q}}\in S$. Assume that there is a point ${\mathbf{q}}\in S$ with $f({\mathbf{q}}) < 0$. By assumption $f = F(\pi_1,\dots,\pi_k)$ for some $F \in {\mathbb{R}}[y_1,\dots,y_k]$. Hence $f$ is negative (and constant) on $X_k({\mathbf{q}})
\subseteq S$. By construction $X_k({\mathbf{q}})$ is $k$-sparse and, since $G$ satisfies Conjecture \[conj:main\], $X_k({\mathbf{q}}) \cap {\mathcal{H}}_k(G) \neq
{\varnothing}$.
The proof of Proposition \[prop:nonneg\] makes use of a key observation: It suffices to consider invariant varieties of the form as any $k$-sparse variety $X$ contains $X_k({\mathbf{q}})$ for all ${\mathbf{q}}\in X$. We call $X_k({\mathbf{q}})$ a [**principal**]{} $k$-sparse variety. Lastly, let us emphasize again that we will work with *real* varieties exclusively. In particular, set-theoretically, every real variety $X
= {\mathrm{V}_{{\mathbb{R}}}}(f_1,\dots,f_m)$ is the set of solutions to the equation $f({\mathbf{x}}) = 0$ for $f = f_1^2 + f_2^2 + \cdots + f_m^2$.
**Acknowledgements.** We are much indebted to Christian Stump for the many helpful discussions regarding the combinatorics of reflection groups and their invariants. We also thank Florian Frick and Christian Haase for an interesting but fruitless afternoon of orbit spaces. We also thank Mareike Dressler for help with GloptiPoly.
The infinite families $A_{n-1},B_n,$ and $D_n$ {#sec:reflection}
==============================================
\[sec:infinite\]
In this section, we prove Theorem \[thm:main\] for the reflection groups of type $A_{n-1},B_n,$ and $D_n$. The symmetric group ${\mathfrak{S}}_n$ acts on ${\mathbb{R}}^n$ but is not essential as it fixes ${\mathbb{R}}\mathbf{1}$. The restriction to $\{ {\mathbf{x}}\in {\mathbb{R}}^n : x_1 + \cdots + x_n = 0\}$ is the essential reflection group of type $\mathbf{A_{n-1}}$. The reflection arrangement ${\mathcal{H}}({\mathfrak{S}}_n)$ was described in the introduction. The $k$-stratum ${\mathcal{H}}_k({\mathfrak{S}}_n)$ is given by the points ${\mathbf{p}}\in {\mathbb{R}}^n$ that have at most $k$ distinct coordinates. A set of basic invariants is given by the [**elementary symmetric polynomials**]{}
$$e_k({\mathbf{x}}) \ := \ \sum_{1 \le i_1 < \cdots < i_k \le n} x_{i_1}\cdots x_{i_k}$$ or, alternatively, by the [**power sums**]{} $${s}_k({\mathbf{x}}) \ := \ x_1^k + x_2^k + \cdots + x_n^k,$$ for $k=1,\dots,n$. The group $\mathbf{B_n} = {\mathfrak{S}}_n \rtimes {\mathbb{Z}}^n_2$ acts on $V = {\mathbb{R}}^n$ by *signed* permutations with reflection hyperplanes $\{x_i
= \pm x_j\}$ and $\{ x_i = 0 \}$ for $1 \le i < j \le n$. A point ${\mathbf{p}}$ lies in ${\mathcal{H}}_i(B_n)$ if and only if $(|p_1|,\dots,|p_n|)$ has at most $i$ distinct nonzero coordinates. A set of basic invariants is given by $\pi_i({\mathbf{x}}) =
{s}_{2i}({\mathbf{x}}) = {s}_i(x_1^2,\dots,x_n^2)$. The index-$2$ subgroup $\mathbf{D_n}$ of $B_n$ given by the semidirect product of ${\mathfrak{S}}_n$ with ‘even sign changes’ yields a reflection group with reflection hyperplanes $\{
x_i = \pm x_j\}$ for $1 \le i < j \le n$. The $k$-stratum of $D_n$ is a bit more involved to describe: denote by $M$ the set of all ${\mathbf{p}}\in {\mathbb{R}}^n$ with exactly one zero coordinate. Then $$\label{eqn:M}
{\mathcal{H}}_k(D_n) \ = \ ({\mathcal{H}}_k(B_n) \setminus M) \cup ({\mathcal{H}}_{k-1}(B_n) \cap M).$$ The invariant that distinguishes $D_n$ from $B_n$ is given by $e_n({\mathbf{x}}) =
x_1x_2\cdots x_n$. A set of basic invariants for $D_n$ are $\pi_1({\mathbf{x}}),\dots,\pi_n({\mathbf{x}})$ with $$\label{eqn:Dn-invs}
\pi_k({\mathbf{x}}) \ := \
\begin{cases}
{s}_{2k}({\mathbf{x}}) & \text{ for } 1 \le k \le
\lfloor\frac{n}{2}\rfloor,\\
e_n({\mathbf{x}}) & \text{ for } k =
\lfloor\frac{n}{2}\rfloor + 1, \text{ and }\\
{s}_{2k-2}({\mathbf{x}}) & \text{ for } \lfloor\frac{n}{2}\rfloor + 1 < k
\le n.\\
\end{cases}$$
We start with the verification of Theorem \[thm:main\] for ${\mathfrak{S}}_n$ which is exactly Theorem \[thm:Sn-degree\]. The proofs for $B_n$ and $D_n$ will rely on the arguments for $A_{n-1}$.
Let us first assume that $\pi_i({\mathbf{x}}) = {s}_i({\mathbf{x}}) = x_1^i + \cdots +
x_n^i$ are the power sums for $i=1,\dots,n$. It suffices to show the claim for a principal $k$-sparse variety $X_k({\mathbf{p}}_0)$ as defined in , i.e. that $X_k({\mathbf{p}}_0) \cap {\mathcal{H}}_k({\mathfrak{S}}_n) \neq
{\varnothing}$ for ${\mathbf{p}}_0 \in X$. Since $\pi_2({\mathbf{p}}) = \|{\mathbf{p}}\|^2$, we conclude that $X_k({\mathbf{p}}_0)$ is compact and $\pi_{k+1}$ attains its maximum over $X_k({\mathbf{p}}_0)$ in a point ${\mathbf{q}}$. At this point, the Jacobian ${\mathrm{Jac}}_{\mathbf{q}}(\pi_1,\dots,\pi_{k+1}) = \bigl(\nabla {s}_1({\mathbf{q}}),\ldots,\nabla
{s}_{k+1}({\mathbf{q}})\bigr)$ has rank $< k+1$. We claim that ${\mathbf{q}}\in{\mathcal{H}}_k$ if and only if the Jacobian $J = {\mathrm{Jac}}_{\mathbf{q}}(\pi_1,\dots,\pi_{k+1})$ has rank $<
k+1$. Indeed, up to scaling columns, $J$ is given by $$\begin{pmatrix}
1 & 1 & \cdots & 1 \\
q_1 & q_2 & \cdots & q_n \\
\vdots & & &\vdots \\
q_1^{k} &q_2^{k}&\cdots & q_n^{k}
\end{pmatrix}.$$ The $k+1$ minors are thus Vandermonde determinants all of which vanish if and only if ${\mathbf{q}}\in {\mathcal{H}}_k$, by the description of $k$-strata for ${\mathfrak{S}}_n$. For an arbitrary choice of basic invariants, the result follows from Lemma \[lem:ind\] below.
\[lem:ind\] Fix a reflection group $G$ acting on $V$. Let $\pi_1,\ldots,\pi_n$ and $\pi_1',\ldots,\pi_n'$ be two sets of basic invariants and let $1 \le k
\le n$ such that $d_{k+1}>d_k$. Then $ {\mathbb{R}}[\pi_1,\ldots,\pi_k] \ = \
{\mathbb{R}}[\pi_1',\ldots,\pi_k'].$ Moreover, $\operatorname{rank}{\mathrm{Jac}}_{\mathbf{p}}(\pi_1,\dots,\pi_k) = \operatorname{rank}{\mathrm{Jac}}_{\mathbf{p}}(\pi_1',\dots,\pi_k')$ for all ${\mathbf{p}}\in V$.
For every $1 \le i \le k$, $\pi_i = F_i(\pi'_1,\dots,\pi'_n)$ for some polynomial $F_i(y_1,\dots,y_n)$. Homogeneity and algebraic independence imply that $F_i \in {\mathbb{R}}[y_1,\dots,y_k]$. This shows the inclusion ${\mathbb{R}}[\pi_1,\dots,\pi_k] \subseteq {\mathbb{R}}[\pi_1',\dots,\pi_k']$. Note that $${\mathrm{Jac}}_{{\mathbf{p}}}(\pi_1,\dots,\pi_k) \ = \ {\mathrm{Jac}}_{\pi'({\mathbf{p}})}(F_1,\dots,F_k) \cdot
{\mathrm{Jac}}_{{\mathbf{p}}}(\pi_1',\dots,\pi_k')$$ for every ${\mathbf{p}}\in V$. The same argument applied to $\pi'_i$ now proves the first claim and shows that ${\mathrm{Jac}}_{\pi'({\mathbf{p}})}(F_1,\dots,F_k)$ has full rank and this proves the second claim.
We proceed to the reflection groups of type $B_n$.
By Lemma \[lem:ind\] and the fact that the degrees $d_i(B_n)$ are all distinct, we may assume that $\pi_i({\mathbf{x}}) = {s}_{2i}({\mathbf{x}})$ for $i=1,\dots,n$. Moreover, we can assume that $X$ is a principal $k$-sparse variety, that is, $$X \ = \ X_k({\mathbf{p}}) \ = \ \{ {\mathbf{x}}\in {\mathbb{R}}^n : {s}_{2i}({\mathbf{x}}) =
{s}_{2i}({\mathbf{p}})
\text{ for } i = 1,\dots,k \}.$$ Since $X_k({\mathbf{p}}) = X_k({\mathbf{q}})$ for all ${\mathbf{q}}\in X_k({\mathbf{p}})$, we can assume that ${\mathbf{p}}= (p_1,\dots,p_r,0,\dots,0) \in X$ with the property that $p_1 \cdots
p_r \neq 0$ and $r$ is minimal.
If $r = n$, then $X$ does not meet any of the coordinate hyperplanes $\{
x_i = 0 \}$. Let ${\mathbf{q}}\in X$ be an extreme point of $\pi_{k+1}$ over $X$. At this point, the Jacobian $J = {\mathrm{Jac}}_{\mathbf{q}}(\pi_1,\dots,\pi_{k+1})$ does not have full rank and hence every maximal minor of $$J \ = \
\begin{pmatrix}
q_1 & q_2 & \cdots & q_n \\
\vdots & & &\vdots \\
q_1^{2k-1} &q_2^{2k-1}&\cdots & q_n^{2k-1}
\end{pmatrix}$$ vanishes. Since $q_i \neq 0$ for all $i=1,\dots,n$, the Vandermonde formula implies that $(q_1^2,q_2^2,\dots,q_n^2)$ has at most $k$ distinct coordinates, which yields the claim.
If $r < n$, we can restrict $X$ to the linear subspace $U = \{ {\mathbf{x}}\in
{\mathbb{R}}^n : x_{r+1} = \cdots = x_n = 0\} \cong {\mathbb{R}}^r$. The set $X' := X \cap U
\subseteq {\mathbb{R}}^r$ is nonempty and, in particular, a $k$-sparse $B_r$-invariant variety that stays away from the coordinate hyperplanes in ${\mathbb{R}}^r$. By the previous case, there is a point ${\mathbf{q}}' \in X'$ such that $(|q'_1|,\dots,|q'_r|)$ has at most $k$ distinct coordinates. By construction, ${\mathbf{q}}= ({\mathbf{q}}',\mathbf{0}) \in X \cap {\mathcal{H}}_k(B_n)$, which proves the claim.
The key to the proof of Theorem \[thm:main\] for $A_{n-1}$ and $B_n$ is the strong connection between the strata ${\mathcal{H}}_k$ and the ranks of the Jacobians ${\mathrm{Jac}}(\pi_1,\dots,\pi_{k+1})$.
\[cor:strongSteinberg\] Let $G \in \{ {\mathfrak{S}}_n, B_n\}$ and $\pi_1,\dots,\pi_n$ a set of basic invariants for $G$. Then a point ${\mathbf{q}}\in V$ lies in ${\mathcal{H}}_k(G)$ for $0 \le k \le n-1$ if and only if ${\mathrm{Jac}}_{{\mathbf{p}}}(\pi_1,\dots,\pi_{k+1})$ has rank at most $k$.
It is tempting to believe that such a statement holds true for all reflection groups and, indeed, necessity follows from a well-known result of Steinberg [@stein]. However, the following example shows that Corollary \[cor:strongSteinberg\] does not hold in general.
\[ex:babacon\] Consider the group $G = D_5$ acting on ${\mathbb{R}}^5$ and the point ${\mathbf{p}}=
(1,1,1,1,0)$. The point lies in ${\mathcal{H}}_2(D_5) \setminus {\mathcal{H}}_1(D_5)$, that is, ${\mathbf{p}}$ lies on exactly $3$ linearly independent reflection hyperplanes. On the other hand, for any choice of basic invariants $\pi_1,\dots,\pi_5$ the gradients $\nabla_{\mathbf{p}}\pi_1,
\nabla_{\mathbf{p}}\pi_2$ are linearly dependent. Indeed, for $\pi_1 = \|x\|^2 =
x_1^2 + \cdots + x_5^2$ and $\pi_2 = x_1^4 + \cdots + x_5^4$, this is easy to check and this extends to all choices of basic invariants using Lemma \[lem:ind\].
\[rem:babacon\] Example \[ex:babacon\] also serves as a counterexample to generalizations of Corollary \[cor:strongSteinberg\] to all finite reflection groups considered in [@giv Lemma 1’] (without a proof) and [@barbancon Statement 3.3]. Moreover, in the language of Acevedo and Velasco [@velasco Definition 7], it is the first example of a reflection group not satisfying the *minor factorization condition*.
The following proof of Theorem \[thm:main\] for type $D_n$ does not rely on an extension of Corollary \[cor:strongSteinberg\].
Let $\pi_1,\dots,\pi_n$ be a choice of basic invariants for $D_n$ and let $X = X_k({\mathbf{q}}) \subseteq {\mathbb{R}}^n$ for some ${\mathbf{q}}\in {\mathbb{R}}^n$ and $1 \le k < n$. If $n$ is odd or if $k \neq \lfloor \frac{n}{2} \rfloor$, then, by Lemma \[lem:ind\], we can assume that the basic invariants are given by . If $n$ is even and $k = \lfloor \frac{n}{2}
\rfloor$, then $\pi_{\lfloor \frac{n}{2} \rfloor}({\mathbf{x}}) \ = \ \alpha
{s}_{n}({\mathbf{x}}) + \beta e_n({\mathbf{x}}), $ for some $\beta \neq 0$. We can also assume that ${\mathbf{q}}= (q_1,\dots,q_l,0,\dots,0)$ with $q_1\cdots q_l \neq 0$ and $l$ maximal among all points in $X_k({\mathbf{q}})$. We distinguish two cases.
**Case $l < n$:** In this case, $e_{n}({\mathbf{x}})$ is identically zero on $X_k({\mathbf{q}})$ and $X' := X_k({\mathbf{q}}) \cap \{ {\mathbf{x}}: x_n = 0 \}$ is nonempty. If $k
\ge \lfloor \frac{n}{2} \rfloor + 1$, then we can identify $$X' \ = \ \{ {\mathbf{x}}' \in {\mathbb{R}}^{n-1} : {s}_{2i}({\mathbf{x}}',{\mathbf{0}})= {s}_{2i}({\mathbf{q}})
\text{ for } i = 1,\dots,k-1 \} .$$ Hence $X'$ is a real variety in ${\mathbb{R}}^{n-1}$ invariant under the action of $B_{n-1}$ and $X'$ is $(k-1)$-sparse. By Theorem \[thm:main\] for $B_{n-1}$, $X' \cap {\mathcal{H}}_{k-1}(B_{n-1}) \neq
{\varnothing}$. The claim now follows the description of ${\mathcal{H}}_k(D_n)$ given in .
If $k < \frac{n}{2}$, consider the Jacobian of $\pi_1 = {s}_{2},\dots,
\pi_k = {s}_{2k}$ and the $(l+1)$-th elementary symmetric polynomial $e_{l+1}({\mathbf{x}})$ at ${\mathbf{q}}$ $$\label{eqn:l_n-1}
J \ = \ {\mathrm{Jac}}_{\mathbf{q}}({s}_2,\dots,{s}_{2k},e_{l+1}) \ = \
\begin{pmatrix}
q_1 & q_2 & \cdots & q_l & 0 & 0 & \cdots & 0 \\
q_1^3 & q_2^3 & \cdots & q_l^3 & 0 & 0 & \cdots & 0 \\
\vdots & \vdots & & \vdots & \vdots & \vdots & \vdots & \vdots \\
q_1^{2k-1} & q_2^{2k-1} & \cdots & q_l^{2k-1} & 0 & 0 &
\cdots & 0 \\
0 & 0 & 0 & 0& q_1\cdots q_l & 0 & \cdots & 0
\end{pmatrix}.$$ We observe that the $(l+1)$-th elementary symmetric function $e_{l+1}({\mathbf{x}})$ is identically zero on $X_k({\mathbf{q}})$ and hence the gradients of $\pi_1,\dots,\pi_k$ and $e_{l+1}$ are linearly dependent on $X_k({\mathbf{q}})$. In particular, the Jacobian $J$ has rank $\le k$. Since $q_1 \cdots q_l \neq 0$, the Vandermonde minors imply $$\prod_{i,j\in I, i<j}{q_i^2-q_j^2} \ = \ 0$$ for any $I \subseteq \{1,\dots,l\}$ with $|I| = k$. This shows that ${\mathbf{q}}\in
{\mathcal{H}}_{k-1}(B_n) \subseteq {\mathcal{H}}_k(D_n)$.
If $k= \frac{n}{2}$, then $n$ is even and $X_k({\mathbf{q}})$ is cut out by ${s}_2,\dots, {s}_{n-2}$ and possibly ${s}_{n}$, since $e_{n}({\mathbf{x}})$ is identically zero on $X_k({\mathbf{q}})$. Thus the argument above remains valid.
**Case $l = n$:** If $k < \frac{n}{2}$, set $f := e_n$. If $k \ge \lfloor \frac{n}{2}
\rfloor +1$, set $f := {s}_{2k}$. For the special case that $n$ even and $k =\frac{n}{2}$, we set $f = e_n$ if $\alpha \neq 0$ and $f =
{s}_{n}$ otherwise. Let ${\mathbf{r}}\in X_k({\mathbf{q}})$ be a maximizer of $|f({\mathbf{x}})|$. In particular, $r_1\cdots r_n \neq 0$. Up to row and column operations, the Jacobian $J = {\mathrm{Jac}}_{\mathbf{q}}({s}_2,{s}_4,\dots,{s}_{2k},f)$ is of the form $$\label{eqn:l_n}
J \ = \ \begin{pmatrix}
r_1 & r_2 & \cdots & r_n \\
r_1^3 & r_2^3 & \cdots & r_n^3 \\
\vdots & \vdots & & \vdots \\
r_1^{2k-1} & r_2^{2k-1} & \cdots & r_n^{2k-1} \\
\widehat{r}_1 r_2\cdots r_n & r_1 \widehat{r}_2\cdots r_n & \cdots
& r_1 r_2\cdots \widehat{r}_n
\end{pmatrix},$$ where $\widehat{r}_i$ is to be omitted from the product. Multiplying the $i$-th column by $r_i$ and dividing the last row by $r_1\cdots r_n$, we get a Vandermonde matrix of rank $\le k$. Hence $(|r_1|,\dots,|r_k|)$ has at most $k$ distinct entries. Since all entries are nonzero, it follows that ${\mathbf{r}}\in {\mathcal{H}}_k(D_n)$.
The proof actually gives stronger implications for the $B_n$-case.
\[cor:stronger\_Bn\] Let $1\le k\le n-2$. Then every nonempty $B_n$-invariant, $k$-sparse variety $X$ meets ${\mathcal{H}}_k(D_n)$.
For $X = X_k({\mathbf{q}})$, we can assume that ${\mathbf{q}}=(q_1,\dots,q_l,0,\dots,0)$ with $q_i\neq 0$ for $1\le i\le l$ and $l$ maximal. If $l \le k$, then $ {\mathbf{q}}\in {\mathcal{H}}_k(D_n)$ and we are done. So assume $k < l \le n$. We distinguish two cases: If $l \le n-1$, let $f = e_{l+1}$ and ${\mathbf{r}}\in X_k({\mathbf{q}})$ arbitrary. If $l = n$, let $f = e_{n}$ and ${\mathbf{r}}\in X_k({\mathbf{q}})$ a maximizer of $|f|$. The corresponding Jacobians and for ${s}_2,\dots,{s}_{2k},f$ at ${\mathbf{r}}$ yield the claim.
Real orbit spaces and reflection arrangements {#sec:codim1}
=============================================
The reflection arrangement ${\mathcal{H}}$ decomposes $V$ into relatively open polyhedral cones. The closure $\sigma$ of a full-dimensional cone in this decomposition serves as a [**fundamental domain**]{}: For every ${\mathbf{p}}\in V$ the orbit $G{\mathbf{p}}$ meets $\sigma$ in a unique point. On the other hand, the basic invariants define an [**orbit map**]{} $\pi : V \rightarrow {\mathbb{R}}^n$ given by $\pi({\mathbf{x}})=(\pi_1({\mathbf{x}}),\ldots,\pi_n({\mathbf{x}}))$. The basic invariants separate orbits, that is, $\pi({\mathbf{p}}) = \pi({\mathbf{q}})$ if and only if ${\mathbf{q}}\in G{\mathbf{p}}$ for all ${\mathbf{p}},{\mathbf{q}}\in
V$. The image ${\mathcal{S}}:=\pi(V)$ is homeomorphic to $V / G$ and, by abuse of terminology, we call ${\mathcal{S}}$ the [**real orbit space**]{}. Since $\pi$ is an algebraic map, ${\mathcal{S}}$ is semialgebraic (with an explicit description given in [@procesi]). Restricted to $\sigma$ the map $\pi|_\sigma:\sigma\to
{\mathcal{S}}$ is a homeomorphism. Moreover, $$\label{eqn:S_boundary}
\pi^{-1}(\partial {\mathcal{S}}) \ = \ {\mathcal{H}}_{n-1}(G),$$ where $\partial {\mathcal{S}}$ denotes the boundary of ${\mathcal{S}}$. Observe that neither the orbit space ${\mathcal{S}}$ nor the fundamental domain $\sigma$ are uniquely determined by $G$.
In terms of the orbit map, Conjecture \[conj:main\] can be put in a more general context. For $J \subseteq [n] := \{1,\dots,n\}$, let us write $\pi_J({\mathbf{x}}) = (\pi_i({\mathbf{x}}) : i \in J)$. For given $J$, we can ask for the smallest $0 \le t \le n$ such that $\pi_J(V) \ = \ \pi_J({\mathcal{H}}_t)$.
\[prop:conj-equiv\] Let $G$ be an irreducible and essential reflection group. Then Conjecture \[conj:main\] is true for $G$ if and only if for $J =
\{1,\dots,k\}$ $$\pi_J(V) \ = \ \pi_J({\mathcal{H}}_k).$$
For ${\mathbf{q}}\in V$, we have $X_k({\mathbf{q}}) = \pi_J^{-1}(\pi_J({\mathbf{q}}))$. Hence, $X_k({\mathbf{q}})
\cap {\mathcal{H}}_k \neq {\varnothing}$ for ${\mathbf{q}}\in V$ if and only if there is some ${\mathbf{p}}\in {\mathcal{H}}_k$ such that $\pi_J({\mathbf{q}}) = \pi_J({\mathbf{p}})$.
A generalization of Theorem \[thm:Sn-degree\] to *$J$-sparse* symmetric polynomials $f \in {\mathbb{R}}[\pi_i : i \in J]$ was considered in [@Riener14]. The correspondence given in Proposition \[prop:conj-equiv\] also shows that the dimensions of strata in Conjecture \[conj:main\] are best possible.
\[prop:lower\_bound\] Let $J \subseteq [n]$ and $0 \le t \le n$ such that $\pi_J(V) \ = \
\pi_J({\mathcal{H}}_t)$. Then $t \ge |J|$.
The set $\pi_J(V)$ is the projection of the real orbit space ${\mathcal{S}}$ onto the coordinates indexed by $J$ and hence is of full dimension $|J|$. By invariance of dimension, this implies that $t = \dim {\mathcal{H}}_t \ge |J|$.
For the next result recall that, by definition, $G \subset O(V)$ and hence $\|{\mathbf{x}}\|^2 = \langle {\mathbf{x}}, {\mathbf{x}}\rangle$ is an invariant of $G$.
\[lem:S\_linefree\] Let $G$ be a finite reflection group and $\pi_1,\dots,\pi_n$ a choice of basic invariants such that $\pi_i({\mathbf{x}}) = \|{\mathbf{x}}\|^2$ for some $i$. Then the orbit space ${\mathcal{S}}= \pi(V)$ is line-free, that is, if $L \subseteq V$ is an affine subspace such that $L \subseteq {\mathcal{S}}$, then $L$ is a point.
Since $\pi_i({\mathbf{x}}) = \|{\mathbf{x}}\|^2 \ge 0$ for all ${\mathbf{x}}\in V$, the linear function $\ell({\mathbf{y}}) = y_i$ is nonnegative on ${\mathcal{S}}\subset {\mathbb{R}}^n$. Hence, if $L
\subseteq {\mathcal{S}}$ is an affine subspace, then $\ell$ is constant on $L$. Let $\sigma \subseteq V$ be a fundamental domain for $G$. Then $L =
{\mathcal{S}}\cap L$ is homeomorphic to $\hat{L} := \{ {\mathbf{p}}\in \sigma : \|{\mathbf{p}}\|^2
= c \}$ for some $c \ge 0$. This implies that $L$ is compact which proves the claim.
\[thm:codim1\] Let $G$ be an essential reflection group with a choice of basic invariants $\pi_1,\dots,\pi_n$. Let $f \in {\mathbb{R}}[V]^G$ be an invariant polynomial such that $f$ is at most linear in $\pi_k$ for some $k$. Then ${\mathrm{V}_{{\mathbb{R}}}}(f)\neq{\varnothing}$ if and only if ${\mathrm{V}_{{\mathbb{R}}}}(f)\cap{\mathcal{H}}\neq{\varnothing}$.
Without loss of generality, we can assume that $f(0) < 0$. Since the arrangement ${\mathcal{H}}$ is path connected, it suffices to show that there is a point ${\mathbf{p}}_+ \in {\mathcal{H}}_{n-1}(G)$ with $f({\mathbf{p}}_+) \ge 0$.
We can assume that $\pi_1 = \|{\mathbf{x}}\|^2$. Indeed, since $G$ is essential, all basic invariants have degree at least $2$ and $\|{\mathbf{x}}\|^2$ is a linear combination of the degree $2$ basic invariants. Let ${\mathbf{p}}\in
{\mathrm{V}_{{\mathbb{R}}}}(f)$ and define $K = \{ {\mathbf{q}}: \pi_1({\mathbf{q}}) = \pi_1({\mathbf{p}})\}$, the sphere centered at the origin that contains ${\mathbf{p}}$. The function $f$ attains its maximum over $K$ in a closed set $M \subseteq K$. We claim that $M \cap {\mathcal{H}}\neq {\varnothing}$. Let ${\mathbf{p}}_0$ be a point in $M$.
We may pass to the real orbit space ${\mathcal{S}}= \pi(V)$ associated to $G$ and $\pi_1,\dots,\pi_n$ and consider the compact set $\overline{K} :=
\pi(K) = \{ {\mathbf{y}}\in {\mathcal{S}}: y_1 = \pi_1({\mathbf{p}})\}$. We can write $f =
F(\pi_1,\dots,\pi_n)$ for some $F \in {\mathbb{R}}[y_1,\dots,y_n]$. In this setting, our assumption states that $F$ is at most linear in $y_k$. If ${\mathbf{p}}_0 \in V \setminus {\mathcal{H}}$, then, by , ${\overline{{\mathbf{p}}}}_0 :=
\pi({\mathbf{p}}_0)$ is in the interior of ${\mathcal{S}}$ and hence in the relative interior of $\overline{K}$. Let $L = \{ {\overline{{\mathbf{p}}}}_0 + t e_k : t \in {\mathbb{R}}\}$ be the affine line through ${\overline{{\mathbf{p}}}}_0$ in direction $e_k$. Restricted to $L$, the polynomial $F$ has degree at most $1$. By Lemma \[lem:S\_linefree\] and our choice of $\overline{K}$, the line $L$ meets $\partial \overline{K}$ in two points ${\overline{{\mathbf{p}}}}_-,{\overline{{\mathbf{p}}}}_+$ and $ F({\overline{{\mathbf{p}}}}_-) \le F({\overline{{\mathbf{p}}}}_0) \le F({\overline{{\mathbf{p}}}}_+)$. This implies that $\pi^{-1}({\overline{{\mathbf{p}}}}_+) \subseteq M$ and, since $\partial
\overline{K} \subseteq \partial {\mathcal{S}}$, equation shows that $\pi^{-1}({\overline{{\mathbf{p}}}}_+) \subseteq {\mathcal{H}}$.
The assumption in Theorem \[thm:codim1\] that $G$ is essential is essential. For example, let $G = B_n$ act on $V = {\mathbb{R}}^n \times {\mathbb{R}}$ by fixing the last coordinate. A set of basic invariants is given by $\pi_1({\mathbf{x}},x_{n+1}) =
x_{n+1}$ and $\pi_i({\mathbf{x}},x_{n+1}) = {s}_{2i-2}({\mathbf{x}})$ for $i = 2,\dots,n+1$. Pick ${\mathbf{p}}\in {\mathbb{R}}^n$ with all coordinates positive and distinct. The variety $$\label{eqn:Bn-counter}
X \ = \ \{ ({\mathbf{x}},x_{n+1}) \in V : {s}_{2i}({\mathbf{x}}) = {s}_{2i}({\mathbf{p}}) \text{
for } i = 1,\dots,n \}$$ is defined over ${\mathbb{R}}[\pi_2,\dots,\pi_{n+1}]$, but is a collection of affine lines that does not meet the reflection arrangement.
If a polynomial $f \in {\mathbb{R}}[V]^G$ depends on all but one basic invariant, then Theorem \[thm:codim1\] applies and directly implies Theorem \[thm:all-but-one\] from the introduction. We give two further applications.
\[cor:codim1apps\] Let $G$ be an essential reflection group and let $J \subset [n]$ with $|J|
= n-1$. For polynomials $f, f_1,\dots,f_m \in {\mathbb{R}}[\pi_i : i \in J]$, the following hold:
1. ${\mathrm{S}}= \{ {\mathbf{p}}: f_1({\mathbf{p}}) \ge 0,\dots, f_m({\mathbf{p}}) \ge 0\}$ is nonempty if and only if ${\mathrm{S}}\cap {\mathcal{H}}_{n-1}(G) \neq {\varnothing}$.
2. $f({\mathbf{q}}) \ge 0$ for all ${\mathbf{q}}\in {\mathrm{S}}$ if and only if $f({\mathbf{q}})
\ge 0$ for all ${\mathbf{q}}\in {\mathrm{S}}\cap {\mathcal{H}}_{n-1}(G)$.
For ${\mathbf{q}}\in {\mathrm{S}}$, it suffices to prove the claim for $$X \ := \ \{ {\mathbf{p}}\in V : \pi_j({\mathbf{p}}) = \pi_j({\mathbf{q}}) \text{ for } j \in J\} \
\subseteq \ S.$$ Claim (i) now follows from Theorem \[thm:codim1\]. As for (ii), assume that ${\mathbf{q}}\in
{\mathrm{S}}\setminus {\mathcal{H}}$ and $f({\mathbf{q}}) < 0$. Then the same argument applied to $X \cap \{ {\mathbf{p}}: f({\mathbf{p}}) = f({\mathbf{q}})\}$ finishes the proof.
If $f \in {\mathbb{R}}[V]^G$ has degree $\deg(f) < 2 \deg(\pi_n)$, then the algebraic independence of the basic invariants implies that Theorem \[thm:codim1\] can be applied to proof the following corollary. Under the assumption that $f$ is homogeneous, the second part of the corollary recovers the main result of Acevedo and Velasco [@velasco].
Let $f \in {\mathbb{R}}[V]^G$ with $\deg(f)<2d_n(G) = 2\deg(\pi_n)$. Then ${\mathrm{V}_{{\mathbb{R}}}}(f)\neq{\varnothing}$ if and only if ${\mathrm{V}_{{\mathbb{R}}}}(f)\cap{\mathcal{H}}\neq{\varnothing}$. In particular, $f \ge 0$ on $V$ if and only if $f \ge 0$ on ${\mathcal{H}}$.
The bound on the degree is tight: For a point ${\mathbf{p}}\in V \setminus
{\mathcal{H}}_{n-1}(G)$, the set of solutions to $$f({\mathbf{x}}) \ := \ \sum_{i=1}^n (\pi_{i}({\mathbf{x}}) - \pi_{i}({\mathbf{p}}))^2 \ = \ 0$$ is exactly $G{\mathbf{p}}$, which does not meet ${\mathcal{H}}_{n-1}(G)$. The defining polynomial $f({\mathbf{x}})$ is of degree exactly $2\deg(\pi_n)$. Theorem \[thm:codim1\] also allows us to prove Theorem \[thm:main\] for groups of low rank.
For $k = \operatorname{rank}(G)$, there is nothing to prove. For $k = 1$, we observe that $X_1({\mathbf{p}})$ is the sphere through ${\mathbf{p}}$, which meets the arrangement ${\mathcal{H}}_1(G)$ of lines through the origin. Thus, the only nontrivial case is $\operatorname{rank}(G) = 3$ and $k= \operatorname{rank}(G) -1 = 2$. This is covered by Theorem \[thm:all-but-one\].
Let $G$ be an essential reflection group of rank $\ge 4$. Since $G$ acts on $V$ by orthogonal transformations, we have that $\pi_1({\mathbf{x}}) = \|{\mathbf{x}}\|^2$ and $X_k({\mathbf{p}})$ is a subvariety of a sphere centered at the origin. Since the basic invariants are homogeneous, we may assume that $\pi_1({\mathbf{p}}) = 1$ and hence $X_k({\mathbf{p}}) \subseteq S^{n-1} = \{ {\mathbf{x}}\in V : \|{\mathbf{x}}\| = 1 \}$. To prove Theorem \[thm:main\] for $k=2$ we can proceed as follows. Let ${\delta_{\min}}$ and ${\delta_{\max}}$ be the minimum and maximum of $\pi_2$ over $S^{n-1}$. Then it suffices to find points ${\mathbf{p}}_{\min},{\mathbf{p}}_{\max} \in {\mathcal{H}}_2(G) \cap S^{n-1}$ with $\pi_2({\mathbf{p}}_{\min}) = {\delta_{\min}}$ and $\pi_2({\mathbf{p}}_{\max}) = {\delta_{\max}}$. Indeed, since ${\mathcal{H}}_2(G)$ is connected (for $\operatorname{rank}(G) \ge 3$), this shows that $\pi_J(V) =
\pi_J({\mathcal{H}}_2(G))$ for $J=\{1,2\}$, which, by Proposition \[prop:conj-equiv\], then proves the claim. For the group $F_4$, we can implement this strategy.
Since $F_4$ is of rank $4$, we only need to consider the case $k=2$ and can use the strategy outlined above. Let ${\delta_{\min}}$ and ${\delta_{\max}}$ be the minimum and maximum of $\pi_2$ over $S^3$. An explicit description of $\pi_2$ for $F_4$ is $$\pi_2({\mathbf{x}}) \ = \ \sum_{1 \le i < j \le 4} (x_i + x_j)^6 + (x_i -
x_j)^6;$$ see, for example, Mehta [@Mehta] or [@IKM Table 5]. The points ${\mathbf{p}}= (1,0,0,0)$ and ${\mathbf{p}}' = (\frac{1}{\sqrt{2}}, \frac{1}{\sqrt{2}}, 0,
0)$ are contained in ${\mathcal{H}}_1(F_4) \subseteq {\mathcal{H}}_2(F_4)$ and takes values $\pi_2({\mathbf{p}}) = 1$ and $\pi_2({\mathbf{p}}') = \frac 3 2$. We claim, that these values are exactly ${\delta_{\min}}$ and ${\delta_{\max}}$, respectively.
Note that $\pi_2({\mathbf{x}}) = g(x_1^2,x_2^2,x_3^2,x_4^2)$ for $$g({\mathbf{y}}) \ = \ 5 {s}_1({\mathbf{y}}) \cdot {s}_2({\mathbf{y}}) - 4 {s}_3({\mathbf{y}}).$$ Let $\Delta_3 = \{ {\mathbf{x}}\in {\mathbb{R}}^4 : x_1,\dots,x_4 \ge 0, x_1 + \cdots + x_4 =
1 \}$ be the standard $3$-simplex. We have that $\rho(S^3) = \Delta_3$ where $\rho(x_1,\dots,x_4) := (x_1^2,\dots,x_4^2)$. Hence, $${\delta_{\max}}\ = \ \max \{ g({\mathbf{p}}) : {\mathbf{p}}\in \Delta_3 \} \quad \text{ and }
\quad {\delta_{\min}}\ = \ \min \{ g({\mathbf{p}}) : {\mathbf{p}}\in \Delta_3 \}.$$ Now, $D_4$ is a subgroup of $F_4$ and $\pi_2 \in
{\mathbb{R}}[{s}_2,{s}_4,{s}_6]$ and does not depend on $e_4({\mathbf{x}})$. By Theorem \[thm:main\] for $D_4$, the varieties $S^3 \cap \{ \pi_2({\mathbf{x}}) =
{\delta_{\min}}\}$ and $S^3 \cap \{ \pi_2({\mathbf{x}}) = {\delta_{\max}}\}$ both meet ${\mathcal{H}}_3(D_4)$. Hence, it suffices to minimize or maximize $g({\mathbf{x}})$ over $$\Delta_3 \cap \{ {\mathbf{x}}\in {\mathbb{R}}^4 : x_1 = x_2 \}.$$ This leaves us with the (standard) task to maximize and minimize a bivariate polynomial $g'(s,t)$ of degree $3$ over a triangle. In the plane, the polynomial has $3$ critical points with values $1,
\frac{11}{9}, \frac{11}{9}$. On the boundary, the extreme values are attained at the points given above.
For the rank-$4$ reflection group $H_4$, the invariant $\pi_2({\mathbf{x}})$ is a polynomial of degree $12$ in four variables; see, for example, [@IKM Table 6]. Since $(B_1)^4$ is a reflection subgroup of $F_4$, $\pi_2$ is a polynomial in the squares $x_1^2,\dots,x_4^2$ and, following the argument in the proof above, we are left with minimizing and maximizing a degree-$6$ polynomial $g({\mathbf{x}})$ over the simplex $\Delta_3$. However, finding the critical points is not easy and an extra computational challenge is the fact that $g({\mathbf{x}})$ is a polynomial with coefficients in ${\mathbb{Q}}(\sqrt{5})$. <span style="font-variant:small-caps;">GloptiPoly</span> [@Glo] *numerically* computes ${\delta_{\min}}= -\frac{5}{16}$ and ${\delta_{\max}}= 1$. These values are attained at ${\mathbf{p}}_{\min} = \frac{1}{\sqrt{2}}(1,1,0,0)$ and ${\mathbf{p}}_{\max} = (1,0,0,0)$, respectively, and both points lie in ${\mathcal{H}}_2((B_1)^4)
\subseteq {\mathcal{H}}_2(H_4)$. This is strong evidence for the validity of Conjecture \[conj:main\] for $H_4$ but, of course, not a rigorous proof.
Strata of higher codimension {#sec:higher_codim}
============================
We have seen in the previous section that every nonempty $(n-1)$-sparse variety meets the hyperplane arrangement ${\mathcal{H}}_{n-1}$. In this section want to extend this result to $k$-sparse varieties for $k<n-1$. This case is considerably more difficult but we can make good use of the techniques and ideas developed in Section \[sec:codim1\].
Let $G$ be an essential finite reflection group acting on $V\cong{\mathbb{R}}^n$. Consider a $G$-invariant $k$-sparse variety $X$ with $k<n$. If $X$ is nonempty, then Theorem \[thm:codim1\] yields that for some reflection hyperplane $H \in {\mathcal{H}}$ the variety $X':=X\cap H$ is nonempty. An inductive argument could now replace $G$ by some other reflection subgroup $G' \subseteq
G$ that fixes $H$. If $X'$ remains sparse with respect to $G'$ we can again apply Theorem \[thm:codim1\] to obtain a point ${\mathbf{p}}\in {\mathcal{H}}_{n-2}(G') \subseteq {\mathcal{H}}_{n-2}(G)$. However, the results obtained using this strategy are far from optimal. We will briefly illustrate this for $G={\mathfrak{S}}_n$: Let $X$ be a nonempty $k$-sparse ${\mathfrak{S}}$-invariant variety for $k<n$. The largest subgroup of ${\mathfrak{S}}_n$ that fixes a given reflection hyperplane $H \in {\mathcal{H}}$ is $G'\cong{\mathfrak{S}}_{n-2}\times{\mathfrak{S}}_2$. Hence Theorem \[thm:codim1\] only applies for $X'=X\cap H$ and $G'$ if $k=d_k(G)<d_n(G')=n-2$, in other words if the original variety $X$ is $(k-3)$-sparse. Inductively, this yields that every nonempty $k$-sparse ${\mathfrak{S}}_n$-invariant variety meets ${\mathcal{H}}_l$ where $l=\lfloor
\frac{n+k}{2}\rfloor$. However, applying the above method to the exceptional types gives nontrivial bounds.
\[prop:fixH\] Let $6 \le n \le 8$. Then every nonempty $2$-sparse $E_n$-invariant variety intersects ${\mathcal{H}}_{n-2}(E_n)$.
We exemplify the argument for the case $n=8$. Let $X$ be a nonempty $2$-sparse $E_8$-invariant variety. By Theorem \[thm:codim1\] we find a point ${\mathbf{p}}\in X \cap {\mathcal{H}}_7(E_8)$. The orbit of ${\mathbf{p}}$ meets every hyperplane in ${\mathcal{H}}(E_8)$ (see [@hum Sect. 2.10]) and hence we may assume that ${\mathbf{p}}$ lies on the hyperplane $H=\{{\mathbf{x}}\in{\mathbb{R}}^8:x_1=x_2\}$. Consider the subgroup $G'\cong D_6\subset E_8$ acting essentially on the coordinates $x_3,\dots,x_8$. Since $d_6(D_6)=10>8=d_2(E_8)$, we can apply Theorem \[thm:codim1\] to finish the proof.
By restricting the class of invariant polynomials, we obtain better bounds than those in Proposition \[prop:fixH\]. In the following, a point ${\mathbf{p}}$ is called [**$\boldsymbol G$-general**]{} if it does not lie on any reflection hyperplane of $G$, and hence $|G{\mathbf{p}}| = |G|$.
For a positive integer $d$, let ${\delta_G}(d)$ be the largest number $\ell$ such that for every $p \in {\mathcal{H}}_{\ell+1}$ there is a reflection subgroup $G' \subseteq G$ such that $p$ is $G'$-general and $2d_n(G') > d$. Moreover, we define ${\sigma_G}(k) := {\delta_G}(2d_k(G))$. That is, ${\sigma_G}(k)$ is the largest $0 \le \ell \le d$ such that for every $p
\in {\mathcal{H}}_{\ell+1}$, there is a reflection subgroup $G' \subseteq G$ such that $p$ is $G'$-general and $d_n(G')> d_k(G)$.
We call an invariant polynomial $f\in{\mathbb{R}}[V]^G$ [**$\boldsymbol G$-finite**]{} if either $V_{\mathbb{R}}(f)={\varnothing}$ or if there is a point ${\mathbf{p}}\in V_{\mathbb{R}}(f)$ such that $f$ has finitely many extreme points restricted to the sphere $K = \{ {\mathbf{q}}\in V :
\|{\mathbf{q}}\| = \|{\mathbf{p}}\|\}$.
\[thm:HalfDeg\]\[thm:DegPrinc\] Let $f \in {\mathbb{R}}[V]^G$ be a $G$-finite polynomial and $X={\mathrm{V}_{{\mathbb{R}}}}(f)$. If $f\in{\mathbb{R}}[\pi_1,\dots,\pi_k]$, then $$X \ \neq \ {\varnothing}\quad \text{ if and only if } \quad
X \cap {\mathcal{H}}_{{\sigma_G}(k)} \ \neq \ {\varnothing}.$$ If $d = \deg(f)$, then $$X \ \neq \ {\varnothing}\quad \text{ if and only if } \quad
X \cap {\mathcal{H}}_{{\delta_G}(d)} \ \neq \ {\varnothing}.$$
We only give a proof for the second result. The proof of the first is analogous. Suppose ${\mathrm{V}_{{\mathbb{R}}}}(f) \neq {\varnothing}$. We may assume that $f(0)
\le 0$ and, since ${\mathcal{H}}_{{\delta_G}(d)}$ is connected, it suffices to show that there is some point ${\mathbf{p}}_+ \in {\mathcal{H}}_{{\delta_G}(d)}$ with $f({\mathbf{p}}_+) \ge 0$.
By assumption, there is a zero ${\mathbf{p}}_0 \in {\mathrm{V}_{{\mathbb{R}}}}(f)$ such that $f$ has only finitely many extreme points restricted to $K = \{ {\mathbf{q}}: \|{\mathbf{q}}\| =
\|{\mathbf{p}}_0\|\}$. Let ${\mathbf{p}}_+ \in K$ be a point maximizing $f$ over $K$ and hence $f({\mathbf{p}}_+) \ge f({\mathbf{p}}_0) \ge 0$. We claim that ${\mathbf{p}}_+ \in
{\mathcal{H}}_{{\delta_G}(d)}$. Otherwise, there is a reflection subgroup $G' \subset G$ such that ${\mathbf{p}}_+ \not \in {\mathcal{H}}(G')$ and $2d_n(G') > d$. Let $\pi'_1,\dots,\pi'_n$ be a choice of basic invariants of $G'$ and, without loss of generality, $\pi'_1({\mathbf{x}}) = \|{\mathbf{x}}\|^2$. Thus, ${\overline{{\mathbf{p}}}}_+ = \pi'({\mathbf{p}}_+)$ is in the interior of ${\mathcal{S}}= \pi'(V)$. We can write $f =
F(\pi'_1,\dots,\pi_n)$ for some $F \in {\mathbb{R}}[y_1,\dots,y_n]$. On the level of orbit spaces, our assumption states that restricted to $\overline{K} =
\pi(L) = \{ {\mathbf{y}}\in {\mathcal{S}}: y_1 = \pi_1({\mathbf{p}}_+)\}$, the polynomial $F$ has only finitely many extreme points. However, $F$ is linear in $y_n$ and thus ${\overline{{\mathbf{p}}}}_+$ is a maximum only if ${\overline{{\mathbf{p}}}}_+ \in \partial \overline{K}
\subseteq \partial {\mathcal{S}}$. This is a contradiction.
Bounds from parabolic subgroups {#sec:StrNum}
===============================
The numbers ${\delta_G}(d)$ and ${\sigma_G}(k)$ defined in Section \[sec:higher\_codim\] are difficult to compute in general. In this section we compute upper bounds on these numbers coming from parabolic subgroups. Following Humphreys [@hum], we represent a reflection group by its Dynkin diagram.
\[lem:leaf\] Fix a finite irreducible reflection group $G$ with Dynkin diagram $D$. Let $D' \subset D$ be a subdiagram obtained by removing a node from $D$ and let $H \in {\mathcal{H}}(G)$ be a reflection hyperplane. Then there is a parabolic subgroup $W \subset G$ with Dynkin diagram $D'$ and $H$ is not a reflection hyperplane of $W$.
Let $W$ be a parabolic subgroup with Dynkin diagram $D'$. Since every parabolic subgroup with Dynkin diagram $D'$ is conjugate to $W$, it suffices to show that there is a $g \in G$ such that $g H \not\in {\mathcal{H}}(W)$.
In terms of the roots $\Phi(G)$ of $G$, this is equivalent to showing that that for every $\alpha \in \Phi(G)$, there is a $g \in G$ such that $g\alpha \not \in \Phi(W)$. Since $G$ acts transitively on roots of the same length (see Humphreys [@hum]) and $\Phi(W) \subsetneq \Phi(G)$ for any proper parabolic subgroup, this yields the claim for all types except $F_4$, $B_n$ and $I_2(2m)$ for $m > 1$. For $F_4$, the possible proper parabolic subgroups are $B_3$ and $A_1 \times A_2$ and the result follows by inspection. For $I_2(2m)$, there are two orbits with each $2m
\ge 4$ roots whereas the only nontrivial proper parabolic subgroup is $A_1$. For $B_n$, this follows from counting the number of elements in each of the two orbits.
The lemma yields the following result about finite reflection groups that might be interesting in its own right.
\[subdiagram\] Let $G$ be a finite irreducible reflection group with Dynkin diagram $D$ acting on a real vector space $V$. For $k \ge 1$, let ${\mathbf{p}}\in {\mathcal{H}}_k
\setminus {\mathcal{H}}_{k-1}$ and $D' \subset D$ be a connected subdiagram on $k$ nodes. Then there is a parabolic subgroup $W \subset G$ with Dynkin diagram $D'$ such that ${\mathbf{p}}$ is $W$-general.
We argue by induction on $s = \dim V - k$. For $s = 0$, ${\mathbf{p}}\in V
\setminus {\mathcal{H}}_{\dim V -1}$ and ${\mathbf{p}}$ is by definition $G$-general. Otherwise, let $D_1
\subset D$ be a subdiagram obtained by removing a leaf such that $D'
\subseteq D_1$ and let $H_1$ be a reflection hyperplane of $G$ containing ${\mathbf{p}}$. We may use Lemma \[lem:leaf\] to obtain a parabolic subgroup $W_1$ with Dynkin diagram $D_1$ and not containing $H_1$ as a reflection hyperplane. In particular, ${\mathbf{p}}$ is contained in precisely $s-1$ linearly independent reflection hyperplanes of $W_1$. By induction, there is a parabolic subgroup $W \subseteq W_1$ with Dynkin diagram $D'$ for which ${\mathbf{p}}$ is $W$-general. In particular, $W$ is a parabolic subgroup of $G$ which concludes the proof.
For a choice of a simple system $\Delta \subseteq \Phi(G)$, let us write $W_I$ for the standard parabolic subgroup generated by $I \subseteq \Delta$. We define $${\tilde{\delta}_G}(d) \ := \ \min\{ |I|-1 : I\subseteq \Delta, 2d_n(W_I) > d \},$$ and, analogously, we define ${\tilde{\sigma}_G}(k) := {\tilde{\delta}_G}(2d_k(G))$. Proposition \[subdiagram\] implies the following bound on ${\delta_G}$.
\[cor:parabolic\_bound\] $ {\delta_G}(d) \le {\tilde{\delta}_G}(d), $ for all $d \ge 0$.
The clear advantage is a simple way to compute upper bounds on ${\delta_G}(d)$ from the knowledge of (standard) parabolic subgroups of reflection groups; cf. [@hum]. The explicit values are given in Table \[tab:parNum\]. However, not every reflection subgroup is parabolic (e.g., $I_2(m) \subseteq I_2(2m)$). Nevertheless, we conjecture that ${\delta_G}(d)$ is attained at a parabolic subgroup.
For any finite reflection group $G$ $${\delta_G}(d) \ = \ {\tilde{\delta}_G}(d)$$ for all $d$.
$$\def\arraystretch{1.2}
\begin{array}{c|r@{{\text{ --- }}}l|c|c|c||r@{{\text{ --- }}}l|c}
G & \multicolumn{2}{c|}{d} & {\tilde{\delta}_G}(d) & W&d_n(W) & {\multicolumn{2}{c|}{k}} & {\tilde{\sigma}_G}(k)\\
\hline
\hline
A_{n-1} / \mathfrak{S}_n & 0 & 2n-1 & \lfloor d/2\rfloor& A_{\lfloor d/2\rfloor}&\lfloor d/2\rfloor+1 & 0 & n-1 & k\\
\hline
B_n & 0 & 4n-1 & \lfloor d/4\rfloor& B_{\lfloor d/4\rfloor+1} & 2(\lfloor d/4\rfloor+1)& 0 & n-1 & k\\
\hline
D_n& 0 & 4n-5 &\lfloor d/4\rfloor+1 & D_{\lfloor d/4\rfloor+2} &2(\lfloor d/4\rfloor+1) & 0 & \lfloor\frac{n}{2}\rfloor & k+1\\
&\multicolumn{1}{c}{}&&&&& \lfloor\frac{n}{2}\rfloor+1 & n & k\\
\hline
I_2(m) & 1 & 2m-1 & 1&I_2(m) & m& {\multicolumn{2}{c|}{1}} & 1\\
\hline
E_6 & 1 & 5 & 1&A_2&3& {\multicolumn{2}{c|}{1}} & 1\\
& 6 & 7 & 2&A_3&4& {\multicolumn{2}{c|}{2}} & 3\\
& 8 & 11 & 3&D_4&6& {\multicolumn{2}{c|}{3}} & 4\\
& 12 & 15 & 4&D_5&8& {\multicolumn{2}{c|}{4}} & 5\\
& 16 & 23 & 5&E_6&12& {\multicolumn{2}{c|}{5}} & 5\\
\hline
E_7 & 1 & 5 & 1&A_2&3& {\multicolumn{2}{c|}{1}} & 1\\
& 6 & 7 & 2&A_3&4& {\multicolumn{2}{c|}{2}} & 4\\
& 8 & 11 & 3&D_4&6& {\multicolumn{2}{c|}{3}} & 5\\
& 12 & 15 & 4&D_5&8& {\multicolumn{2}{c|}{4}} & 5\\
& 16 & 23 & 5&E_6&12& {\multicolumn{2}{c|}{5}} & 6\\
& 24 & 35 & 6&E_7&18& {\multicolumn{2}{c|}{6}} & 6\\
\hline
E_8 & 1 & 5 & 1&A_2&3& {\multicolumn{2}{c|}{1}} & 1\\
& 6 & 7 & 2&A_3&4& {\multicolumn{2}{c|}{2}} & 5\\
& 8 & 11 & 3&D_4&6& {\multicolumn{2}{c|}{3}} & 6\\
& 12 & 15 & 4&D_5&8& {\multicolumn{2}{c|}{4}} & 6\\
& 16 & 23 & 5&E_6&12& {\multicolumn{2}{c|}{5}} & 7\\
& 24 & 35 & 6&E_7&18& {\multicolumn{2}{c|}{6}} & 7\\
& 36 & 59 & 7&E_8&30& {\multicolumn{2}{c|}{7}} & 7\\
\hline
F_4 & 1 & 7 & 1&B_2&4& {\multicolumn{2}{c|}{1}} & 1\\
& 8 & 11 & 2&B_3&6& {\multicolumn{2}{c|}{2}} & 3\\
& 12 & 23 & 3&F_4&12& {\multicolumn{2}{c|}{3}} & 3\\
\hline
H_3 & 1 & 9 & 1&I_2(5)&5& {\multicolumn{2}{c|}{1}} & 1\\
& 10 & 19 & 2&H_3&10& {\multicolumn{2}{c|}{2}} & 2\\
\hline
H_4 & 1 & 9 & 1&I_2(m)&5& {\multicolumn{2}{c|}{1}} & 1\\
& 10 & 19 & 2&H_3&10& {\multicolumn{2}{c|}{2}} & 3\\
& 20 & 59 & 3&H_4&30& {\multicolumn{2}{c|}{3}} & 3\\
\end{array}
\quad \quad \quad
$$
Test sets for Lie groups {#sec:lie}
========================
In this last section, we extend some of our results to polynomials invariant under the action of a Lie group. More precisely, we consider the case of a real simple Lie group $G$ with the adjoint action on its Lie algebra ${\mathfrak{g}}$.
We illustrate our results for the case $G = {\mathrm{SL}}_n$. Its Lie algebra ${\mathfrak{sl}}_n$ is the vector space of real $n$-by-$n$ matrices of trace $0$. The adjoint action of ${\mathrm{SL}}_n$ on ${\mathfrak{sl}}_n$ is by conjugation: $g \in {\mathrm{SL}}_n$ acts on $A \in {\mathfrak{sl}}_n$ by $g\cdot A := gAg^{-1}$. The following description of its ring of invariants is well-known. We briefly recall the standard proof which immediately suggests a connection to our treatment of reflection groups.
\[thm:GL\] For $G = {\mathrm{SL}}_n$ $${\mathbb{R}}[{\mathfrak{sl}}_n]^G \ = \ {\mathbb{R}}[{s}_2,\dots,{s}_n],$$ where ${s}_k(A) = {\mathrm{tr}}(A^k)$ for $k=2,\dots,n$. Moreover, ${s}_2,\dots,{s}_n$ are algebraically independent.
We write $D \subset {\mathfrak{sl}}_n$ for the set of diagonalizable matrices and we denote by $\lambda(A) = (\lambda_1,\dots,\lambda_n)$ the eigenvalues of $A
\in D$. Then for any $A \in D$ $${s}_k(A) \ = \ {s}_k(\lambda(A)) \ = \ \lambda_1(A)^k + \lambda_2(A)^k +
\cdots + \lambda_n(A)^k$$ and ${s}_2,\dots,{s}_n$ are simply the power sums restricted to the linear subspace $\Delta \subset D$ of diagonal matrices. This shows that ${s}_2,\dots,{s}_n$ are algebraically independent. Now for a polynomial $f({\mathbf{X}}) \in {\mathbb{R}}[{\mathfrak{sl}}_n]$ invariant under the action of ${\mathrm{SL}}_n$, the restriction to $\Delta \cong {\mathbb{R}}^{n-1}$ is a polynomial $f({\mathbf{x}})$ that is invariant under $A_{n-1}$. Hence $f({\mathbf{x}}) = F({s}_2({\mathbf{x}}),\dots,{s}_n({\mathbf{x}}))$ for some $F \in {\mathbb{R}}[y_2,\dots,y_n]$. The polynomial $\tilde f({\mathbf{X}}) =
F({s}_2({\mathbf{X}}),\dots,{s}_n({\mathbf{X}}))$ is invariant under ${\mathrm{SL}}_n$ and agrees with $f$ on $D$. Since $D$ contains a nonempty open set, $f = \tilde f$ as required.
For the special orthogonal group ${\mathrm{SO}}_{n}$, its Lie algebra ${\mathfrak{so}}_n \subset
{\mathfrak{sl}}_n$ is the vector space of skew-symmetric $n$-by-$n$ matrices on which ${\mathrm{SO}}_{n}$ acts by conjugation. If $n = 2k+1$, then the corresponding Weyl group is $B_k$ and $D_k$ if $n=2k$. Hence ${\mathbb{R}}[{\mathfrak{so}}_{2k+1}]^{{\mathrm{SO}}_{2k+1}}$ is generated by ${s}_2({\mathbf{X}}),{s}_4({\mathbf{X}}),\dots,{s}_{2k}({\mathbf{X}})$. For $n=2k$, a minimal generating set is given by ${s}_2({\mathbf{X}}), {s}_4({\mathbf{X}})$, ..., ${s}_{2n-2}({\mathbf{X}})$ and the Pfaffian ${\mathrm{pf}}({\mathbf{X}}) = \sqrt{\det {\mathbf{X}}}$.
Analogously to the case of reflection groups, we call a $G$-invariant variety $X\subseteq{\mathfrak{g}}$ $k$-sparse if $X = {\mathrm{V}_{{\mathbb{R}}}}(f_1,\dots,f_m)$ for some $f_1,\dots,f_M \in {\mathbb{R}}[\pi_1,\dots,\pi_k]$ The discriminant locus ${\mathcal{D}}\subset
{\mathfrak{g}}$ of the Lie group $G$ are the points with nontrivial stabilizer. This is exactly the orbit of the corresponding reflection arrangement ${\mathcal{H}}\subset {\mathfrak{g}}$ under $G$. Together with Theorem \[thm:GL\] and the result by Steinberg [@stein], this is a real $G$-invariant hypersurface. This yields a stratification of ${\mathfrak{g}}$ by defining ${\mathcal{D}}_i$ to be the orbit of ${\mathcal{H}}_i$, which corresponds to the points for which the discriminant vanishes up to order $n-i$. Hence, the results from the previous sections generalize to Lie groups. Theorem \[thm:main\] yields the following.
\[thm:Lie\_codim1\] Let $G \in \{ {\mathrm{SL}}_k, {\mathrm{SO}}_k : k \in {\mathbb{Z}}_{\ge 1} \}$ and let $X\subseteq{\mathfrak{g}}$ be $G$-invariant and $k$-sparse. If $X$ is nonempty it intersects ${\mathcal{D}}_k$.
For suitable Lie groups $G$, Theorem \[thm:Lie\_codim1\] gives a first relation between real varieties invariant under the action of $G$ and the discriminant locus and Conjecture \[conj:main\] is reasonable for this setting. It would be very interesting to explore this connection further.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Andrew S. Lan\
\
Christoph Studer\
\
Richard G. Baraniuk\
\
bibliography:
- 'sparfaclustbib.bib'
title: Quantized Matrix Completion for Personalized Learning
---
Personalized learning, learning analytics, content analytics, factor analysis, matrix completion, convex optimization.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We extend our recent study of chiral magnetic effect in relativistic heavy ion collisions based on an anomalous transport model by including also the chiral vortical effect. We find that although vorticities in the chirally restored quark matter, which result from the large angular momentum in non-central collisions, can generate an axial charge dipole moment in the transverse plane of a heavy ion collision, it does not produce a difference in the eccentricities of negatively and positively charged particles. As a result, including the chiral vortical effect alone cannot lead to a splitting between the elliptic flows of negatively and positively charged particles. On the other hand, negatively and positively charged particles do develop a splitting in their elliptic flows if the effect due to a strong and long-lived magnetic field is also included. However, to have a positive slope in the dependence of the elliptic flow splitting on the charge asymmetry of the quark matter, as seen in experiments, requires the neglect of the effect of the Lorentz force. In this case, an elliptic flow splitting appears even at vanishing charge asymmetry.'
author:
- Yifeng Sun
- Che Ming Ko
bibliography:
- 'ref.bib'
title: Study of chiral vortical and magnetic effects in the anomalous transport model
---
introduction
============
Novel transport phenomena induced by the chiral anomaly have recently generated a lot of interests in various areas of physics [@Charbonneau:2009ax; @PhysRevB.86.115133; @PhysRevB.89.035142]. One such phenomenon is the chiral magnetic (separation) effect (CME/CSE) in which vector (axial) charges of massless fermions become spatially separated in the presence of a magnetic field if the axial (vector) charge chemical potential is nonzero [@PhysRevD.70.074018; @PhysRevD.72.045011; @Kharzeev2008227; @PhysRevD.78.074033; @Kharzeev2010205]. Another is the chiral vortical effect (CVE) in which a vector current appears in a rotating system when both the vector and axial chemical potentials are nonzero [@PhysRevLett.103.191601]. Moreover, through the interplay between the vector and axial charge fluctuations, two gapless collective excitations called the chiral magnetic wave (CMW) and the chiral vortical wave (CVM) can be induced [@PhysRevLett.107.052303; @PhysRevD.92.071501].
In relativistic heavy ion collisions, a quark-gluon plasma (QGP) consisting of de-confined quarks and gluons can be produced. Because of the restoration of chiral symmetry in the QGP, quarks and antiquarks become essentially massless. Non-central heavy ion collisions, in which both a strong magnetic field and a large vorticity are produced in the QGP as a result of the large initial orbital angular momentum [@PhysRevC.83.054911; @PhysRevC.85.044907], thus provide the opportunity to study above mentioned phenomena. By solving the wave equation for the charge densities of particles of right- and left-chiralities in a schematic model [@PhysRevLett.107.052303; @PhysRevD.92.071501], it has been found that an electric or baryon charge quadrupole moment is produced, which can then lead to a splitting between the elliptic flows of positively and negatively charged particles as observed in experiments. More realistic studies of CMW and CVW in heavy ion collisions have been carried out using the anomalous hydrodynamics, which extends the normal hydrodynamics by including also the axial charge current in QGP [@PhysRevLett.103.191601]. Also, chiral kinetic equations, which include the anomalous transport of massless fermions and the non-equilibrium effect, have been developed [@PhysRevLett.109.162001; @Son:2012wh; @Son:2012zy; @PhysRevLett.109.232301; @PhysRevLett.110.262301; @Manuel:2014dza]. However, different methods have been proposed in including the chiral vortical effect in the chiral kinetic approach. For example, in Refs. [@PhysRevLett.109.232301; @PhysRevLett.110.262301], the vorticity effect is included in the equations of motion for massless fermions, while in Refs. [@PhysRevLett.113.182302; @PhysRevLett.115.021601], the chiral vortical effect is taken into account via the jump current resulting from the side jump that is introduced in the scattering of massless fermions to conserve the angular momentum.
Using the anomalous transport model for massless quarks and antiquarks, we have recently studied the effect of a magnetic field on the elliptic flows of quarks and antiquarks in relativistic heavy ion collisions [@PhysRevC.94.045204] . With initial conditions from a blast wave model and assuming that the strong magnetic field produced in non-central heavy ion collisions can last for a sufficiently long time, we have obtained an appreciable electric quadrupole moment in the transverse plane of a heavy ion collision. The electric quadrupole moment subsequently leads to a splitting between the elliptic flows of quarks and antiquarks during the expansion of the system. The slope in the charge asymmetry dependence of the elliptic flow difference between negatively and positively charged particles is found to be positive as expected from the chiral magnetic wave in the QGP and observed in experiments at the BNL Relativistic Heavy Ion Collider (RHIC). This result is, however, obtained only if we neglect the Lorentz force acting on the charged particles and assume that the quark-antiquark scattering is dominated by the chirality-changing channel. In the present study, we extend our previous study by including also the vorticities in the partonic matter and investigate how our previous results, particularly the splitting between the elliptic flows of negatively and positively charged particles, are affected.
This paper is organized as follows. In the next section, we describe the chiral equations of motion that include both effects of the magnetic field and the finite vorticity in the QGP. We then give in Sec. III the initial conditions for the partonic matter produced in non-central heavy ion collisions at the top energy of RHIC. In Sec. IV, we present and discuss the results. Finally, a summary is given in Sec. V.
The chiral equations of motion
==============================
For the chiral kinetic equation that includes both the magnetic and vorticity fields, we follow that obtained in Ref. [@PhysRevLett.110.262301] using the covariant Wigner function approach for massless spin-1/2 fermions in four dimensions. After integrating over the energy in the Lorentz covariant chiral kinetic equation, a chiral kinetic equation in three dimensions is obtained. The resulting chiral equations of motion for massless quarks and antiquarks are given by $$\begin{aligned}
&&\sqrt{G}\dot{\mathbf{r}}=\hat{\mathbf{p}}+Qh(\hat{\mathbf{p}}\cdot\mathbf{b})\mathbf{B}+h\frac{\boldsymbol{\omega}}{p},
\label{CKM}
\\&&\sqrt{G}\dot{\mathbf{p}}=Q\hat{\mathbf{p}}\times\mathbf{B},
\label{LF}
\\&&\sqrt{G}=1+Qh\mathbf{b}\cdot\mathbf{B}+4hp(\mathbf{b}\cdot\boldsymbol{\omega}).
\label{phase}\end{aligned}$$ In the above, $\mathbf{b}=\frac{\mathbf{p}}{2p^3}$ is the Berry curvature due to a vector potential in the momentum space, $h$ is the helicity of the particle, and $\boldsymbol{\omega}=\frac{1} {2}\boldsymbol{ \nabla}\times\mathbf{u}$ is the vorticity with ${\bf u}$ being the velocity field.
As shown in our previous study of CMW based on the anomalous transport model [@PhysRevC.94.045204], the chirality-changing scattering (CCS) between quark and antiquark is essential for generating the elliptic flow difference between negatively and positively charged particles. We thus also allow in the present study massless quarks and antiquarks of same chiralities to undergo the CCS scattering. We neglect, however, the effect from the change of phase space, i.e., Eq.(\[phase\]), due to the magnetic field and/or the vorticity field as in Ref. [@PhysRevC.94.045204].
Initial conditions of non-central heavy ion collisions
======================================================
For the initial conditions of a heavy ion collision, we use same ones in our previous study [@PhysRevC.94.045204], i.e., a fireball with geometry corresponding to collisions of Au+Au at $\sqrt{s_{NN}}=200$ GeV at centralities of 30-40$\%$. Specifically, the number density distribution is taken to have the Woods-Saxon form $$\begin{aligned}
\rho(x,y)=\frac{\rho_0}{1+e^{\frac{\sqrt{x^2+y^2/c^2}-R}{a}}},\end{aligned}$$ where $\rho_0=13$ fm$^{-3}$ is the central density of quarks and antiquarks, $c=1.5$ describes the spatial anisotropy of produced partonic matter in the transverse plane of non-central heavy ion collisions, $R=3.5$ fm is the radius, and $a=0.5$ fm is the surface thickness. The transverse momentum distributions of quarks and antiquarks are given by the Boltzmann distributions of temperature $T$ and charge chemical potential $\mu$ with $\mu/T$ being uniform in space as in our previous study. We determine the temperature from the density of quarks and antiquarks by assuming they are noninteracting and the charge chemical potential via $\mu/T=A_\pm$, where $A_\pm=\frac{N_{+}-N_{-}}{N_{+}+N_{-}}$ is the charge asymmetry of the partonic matter with $N_+$ and $N_-$ being the numbers of positively and negatively charged quarks, respectively.
For both the magnetic field and the vorticity field produced in these collisions, we take their directions along the $y$-axis perpendicular to the reaction plane, and assume that their strengths are uniform in space [@PhysRevLett.109.202303] and have the following time dependence: $$\begin{aligned}
eB&=&\frac{eB_0}{1+(t/\tau_B)^2}\\
\omega&=&\frac{\omega_0}{1+(t/\tau_{\omega})^2}.\end{aligned}$$ We choose $eB_0=7m_{\pi}^2$ and $\tau_B=6$ fm/$c$ for the magnetic field as in Ref. [@PhysRevC.94.045204], and $\omega_0=0.1$ fm$^{-1}$ and $\tau_{\omega}=2.7$ fm/$c$ for the vorticity field. The latter is similar to those of Ref. [@PhysRevC.94.044910] based on a multipase tranport (AMPT) model [@Lin:2004en] for Au+Au collision at 200 GeV and centrality 30-40$\%$, except that we take the vorticity field in heavy ion collisions as a spatially uniform external field.
In modeling the time evolution of the partonic matter, massless quarks and antiquarks not only follow the chiral equation of motion \[Eq. (\[CKM\])\] and subject to the Lorentz force \[Eq. (\[LF\])\] but also undergo scattering, particularly the chirality-changing scattering between quark and antiquark of same chirality. As in our previous study, we take these scattering cross sections to have the temperature dependence $\sigma=\sigma_0(T_0/T)^3$ [@Ghosh:2015mda] in order to model the specific viscosity $\eta/s$ of the QGP, where $\eta$ and $s$ are the viscosity and entropy, and its temperature dependence obtained from the lattice QCD and experiments [@Heinz:2013th]. However, when the temperature of the partonic matter drops below the phase transition temperature $T_C=150$ MeV, we adopt the duality ansatz as in our previous study [@PhysRevC.94.045204] by relabeling each parton as a pion of same sign in charge and evolve the resulting pionic matter using the normal equations of motion and the empirical pion-pion scattering cross sections [@Li:1995pra] until the kinetic freeze out temperature $T_f=120$ MeV.
Results
=======
In the present section, we study the time evolution of the partonic matter by following the trajectories of massless quarks and antiquarks according to the chiral equations of equations \[Eqs. (\[CKM\]) and (\[LF\])\]. We consider the three cases of including only the vorticity field (CVE), both magnetic and vorticity fields without the Lorentz force (CVE+CME), and both magnetic field and vorticity fields with the Lorentz force (CVE+CME+LF). In all cases, the chirality-changing scattering (CCS) between quark and antiquark of same chirality is included, although it has no effect in the case of CVE only.
Pion elliptic flow
------------------
![(Color online) Elliptic flow of kinetically freeze-out pions in midrapidity ($|y|\le1$) as a function of transverse momentum for the three cases of including both chiral vortical and magnetic effects and the effect due to the Lorentz force (CVE+CME+LF), without the Lorentz force (CVE+CME), and with only the vortical effect (CVE). Experimental data (solid circles) are from Ref. [@PhysRevC.72.014904].[]{data-label="pion"}](ptv2.eps){width="50.00000%"}
We first show in Fig. \[pion\] the elliptic flow of kinetically freeze-out pions in midrapidity ($|y|\le1$) as a function of transverse momentum. The dashed line is obtained with the coefficient $\sigma_0=13.6$ mb in the parton scattering cross section for the case of including the chiral effects due to both the vorticity and magnetic fields as well as the Lorentz force (CVE+CME+LF). The result obtained without the Lorentz force (CVE+CME) and using $\sigma_0=15.5$ mb is shown by the solid line, while the result including only the chiral vortical effect (CVE) and using $\sigma_0=13.3$ mb is shown by the dotted line. The transverse momentum dependence of the pion elliptic flow is seen to agree with the experimental data from the STAR Collaboration [@PhysRevC.72.014904] for all three cases. The integrated $v_2$ of pions with transverse momenta in the range of $0.15\le p_T\le0.5$ GeV/$c$ is 0.036 in all three cases, which also agrees with the experimental value.
Effect of the vorticity field
-----------------------------
In Ref. [@PhysRevD.92.071501], it is argued that the chiral vortical wave (CVW) generated in a rotating fluid system can lead to a splitting between the elliptic flows of baryons and anti-baryons if the baryon chemical potential of the system is non-zero. Similarly, for non-zero charge chemical potential, we expect the CVW to lead to a splitting between the elliptic flows of negatively and positively charged particles. In this section, we use the anomalous transport model to study the effect of the vorticity field on the elliptic flow splitting of $\pi^+$ and $\pi^-$ for different values of charge chemical potential by taking the magnetic field to be zero, i.e., $B_0=0$.
![(Color online) Eccentricity and elliptic flow difference between negatively and positively charged particles as functions of time for total charge asymmetry of quark matter $A_{\pm}=0.16$.[]{data-label="cvw"}](delta.eps){width="50.00000%"}
Figure \[cvw\] shows the time evolution of the differences $\Delta \epsilon_2=\epsilon_{2-}-\epsilon_{2+}$ between the eccentricities \[$\epsilon_2=\langle (x^2-y^2)/(x^2+y^2)\rangle$\] and $\Delta v_2=v_{2-}-v_{2+}$ between the elliptic flows \[$v_2=\langle (p_x^2-p_y^2)/(p_x^2+p_y^2)\rangle$\] of negatively and positively charged particles for the total charge asymmetry $A_{\pm}=0.16$ of these particles. The transverse momenta are again in the range of $0.15\le p_T\le 0.5$ GeV/$c$. These differences are consistent with zero, indicating that the CVW alone does not lead to any eccentricity and elliptic flow differences between charged particles.
The above result can be understood as follows. According to the chiral equation of motion \[Eqs. (\[CKM\])\] for the case without magnetic field, the vorticity field modifies the velocities of particles of opposite helicities by the same amount but with opposite sign. As a result, nonzero positive and negative axial charge chemical potentials $\mu_5$ appear in the positive and negative $y$ regions of the transverse plane, respectively. However, the average change in the velocity of positively charged particles is the same as that of negatively charged particles with both given by $\Delta v=v\frac{e^{\mu_R/T}-e^{\mu_L/T}}{e^{\mu_R/T}+e^{\mu_L/T}}$, where $v$ is the magnitude of average modified velocity of particles of right or left helicity, and $\mu_R$ and $\mu_L$ are chemical potentials of right and left chiralities, respectively. Negatively and positively charged particles thus have same spatial distributions if their initial spatial distributions are the same. Therefore, no eccentricity difference between negatively and positively charged particle appears even though an additional charge quadrupole moment is generated in the transverse plane by the CVW. The effect of CVW is thus different from that of CMW, which causes the average change in the velocities of negatively and positively charged particles to have opposite sign given by $\Delta v$ and $-\Delta v$, respectively, and can thus result in a larger eccentricity for negatively charged particles than for positively charged particles.
Our result obtained with only the vorticity field differs from that of Ref. [@PhysRevD.92.071501] based on schematic considerations. Using the argument for the effect of the chiral magnetic wave introduced in Ref. [@PhysRevLett.107.052303] based on consideration of the net charge distribution, it is argued in Ref. [@PhysRevD.92.071501] that the additional charge quadrupole moment generated by the vorticity field would lead to an elliptic flow splitting between negatively and positively charged particles. As discussed in the previous paragraph, since the spatial distributions of negatively and positively charged particles remains the same, the additional charge quadrupole moment does not lead to different eccentricities between negatively and positively charged particles and thus cannot generate a splitting between their elliptic flows. This is different from the case with only the magnetic field, where the additional quadrupole momentum generated by the chiral magnetic wave can indeed lead to different eccentricities and thus elliptic flows between negatively and positively charged particles as shown in our previous study [@PhysRevC.94.045204].
![(Color online) Axial charge chemical potential $\mu_5/T$ distribution in the transverse plane $z=0$ at time $t=5$ fm/$c$ for partonic matter of zero charge asymmetry.[]{data-label="dipole"}](u5t.eps){width="50.00000%"}
Although the CVW does not lead to the eccentricity and elliptic flow splittings between negatively and positively charged particles, it does result in a large axial charge dipole moment in the transverse plane even for quark matter of vanishing charge asymmetry $A_{\pm}=0$. As shown in Fig. \[dipole\], the value is even larger than that in Ref. [@PhysRevC.94.045204] due to the CMW based on a magnetic field of strength $eB=7m_{\pi}^2$. To understand this, we note that the axial charge current induced by the magnetic and vorticity fields is [@PhysRevLett.103.191601] $$\begin{aligned}
\mathbf{j}_5\propto\left(\frac{1}{6}T^2+\frac{1}{2\pi^2}(\mu^2+\mu_5^2)\right)\boldsymbol{\omega}+\frac{Q}{2\pi^2}\mu\mathbf{B}.\end{aligned}$$ With $T=0.3$ GeV, $\mu=\mu_5=0$, $\omega=0.02$ GeV in the present study, the axial charge current is 0.0375/fm$^3$, which is larger than the value 0.02/fm$^3$ obtained with $\mu/T=0.16$ and $QB=3.5m_{\pi}^2$ used in our previous study at same temperature [@PhysRevC.94.045204].
**Effect of vorticity field plus magnetic field without the Lorentz force**
---------------------------------------------------------------------------
![(Color online) Eccentricity difference between negatively and positively charged particles as a function of time for partonic matter of charge asymmetries $A_{\pm}=0$ and $A_{\pm}=0.16$.[]{data-label="cmw1"}](de21.eps){width="50.00000%"}
Including also the magnetic field but still neglecting the Lorentz force in the chiral equations of motion, we have studied the time evolution of the difference between the eccentricities of negatively and positively charged particles for partonic matter of charge asymmetries $A_{\pm}=0$ and $A_{\pm}=0.16$. Results for particles of momenta in the range $0.15\le p_T\le0.5$ GeV/$c$ are shown in Fig. \[cmw1\]. It is seen that the eccentricity difference increases with time in both cases and is nonzero even for zero charge asymmetry, indicating that effects of the vorticity field and the magnetic field are not additive because the eccentricity difference is zero in both cases for zero charge asymmetry. The latter is due to the finite axial charge current induced by the vorticity field, which leads to an axial charge dipole moment in the transverse plane, characterized by positive and negative axial charge chemical potentials in the positive and negative $y$ regions of the transverse plane, respectively. Because of the chiral magnetic effect, there is a vector charge current along the magnetic field in the positive $y$ region and opposite to the magnetic field in the negative $y$ region. This then leads to a vector charge quadrupole moment in the transverse plane even when the charge asymmetry is zero.
![(Color online) Same as Fig. \[cmw1\] for the elliptic flow difference between positively and negatively charged particles as a function of time.[]{data-label="cmw2"}](v2t1.eps){width="50.00000%"}
We further show in Fig. \[cmw2\] the time evolution of the elliptic flow difference between negatively and positively charged particles due to both the vorticity and the magnetic field but with the neglect of the Lorentz force for quark matter of charge asymmetries $A_{\pm}=0$ and $A_{\pm}=0.16$. A finite positive elliptic flow difference is seen for zero charge asymmetry as a result of the finite eccentricity difference shown in Fig. \[cmw1\]. Compared to the case of having only the magnetic field [@PhysRevC.94.045204], the elliptic flow difference in the present case of having also the vorticity field appears earlier in time, and this is because the vorticity field helps to generate different average velocities for negatively and positively charged particles more quickly, which is in contrast to the case with only the magnetic field, where this difference is proportional to the axial charge chemical potential and thus takes time to develop.
Effect of vorticity field plus magnetic field with the Lorentz force
--------------------------------------------------------------------
![(Color online) Eccentricity and elliptic flow differences between negatively and positively charged particles as a function of time for different scenarios of parton dynamics when the total charge asymmetry of the quark matter is $A_{\pm}=0.16$.[]{data-label="lorentz"}](de22.eps){width="50.00000%"}
Effect of the Lorentz force on the eccentricity and elliptic flow differences between negatively and positively charged particles is shown in Fig. \[lorentz\] for partonic matter of charge asymmetry $A_{\pm}=0.16$. Compared to that shown in Fig. \[cmw1\] without the Lorentz force, the eccentricity difference between negatively and positively charged particles with transverse momenta in the range $0.15\le p_T\le0.5$ GeV/$c$ is slightly larger after including the Lorentz force, and the difference decreases after reaching a maximum value. For the time evolution of the elliptic flow difference between negatively and positively charged particles, Fig. \[lorentz\] shows that the combined effects of the vorticity and magnetic fields with the inclusion of the Lorentz force lead to an initial increase of the elliptic flow difference, which then quickly decreases and becomes negative. The Lorentz force thus cancels the chiral effects due to the vorticity and magnetic fields shown in Fig. \[cmw2\] and even leads to an opposite effect on the elliptic flow difference between negatively and positively charged particles.
Charge asymmetry dependence of the elliptic flow difference
-----------------------------------------------------------
![(Color online) Elliptic flow difference as a function of charge asymmetry $A_{\pm}$ for different scenarios of parton dynamics.[]{data-label="elliptic"}](asy.eps){width="50.00000%"}
In Fig. \[elliptic\], we show the charge asymmetry dependence of the elliptic flow difference between negatively and positively charged particles for different scenarios of parton dynamics. The dotted line shows that the elliptic flow difference is consistent with zero for the case when only the vorticity field is present (CVE), and this is because the vorticity field has same effects on negatively and positively charged particles. When both the vorticity and the magnetic field are present but without including the Lorentz force (CVE+CME), the elliptic flow difference increases linearly with the charge asymmetry of the partonic matter with a slope parameter 0.0108, which is comparable to the results obtained in Refs. [@PhysRevC.89.044909; @PhysRevC.94.045204] based on the CMW. However, the inclusion of the vorticity field leads to a finite intercept of 4.9$\times 10^{-3}$ at zero charge asymmetry, and this is due to fact that the axial charge dipole moment in the transverse plane generated by the vorticity field can subsequently induce by the magnetic field a vector charge quadrupole moment of different eccentricities for negatively and positively charged particles. With the inclusion of the Lorentz force, which is shown by the dashed line, chiral effects due to the vorticity and magnetic fields disappear, and the slope parameter in the charge asymmetry dependence of the elliptic flow difference becomes negative with a magnitude of 0.0139.
Summary
=======
Based on the anomalous transport model, which includes the propagation of massless quarks and antiquarks according to the chiral equations of motion and the chirality-changing scattering, we have studied the elliptic flows of charged particles in non-central heavy ion collisions at relativistic energies. Using initial conditions from a blast wave model and assuming the presence of only the vorticity field, which is modeled according to that from the AMPT model [@Lin:2004en], we find that the CVW does not lead to an elliptic flow splitting of negatively and positively charged particles when the charge asymmetry of the partonic matter is nonzero. On the other hand, including also a strong and long-lived magnetic field but neglecting the Lorentz force can lead to a splitting between the elliptic flows of negatively and positively charged particles. However, the slope parameter in the charge asymmetry dependence of the elliptic flow splitting is smaller than the experimental data, while the positive intercept at zero charge asymmetry is larger than the experimental value [@Wang2013248c]. Unfortunately, as in our previous study including only the magnetic field [@PhysRevC.94.045204], the inclusion of the Lorentz force cancells the chiral effects due to the magnetic and vorticity fields and leads instead to a negative slope parameter in the charge symmetry dependence of the elliptic flow splitting of negatively and positively charged particles, contrary to that observed in experiments. Understanding this experimental results in terms of the chiral effects thus remains a challenge. On the other hand, the large axial charge dipole moment generated in the transverse plane of a heavy ion collision by the CVW can lead to the polarization of particles with spins, which may be relevant to the observed finite polarization of lambda hyperons at RHIC [@1742-6596-736-1-012016]. We plan to investigate this quantitatively in a future study.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported in part by the US Department of Energy under Contract No. DE-SC0015266 and the Welch Foundation under Grant No. A-1358.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Attempts to understand the recent observation of an excess of events in the neutral and charged current channels at high$-Q^2$ at HERA has provided an excellent example of how experiments at both low and high energies can be used to simultaneously constrain scenarios which predict new physics beyond the Standard Model. In this talk I will discuss this subject from the point of view of the construction of new models of leptoquarks.'
---
psfig.sty -0.5in 6.5in 8.5in 4.0=0.5in -0.5in \#1[\_[.25ex]{}]{} \#1[\_[-.1em\#1]{}]{} \#1[\_[-.2em\#1]{}]{} \#1[\_[-.1em.25ex]{}]{} \#1[\_[.25ex]{}]{} \#1[\_[.4ex]{}]{} \#1[\^[\#1]{}]{} \#1[\^[\#1]{}]{} / plus0pt minus0pt \#1\#2[0.7ex]{} \#1[$\underline{\smash{\vphantom{y}\hbox{#1}}}$]{}
\#1[$^{#1}$ ]{} \#1
\#1
\#1[\[\#1\]]{}
[**Leptoquarks: Pride and Prejudice**]{} [^1]
THOMAS G. RIZZO\
\
0.95in
Talk given at the [*Workshop on Physics Beyond the Standard Model: Beyond the Desert: Accelerator and Nonaccelerator Approaches*]{}, Ringberg Castle, Germany, June 8-14, 1997.
Introduction: Physics Beyond the Standard Model
===============================================
It is a widely held belief that new physics must exist beyond that predicted by the Standard Model(SM) if for no other reason than that it leaves us with too many unanswered questions and too many free parameters. Just when, where and how new physics will make its first appearance has been–and continues to be–a matter of some speculation. In the past few years we have seen a number of potential new physics signatures vanish as either more statistics was accumulated or the data and analysis thereof improved. These signatures have each in their turn induced some excitement in the community with the hope that a new window beyond the SM was finally opening. Though later vanishing, these affects have each taught us something new about what kinds of models can be constructed, even if they were not necessarily realized in nature, given the ever-tightening constraints provided by experiment. Still, new physics is out there somewhere waiting to be discovered; it is only a question of looking. We should learn to expect the unexpected.
Searching for new physics is a multi-pronged attack on the unknown. While the production of a non-SM particle at a collider would be the most obvious and undeniable signature that one can imagine, the first sign of something new maybe more subtle. For example, one can imagine a significant deviation from SM expectations in a precision measurement, , the $W$ mass or polarized forward-backward asymmetry for $b-$quarks, $A_b$. Instead, one might imagine the observation of a process forbidden by the SM, such as $\mu \to e\gamma$. However, as is well known, new physics rarely contributes to only one of these scenarios. For example, SUSY leads to new particle production at colliders, a potentially observable shift in the $W$ mass and/or $\sin^2\theta_{eff}$ and can enhance many rare decay processes beyond their SM expectations. The same is of course true with other forms of new physics. In fact, it likely that once new physics is found all three types of experiments will be necessary to unravel its detailed nature. Attempts to understand the excess of events at high$-Q^2$ recently observed at HERA in both the neutral current(NC) and charged current(CC) channels[[@hera]]{} provides just such an example of the strong interplay between the various new physics search scenarios and the constraints that arise from the three classes of experiments.
Leptoquark Model Requirements
=============================
If the HERA excesses are real and non-resonant then possible explanations include, , the presence of higher dimensional operators[[@vb]]{} signalling compositeness or exotic modifications in the parton densities at large $x$[[@parton]]{}; both these proposed scenarios face some very serious difficulties. Instead, if the excess is resonant a popular[[@big; @old]]{} explanation is the $s-$channel production of a $\simeq 200-220$ GeV scalar(, spin-0) leptoquark(LQ) with fermion number($F$) equal to zero–the subject of the present work. (Given the most recent results[[@straub]]{} it appears that the H1 excess in the NC channel is apparently clustered at $201\pm 5$ GeV while that for ZEUS is at $219\pm 9$ GeV. Whether this is reconcilable with a single resonance is still unknown.) How did we so quickly deduce the LQ spin and $F$ quantum number from the data?
Any discussion of LQ models has been historically based on the classic work by Buchm" uller, R" uckl and Wyler(BRW)[[@brw]]{}. Those authors provided not only a set of assumptions under which consistent LQ models can be constructed but then classified them according to their possible spins and fermion number thus leading to the 10 states displayed in Table 1. These assumptions may be stated as follows:
\(a) LQ couplings must be invariant with respect to the SM gauge interactions
\(b) LQ interactions must be renormalizable
\(c) LQs couple to only a single generation of SM fermions
\(d) LQ couplings to fermions are chiral
\(e) LQ couplings separately conserve Baryon and Lepton numbers
\(f) LQs only couple to the SM fermions and gauge bosons
[l|ccrccc]{} & Leptoquark
------------------------------------------------------------------------
& SU(5) Rep & & $Q$ & Coupling & $B_\ell$\
Scalars & & & & & &\
$F=-2$
------------------------------------------------------------------------
& $S_{1L}$ & [**5**]{} & & $1/3$ & $\lambda_L\, (e^+\bar u)$, $\lambda_L\, (\bar\nu\bar d)$ & $1/2$\
& $S_{1R}$ & [**5**]{} & & $1/3$ & $\lambda_R\, (e^+\bar u)$ & 1\
& $\widetilde S_{1R}$ & [**45**]{} & & $4/3$ & $\lambda_R\, (e^+\bar d)$ & 1\
& & & & $4/3$ & $-\sqrt 2\lambda_L\, (e^+\bar d)$ & 1\
& $S_{3L}$ & [**45**]{} & $\left\{\rule{0pt}{30pt}\right.$ & $1/3$& $-\lambda_L\, (e^+\bar u)$, $-\lambda_L\, (\bar\nu\bar d)$ & $1/2$\
& & & & $-2/3$& $\sqrt 2\lambda_L\, (\bar\nu\bar u)$ & 0\
\[0pt\][$F=0$]{} & \[0pt\][$R_{2L}$]{} & \[0pt\][[**45**]{}]{} & \[0pt\][$\left\{\rule{0pt}{20pt}\right.$]{} & $5/3$ & $\lambda_L\, (e^+u)$ & 1\
& & & & $2/3$ & $\lambda_L\, (\bar\nu u)$ & 0\
& \[0pt\][$R_{2R}$]{} & \[0pt\][[**45**]{}]{} & \[0pt\][$\left\{\rule{0pt}{20pt}\right.$]{} & $5/3$ & $\lambda_R\, (e^+u)$ & 1\
& & & & $2/3$ & $-\lambda_R\, (e^+d)$ & 1\
& \[0pt\][$\widetilde R_{2L}$]{} & \[0pt\][[**10/15**]{}]{} & \[0pt\][$\left\{\rule{0pt}{20pt}\right.$]{} & $2/3$ &$\lambda_L\, (e^+d)$ & 1\
& & & & $-1/3$& $\lambda_L\, (\bar\nu d)$ & 0\
Vectors & & & & &\
------------------------------------------------------------------------
\[0pt\][$F=-2$]{} & \[0pt\][$V_{2L}$]{} & \[0pt\][[**24**]{}]{} & \[0pt\][$\left\{\rule{0pt}{20pt}\right.$]{} & $4/3$ & $\lambda_L\, (e^+\bar d)$ & 1\
& & && $1/3$ & $\lambda_L\, (\bar\nu\bar d)$ & 0\
& \[0pt\][$V_{2R}$]{} & \[0pt\][[**24**]{}]{} & \[0pt\][$\left\{\rule{0pt}{20pt}\right.$]{} & $4/3$ & $\lambda_R\, (e^+\bar d)$ & 1\
& & & & $1/3$ & $\lambda_R\, (e^+\bar u)$ & 1\
& \[0pt\][$\widetilde V_{2L}$]{} & \[0pt\][[**10/15**]{}]{} & \[0pt\][$\left\{\rule{0pt}{20pt}\right.$]{} & $1/3$ & $\lambda_L\, (e^+\bar u)$ & 1\
& & & & $-2/3$ & $\lambda_L\, (\bar\nu\bar u)$ & 0\
$F=0$ & $U_{1L}$ & [**10**]{} && $2/3$ & $\lambda_L\, (e^+d)$, $\lambda_L\, (\bar\nu u)$ & $1/2$\
& $U_{1R}$ & [**10**]{} && $2/3$ & $\lambda_R\, (e^+\bar d)$ & 1\
& $\widetilde U_{1R}$ & [**75**]{} && $5/3$ & $\lambda_R\, (e^+u)$ & 1\
& & & & $5/3$ & $\sqrt 2\lambda_L\, (e^+u)$ & 1\
& $U_{3L}$ & [**40**]{} & $\left\{\rule{0pt}{30pt}\right.$ & $2/3$ & $-\lambda_L\, (e^+d)$, $\lambda_L\, (\bar\nu u)$ & $1/2$\
& & & & $-1/3$ & $\sqrt 2\lambda_L\, (\bar\nu d)$ & 0\
If we strictly adhere to these rules then the requirements of gauge invariance and renormalizability fix all of the spin-1 LQ couplings and thus its production cross section at the Tevatron[[@hrphp]]{} is simply a function of its mass. The possibility that such particles can exist in the mass range below approximately 350 GeV can then be excluded based on the direct searches by both CDF and D0[[@cdfd0]]{}. (As noted by Blümlein[[@big]]{}, the introduction of non-renormalizable anomalous couplings for the LQ may allow us to somewhat soften this conclusion but spin-1 LQs are still found to be excluded in the range of interest for the HERA excess.) If the LQ is a scalar but is of the $F=2$ type then we would have expected to see an event excess show up in the $e^-p$ NC channel and not the $e^+p$ channel as is the case. This is demonstrated in Table 2 which shows the event rate for each of the BRW scalar LQs assuming a Yukawa coupling strength of $\tilde \lambda=\lambda/e=0.1$ and a mass of 200 GeV normalized to a luminosity of 100 $pb^{-1}$. Even with the great disparity in integrated luminosity collected by the experiments in both channels and allowing for the free adjustment of the strength of the Yukawa coupling we must conclude that the LQ is of the $F=0$ type. We note from Table 1 that all $F=0$ scalar LQs must have $B_\ell=1$ and lie in $SU(2)_L$ doublets, implying that more than one type of LQ must exist. Given the recent strengthening of the Tevatron search reach and the possible CC excess at HERA, this poses a serious challenge to the scalar LQ interpretation, although not as serious as was found in the case of vector LQs. Using the next-to-leading order cross section formulae of Krämer , the $95\%$ CL lower limit on the mass of a $B_\ell=1$ scalar LQ is found by D0 to be 225 GeV. D0 has also performed a combined search for first generation leptoquarks by using the $eejj$, $e\nu jj$ and $\nu \nu jj$ channels. For fixed values of the leptoquark mass below 225 GeV, these search constraints can be used to place an upper limit on $B_\ell$. For $M_{LQ}$=200(210,220) GeV, D0 obtains the constraints $B_\ell \leq 0.45(0.62,0.84)$ at $95\%$ CL. Of course if CDF and D0 combine their searches in the future, then the 225 GeV bound may rise to $\simeq 240$ GeV, in which case even stronger upper bounds on $B_\ell$ will be obtained. Allowing the LQ to have decays into the $\nu j$ final state with a reasonable branching fraction would solve this problem and would yield the desired CC signal at HERA. However the models in Table 1 do not allow for this possibility.
Leptoquark $N_L^-$ $N_R^-$ $N_L^+$ $N_R^+$
--------------------- --------- --------- --------- ---------
SM background 51.7 28.7 9.98 20.0
$S_{1L}$ 121. 28.7 9.98 20.4
$S_{1R}$ 51.7 167. 10.8 20.0
$\widetilde S_{1R}$ 51.7 63.0 11.5 20.0
$S_{3L}$ 190. 28.7 9.98 23.5
$R_{2L}$ 52.4 28.7 9.98 158.
$R_{2R}$ 51.7 29.4 148. 20.0
$\widetilde R_{2L}$ 53.2 28.7 9.98 54.4
: Expected number of events per $100^{-1}pb$ for each electron charge and state of polarization for a 200 GeV scalar leptoquark at HERA assuming $0.4<y<1$, $\tilde \lambda=0.1$, and an electron-jet invariant mass $M_{ej}=200\pm 20$ GeV. These results have been smeared with a detector resolution of $5\%$ in $M_{ej}$.
How do we interpret these conflicting demands? It is clear that the BRW structure must be too restrictive and so conditions (a)-(f) must be critically re-examined. While the assumptions of gauge invariance and renormalizability are unquestionable requirements of LQ model building, it is possible that the other conditions one usually imposes are much too strong–unless they are specifically demanded by data. This observation implies that for LQs to be experimentally accessible now, or anytime soon, their couplings to SM fermions must be essentially purely chiral and must also separately conserve both Baryon and Lepton numbers. The condition that LQs couple to only a single SM generation is surely convenient by way of avoiding the numerous low energy flavor changing neutral current constraints[[@rev]]{} but is far from natural in the mass eigenstate basis. A short analysis indicates[[@old]]{} that the natural imposition of this condition in the original weak basis for the first generation LQ and then allowing for CKM-like intergenerational mixing does not obviously get us into any trouble with experimental constraints especially in lepton generation number is at least approximately conserved. However, this does not give us the flexibility we need to avoid the Tevatron bounds or to induce an excess in the CC channel. Clearly then, to obtain a new class of LQ models the LQs themselves must be free to couple to more than just the SM fermions and gauge fields. Note that assumption (f) effectively requires that the LQ be the only new component added to the SM particle spectrum which seems quite unlikely in any realistic model; this assumption must be dropped.
What kind of LQ interaction do we want? In order to satisfy the HERA and Tevatron constraints it is clear that we need to have an $F=0$ scalar LQ as before, preferably an isosinglet so that we do not have several LQ states of various masses to worry about, but now with an effective coupling to SM fermions such as $${\cal L}_{wanted} = [\lambda_u \nu u^c+\lambda_d ed^c]\cdot LQ +h.c.\,,$$ with comparable values of the effective Yukawa couplings $\lambda_u$ and $\lambda_d$ thus fixing the LQ’s electric charge, $Q(LQ)=\pm 2/3$. An alternative possibility, allowing for either Dirac neutrinos or a $\nu^c$ which is light and appears as missing $p_T$ in a HERA or Tevatron detector, is the interaction $${\cal L}_{wanted}' = [\lambda_u' \nu^c u+\lambda_d' e^cd]\cdot LQ' +h.c.\,.$$ \[It is important for later analyses to note that we cannot have these two interactions simultaneously as we would then strongly violate assumption (d).\] It is easy to see that in either case the LQs of any other charge assignment cannot simultaneously couple to both $ej$ and $\nu j$ as is required by the HERA and Tevatron data. Unfortunately, either of the above Lagrangians as they stand violate assumption (a) in that they are not gauge invariant with respect to $SU(2)_L$. This implies that the desired Yukawa couplings are only effective ones and must arrived at from some more fundamental theory. Even if we are successful in obtaining one or both of these Lagrangians, is it clear that we can find values of $\lambda_u$ and $\lambda_d$ which are compatible with all of the data?
Constraining Leptoquark Couplings
=================================
What are the existing constraints on LQ Yukawa couplings? If the LQ has only the couplings described by one of the above Lagrangians then we can, , trade in $\lambda_u$ for $B_\ell=\lambda_d^2/(\lambda_d^2+\lambda_u^2)$, since we are assuming that the LQ has no other decay modes. As discussed above, the Tevatron searches place a $\lambda_d$-independent constraint on $B_\ell$ for any fixed value of the LQ mass. Similarly, as promised, low energy measurements play an important role here as well. The recent constraints on the size of any allowed deviation of the weak charge from its SM value in Atomic Parity Violation(APV) in Cesium[[@wood]]{}, $\Delta Q_W=1.09\pm 0.93$, places $B_\ell$-independent bounds on $\lambda_d$[[@ros]]{} for fixed $M_{LQ}$. Similarly, $\mu-e$ universality in $\pi$ decay, expressed through the ratio $R=\Gamma(\pi \to e\nu)/\Gamma(\pi \to \mu \nu)=0.9966\pm 0.0030$, constrains the product of couplings $\lambda_u \lambda_d$[[@pdg]]{}. The observed rate of NC events at HERA itself essentially constrains instead the product $\lambda_d^2B_\ell$; in the later case QCD and efficiency corrections are quite important[[@ks]]{}. Putting all of these together defines an approximate allowed region in the $B_\ell-\tilde \lambda_d$ plane shown in Fig.1 for different values of $M_{LQ}$. Here, we define $\tilde \lambda=\lambda/e$, with $e$ the conventional proton charge. (This scaling of the coupling to $e$ follows earlier tradition[[@phyrep]]{}.) We note that these allowed regions are compatible with the cross section required to explain the HERA CC excess.
There are other means to probe LQ couplings but they are somewhat more indirect. There are several ways in which LQs may make their presence known in $e^+e^-$ collisions[@old; @us]. At center of mass energies below the threshold for pair production, the existence of LQs can lead to deviations in both the cross section and angular distributions for $e^+e^- \to q\bar q$. This may be particularly relevant when $\sqrt s$ is comparable to the leptoquark mass as would be the case at LEP II if a 200-220 GeV LQ did exist. The origin of these modifications is due to the $t-$channel LQ exchange and is thus proportional in amplitude to the square of the unknown Yukawa coupling. However, in [[@old]]{} it was shown that even with large data samples it is unlikely that LEP II will have the required sensitivity to probe couplings as small as those suggested by the HERA data as shown in Fig.2. The OPAL Collaboration[[@opal]]{} has recently performed this analysis with real data at somewhat lower energies but with comparable results.
Turning this process around, we can imagine that the Drell-Yan production of either $e^+e^-$ or $e^\pm \nu$ channel at the Tevatron Main Injector may show some sensitivity[[@dryan]]{} to LQ exchange in the $t-$channel. In the $e^+e^-$ case the observables are the invariant mass distribution and the forward-backward asymmetry. In the $e^\pm \nu$ channel, the corresponding observables are the transverse mass distribution on the electron rapidity asymmetry. Figs. 3 and 4 show the result of these considerations; neither channel has the sensitivity to probe Yukawas in the desired range if only 2 $fb^{-1}$ of luminosity is available.
Leptoquarks can also be produced singly at hadron colliders through their Yukawa couplings. Compared to pair production, this mechanism has the advantage of a larger amount of available phase space, but has the disadvantage in that it is directly proportional to the small Yukawa coupling. For a general 200 GeV scalar leptoquark with a coupling strength of $\tilde \lambda=0.15$ calculations show that we would obtain approximately $\sim 88\,, 196$ events from $gd+g\bar d\,, gu+g\bar u$ fusion, respectively, assuming $10$ fb$^{-1}$ of integrated luminosity at the Main Injector/TeV33. This event rate should be marginally sufficient to provide a very rough determination of the value of Yukawa coupling $\tilde \lambda_d$ for the models of specific interest.
Additional LQ coupling information can possibly be obtained from the sum of the squares of the first row of the CKM matrix, $\sum_i|V_{ui}|^2$. This involves combining experiments at low, medium and high energies and it an essential test of quark-lepton universality and CKM unitarity. In the SM this sum is, of course, unity, but LQ exchange can yield either an apparent upward or downward shift in the extracted value of $|V_{ud}|$: $$|V_{ud}|^2_{eff}\simeq |V_{ud}|^2_{true}-1.52\times 10^{-3}\left( {200~GeV
\over M_{LQ} }\right)^2\left( {\tilde\lambda_u\over 0.15}\right)
\left( {\tilde\lambda_d\over 0.15}\right) \,,$$ so that it would appear experimentally as if a unitarity violation were occurring. Interestingly, the value of the above sum has recently been discussed by Buras[@ajb], who reports $\sum_i|V_{ui}|^2=0.9972\pm 0.0013$, which is more than $2\sigma$ below the SM expectation. Clearly, if $\tilde\lambda_u\tilde\lambda_d>0$, the LQ exchange provides one possible additional contribution which, for $\tilde\lambda_u=\tilde\lambda_d=0.15$ and $M_{LQ}=200$ GeV, would increase the sum to $0.9987$, now only $1\sigma$ low. This “same sign” possibility is clearly preferred by the combined set of present data as shown in Fig.2. This situation requires watching in the future.
A last possibility is the observation of LQ exchange through radiative corrections. Unfortunately, it has been shown[[@old]]{} that LQs rapidly decouple and give only tiny contributions to the oblique parameters[[@obl]]{}. In addition it has also been shown that LQ exchange does not significantly modify $Z$-pole physics[[@oscar]]{} through vertex corrections.
New Leptoquark Model Building
=============================
We now turn to dealing with the construction of models that lead to one of the Lagrangians above. Given the fixed gauge structure of the SM the most likely new interactions that LQs may possess are with the Higgs field(s) responsible for spontaneous symmetry breaking and with new vector-like fermions that are a common feature in many extensions of the SM. In this lecture we will consider and classify all models wherein heavy vector-like fermions(VLFs) are used to generate the effective interactions ${\cal L}_{wanted}$ or ${\cal L}_{wanted}'$ at low energies. The emphasis of our approach will be to keep the VLFs as active participants in our models and not auxiliary devices to produce the desired coupling structure. We will assume that the fundamental LQ coupling is that between a VLF and a SM fermion. a LQ couples to both a VLF and that ordinary SSB induces a a mixing between the two sets of fermions. SSB is thus the true source of the desired LQ interactions and explains how effective interactions can arise that are not obviously gauge invariant. The small size of the effective Yukawa couplings in the above Lagrangians, ${\cal L}_{wanted}$ or ${\cal L}_{wanted}'$, will then be subsequently explained by the same mechanism that produces the ordinary-exotic fermion mixing and sets the scale of the VLF masses in the TeV region. We note that the use of VLFs in this role is particularly suitable since in their unmixed state they make essentially no contribution to the oblique parameters[[@obl]]{}, they are automatically anomaly free and they can have bare mass terms which are SM gauge invariant. (Alternatively their masses can be generated by the vacuum expectation value of a SM singlet Higgs field.)
To proceed with the analysis, we first construct the six ‘skeleton’ models that are obtainable by simply coupling one of the usual SM fermion representations, $(L,Q,u^c,d^c,e^c,\nu^c)$, with an appropriate VLF, $X_i$(or $X_i^c$), and the LQ field. \[We use $X_i(X_i^c)$ to denote the VLF fields with $F>(<)0$.\] Note that we have allowed for the possibility if right-handed neutrinos. To this we must add the bare mass term for $X_i$ as well as any gauge invariant terms that can be constructed using the remaining SM fermion fields, $X_i^c$(or $X_i$), and the SM Higgs doublet fields $H$ and $H^c$. In each case gauge invariance tells us the the quantum numbers of the VLFs under the assumption that they are either in singlets or in fundamental representations under $SU(2)_L$ and $SU(3)_C$. We strictly adhere to BRW constraints (a)-(e) in forming our constructions. These six ‘skeletons’ are: $$\begin{aligned}
{\cal L}_A & = & \lambda_A LX_1^c\cdot LQ+a_uX_1u^cH+a_dX_1d^cH^c-M_1X_1X_1^c
-M_Q'QX_1^c\,, \nonumber\\
{\cal L}_B & = & \lambda_B QX_2^c\cdot LQ+a_eX_2e^cH^c+a_{\nu} X_2 \nu^cH
-M_2X_2X_2^c-M_L'LX_2^c\,, \nonumber\\
{\cal L}_C & = & \lambda_C X_3u^c\cdot LQ+a_1LX_3^cH
-M_3X_3X_3^c-M_N'N\nu^c\,, \nonumber\\
{\cal L}_D & = & \lambda_D X_4d^c\cdot LQ+a_2LX_4^cH^c
-M_4X_4X_4^c-M_E'Ee^c\,, \\
{\cal L}_E & = & \lambda_E X_5e^c\cdot LQ+a_3QX_5^cH^c
-M_5X_5X_5^c-M_D'Dd^c\,, \nonumber\\
{\cal L}_F & = & \lambda_F X_6\nu^c\cdot LQ+a_4QX_6^cH
-M_6X_6X_6^c-M_U'Uu^c\,, \nonumber\end{aligned}$$ We will assume that all of the Yukawa couplings that appear in these ‘skeletons’ are of order unity and that no fine-tuning is present. These constructs have a number of obviously desirable features but they do not yet have all the necessary ingredients. For example, the LQ in ‘skeleton’ C(D) only couples to the $u(d)$ quark while that in ‘skeleton’ E(F) only couples to $e(\nu)$. For ‘skeletons’ A and B we see that the LQs couple to both $u,d$ and $e,\nu$, respectively. The solution to this problem is to combine the various ‘skeletons’ into full models that have all of the desired couplings. This procedure is straightforward but when doing so we must take care not to violate the assumption that the LQ couplings are chiral. Given this very strong constraint, the entire list of models that can be constructed in this fashion are only ten in number: A, B, CD, EF, AC, AD, ACD, BE, BF and BEF. The combination of letters symbolizes that we add the respective Lagrangians and identify the LQ field as common. We note that models A, CD, AC, AD and ACD produce the interaction ${\cal L}_{wanted}$, while models B, EF, BE, BF and BEF produce instead ${\cal L}_{wanted}'$. As we will see below in each case the color, isospin and electric charge quantum numbers are completely fixed by gauge invariance and the assumption that the LQ is a $|Q|=2/3$ isosinglet with $F=0$. It is very important to remember that the five models leading to ${\cal L}_{wanted}'$ would be [*excluded*]{} if the neutrino is not a Dirac field of if $\nu^c$ cannot appear as missing energy or $p_T$ in a detector.
Having said all this we have yet to explicitly see how even one of these models works so we now examine model A in detail. Here, we have coupled an exotic fermion, denoted as $X_1$, to $L$ plus a leptoquark. In this case gauge invariance requires that $X_1$ be an isodoublet, with member charges of $2/3,-1/3$ since the leptoquark charge and fermion number are fixed, as well as an $SU(3)_C$ triplet. We can thus write $X_1^T=(U^0,D^0)$, where the superscript denotes the weak eigenstate fields. When $H$ and $H^c$ receive vevs (which we denote by $v$ and $v^c$, respectively), the $a_{u,d}$ terms in the above Lagrangian induce off-diagonal couplings in both the $Q=-1/3$ and $Q=2/3$ quark mass matrices. Neglecting the $u$- and $d$-quark masses, these are given in the $\bar \psi_L^0 M \psi_R^0$ weak eigenstate basis by $$\begin{aligned}
\bar \psi_L^0 M_{u}\psi_R^0 & = & (\bar u^0,\bar U^0)_L \left( \begin{array}{cc}
0 & -M_Q' \\
a_uv & -M_1
\end{array} \right)\left( \begin{array}{c}
u^0 \\
U^0
\end{array} \right)_R \,, \\
\bar \psi_L^0 M_{d}\psi_R^0 & = &(\bar d^0,\bar D^0)_L \left( \begin{array}{cc}
0 & -M_Q' \\
a_dv^c & -M_1
\end{array} \right) \left( \begin{array}{c}
d^0 \\
D^0
\end{array} \right)_R \,.\end{aligned}$$ Both $M_{u,d}$ can be diagonalized by a bi-unitary transformation which becomes simply bi-orthogonal under the assumption that the elements of $M_{u,d}$ are real, resulting in the diagonal mass matrices $M_{u,d}^{diag}=U_L(u,d)M_{u,d}U_R(u,d)^\dagger$. Since $U_{L,R}(u,d)$ are simple $2\times 2$ rotations they can each be parameterized by a single angle $\theta_{L,R}^{u,d}$. Assuming that both $M_1,M_Q'$ are large in comparison to either $a_uv$ or $a_dv^c$ we find $\theta_R^{u,d}\simeq a_{u,d}v(v^c)M_1/(M_1^2+M_Q'^2)$. With $M_1,M_Q'$ of order 1 TeV, $v,v^c$ of order 100 GeV and $a_{u,d}$ of order unity this implies that $\theta_R^{u,d}\simeq 0.05$. Writing $U^0\simeq U+\theta_R^u u$ in terms of the mass eigenstate fields, and similarly for $D^0$, the interaction involving the SM fermions and the leptoquark thus becomes $${\cal L}_{light} = \left[\left({\lambda_A a_u vM_1\over (M_1^2+M_Q'^2)}
\right)~\nu u^c+\left({\lambda_A a_dv^cM_1\over (M_1^2+M_Q'^2)}\right)
~ed^c\right]\cdot LQ +h.c.\,,$$ which is the exact form we desired in Eqn. (1). For $\lambda_A$ again of order unity this naturally leads to a reasonable relative branching fraction for the $LQ\to \nu j$ decay mode, and gives acceptable values for $\lambda_{u,d}$ in Eqn. (1) for $M_1,M_Q'$ in the TeV range. Note that $\theta_R^{u,d}\neq 0$ leads to a modification of both the $u$ and $d$ quark couplings to the $Z$ and induces $W$-mediated right-handed charged current interactions as well.
What about $\theta_L^{u,d}$? $\theta_L^{u,d}\neq 0$ does not contribute to the LQ couplings or influence $Z$ couplings of $u,d$ since the left-handed SM fermions and the VLFs have the same quantum numbers. With $M_1$ and $M_Q'$ of comparable size both $\theta_L^{u,d}$ are found to be large and of almost identical magnitude. However, modifications to the left-handed CC couplings of $u$ and $d$ to the $W$ are only sensitive to the deviation $\Delta=1-\cos(\theta_L^u-\theta_L^d)$. Here, for $M_1=M_Q'$, the difference $\theta_L^u-\theta_L^d\approx (a_u^2v^2-a_d^2v^{c2})/4M_1^2$ is found to be very small, of order $\sim \theta_R^2\simeq (0.05)^2$. This implies that $\Delta$ itself is of only order $10^{-5}$ or less–practically invisible.
The other models above work more or less in a similar fashion except that in most cases mixing is taking place between a number of different SM fermions and their VLF partners. The effective couplings in Eqns. 1 and 2 then derive from more than a single source. For completeness Table 3 identifies the VLFs which are present in each of these models.
------------------------------------------------------------
Model Vector-like Fermions
------------- ----------------------------------------------
A $\left( \begin{array}{c}
U \\
D
\end{array} \right)_{L,R}$
\[4ex\] CD $N_{L,R}; E_{L,R}$
\[2ex\] AC $\left( \begin{array}{c}
U \\
D
\end{array} \right)_{L,R}; N_{L,R}$
\[4ex\] AD $\left( \begin{array}{c}
U \\
D
\end{array} \right)_{L,R}; E_{L,R}$
\[4ex\] ACD $\left( \begin{array}{c}
U \\
D
\end{array} \right)_{L,R}; N_{L,R}; E_{L,R}$
\[4ex\]
B $\left( \begin{array}{c}
N \\
E
\end{array} \right)_{L,R}$
\[4ex\] EF $U_{L,R}; D_{L,R}$
\[2ex\] BE $\left( \begin{array}{c}
N \\
E
\end{array} \right)_{L,R}; D_{L,R}$
\[4ex\] BF $\left( \begin{array}{c}
N \\
E
\end{array} \right)_{L,R}; U_{L,R}$
\[4ex\] BEF $\left( \begin{array}{c}
N \\
E
\end{array} \right)_{L,R}; U_{L,R}; D_{L,R}$
\[4ex\]
------------------------------------------------------------
: Listing of models and the new vector-like fermions which are contained in them.[]{data-label="vlf"}
Tests
=====
To directly test the proposed models we must look for new physics signatures [*beyond*]{} those suggested by the LQ interactions described Eqns. 1 and 2. One possibility is to probe for the additional interactions between the VLFs and the LQ; this is obviously difficult since the VLFs are so massive. A second possibility is to directly look for the influences of the VLFs themselves. For the time being this must be an indirect search since the LHC will be required to directly produce the VLFs in the range of interest to us here. Once the LHC is available, however, a 1 TeV color triplet will produce 1000 events/yr even with a luminosity of 10 $fb^{-1}$.
In the first category a potential new process of interest is the pair production of like sign LQs at the Tevatron through $t-$ and $u-$channel $N$ exchange (in those models where it is present) when $N$ is a Majorana field: $uu\to 2LQ$. This rate for this process goes as the fourth power of the $\lambda_B$ or $\lambda_C$ Yukawa coupling but these are assumed to be of order unity so that potentially large cross sections are obtainable. One finds the subprocess cross section to be $${d\sigma\over {d\hat t}} ={\lambda^4\over {64\pi \hat s}}
\left[{M_N(\hat t+\hat u-2M_N^2)
\over {(\hat t-M_N^2)(\hat u-M_N^2)}}\right]^2\,,$$ where $M_N$ is the mass of the $N$. Note that as $M_N \to 0$ the rate vanishes as one might expect for a Majorana fermion induced process. The cross section for this reaction at the Tevatron Main Injector for $\lambda=1$ and $M_{LQ}=200$ GeV is approximately 50 fb for $M_N$=1 TeV and falls off quickly with increasing as $M_N$ increases. Since the signature for this process is 2 jets plus like-sign leptons there is little SM backgrounds and so it may be observable during Run II.
The best indirect tests for the presence of VLFs which mix with the conventional fermions are searches for new physics associated with deviations in couplings from SM expectations[[@phyrep; @dpf]]{}. Two of the best tests here are our old friends quark-lepton universality in the guise of $V_{ud}$ and the leptonic decays of the $\pi$ discussed above. To clarify this point we note that in models where, , the $u$ and $d$ mix with isodoublet VLFs the $W$ couples as $\bar ud_L\cos(\theta_L^u-\theta_L^d)+\bar ud_R\sin \theta_R^u\sin \theta_R^d$ whereas if the VLFs are isosinglets the corresponding coupling is $\bar ud_L\cos \theta_L^u\cos \theta_L^d$. (The corresponding couplings can be also written down for the case of leptonic mixing.) We note also the general feature we find is that in models where the VLFs are in isodoublets $\theta_R's \sim 0.05$ and differences in $\theta_L's \sim (0.05)^2$, whereas the converse is true when the relevant VLFs are isosinglets.
Leptonic mixing will never show up as a shift in the value of $|V_ud|^2$ since the modification in the amplitude occurs not only in the process $n\to pe\bar \nu$ but also in $\mu$ decay so that it is absorbed into the definition of $G_F$. On the otherhand if quark mixing occurs $|V_ud|^2$ will experience an apparent small shift(to leading order in the mixing angles) $\sim -(\theta_L^u-\theta_L^d)^2+(\theta_R^u \theta_R^d)^2$ in models with isodoublet VLFs and $\sim -(\theta_L^u)^2-(\theta_L^d)^2$ in the isosinglet VLF case. For the ratio $R$ in $\pi$ decay any modification of the hadronic matrix element will factor out so that there is no sensitivity to quark mixing. However, leptonic mixing no longer factorizes and we find a shift in $R$ by an amount $\sim -(\theta_L^\nu-\theta_L^e)^2-(\theta_R^\nu \theta_R^e)^2$ in the case of isodoublet mixing and $\sim -(\theta_L^\nu)^2-(\theta_L^e)^2$ for isosinglets. Recall that isodoublet leptonic mixing is only viable in models where the neutrino is Dirac or the right-handed neutrino appears as an ordinary neutrino. In these same isodoublet models, the presence of both left- and right-handed CC couplings and heavy VLFs can lead to a contribution to the $g-2$ of the electron and $\nu$ which are typically both or order $a ~few \cdot 10^{-11}$, neither of which are far from the present level of sensitivity.
The mixing of the SM fermions with the VLF modify their couplings to the $Z$. These are difficult to observe particularly in the case of quarks due to the small size of the effect and QCD correction uncertainties. In the case of leptonic mixing there is not only the contribution due to mixing but there is an overall normalization change in the couplings due to our redefinition of $G_F$. This mixing can lead to a shift in $Z\to e^+e^-$ width by $\simeq 0.2$ MeV and an apparent shift in $\sin ^2 \theta_{eff}$ from the asymmetries of $\simeq 0.0006$. Again, shifts of this size are near the present limit of experimental sensitivity. In a similar manner SM expectations for $Q_W$ in APV measurements may also be modified if SM mixing with the VLF occurs. However, in this case it is easy to show that the fractional change in $Q_W$ due to these effects is only at the level of $\sim 10^{-3}$.
Grand Unification with Leptoquarks
==================================
If LQs are indeed real and we also believe that there is experimental evidence for coupling constant unification then we must begin to examine schemes which contain both ingredients as pointed out in[[@old]]{}. In the scenarios presented here the SM quantum numbers of the LQ are fixed but new VLFs have now been introduced as well, all of which will alter the usual RGE analysis of the running couplings.
Before discussing SUSY models we note with some curiosity that coupling unification [*can*]{} occur in LQ models containing exotic fermions even if SUSY is not introduced as was shown many years ago in [[@my; @oldt]]{}. Of course in the work of Murayama and Yanagida [[@my]]{}, the LQ was an isodoublet and one of the particular models on the BRW list, now excluded by the combined HERA and Tevatron data. In the scenarios presented above the LQ is now a $Q=2/3$ isosinglet so that the Murayama and Yanagida analysis does immediately apply. Fortunately, we see from the results of Ref. [[@oldt]]{} that a second possibility does exist for just this case: one adds to the SM spectrum the LQ and its conjugate as well as a vector-like pair of color-triplet, isodoublets together with the field $H^c$. This is the just particle content of the model A. To verify and update this analysis, let us assume for simplicity that all the new matter fields are introduced at the weak scale and take $\sin ^2 \theta_w=0.2315$ as input to a two-loop RGE analysis. We then obtain the predictions that coupling unification occurs at $3.5 \times 10^{15}$ GeV and $\alpha_s(M_Z)$ is predicted to be 0.118. If unification does indeed occur we can estimate the proton lifetime[[@mrbill]]{} to be $\tau_p=1.6\times 10^{34\pm 1}$ years, safely above current constraints[[@pdg]]{}. We find this situation to be rather intriguing and we leave it to the reader to further ponder.
Of course there are other reasons to introduce SUSY beyond that of coupling constant unification. This subject has been discussed at some length in [[@old]]{} from which we extract several important observations: ($i$) To [*trivially*]{} preserve the successful unification of the SUSY-SM, only complete $SU(5)$ representations can be added to the MSSM spectrum. As is well-known, the addition of extra matter superfields in complete $SU(5)$ representations delays unification and brings the GUT scale closer to the string scale. Of course, there still remains the rather unnatural possibility of adding incomplete, but ‘wisely chosen’, split representations. Employing split representations certainly allows for more flexibility at the price of naturalness but still requires us to choose sets of $SU(3)_C\times SU(2)_L \times U(1)_Y$ representations which will maintain asymptotic freedom and perturbative unification. An example of this rather bizarre scenario is the possibility of adding a $(2,3)(1/6)$ from a [**15**]{} and a $(1,1)(1)\oplus (1,\bar 3)(-2/3)$ from a [**10**]{} to the low energy spectrum[[@old]]{}. Here the notation refers to the $(SU(3)_C,SU(2)_L)(Y/2)$ quantum numbers of the representation. We remind the reader that the LQ itself transforms as $(1,3)(2/3)$; the smallest standard $SU(5)$ representation into which the ${\mbox{LQ}}+{\mbox{LQ}}^c$ can be embedded is a [**10**]{}$\oplus\overline{\mbox{\bf 10}}$ while in flipped-$SU(5)\times U(1)$[[@flip]]{}, it can be placed in a [**5**]{}$\oplus\overline{\mbox{\bf 5}}$. ($ii$) Since we are using VLFs in these models, it is clear that only pairs of representations, [**R**]{}$+\overline{\mbox{\bf R}}$, can be added to the MSSM spectrum in order to maintain anomaly cancelation. Of course this is also true for the LQ superfield in that both LQ and $LQ^c$ fields must now be present as discussed above. ($iii$) To preserve perturbation theory and asymptotic freedom up to the GUT scale when adding complete representations, at most one [**10**]{}$+\overline{\mbox{\bf 10}}$ or three [**5**]{}$+\overline{\mbox{\bf 5}}$ can be appended to the low energy spectrum of the MSSM apart from SM singlets. The reason for this is the general observation that if one adds more than three, vector-like, color triplet superfields to the MSSM particle content then the one-loop QCD beta function changes sign. Recall that the LQ itself [*already*]{} accounts for one of these color triplets. This same consideration also excludes the introduction of light exotic fields in higher dimensional $SU(3)_C$ representations. Complete $SU(5)$ representations larger than [**10**]{}$+\overline{\mbox{\bf 10}}$ are found to contribute more than this critical amount to the running of the QCD coupling.
These are highly restrictive constraints on the construction of a successful GUT scenario containing both VLFs and LQs and we see than none of the models discussed above can immediately satisfy them [*unless*]{} the LQ and VLF superfields can be placed into a single $SU(5)$ representation. In the standard $SU(5)$ picture, we can then place $(U,D)^T$, an isosinglet $E^c$ and $LQ^c$ into a single [**10**]{} with the corresponding conjugate fields in the $\overline{\mbox{\bf 10}}$. This would form a hybrid of model A with the ‘skeleton’ model D, which we’ve denoted by AD above. Of course we pay no penalty for also including ‘skeleton’ model C here as well, which then yields model ACD. Instead, when we consider the flipped-$SU(5)\times U(1)$ case, it would appear that we can place $(N,E)^T$ and $LQ^c$ into a $\overline{\mbox{\bf 5}}$ with the conjugate fields in the [**5**]{}; this is exactly model B. It would also seem that no penalty is paid as far as unification is concerned for including the ‘skeleton’ model C here as well [*except*]{} that this would violate assumption ($iii$) about the chirality of LQ couplings to fermions. However, this model is no longer truly unified since the hypercharge generator is not fully contained within the $SU(5)$ group itself and lies partly in the additional $U(1)$. While the $SU(3)_C$ and $SU(2)_L$ couplings will unify, $U(1)_Y$ will not join them even when arbitrary additional vector-like singlet fields are added. Thus unification no longer occurs in this scenario so that this possibility is now excluded.
The LQ embedding situation becomes more perplexing if the LQ and VLFs cannot occupy the same GUT multiplet. In this case unification and asymptotic freedom constraints become particularly tight and we are forced to consider the split multiplet approach mentioned above. This means that we add the fields $(2,3)(1/6)\oplus (1,1)(1)\oplus (1,\bar 3)(-2/3)$ and their conjugates at low energies but constrain them to be from different $SU(5)$ representations. In this case the combination $(1,3)(2/3)\oplus (1,\bar 3)(-2/3)$ corresponds to the isosinglet LQ and its conjugate so what remains can only be the VLF fields. Note that we have again arrived back at models AD and ACD. Are these the only solutions? We have performed a systematic scan over a very large set of VLFs with various electroweak quantum numbers under the assumption that they are either color singlets or triplets, demanding only that ($i$) QCD remains asymptotically free and ($ii$) the model passes the so-called “B-test”[[@mpeskin]]{} which is highly non-trivial to arrange. Essentially the B-test takes advantage of the observation that if we know the couplings at the weak scale and we demand that unification takes place [*somewhere*]{} then the values of the one-loop beta functions must be related. Note that it is a necessary but not sufficient test on our choice of models but is very useful at chopping away a large region of parameter space. Using the latest experimental data[[@moriond]]{}, we find that $$B = {b_3-b_2\over {b_2-b_1}} = 0.720\pm 0.030 \,,$$ where the $\pm 0.030$ is an estimate of the corrections due to higher order as well as threshold effects and the $b_i$ are the one-loop beta functions of the three SM gauge groups. Note that $B_{MSSM}=5/7 \simeq 0.714$ clearly satisfies the test. If we require that ($i$) and ($ii$) be satisfied and also require that the unification scale not be too low then only the solutions described above survive after examining $>7\times 10^{7}$ combinations of matter representations. While not completely exhaustive this search indicates the solutions above are fairly unique. It is interesting to observe that models constructed around model A produce successful grand unification both with and without SUSY.
Conclusion and Outlook
======================
In this talk we have seen how a wealth of data from low and medium energy experiments as well as high energy colliders can be combined to point us in a fixed direction for LQ model building. I have also discussed a general framework for the construction of new $F=0$ scalar LQ models which go beyond the original classification by Buchmüller, Rückl and Wyler. This approach is based on the observation that in any realistic extension of the SM containing LQs it is expected that the LQs themselves will not be the only new ingredient. This construction technique is, of course, far more general than that required to address the specific issue of the HERA excess. While the assumptions of gauge invariance and renormalizability are unquestionable requirements of model building, it is possible that the other conditions one usually imposes are much too strong–unless they are clearly demanded by data. This observation implies that for LQs to be experimentally accessible their couplings to SM fermions must be essentially chiral and separately conserve both Baryon and Lepton numbers. The assumption that LQs couple to only a single SM generation is surely a convenient way of avoiding numerous low energy flavor changing neutral current constraints but is far from natural in the mass eigenstate basis. What is required to obtain a new class of LQ models is that the LQs themselves must be free to couple to more than just the SM fermions and gauge fields.
Given the fixed gauge structure of the SM the most likely new interactions that LQs may possess are with the Higgs field(s) responsible for spontaneous symmetry breaking and with new VLFs that are a common feature in many extensions of the SM. In the discussion above it has been shown how two new forms of the effective interactions of LQs with the SM fermions, consistent with Tevatron searches, the HERA excess in both the NC and CC channels and low-energy data, can arise through the action of VLFs and ordinary symmetry breaking. The typical VLF mass was found to lie in the low TeV region and they could thus be directly produced at future colliders with known rates.
We saw that we could construct ten new models which fell into two broad classes according to the chirality of the resulting LQ couplings to the SM fermions. The VLFs themselves were shown to lead to a number of model-dependent effects which are close to the boundary of present experimental sensitivity. LQs within the framework of models containing VLFs were also shown to be consistent with Grand Unification in both a supersymmetric [*and*]{} non-supersymmetric context. The common feature of both schemes is the structure associated with model A, , the VLFs are color triplet, weak isodoublets in a $(2,3)(1/6)$ representation and both $H$ and $H^c$ Higgs fields are required to be present as is $LQ^c$ field. In both scenarios the GUT scale is raised appreciably from the corresponding model wherein LQs and vector-like fermions are absent. In the SUSY case a $(1,1)(1)$ field is also required with the optional addition of a SM singlet, corresponding to models AD and ACD. In some sense, ACD is the “anti-$E_6$” model in that the color triplet VLFs are in isodoublets while the color singlet fields are all isosinglets. Interestingly, in this scenario there is a vector-like fermion corresponding to every type of SM fermion.
Realistic LQ models provide a rich source of new physics beyond the Standard Model.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author would like to thanks J.L. Hewett for collaboration on the work on which this talk is based. The author also appreciates comments, discussions and input from S. Eno(D0), G. Landsberg(D0), J. Conway(CDF) and H. Frisch(CDF) regarding the current Tevatron constraints. In addition, he thanks Y. Sirois(H1) and D. Krakauer(ZEUS) for discussions of the HERA data.
\#1 \#2 \#3 [Mod. Phys. Lett. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Nucl. Phys. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Phys. Lett. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Phys. Rep. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Phys. Rev. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Phys. Rev. Lett. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Rev. Mod. Phys. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Z. Phys. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Int. J. Mod. Phys. [**\#1**]{}, \#2 (\#3)]{}
[99]{}
C. Adloff , H1 Collaboration, C74 191 1997 ; J. Breitweg , ZEUS Collaboration, C74 207 1997 . For a comprehensive analysis of this possibility and original references, see V. Barger , hep-ph/9707412. The analysis by these authors apparently excludes this possibility as the source of the HERA excess in the NC channel. S. Kuhlman, H.L. Lai and W.K. Tung, hep-ph/9704338; K.S. Babu, C. Kolda and J. March-Russell, hep-ph/9705399; J.F. Gunion and R. Vogt, hep-ph/9706252; S. Rock and P. Bosted, hep-ph/9706436; W. Melnitchouk and A.W. Thomas, hep-ph/9707387. G. Altarelli , hep-ph/9703279; J. Blümlein, hep-ph/9703287; J. Kalinowski , hep-ph/9703288; K.S. Babu , hep-ph/9703299; M. Suzuki, hep-ph/9703316; G.K. Leontaris and J.D. Vergados, hep-ph/9703338; I. Montvay, hep-ph/9704280; S.F. King and G.K. Leontaris, hep-ph/9704336; J. Elwood and A. Faraggi, hep-ph/9704363; B. Dutta, R.N. Mohapatra and S. Nandi, hep-ph/9704428; M. Heyssler and W.J. Stirling, hep-ph/9705229; J. Ellis, S. Lola and K. Sridhar, hep-ph/9705416; S. Jadach, B.F.L. Ward and Z. Was, hep-ph/9705429; A. Blumhofer and B. Lampe, hep-ph/9706454; E. Keith and E. Ma, hep-ph/9707214; N.G. Deshpande and B. Dutta, hep-ph/9707274. J.L Hewett and T.G. Rizzo, SLAC-PUB-7430, hep-ph/9703337, to appear in Physical Review [**D**]{} and SLAC-PUB-7549, hep-ph/9708419. B. Straub, ZEUS Collaboration, talk given at the [*XVIII International Symposium on Lepton and Photon Interactions*]{}, July 28-August 1, 1997, Hamburg, Germany. For the most recent results, see E. Elsen, H1 Collaboration, talk given at the [*International Europhysics Conference on High Energy Physics*]{}, Jerusalem, August 19-26, 1997. W. Buchm" uller, R. R" uckl, and D. Wyler, B191 442 1987 . J.L. Hewett, T.G. Rizzo, S. Pakvasa, H.E. Haber, and A. Pomarol, in [*Proceedings of the Workshop on Physics at Current Accelerators and the Supercollider*]{}, June 1993, ed. J.L. Hewett, A. White, and D. Zeppenfeld. F. Abe , CDF Collaboration, hep-ex/9708017, submitted to Phys. Rev. Lett.; B. Abbott , D0 Collaboration, hep-ex/9707033, submitted to Phys. Rev. Lett. .For the most recent results, see B. Klima, D0 Collaboration, talk given at the [*International Europhysics Conference on High Energy Physics*]{}, Jerusalem, August 19-26, 1997. M. Krämer, T. Plehn, M. Spira and P. Zerwas, hep-ph/9704322. S. Davidson, D. Bailey, and B.A. Campbell, C61 613 1994 ; M. Leurer, D50 536 1994 , and [**D49**]{}, 333 (1994). C.S. Wood , Science [**275**]{} 1759 (1997). For a recent review and updated, see J.L. Rosner, hep-ph/9704331. R.M. Barnett , (Particle Data Group), D54 1 1996 . Z. Kunszt and W.J. Stirling, hep-ph/9703427; T. Plehn, H. Spiesberger, M. Spira and P.M. Zerwas, hep-ph/9703433. For an updated discussion on the effects of QCD and initial and final state radiation, see C. Freiberg, E. Norrbin and T. Sjöstrand, hep-ph/9704214. J.L. Hewett and T.G. Rizzo, 183 193 1989 . J.L. Hewett and T.G. Rizzo, D36 3367 1987 ; J. Bl" umlein and R. R" uckl, B304 337 1993 ; J. Blümlein and E. Boos, [*Nucl. Phys.*]{}(Proc. Suppl.) [*37B*]{}, 181 (1994) ; J. Bl" umlein, E. Boos, and A. Kryukov, B392 150 1997 ; D. Choudhury, B346 291 1995 ; J.E. Cieza-Montalvo and O.J. Eboli, D47 837 1993 ; T.G. Rizzo, D44 186 1991 ; H. Dreiner , 3A 443 1988 ; G. Bhattacharyya, D. Choudhury, and K. Sridhar, B349 118 1995 ; M.S. Berger, hep-ph/9609517. K. Ackerstaff , OPAL Collaboration, hep-ex/9708024. R. Rückl and P. Zerwas, in [*Proceedings of the LaThuile Workshop on Physics at Future Accelerators*]{}, LaThuile, Italy, 1987, ed. J.H. Mulvay, CERN Report 87-07 (1987). A. Buras, talk given at the [*Seventh International Symposium on Heavy Flavor Physics*]{}, University of California, Santa Barbara, July 7-11, 1997. This analysis relies heavily on the work presented in E. Hagberg , nucl-ex/9609002. M. Peskin and T. Takeuchi, 65 964 1990 , and D46 381 1992 ; W. Marciano and J. Rosner, 65 2963 1990 ; G. Altarelli and R. Barbieri, B253 161 1990 ; D. Kennedy and P. Langacker, 65 2967 1990 , and D44 1591 1991 ; I. Maksymyk, C.P. Burgess, and D. London, D50 529 1994 ; C.P. Burgess , B326 276 1994 . O. Eboli, hep-ph/9508342; O.J. Eboli, M.C. Gonzalez-Garcia, and J.K. Mizukoshi, hep-ph/9612254, and B443 20 1995 ; G. Bhattacharyya, J. Ellis, and K. Sridhar, B336 100 1994 , and Erratum [**B338**]{}, 522 (1994). See, A. Djouadi, J. Ng, and T.G. Rizzo, in [*Electroweak Symmetry Breaking and Beyond the Standard Model*]{}, ed. T. Barklow , (World Scientific, Singapore), hep-ph/9504210, and references therein. H. Murayama and T. Yanagida, A7 147 1992 . T.G. Rizzo, D45 3903 1992 . W.J. Marciano, talk given at the [*1983 International Symposium on Lepton and Photon Interactions at High Energies*]{}, Ithaca, NY, August 1983. I. Antoniadis, J. Ellis, J.S. Hagelin and D.V. Nanopolous, B194 321 1987 . M. Peskin, SLAC-PUB-7479, 1997. A. Bohm, L3 Collaboration, talk presented at the [*$32^{nd}$ Recontres de Moriond: Electroweak Interactions and Unified Theories*]{}, Le Arces, France, 15-22 March, 1997.
[^1]: Work supported by the Department of Energy, Contract DE-AC03-76SF00515
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A scheme to produce time-bin entangled pairs of electrons and holes is proposed. It is based on a high frequency time-resolved single-electron source from a quantum dot coupled to one-dimensional chiral channels. Operating the device in the weak tunneling regime, we show that at the lowest order in the tunneling rate, an electron-hole pair is emitted in a coherent superposition state of different time bins determined by the driving pulse sequence.'
author:
- 'Luca Chirolli, Vittorio Giovannetti, Rosario Fazio'
- Valerio Scarani
title: 'Time-bin entanglement of quasi-particles in semiconductor devices'
---
Introduction
============
Because of the very long coherence times, electronic states at the edge of a two-dimensional electron gas in the integer quantum Hall effect (IQHE) regime [@DATTA; @BEEN] are ideal systems for designing of coherent electronics circuitry or to implement quantum information processing. In particular, electronic versions of several optical interferometers have been realized [@INTERF; @NederNature07; @NederPRL06; @NederPRL07; @RoulleauPRL08; @RoulleauPRBr07; @RoulleauPRL09; @LitvinPRB07; @BieriPRB09] using continuous electron sources. Recently it was also shown that high frequency gate modulation can realize single electron sources (SESs) [@GlattliScience07; @MahePRB; @Blumenthal], that allow one to inject in a controlled and coherent way single electrons and holes onto an edge state. Exploiting the beam splitting technique via quantum point contact (QPC), the electronic wave packet produced by the SES could also be split, recombined, or coherently guided toward different paths, via the application of external gate voltages [@JAN1; @JAN]. Altogether, SESs, QPCs, and electronic waveguides represent the necessary toolbox toward the realization of an electronic version of the numerous protocols developed in quantum optics [@Tittel]. The only qualitative difference is in the measurement process, in that photons are usually absorbed by photodetectors and are studied in terms of $n$-photon coincidence correlation functions, whereas electrons are characterized by currents and higher-order momenta. In Ref. \[\] this formal equivalence was used to characterize two-particle non-local effects originating via collision and proper post-selection from two independent SESs whose output states were coherently mixed at a QPC.
A full exploitation of the capabilities of quantum information however requires the production and manipulation of entangled states. Several schemes to generate entangled states in multi-terminal mesoscopic conductors have been proposed so far (see \[\] and references therein). The realization of the SES [@GlattliScience07] opens possibilities to realize entangled states. Here we exploit the coherence of time-resolved single-electron wave packets at the output of a single SES, operated in the weak tunneling regime to generate superpositions of an electron-hole ([*e*]{}-[*h*]{}) pair produced at different times. Differently from Ref. \[\] our scheme does not rely on collisional mechanisms followed by post-selection: As a consequence, in our case the [*e*]{}-[*h*]{} pair emerges from the device in an entangled state without the need of any filtering processes. The resulting output closely reminds us of the bi-photon state produced by a down-conversion nonlinear crystal (see e.g. Ref. \[\]) or, more precisely, the entangled photon holes states [@Beenakker03; @FRANSON1] produced via two-photon absorbing processes. It admits a representation in terms of time-bin entanglement [@TIMBIN; @ScaraniPRL04] whose two-particle correlations we characterize by performing current cross-correlation measurements at the output of a Franson interferometer [@Franson] (the latter being an interferometric setting which is specifically designed to detect coherence properties of sources emitting pairs of correlated particles).
The paper is organized as follows. In Sec. \[sec:model\] we present the setup and describe in brief the basic idea of the proposal. The dynamics of the device is discussed in Sec. \[sec:dyn\] and in Sec. \[sec:nonidealcase\] we discuss deviation from the ideal case. Conclusions and remarks are given in Sec. \[sec:conclusion\].
![(Color online) (a) Driving sequence of the dot level showing the first and the second time-bin. (b) Pictorial representation of the state of the system at first order in the tunneling rate $\Gamma$. Here the red (gray) and blue (dark gray) dots represent, respectively, the electron and hole emitted from the SES, while the slots inside the brackets define the various time bins of the sequence. (c) Implementation of a Franson interferometer for an IQHE architecture at $\nu=1$. A quantum dot driven by a gate potential $U(t)$ generates an [*e*]{}-[*h*]{} pair on the edge channel at position $x=0$. The electron and hole propagate freeley until they reach the QPC driven by the time-dependent voltage $V(t)$ that splits them and send the electron toward the right-hand MZI [@INTERF] and the hole toward the left-hand one. The black elements are ohmic contacts to different reservoirs, all kept grounded.\[Fig1\]](FigureAPS){width="8cm"}
The setup {#sec:model}
=========
We start with an intuitive description of the model whose validity will be checked later. As shown in Fig. \[Fig1\] (c), the SES is described as a quantum dot whose energy levels are externally controlled via a time-dependent voltage gate $U(t)$, and which is connected to an edge state of the IQHE effect at a filling factor $\nu=1$ via a QPC characterized by a tunneling amplitude $A$ which can be externally modulated [@GlattliScience07; @MahePRB]. In this setup the linear dispersion relation of the edge state close to the Fermi energy gives rise to a constant velocity of propagation $v_F$ for the electronic wave packets: Consequently, a particle generated by the SES at the position $x=0$ of the edge will be found translated by an amount $v_Ft$ after time $t$ has elapsed. As in Refs. \[ and \[\], we consider then an elementary driving sequence of duration $2\tau$ consisting of two subsequent movements in which we first rapidly rise the dot energy level above the Fermi sea and keep it there for a time $\tau$, and then we rapidly lower it below the Fermi sea and keep it there for a time $\tau$. Assuming $\tau$ to be much smaller than the dot escaping time $1/\Gamma$ ($\Gamma\propto A^2$), a pair of well-separated time-resolved electron and hole can be created on the edge channel with probability $(2\Gamma\tau)^2\ll 1$, the electron being localized in the first half of the time-bin and the hole being localized in the second half (other processes in which no excitation or just a single excitation is emitted can be neglected since they do not contribute to the current cross-correlation measurements). This weak tunneling regime is crucial for the results we present and is actually opposite to the one usually considered in other proposals [@JAN1; @JAN] (to stress this fact from now on we will refer to our source as a weak tunneling SES, or wt-SES in brief). Still it is well within the reach of current experimental capabilities: For instance, considering that $1/\Gamma$ of the order of $10$ ns can be achieved while preserving the coherence of the process [@GlattliScience07; @MahePRB], the rising time of the dot energy level can be set to be of order of few ps, while $\tau$ can be taken to be of the order of a fraction of ns. Notice also that during the driving we never leave the dot at resonance with the Fermi level for an extended period of time: As experimentally verified in Ref. \[ and \[\], this allows us to avoid the collateral generation of pairs induced by resonance effect between the Fermi level and the dot [@Levitov08].
Consider hence the case in which the dot is initially charged and the elementary driving sequence is repeated several times, say, twice, as depicted in Fig. \[Fig1\](a). In the weak tunneling regime the global state of the system \[dot + one-dimensional (1D) line\] at the lowest order in the tunneling amplitude $A$ contains three contributions. First, we have an unperturbed component in which the dot is still charged and no excitations are produced in the 1D channel (zero-order contribution in $A$). Then there is a contribution proportional to $A$ in which the dot electron has emerged from the dot but no holes have been subsequently produced by the driving sequence. Last, we have a term proportional to $A^2$ describing the case in which no charge is trapped in the dot and the 1D line contains a delocalized [*e*]{}-[*h*]{} pair: As anticipated, this is the only component of the state which can contribute to the current cross correlations we perform at the output of the setup, the probability of the event being proportional to $|A|^4$ – see below. If the driving process is kept coherent, it is described as a coherent superposition of an [*e*]{}-[*h*]{} pair emitted in the first time bin, an [*e*]{}-[*h*]{} pair emitted in the second time bin and an [*e*]{}-[*h*]{} pair with the electron emitted in the first time bin and a hole emitted in the second time bin, – see Fig. \[Fig1\](b). We can hence represent it as the following vector: $$\label{Eq:ehTimeBin}
|\Psi\rangle\,\propto\,
|1,0\rangle_e|1,0\rangle_h+|0,1\rangle_e|0,1\rangle_h+|1,0\rangle_e|0,1\rangle_h\;,$$ where $|1,0\rangle_{e,h}$ is the state with one electron (hole) in the first time bin (first label in the ket) and none in the second time bin (second label in the ket), while $|0,1\rangle_{e,h}$ is the state with no electron (hole) in the first time bin and one electron (hole) in the second time bin. We stress that $|\Psi\rangle$ is only the second order contribution in the tunneling amplitude $A$ to the real full state of the system, which we can use to evaluate the current cross correlations at the output of the device (in other words, it is the component of the full state that gets post-selected by our measuring apparatus).
Equation (\[Eq:ehTimeBin\]) represents an entangled state of two qubits: For instance, it violates the Clauser-Horne-Shimony-Holt (CHSH) inequality [@Bib:CHSH] up to $$\begin{aligned}
\mathrm{CHSH}(\Psi)\,=\,2\sqrt{13}/3\,\approx\,2.404\;,\label{CHSH}\end{aligned}$$ for suitable measurements [@HorodeckiPLA95]. Specifically, identifying $|1,0\rangle$ and $|0,1\rangle$ with the eigenstates of $\sigma_z$, the value in (\[CHSH\]) is achieved by using the following set of local observables $$\begin{aligned}
\Sigma_e^{(0)}&=&(\sigma_z+2\sigma_x)/\sqrt{5}\;,
\nonumber \\
\Sigma_e^{(1)}&=&(\sigma_x-2\sigma_z)/\sqrt{5}\;, \label{SIGMAe}\end{aligned}$$ for the electronic part of Eq. (\[Eq:ehTimeBin\]), and $$\begin{aligned}
\Sigma_h^{(0)}&=&(4\sigma_x-7\sigma_z)/\sqrt{65}\;,\nonumber \\
\Sigma_h^{(1)}&=&(\sigma_z+8\sigma_x)/\sqrt{65}\;, \label{SIGMAh}\end{aligned}$$ for the hole part.
The time-bin entanglement of the state in Eq. (\[Eq:ehTimeBin\]) can be detected by means of a Franson interferometer [@Franson]. The two sub-systems (here, the electron and the hole) must be separated and sent to unbalanced Mach-Zehnder interferometers (MZIs), at the output of which coincidences are recorded. Furthermore, on each side, single-qubit measurements can be implemented probabilistically by adapting the phase delays and the transmittivity of the second beam-splitter (BS) of the MZI. Therefore, in principle, in the Franson setup the value of CHSH or of any other entanglement witness can be measured. This is well known and we refer the reader to the optical implementations for details [@optics]. Here, we rather have to discuss how a Franson setup can be realized for our [*e*]{}-[*h*]{} pairs generator. The electron and the hole in each time bin are separated by means of a time-dependent QPC that acts as a switch, sending the electron and the hole, respectively, toward different MZIs, which are implemented along the lines of Ref. \[\], as depicted in Fig. \[Fig1\]. This separation is a challenging task: The potential $V(t)$ of the QPC has to rise on a time scale equal or smaller to the inverse frequency of the gate potential $U(t)$ on the dot, in such a way that the switch perfectly “cuts" the wave function. The difference $\ell$ between the long and the short path in each MZI is chosen equal to $2v_F \tau$. On the one hand, this guarantees that no single-particle interference arises within the MZIs which then operate as effective, probabilistic (but coherent) delay lines. On the other hand, this allows us to align the time slot associated with the long path of a particle belonging to the first time bin with the time slot of the short path associated with second time bin. The current cross correlations are finally measured at the outputs of the setup, i.e., $$\begin{aligned}
\delta{\cal C}_{R_j,L_i}=\langle\delta I_{R_j}(t')\delta I_{L_i}(t)\rangle\;,\end{aligned}$$ with $\delta I_{\alpha_i}(t)= I_{\alpha_i}(t)-\langle I_{\alpha_i}(t)\rangle$, where for $\alpha=L,R$ and $i,j=1,2$, $I_{\alpha_i}(t)$ is the current operator at the $\alpha_i$ port evaluated at time $t$, and where $\langle \cdots \rangle$ stands for the expectation value on the initial state of the system (notice that different times $t\neq t'$ have to be chosen since the electron and hole are time shifted within the same time bin). This allows to post-select the events we are interested in, discarding single-particle currents.
The transmittivities of the BSs can be engineered as in Ref. \[\] and do not require special discussion. For the sake of simplicity, we first assume to fix them at $50\%$ as in the original proposal by Franson [@Franson]: Indeed, this choice is sufficient to detect a signal which is sensitive to the superposition of Eq. (\[Eq:ehTimeBin\]) \[instead to recover $\mathrm{CHSH}(\Psi)$ one would need to adjust the transmittivities as detailed at the end of the section\]. In order to have a phase difference between the two arms of a MZI, we invoke instead the Aharonov-Bohm (AB) effect. Accordingly, the statistics of the coincidence events will be sensitive to a non-local two-particle AB phase [@ButtikerPRL04] that gives rise to interference fringes that can witness entanglement in the [*e*]{}-[*h*]{} pair. To see this, consider that each of the three input terms of Eq. (\[Eq:ehTimeBin\]) is mapped after the MZIs into the sum of four states. Out of the 12 total contributions, three will be indistinguishable when revealing the proper coincidence counts: The event where the [*e*]{}-[*h*]{} pair is generated in the first time-bin and both particles choose the long paths in the MZIs, i.e., $$\begin{aligned}
|1,0\rangle_e |1,0\rangle_h \rightarrow e^{i\varphi_R-i\varphi_L}|0,1\rangle_e|0,1\rangle_h+\cdots\;,\end{aligned}$$ (here $\varphi_{R,L}$ are the AB phases computed with a gauge for which only particles traveling along the long arms of the MZIs acquire a phase, the difference in sign arising from the opposite charge for electron and the hole); the event where the [*e*]{}-[*h*]{} pair is generated in the second time bin and both the particles choose the short paths of the MZIs, i.e., $$\begin{aligned}
|0,1\rangle_e|0,1\rangle_h\rightarrow |0,1\rangle_e|0,1\rangle_h + \cdots\;;
\end{aligned}$$ finally, the event where the electron is generated in the first time bin, the hole in the second time bin and choose, respectively, the long and the short paths, i.e., $$\begin{aligned}
|1,0\rangle_e|0,1\rangle_h\rightarrow e^{i\varphi_R} |0,1\rangle_e|0,1\rangle_h + \cdots\;.
\end{aligned}$$ Measuring then coincidence events in the second time-bin at the output of the MZIs (i.e., $|0,1\rangle_e|0,1\rangle_h$) will then produce a signal which is sensitive to the coherent superposition of Eq. (\[Eq:ehTimeBin\]), i.e., $$\begin{aligned}
\delta{\cal C}_{R1,L1}&=&-({\Gamma}/{4})^2\; |1+e^{i\varphi_L}+e^{i\varphi_L-i\varphi_R} |^2\;. \label{COINC}\end{aligned}$$ The three terms correspond to the three vectors which compose Eq. (\[Eq:ehTimeBin\]) and display coherent oscillations. In particular, a dependence on a non-local two-particle AB phase appears as the difference of the AB phases of the left- and right-hand MZIs.
Measuring CHSH correlations {#sec:CHSH}
---------------------------
We analyze in more details how to use our Franson setting to measure the value of CHSH$(\Psi)$. We remind that to do so we need to perform local measurements on the two subsystems (electron and hole separately). As anticipated, this can be done in a probabilistic fashion by properly adjusting the transmittivities of the BS of the setup. Indeed, let us consider the effect of an asymmetric MZI, characterized by transmission $\sqrt{s}$ and reflection $i\sqrt{r}$ amplitudes of the second BS, while keeping the first BS symmetric. Let us suppose also that the generic state of, say, the electron before the right-hand MZI can be written as $|\phi\rangle_{\rm in}=\alpha|1,0\rangle_e+\beta|0,1\rangle_e$. By focusing on the second time bin at the outputs of the MZI, the electron will come out with an amplitude $$\begin{aligned}
S_1=-(\alpha e^{i\varphi_R}\sqrt{r}-\beta\sqrt{s})/\sqrt{2}\;,\end{aligned}$$ from the output 1 and with amplitude $$\begin{aligned}
S_{2}=i(\alpha e^{i\varphi_R}\sqrt{s}+\beta\sqrt{r})/\sqrt{2}\;,\end{aligned}$$ from output 2. Defining the rotated input qubit states $$\begin{aligned}
|u_+\rangle&=&e^{-i\varphi_R}\sqrt{s}|1,0\rangle_e+\sqrt{r}|0,1\rangle_e\nonumber \\
|u_-\rangle&=&e^{-i\varphi_R}\sqrt{r}|1,0\rangle_e-\sqrt{s}|0,1\rangle_e\end{aligned}$$ it follows that $\langle u_+|\phi\rangle_{\rm in}=-i\sqrt{2}S_2$ and $\langle u_-|\phi\rangle_{\rm in}=-\sqrt{2}S_1$, and it becomes clear that in the second time-bin outputs one reads the results of the measurements of $|\phi\rangle_{\rm in}$ on the eigenstates of $$\begin{aligned}
(s-r)\sigma_z+2\sqrt{sr}(\cos(\varphi_R)\sigma_x+\sin(\varphi_R)\sigma_y)\;,\end{aligned}$$ where the Pauli matrices are written in the basis $|1,0\rangle_e$ and $|0,1\rangle_e$. The probability of success is $1/2$, provided that $|\alpha|^2+|\beta|^2=1$. According to this analysis we can hence realize the observables (\[SIGMAe\]) by taking $\varphi_R=0$ and $s_{0}=(1+1/\sqrt{5})/2$, $s_{1}=(1-2/\sqrt{5})/2$ for the MZI on the right-hand side of Fig. \[Fig1\], and the observables (\[SIGMAh\]) by taking instead $\varphi_L=0$ and $s_{0}=(1-7/\sqrt{65})/2$, and $s_{1}=(1+1/\sqrt{65})/2$ for the MZI on the left-hand side.
Microscopic model {#sec:dyn}
=================
We now turn to a more quantitative model where the dynamics of the wt-SES is described via a time-dependent Hamiltonian [@Levitov08] of the form $$\begin{aligned}
H(t)=\varepsilon_D(t)d^{\dag}d+\sum_{k}\varepsilon_kc^{\dag}_{k}c_{k} + H_{\rm tun}(t)\;,\end{aligned}$$ where $d$ and $c_k$ are, respectively, the Fermionic annihilation operators of the dot energy level and of the 1-D free-electron modes associated with the chiral IQHE edge channel at $\nu=1$, while $H_{\rm tun}(t)$ is the tunneling term. We take the dot energy $\varepsilon_D(t)$ as in Fig. \[Fig1\](a) while, assuming linear dispersion around the Fermi energy (set to zero), we write the energy levels of the $c_k$ modes as $\varepsilon_k=\hbar v_F k$ (nonlinear corrections being typically negligible in IQHE systems for small bias voltages). In the weak-coupling regime we consider a (time dependent) tunneling amplitude peaked around the resonance value $k_D(t) =\varepsilon_D(t)/\hbar v_F$ associated with the instantaneous dot energy [@MahePRB] within a bandwidth $BW=\hbar v_F\Delta k$, i.e., $$\begin{aligned}
H_{\rm tun}(t) =A\sum_{k\in[k_D(t) ,\Delta k]}(d^{\dag}c_k+c^{\dag}_kd)
\;, \end{aligned}$$ and assume $k_D>\Delta k$. The spread $\Delta k$ is associated with the uncertainty in the emission position of the electron on the 1D channel, taking into account the not perfect point-like coupling between the dot and the edge, allowing hence the tunneling hopping to extend over a range $\Delta x\simeq 2\pi /\Delta k$. A natural bound on $\Delta x$ can be set as $\Delta x\leq R_D$, where $R_D$ is the linear dimension of the quantum dot. The latter, however, is directly related with the dot energy-level spacing $\Delta=2 \varepsilon_D$ via the expression $\Delta\simeq \hbar^2(2\pi/R_D)^2/2m^*$, $m^*$ being the effective mass of the electron ($m^*=0.068~m_0$ for GaAs heterostructures). It follows a minimum bandwidth on order $$\begin{aligned}
BW_0=2\pi \hbar v_F/R_D=\sqrt{2m^*v_F^2\Delta}\;.\end{aligned}$$ To put some number we notice in the experiment of Ref. \[\] one has $\Delta\simeq 4.4~{\rm K}$ ($\Delta\simeq 0.36~{\rm meV}$). Choosing $v_F\simeq 10^{4}~{\rm m/s}$, we get $BW_{0}\simeq~0.2~{\rm meV}$. Therefore, we see that the validity of the condition $k_D>\Delta k$, ensured if $BW_0/\Delta\leq BW/\Delta<1$, can be satisfied.
In the interaction picture with respect to free evolution of the system, the Hamiltonian becomes hence $$\begin{aligned}
H_I=A\sqrt{L_c}\,e^{i\varphi_D(t)}\,d^{\dag}\,\psi_{k_D(t)}(-v_Ft)+{\rm h. c.}\;,\end{aligned}$$ where $$\begin{aligned}
\varphi_D(t)=\frac{1}{\hbar}\int^tdt'\varepsilon_D(t')\;, \end{aligned}$$ is the dot dynamical phase, $L_c$ is the channel length, and $$\begin{aligned}
\psi_{k_D}(x)=\frac{1}{\sqrt{L_c}}\sum_{k\in[k_D,\Delta k]}e^{ikx}c_k\;,\end{aligned}$$ is the field operator in position space of an electron propagating with mean momentum $k_D$. When applied to the Fermi sea, the operator $\psi^{\dag}_{k_D}(x)$ adds an electron above the Fermi sea only if $k_D>0$ ($k_F=0$). If the initial dot energy is negative, $\varepsilon_D<0$, the electronic field operator $\psi_{-|k_D|}(x)$, when applied to the Fermi sea, creates a hole of average momentum $|k_D|$. Field operators at positions $x$ and $x'$ satisfy canonical anti-commutation rules if $|x-x'|\gg \Delta x$, which amounts to a coarse graining of the position resolution. Due to the linearity of the channel dispersion, we have a one-to-one mapping between position and time, which implies a coarse graining in the time coordinate: Fields at times $t$ and $t+\delta t$, with $\delta t\lesssim 2\pi/(v_F\Delta k)$, are indistinguishable and the anti-commutation relations at times differing by $\delta t$ have to be understood as at equal time.
Consider now the time evolution of the input state $|1\rangle_D|\Theta_F\rangle$, where $|1\rangle_D$ describe the charged state of the dot and $|\Theta_F\rangle$ is the free Fermi-sea state of the edge. In the interaction picture the first correction that gives rise to nonzero current cross correlations at the output is described by the vector
$$\begin{aligned}
\int_0^t dt'\int_0^{t'} dt''e^{i\varphi_D(t',t'')}\psi_{k_D(t')}(-v_Ft')\psi^{\dag}_{k_D(t'')}(-v_Ft'')
|\Theta_F\rangle\;,\end{aligned}$$
which exhibits an electron of momentum $k_D(t'')$ at time $t''$ and a hole with momentum $-k_D(t')$ at time $t'>t''$. Assuming piece-wise constant $k_D$, sharp transitions, and neglecting transient effects, we can set $k_D>0$ for $0<t<\tau$ and $k_D<0$ for $\tau<t<2\tau$, as shown in Fig. \[Fig1\](a). Since the two consecutive time slots do not overlap, one can create a state with the electron localized within $\ell/2<x<\ell$ and a hole localized within $0<x< \ell/2$ (here $\ell\equiv 2v_F\tau$). Defining thus the electron and hole field operators $\psi_e(x)\equiv\psi_{k_D}(x)$, $\psi_h^{\dag}(x)\equiv\psi_{-k_D}(x)$, we can write the above vector after time $t=4\tau$ as $$\label{Eq:eh-state}
|\Psi\rangle=-\frac{\Gamma}{v_F}\int dx_1dx_2\phi^{(2)}(x_1,x_2)\psi^{\dag}_h(x_1)\psi_e^{\dag}(x_2)|\Theta_F\rangle,$$ with the escaping rate $\Gamma=A^2L_c/\hbar^2v_F$ and the [*e*]{}-[*h*]{} wave function $$\begin{aligned}
\phi^{(2)}(x_1,x_2)&=&{\rm Step}(x_1,x_2)+e^{-ik_D\ell}{\rm Step}(x_1+\ell,x_2)\nonumber \\
&&+e^{-ik_D\ell}{\rm Step}(x_1+\ell,x_2+\ell)\;,\end{aligned}$$ expressed in terms of the Heaviside distribution $\theta(x)$ through the identity $$\begin{aligned}
{\rm Step}(x_1,x_2)&=&e^{ik_D(x_1+x_2)}\theta(x_1+\ell)\theta(-x_1-\ell/2)\nonumber \\
& &\times \theta(-x_2)\theta(x_2+\ell/2)\;.\end{aligned}$$ This is the second quantization representation of the state of Eq. (\[Eq:ehTimeBin\]) which defines, in the interaction picture, the electron-hole distribution on the 1-D channel that connects the dot with the time-dependent QCP. The propagation through the latter can then be described in the Landauer-Büttiker formalism [@LandauerButtiker] by introducing a time-dependent transmission amplitude toward the right-hand MZI, $s_R(t)$, and reflection amplitude toward the left-hand MZI, $s_L(t)$, with $|s_L|^2+|s_R|^2=1$, ${\rm Re}~s^*_Ls_R=0$. We assume sharp transitions and neglect transient effects, such that the QPC is totally transmitting during the first half of the time bin and it is totally reflecting during the second half of the time bin (i.e., $|s_R(t)s_L(t+\tau)|=1$). For $\alpha = R,L$ and $j=1,2$, the field at the $\alpha_j$ output of the setup can then be expressed as $$\begin{aligned}
\psi_{\alpha_j}(x,t)&=&i^{j-1}s_{\alpha}(t)\left[\psi(x-v_Ft)\right.\nonumber\\
&+&\left.(-)^{j-1}e^{i\varphi_{\alpha}}\psi(x+\ell- v_F t)\right]/2\;, \end{aligned}$$ where $\varphi_{\alpha}$ the AB phases and where $\psi(x-v_Ft)$ is the drifting full free field operator $$\begin{aligned}
\psi(x)=\frac{1}{\sqrt{L_c}} \sum_k e^{ikx} c_k\;,\end{aligned}$$ of the channel that connects the dot with the time-dependent QCP which, neglecting the Fermi-sea contribution, in our case can be approximated as $\psi(x)\approx\psi_e(x)+\psi^{\dag}_h(x)$. The correlation $\delta C_{R_1,L_1}$ of Eq. (\[COINC\]) is then computed by observing that the current density operator at position $x$ in the 1D channel associated with the output port $\alpha_j$ can be expressed as $$\begin{aligned}
{I}_{\alpha_j} (x,t)=v_F{\psi}_{\alpha_j}^{\dag}(x,t){\psi}_{\alpha_j}(x,t)\;.\end{aligned}$$ Setting $x_e=x-v_Ft$, $y_h=x_e-v_F\tau$ to compensate for the time shift between the electron and the hole, the resulting expression for the cross-correlator results in $$\begin{aligned}
\delta C_{R_1,L_1}= -\frac{v_F^2}{16} \Big| \sum_{a,b=0,1} {\cal G}(x_e+a \ell, y_h + b \ell) e^{i a \varphi_R - i b \varphi_L}\Big|^2,\nonumber\\\end{aligned}$$ in terms of the Green’s function ${\cal G}(x,y)\equiv\langle\psi^{\dag}(x,t)\psi(y,t)\rangle$, which at the lowest order in $\Gamma$ is $$\begin{aligned}
{\cal G}(x_e, y_h) &=&\langle\Theta_F| \psi^{\dag}(x_e)\psi(y_h)|\Psi\rangle\nonumber\\
&+&\langle\Psi| \psi^{\dag}(x_e)\psi(y_h)|\Theta_F\rangle\;.
\end{aligned}$$ Choosing then $k_D\ell=2\pi$ and measuring the currents at the central peak $-3/2<x_e/\ell<-1$ one finally gets Eq. (\[COINC\]).
Non ideal case {#sec:nonidealcase}
==============
Until now we have considered an ideal situation of a linear dispersion channel, a zero-temperature working regime, and no dephasing process has been taken into account. Quadratic deviation from a linear dispersion implies that each $k$ state propagates at his own speed, yielding a spread of the wave function. Typically, electron-like states with energy above the Fermi energy propagate slightly faster than hole-like states with energy below the Fermi energy. Consequently, the time bin becomes smeared and adjacent time bins develop an overlap as the electron-hole pair propagates along the channel, making the synchronization of the electron and hole more difficult.
After the state Eq. (\[Eq:eh-state\]) has been produced at position $x=0$, it propagates into the channel and interaction with the environment reduces the degree of time-bin entanglement of the electron-hole pair. Typically, in the IQHE the single-particle coherence length has been proved to be very long, on the scale of hundreds of micrometers. For dc-biased MZI interferometers a direct measurement of the single-particle coherence length has been reported in Ref. \[\] by monitoring the decrease of the visibility of single-particle Aharonov-Bohm oscillation in MZIs of different sizes. An observed $1/T$ dependence versus the temperature has been attributed to thermal noise of the dissipative part of the finite frequency coupling impedance between the environment and the reservoirs. Short- and long-range interactions as well as curvature of the fermion dispersion have been ruled out due to an expected different dependence on temperature. As far as two-particle processes are concerned, an experimental measurement of two-particle Aharonov-Bohm phase has been performed by Neder [*et al.*]{} \[\], where a visibility on order of $70\%$ for a dc-biased case has been reported.
In the case that we consider, the electronic reservoirs are kept at the same bias and the non-equilibrium nature of the excitation is entirely due to the dot driving. At the moment no experimental test of single- or two-particle interferometry with SESs or wt-SESs has been reported to our knowledge. On the other hand, a theoretical study has shown that, for a single-electron wave packet injected in chiral Luttinger liquid, the deformation of the wavepacket due to electron-electron interaction can be partly undone by a suitable voltage pulse \[\], whereas in a real device a capacitive Coulomb interaction may add dissipation into the system \[\].
Conclusions {#sec:conclusion}
===========
By operating a SES in the weak tunneling regime, we have proposed a scheme to generate a time-bin entangled state of an [*e*]{}-[*h*]{} pair which can be detected via current cross-correlation measurements at the output of a Franson interferometer. Within the range of validity of our approximations the only higher-order excitations produced in a two time-bin cycle are $|1,1\rangle_e |1,0\rangle_h$ and $|1,1\rangle_e|1,1\rangle_h$, which are clearly discriminable since they bring about a different charge (hence, by monitoring all the four outputs of the Franson interferometer, it is possible, in principle, to discard their contribution). Transient effects, together with all higher-order terms, also contribute to the shape of the electron-hole wave function, which within our first-order approximation only amounts to a phase, resulting in a delocalization of the electron and hole in the two time bins and a degradation of the signal.
We acknowledge financial support by the FIRB-IDEAS project, RBID08B3FM, EU Project IP-SOLID, and by the National Research Foundation and the Ministry of Education of Singapore.
[99]{}
S. Datta, [*Electronic Transport in Mesoscopic Systems*]{} (Cambridge University Press, Cambridge U. K., 1997).
C. W. J. Beenakker, in [*Quantum Computers, Algorithms and Chaos*]{}, Proceedings of the International School Physics “Enrico Fermi”, Course CLXII, edited by G. Casati, D. L. Shepelyansky, P. Zoller, and G. Benenti (IOS, Amsterdam, 2006), Vol. 162.
Y. Ji, [*et al.*]{}, Nature (London) **422**, 415 (2003);
I. Neder, N. Ofek, Y. Chung, M. Heiblum, D. Mahalu, and V. Umansky, Nature (London) [**448**]{}, 333 (2007).
I. Neder, M. Heiblum, Y. Levinson, D. Mahalu, and V. Umansky, Phys. Rev. Lett. [**96**]{}, 016804, (2006).
I. Neder, M. Heiblum, D. Mahalu, and V. Umansky, Phys. Rev. Lett. [**98**]{}, 036803 (2007).
P. Roulleau, F. Portier, P. Roche, A. Cavanna, G. Faini, U. Gennser, and D. Mailly, Phys. Rev. Lett. [**100**]{}, 126802 (2008).
P. Roulleau, F. Portier, D.C. Glatti, P. Roche, A. Cavanna, G. Faini, U. Gennser, and D. Mailly, Phys. Rev. B [**76**]{}, 161309(R) (2007).
P. Roulleau, F. Portier, P. Roche, A. Cavanna, G. Faini, U. Gennser, D. Mailly, Phys. Rev. Lett. [**102**]{}, 236802 (2009).
L. V. Litvin, H.-P. Tranitz, W. Wegscheider, and C. Strunk, Phys. Rev. B [**75**]{}, 033315 (2007); [*ibid.*]{} Phys. Rev. B [**78**]{}, 075303 (2008).
E. Bieri, M. Weiss, O. Götkas, M. Hauser, C. Schönenberger, and S. Oberholzer, Phys. Rev. B [**79**]{}, 245324 (2009).
F. Feve, [*et al.*]{}, Science [**316**]{}, 1169 (2007); Physica E, [**40**]{}, 954 (2008).
M. D. Blumenthal, [*et al.*]{}, Nature Phys. [**3**]{}, 343 (2007).
A. Mahé, [*et al.*]{}, Phys. Rev. B [**82**]{}, 201309(R) (2010).
S. Ol’khovskaya, [*et al.*]{} , Phys. Rev. B [**78**]{}, 205110 (2008); Phys. Rev. Lett. [**101**]{}, 166802 (2008).
J. Splettstoesser, M. Moskalets, and M. Büttiker, Phys. Rev. Lett. [**103**]{}, 076804 (2009).
W. Tittel and G. Weihs, Quant. Inf. Comput. [**1**]{}(2), 3 (2001).
P. G. Kwiat, K. Mattle, H. Weinfurter, A. Zeilinger, A. V. Sergienko, and Y. Shih, Phys. Rev. Lett. [**75**]{}, 4337 (1995).
C. W. J. Beenakker, [*et al.*]{}, Phys. Rev. Lett. [**91**]{}, 147901 (2003).
J. D. Franson, Phys. Rev. Lett. [**96**]{}, 090402 (2006).
J. Brendel, [*et al.*]{}, Phys. Rev. Lett. [**82**]{}, 2594 (1999); R. T. Thew, [*et al.*]{}, Phys. Rev. A [**66**]{}, 062304 (2002).
V. Scarani, N. Gisin, S. Popescu, Phys. Rev. Lett. [**92**]{}, 167901 (2004).
J. D. Franson, Phys. Rev. Lett. [**62**]{}, 2205 (1989); J. D. Franson, in [*Proceedings of the Ninth Rochester Conference on Coherence and Quantum Optics*]{}, edited by N. P. Bigelow, J. H. Eberly, and C. R. Stroud, Jr., (Optical Society of America, Washington, DC, 2008), p. 178.
M. Albert, C. Flindt, and M. Büttiker, Phys. Rev. B [**82**]{}, 041407(R) (2010).
J. Keeling, A. V. Shytov, and L. S. Levitov, Phys. Rev. Lett. [**101**]{}, 196404 (2008).
J. F. Clauser, [*et al.*]{}, Phys. Rev. Lett. [**23**]{}, 880 (1969).
R. Horodecki, P. Horodecki, M. Horodecki, Phys. Lett. A [**200**]{}, 340 (1995).
J. Brendel, E. Mohler, and W. Martienssen, Phys. Rev. Lett. **66**, 1142 (1991); W. Tittel, [*et al.*]{}, ibid. **81**, 3563 (1998); I. Marcikic, [*et al.*]{}, Phys. Rev. A **66**, 062308 (2002).
P. Samuelsson, E. V. Sukhorukov and M. Büttiker, Phys. Rev. Lett. [**92**]{}, 026805 (2004).
M. Büttiker, [*et al.*]{}, Phys. Rev. B [**31**]{}, 6207 (1985); M. Büttiker, ibid. [**38**]{}, 9375 (1988).
A. V. Lebedev and G. Blatter, Phys. Rev. Lett. [**107**]{}, 076802 (2011).
P. Degiovanni, Ch. Grenier, and G. Feve, Phys. Rev. B [**80**]{}, 241307(R) (2009).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Online symptom checkers have significant potential to improve patient care, however their reliability and accuracy remain variable. We hypothesised that an artificial intelligence (AI) powered triage and diagnostic system would compare favourably with human doctors with respect to triage and diagnostic accuracy. We performed a prospective validation study of the accuracy and safety of an AI powered triage and diagnostic system. Identical cases were evaluated by both an AI system and human doctors. Differential diagnoses and triage outcomes were evaluated by an independent judge, who was blinded from knowing the source (AI system or human doctor) of the outcomes. Independently of these cases, vignettes from publicly available resources were also assessed to provide a benchmark to previous studies and the diagnostic component of the MRCGP exam. Overall we found that the Babylon AI powered Triage and Diagnostic System was able to identify the condition modelled by a clinical vignette with accuracy comparable to human doctors (in terms of precision and recall). In addition, we found that the triage advice recommended by the AI System was, on average, safer than that of human doctors, when compared to the ranges of acceptable triage provided by independent expert judges, with only a minimal reduction in appropriateness.'
author:
- '[ Salman Razzaki$^*$, Adam Baker$^*$, Yura Perov$^*$, Katherine Middleton$^*$, Janie Baxter$^*$, Daniel Mullarkey$^*$, Davinder Sangar$^*$, Michael Taliercio$^*$, Mobasher Butt$^*$, Azeem Majeed$\textsuperscript{\textdagger}$, Arnold DoRosario$\textsuperscript{\textdaggerdbl}$, Megan Mahoney$\textsuperscript{\textsection}$ and Saurabh Johri$\textsuperscript{*,\textparagraph}$ ]{}'
bibliography:
- 'references.bib'
title: A comparative study of artificial intelligence and human doctors for the purpose of triage and diagnosis
---
INTRODUCTION {#sec:intro}
============
Online symptom checkers are a convenient and valuable resource for users to better understand the underlying cause(s) of their symptoms and to receive advice on the most appropriate point of care. Typically, symptom checkers cater to three healthcare needs of a patient. First is the provision of information, wherein a patient may seek to know more about the symptoms or conditions that they know or think they have. Secondly, a patient may want to know whether their symptoms require treatment or further investigation; this is medical triage and involves directing patients to the most suitable location within an appropriate time frame. The appropriate action depends on the nature and urgency of the symptoms or their underlying cause, which might require further investigation. Finally, patients may want to understand the conditions that might be responsible for their symptoms. This corresponds to diagnosis or “differential diagnosis” and is typically performed by an experienced medical practitioner.
Symptom checkers have the potential to alleviate the pressure on overly burdened healthcare systems. For this to happen, healthcare professionals and the wider public must have confidence in the performance of symptom checkers and applications of AI to medicine more generally. Previous work has investigated the diagnostic and triage accuracy of competing symptom checkers and highlighted significant variation in terms of clinical accuracy [@semigran2015evaluation]. Whilst providing a useful benchmark, that study did not assess the accuracy of symptom checkers against the gold-standard performance of human doctors. This was assessed in a follow-up study, where the authors noted that doctors significantly outperform symptom checkers, providing a valuable contribution to our understanding of comparative diagnostic performance [@semigran2016comparison]. However, the method used in this follow-up study did not adequately assess the information gathering process through which patients typically interact with symptom checkers or doctors, and so the conclusions are not based on a fair or realistic comparison. Diagnostic accuracy is not routinely measured in clinical practice, but a wide range of studies have attempted to estimate the incidence of diagnostic error. Irrespective of whether the true error rate lies closer to the 10-20% found in autopsy studies [@graber2013incidence] or the 44% found in a study analysing the correlation of diagnostic accuracy with doctor confidence [@meyer2013physicians], it is critical to perform a fair assessment of how a doctor takes a history and establishes a diagnosis when comparing against symptom checkers.
In this study we adopt a semi-naturalistic, role-play paradigm that simulates a realistic consultation between a patient and either our Triage and Diagnostic System or human doctor. Based on the assessment technique used throughout medical school and post-graduate medical qualifications (Objective Structured Clinical Examinations \[OSCE\]), this protocol was designed to assess not only the clinical (diagnostic and triage) accuracy, but also the ability to gather all of the relevant information from the patient i.e. to take a history.
The Babylon Triage and Diagnostic System {#sec:babylon_td_system}
========================================
The Babylon Triage and Diagnostic System[^1][^2] is designed to provide users with triage advice alongside an explanation of why this action has been suggested; this consists of any reported symptoms that require urgent attention, and/or a list of possible causes for the user’s symptoms. A comprehensive description of the system that powers the Babylon Triage and Diagnostic System is outside of the scope of this paper, however we provide a brief summary of this system by way of background.
The Babylon Triage and Diagnostic System – a new implementation after the previous generation [@middleton2016sorting] – is based on a Probabilistic Graphical Model (PGM) [@koller2009probabilistic] of primary care medicine, which models the prior probabilities of diseases and the conditional dependencies between diseases, symptoms and risk factors via a directed acyclic graph. The structure of the graph (i.e., the connections between diseases, symptoms and risk factors) is created by medical experts and reviewed from a modelling perspective. The prior probabilities of diseases and risk factors are obtained from epidemiological data, where available. Conditional probabilities (for example, the probability of a symptom occurring given a disease) are obtained through elicitation from multiple independent medical experts.
Once constructed and parameterised, the model is used to reason about the possible underlying diseases that cause the user-entered symptoms and risk factors, using novel Bayesian inference methods [@douglas2017universal; @cheng2000ais; @wainwright2008graphical; @gu2015neural]. This allows the AI powered Triage and Diagnostic System to output the most likely causes of the symptoms entered by a user, and also generate follow up questions that provide the most information to confirm or rule out the disease under consideration.
The triage capability is based on an extension of the PGM with a utility model [@koller2009probabilistic]. This combines the likelihood of each disease with the potential harm caused by that disease, under all possible triage decisions. The triage decision that is recommended is the one that minimises the expected harm to the patient, while also penalising overtriaging. To guarantee the safe triage of patients with symptoms that require urgent treatment (regardless of their underlying cause), the utility model is augmented with a set of rules that dictate a specific triage action where a particular combination of symptoms (so-called “red-flag” symptoms) are present. The system is designed to identify one of six triage actions: “call an ambulance”, “go to A&E/ER”, “urgent GP” (i.e., within 6 hours), “non-urgent GP” (i.e. within a week), “pharmacy” and “self-care”.
Iterative validation and development {#sec:val_and_development}
====================================
During development, the performance of the model is continuously assessed through an internal validation process to identify areas of weakness that could be improved. Validation is performed against a set of simulated clinical vignettes that are created by medical experts within an internal web tool. Each clinical vignette is written to reflect a realistic presentation of a patient with a particular disease or condition, containing the patient’s symptoms, past medical history and basic demographic information such as age and sex. The author of the vignette is instructed to also enter any absent symptoms or risk factors that a doctor would be likely to enquire about during a typical consultation. Where possible, symptoms and risk factors match those in the model to allow the system to recognise these entities automatically. However, to avoid the bias of only including those entities that are present in the model, the author of the clinical vignette is allowed to enter any other symptoms or risk factors via free text.
All clinical vignettes are assessed by our model in two modes: 1) by only providing to the model those symptoms that are elicited by the chatbot, and 2) by providing to the model all symptom and risk-factor entities listed on the vignette. This helps evaluate not only the accuracy of the model, but also the history-taking ability of the system; because the chatbot must choose the appropriate questions to ask it is not guaranteed that all entities that exist on the vignette will be available to the model.
The clinical vignettes are also assessed by doctors. Each doctor independently reviews the vignette and enters their own differential diagnosis and triage outcomes. This allows the predictions from the model (the list of possible causes and recommended triage action) to be compared not only against the disease modelled by the clinical vignette but also against the consensus of multiple doctors.
Those vignettes against which the model performs poorly are analysed in depth by a combination of doctors and scientists to identify parts of the model that require review, and this process of validation and iterative improvement results in continuous improvement of the model. As is standard practice in machine learning, any vignettes used to train or otherwise inform the structure or parameters of the model are completely separate from the test vignettes we used for the experiments outlined below.
Methods {#sec:methods}
=======
We compared the accuracy and safety of the Babylon Triage and Diagnostic System against that of human doctors. Accuracy was assessed in terms of the relevance of the suggested conditions, and the appropriateness of the recommended triage action. Triage safety was assessed in terms of whether the suggested triage action was deemed safe (even if it was overly cautious).
Outline
-------
The evaluation was performed using a semi-naturalistic role-play scenario that involved mock consultations between a patient and either a human doctor or the chatbot, based on realistic clinical vignettes. The role of doctors was played by general practitioners (GPs) who were hired on a locum basis for the experiment and who were not involved with the development of the model. Patients were played by GPs, some of whom were employees of Babylon, but none of whom were involved with the development of the model. We opted to use GPs to play the patients instead of professional actors as in a previous study [@middleton2016sorting], to prioritise the accuracy of interpreting the information on the vignette over the realism of a layperson. One hundred clinical vignettes were created by independent medical practitioners who were not involved in the role-play experiment. Each vignette was designed to simulate a medical condition from the list of all conditions currently modelled by the Triage and Diagnostic System[^3], in a patient of at least 16 years of age. The vignettes contained information about the patient, their initial complaint(s), information about their symptoms and past medical history that should be offered on open questioning, and information that should only be reported on direct questioning. An example can be found in Supplementary Figure \[fig:example\_vignette\].
Testing paradigm
----------------
The study was conducted in four rounds over consecutive days. In each round, there were up to four “patients” and four doctors. Each patient was assigned a vignette as their presenting history and had independent consultations with each doctor and the Babylon Triage and Diagnostic System. This protocol was designed in the OSCE format to assess both history taking and diagnostic and triage accuracy. After each consultation the differential diagnosis and recommended triage produced by the doctor or Triage and Diagnostic System was recorded. In order to maintain blinding in the judging process, doctors selected their differential diagnoses from a list of all conditions covered by the Triage and Diagnostic System. Once the patient had completed consultations with all doctors and the Triage and Diagnostic System, they were assigned a new vignette and the process was repeated.
Results {#sec:results}
=======
Accuracy of differential diagnosis against vignette modelled disease
--------------------------------------------------------------------
We assessed the precision (also called positive predictive value) and recall (also called sensitivity) of the Babylon Triage and Diagnostic System and doctors against the condition modelled by the vignette. Recall is the proportion of relevant diseases that are included in the differential. When considering only the single disease modelled by the vignette, this corresponds to the proportion of differentials that contained the modelled disease, over all vignettes. Precision is the proportion of the diseases in the differential that are relevant. A precision of one hundred percent would be achieved if the differential diagnosis contained only the disease modelled by the vignette. In general this level of certainty is unlikely and even undesirable, given only the information provided on the vignette (i.e. in the absence of diagnostic tests), but penalises overly long differentials that would result in a higher recall.
In this study, the Babylon Triage and Diagnostic System was able to produce differential diagnoses with precision and recall comparable to that of doctors, and in some cases exceeded human level performance (Table \[tab:doctor\_performance\]). The average recall of doctors was found to be 83.9%, (64.1–93.8%), meaning that doctors failed to include the vignette modelled disease in their differential in sixteen percent of cases on average.
The Babylon Symptom Selector is based on a Bayesian model, which can calculate the posterior probabilities of all conditions in the model given the evidence known about a patient. Whether particular conditions are displayed to the user depends on whether they meet internal thresholds, defined by a combination of the probability and severity of these conditions. The threshold parameters used in the model are selected based on independent training vignettes but may be varied to allow a trade-off to be made between recall and precision. It is interesting to observe that different parameters can move the model’s result closer to those of different doctors, for example towards Doctor D or E (Figure \[fig:main\_PR\_curve\]), perhaps emulating the variability in individual doctors’ preference for shorter, more precise differentials or longer, more exhaustive ones.
{width="\textwidth"}
Expert rating of differential diagnoses
---------------------------------------
In addition to assessing the precision and recall compared to the disease modelled by the vignette, we also evaluated the overall differential diagnosis qualitatively. This was based on the intuition that, to be useful, a differential diagnosis must not only include the relevant diseases but also exclude diseases that are of little relevance to the patient’s symptoms. To this end, we asked a senior medical practitioner[^4], who was not part of the role play experiment, to serve as a judge and to rate the quality of the differentials produced both by the Babylon Triage and Diagnostic System and by doctors. The judge first reviewed the vignette and then rated all the differentials for this vignette on a four point scale *(poor, okay, good, excellent)*. A differential was rated “excellent” if the judge could not find any issues with it, “good” if it had minor issues (such as the omission of a slightly irrelevant conditions, or if the order of the differential was deemed imperfect), “okay” if the list of conditions was generally acceptable, and “poor” if it was unacceptable (such as the omission of the most important conditions, or the inclusion of diseases completely unrelated to the presenting symptoms). The differentials were shown in random order and the judge was blinded to whether the differential had been produced by a human or the Babylon Triage and Diagnostic System. We also repeated this process with two in-house GPs.
We found that there was considerable disagreement between the medical practitioners’ subjective assessment of the differentials (see Figure \[fig:fig\_all\_judges\]; Supplementary Tables \[tab:differential\_diag\_perf\], \[tab:differential\_diag\_perf\_gp1\] and \[tab:differential\_diag\_perf\_gp2\]). For the judge, the lists of diseases output by the Babylon Triage and Diagnostic System were found to be of comparable quality to those produced by doctors (83.0% rated “okay” or better, compared to 78.2–97.9%). The same was the case for one of the GPs (GP-2), who was generally harsher on the evaluation (53.0% rated “okay” or better, compared to 51.3–82.4%). However, another GP (GP-1) rated the quality of differentials of the Babylon Triage and Diagnostic System lower than those of doctors (52.0% rated “okay” or better, compared to 76.9–93.8%).
We considered that the disparity in the qualitative evaluation of differential diagnoses might be due to a difference in interpretation and that some medical practitioners might be less tolerant of disease lists that are long or contain less relevant diseases, even if the relevant conditions are included. Although we don’t have sufficient statistical power to test this hypothesis, we repeated the experiment with the Babylon Triage and Diagnostic System tuned to provide higher precision at the expense of lower recall (Supplementary Figure \[fig:main\_PR\_curve\]). This mode resulted in a marked improvement in the ratings of the GPs, which anecdotally suggests a preference for more concise differentials for these individuals (Figure \[fig:fig\_all\_judges\_new\_thresholds\]).
Assessment of triage safety and appropriateness
-----------------------------------------------
In addition to rating the quality of doctors differential diagnoses, the expert judge was also asked to specify a range of safe and appropriate triage outcomes for each vignette. Providing a range of acceptable triage recommendations was motivated by the fact that doctors often disagree on the most appropriate triage recommendation (Supplementary Figure \[fig:doctor\_vs\_doctor\_confusion\]), however it is not necessarily the case that any of these different opinions are inappropriate or unsafe [@cathain2003]. By providing the minimum and maximum appropriate triage, our judge indicates the range of recommendations that are neither unsafe nor overly cautious.
We compared the triage recommendations of doctors and the Babylon Triage and Diagnostic System against the judge’s “gold standard” range. We define a “safe” triage as any recommendation which was of equal or greater urgency than the judge’s minimum triage, and an “appropriate” triage as any recommendation that fell within the judge’s range of acceptable recommendations. In this study, we found that the Babylon Triage and Diagnostic System provided a safer triage recommendation than doctors on average (97.0% versus 93.1%), at the expense of a marginally lower appropriateness (90.0% versus 90.5%; see Table \[tab:triage\_safety\]).
We repeated this process with three in-house GPs and found the triage safety and appropriateness of the Babylon Triage and Diagnostic System relative to the doctors to be consistent with those of the judge, although the scores from the GPs were found to be lower for both the Babylon Triage and Diagnostic System and the doctors (Table \[tab:triage\_recommendations\]).
Performance against publicly available case vignettes
-----------------------------------------------------
In order to provide a benchmark against previous work, as well as to the diagnostic accuracy that is expected for human practitioners, we assessed the performance of the Babylon Triage and Diagnostic System against three sets of publicly available case vignettes. These were case vignettes used in a previous study [@semigran2015evaluation], and from preparation materials for the MRCGP Clinical Skills Assessment (CSA) and Applied Knowledge Test (AKT), focusing on the diagnostic component of the curriculum (in contrast to content assessing management).
### @semigran2015evaluation Vignettes
The methodology described previously was repeated for 30 vignettes from a previous study by [@semigran2015evaluation]. We excluded vignettes from the original study that were based on conditions that are outside of the scope of the Babylon Triage and Diagnostic System, consistent with the original methodology. Specifically, these included paediatric and dermatological conditions, and tetanus which is not currently in the model yet based on its very low incidence rate in the United Kingdom. These vignettes were tested against both the Babylon Triage and Diagnostic System and three doctors. As per the original study, we report the recall of the condition modelled by the vignette for the top 1 and top 3 conditions listed in the differential. The Babylon Triage and Diagnostic System identified the modelled condition as its top 1 in 21 out of 30 vignettes (70.0%) and in its top 3 in 29 out of 30 vignettes (96.7%). On average, doctors identified the modelled condition in their top 1 in 75.3% of vignettes and in their top 3 in 90.3% of vignettes. This demonstrates a significant improvement relative to other symptom checkers evaluated in the original study.
### MRCGP Applied Knowledge Test (AKT)
The Babylon Triage and Diagnostic System was tested against 15 AKT vignettes (based on @RCGPLearning) by providing all available information about the patient to the model, consistent with how an AKT exam participant would see the exam question (since there is no history-taking component). The correct condition appeared in the top 3 conditions suggested by the Babylon Triage and Diagnostic System in 13 of 15 vignettes (86.7%).
### MRCGP Clinical Skills Assessment (CSA)
The Babylon Triage and Diagnostic System was tested against 36 CSA vignettes (translated from @Wadhera2011; and @Stannett2016) by first providing the main presenting complaint, then simulating a user’s interaction with the Babylon Triage and Diagnostic System by only providing information to the model if the Triage and Diagnostic System would directly ask about it. This is consistent with the CSA exam which requires the participant to take a medical history. The modelled disease appeared in the top 3 conditions suggested by the Triage and Diagnostic System in 27 of 36 vignettes (75.0%).
Discussion {#sec:discussion}
==========
We performed a prospective validation study of the accuracy and safety of an AI powered Triage and Diagnostic System, using an experimental paradigm designed to simulate realistic consultations. Overall we found that the Babylon Triage and Diagnostic System was able to identify the condition modelled by a clinical vignette with accuracy comparable to human doctors (in terms of precision and recall). We also found that the triage advice recommended by the Babylon Triage and Diagnostic System was safer on average than human doctors, when compared to the ranges provided by independent expert judges, with only minimal reduction in appropriateness. In other words, the AI system was able to safely triage patients without reverting to overly pessimistic fallback decisions.
We adopted a test protocol using simulated clinical vignettes which allowed us to evaluate a combination of common and rare conditions, the latter of which would be difficult to evaluate without a clinical trial with a sample size large enough to contain diseases with low incidence rates. While this might be considered a strength of our study, since it is not biased towards common presentations, our results cannot be directly interpreted with respect to real-world accuracy and safety. To illustrate the differences that might be expected in a real-world study, we reweighted our results by the annual incidence of the modelled disease for each vignette. We found that the accuracy and rating of differentials produced by the Babylon Triage and Diagnostic System improved compared to those of doctors after accounting for disease incidence (Supplementary Table \[tab:modelled\_disease\_perf\] and Supplementary Figure \[fig:all\_judges\_modelled\_diseases\]). Surprisingly, we found that the accuracy and rating of some doctors decreased considerably after reweighting. This is likely due to the fact that the most common conditions carry substantially more weight than the rarer ones; thus the results will be highly sensitive to a few vignettes (in particular, Doctor A did not include a modelled disease in their differential for a vignette, where that modelled disease was very common and hence had high weight). Further work will be required to more rigorously investigate the diagnostic accuracy in a real-world clinical setting.
One source of bias in this study derives from the limitation imposed on doctors to only select diseases that are modelled in the Babylon Triage and Diagnostic System. As the “correct” disease for each vignette was always from this list, this may have provided human doctors with some advantage in terms of precision and recall compared to free text entry. However, it would have also constrained them from providing a fuller and more nuanced differential diagnosis overall, which may have disadvantaged them in terms of judge rating of overall differential quality. The intention in assigning this limitation as part of the testing protocol was to ensure blinding when the judges assessed the quality of the differential diagnosis.
The Babylon Triage and Diagnostic System listed the correct disease among the top 3 differentials in 86.7% of AKT cases and 75.0% of CSA cases, for the limited question set that was tested. However, it is important to note that the components of the MRCGP examination used to assess the Babylon Triage and Diagnostic System were limited to those based on history taking and diagnostics. The full MRCGP examination (a combination of the AKT and CSA) also assesses a wide range of other skills not examined in this study. Therefore, the performance of the Babylon Triage and Diagnostic System in this study cannot be taken as a demonstration of an ability to pass the examination in full, but on the diagnostic component in isolation, it achieved accuracy rates above 72%, the average pass mark for the past 5 years [@RCGP2018one; @RCGP2018two] for the entire MRCGP exam.
Another possible limitation of our study is that we evaluated only clinical cases that were based on a single underlying condition (although we did include past medical history and pre-existing conditions). In reality, patients may have multiple undiagnosed diseases. However, one of the strengths of our approach, which uses a Bayesian model, is that it is able to reason about multiple causes of a patient’s presenting symptoms. It would be useful to test whether the performance relative to doctors is different in cases where multiple diseases must be diagnosed.
Finally, this study emphasises the difficulty in objectively evaluating the accuracy of a differential diagnosis. Even when the true underlying condition is identified, the quality of the overall differential may be poor due to the omission of important alternative explanations for a patient’s symptoms, or the inclusion of irrelevant diseases. By evaluating differential diagnoses qualitatively using independent judges, we found that considerable disagreement exists in the subjective rating by different individuals, including differential diagnoses of human doctors. This may be due to the fact that a judge’s rating is itself based on personal assessment of the clinical case, which may be prone to error, or due to differences in personal preference for longer or shorter differential diagnoses. Ultimately, there is likely no adequate “gold standard” differential diagnosis, and future work would benefit from assessing the inter-rater agreement between a larger sample of doctors.
Conclusion {#sec:conclusion}
==========
Artificial intelligence powered symptom checkers have the potential to provide diagnostic and triage advice with a level of accuracy and safety approaching that of human doctors. Such systems may hold the promise of reduced costs and improved access to healthcare worldwide, but realising this requires greater levels of confidence from the medical community and the wider public. Key to this confidence is a better understanding of the strengths and weaknesses of human doctors, who do not always agree on the cause of a patient’s symptoms or the most appropriate triage outcome, and an improved awareness of the accuracy and safety of AI powered systems. Further studies using larger, real-world cohorts will be required to demonstrate the relative performance of these systems to human doctors.
Acknowledgements {#sec:acknowledgements}
================
Independent medical practitioners helped us create the testing vignettes for the main experiment of our paper. They were: Prof. Megan Mahoney from Stanford University, Dr Benjamin A. White from Harvard Medical School/Massachusetts General Hospital, Department of Emergency Medicine. The authors would like to thank Dr Benjamin A. White, Dr Hannah Allen, Dr Marieke Reddingius for their help rating the differentials and triage.
![Percentage of differential diagnoses rated as “okay” or better by the judge and the two GPs for doctors and the Babylon Triage and Diagnostic System (Babylon AI). There is considerable disagreement between the three ratings, suggesting the qualitative assessment of differential diagnoses might be influenced by personal preference.[]{data-label="fig:fig_all_judges"}](figures/all_judges.pdf){width="\columnwidth"}
![Percentage of differential diagnoses rated as “okay” or better by the judge and the two GPs for doctors and the Babylon Triage and Diagnostic System (Babylon AI), where the latter is tuned to provide higher precision (at the expense of recall). The differentials provided by the Babylon Triage and Diagnostic System were rated to be of comparable quality to those of doctors.[]{data-label="fig:fig_all_judges_new_thresholds"}](figures/all_judges_new_thresholds.pdf){width="\columnwidth"}
[^1]: As of the current date (June 2018), the model of Babylon Triage and Diagnostic System being evaluated in the current study is not released yet.
[^2]: It should be noted that, this paper is for general information and academic purposes, and to analyse the use of online symptom checkers in healthcare. The Babylon Triage and Diagnostic System is referred to in this paper to facilitate discussion on this topic, but this paper is not designed to be relied upon for any other purpose.
[^3]: The list of conditions modelled by the Triage and Diagnostic System includes the majority of those encountered in General Practice in the United Kingdom, but does not include skin conditions, pregnancy-related conditions or paediatric conditions.
[^4]: Dr Benjamin A. White.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The lowest metallicity massive stars in the Local Universe with $Z \sim$($Z$[$_{\odot}$]{}/50-$Z$[$_{\odot}$]{}/30) are the crucial objects to test the validity of assumptions in the modern models of very low-metallicity massive star evolution. These models, in turn, have major implications for our understanding of galaxy and massive star formation in the early epochs. DDO68-V1 in a void galaxy DDO68 is a unique extremely metal-poor massive star. Discovered by us in 2008 in the HII region Knot3 with $Z = Z$[$_{\odot}$]{}/35 \[12+$\log$(O/H)$\sim$7.14\], DDO68-V1 was identified as an LBV star. We present here the LBV lightcurve in V band, combining own new data and the last archive and/or literature data on the light of Knot3 over the 30 years. We find that during the years 2008-2011 the LBV have experienced a very rare event of ‘giant eruption’ with V-band amplitude of 4.5 mag ($V \sim24.5^m -20^m$).'
---
Introduction
============
Luminous Blue Variable (LBV) stars represent a short (about or less than 0.1 Myr) transient phase of massive star evolution from the main sequence hydrogen burning O stars to the core-helium burning Wolf-Raye (WR) stars.
Evolution of massive stars with the lowest known metallicities is crucial for understanding the early galaxy formation and evolution at high redshifts due to their great energy release/feedback (e.g., [@Barkana01 Barkana & Loeb (2001)]).
The most metal-poor [**massive**]{} stars are currently identified in several extremely metal-poor ($Z \sim Z$[$_{\odot}$]{}/45– $Z$[$_{\odot}$]{}/35) dwarf galaxies. Most of these extreme galaxies are found in nearby voids.
Stellar evolution models (including those with the fast rotation) have substantially advanced during the last decade. However, the direct comparison of the model predictions with the properties of real extremely metal-poor massive stars is still absent. Such studies should await for the next generation extremely large telescopes.
Overview
========
DDO68, at the distance D=12.75 Mpc, is one of the most metal-poor galaxies ($Z \sim Z$[$_{\odot}$]{}/35) residing in the nearby Lynx-Cancer void. DDO68 is a merger of low-mass gas-rich components ([@Ekta08 Ekta, Chengalur, Pustilnik (2008)], [@Makarov17 Makarov et al. (2017)]). Its very low-Z gas was identified with BTA spectra in 2005. Most of SF regions are found at the periphery, mainly in the ‘Northern ring’ and the ‘Southern tail’ ([@DDO68 Pustilnik, Kniazev, & Pramskij (2005)], [@IT07 Izotov & Thuan (2007)]).
In 2008 we discovered in its SF Knot 3 (Fig. \[fig1\]) a transient which was identified with an LBV ([@LBV Pustilnik et al. (2008)], see also [@Izotov09 Izotov & Thuan (2009)]). Hubble Space Telescope (HST) images of DDO68 were obtained in May 2010 with $ACS$ for Proposal GO 11578 (PI A.Aloisi) and presented in papers by [@Sacchi2016 Sacchi et al. (2016)] and by [@Makarov17 Makarov et al. (2017)].
![ The part of the HST image of DDO68 in $W606$ ($V$) band centered on the region Knot 3 with the used aperture superimposed ($D_{\rm aper} =$ 5$''$). DDO68-V1 is in the center of the aperture. []{data-label="fig1"}](./fa_neg.eps){width="5.3in"}
The lightcurve of Knot 3 (Fig. \[fig2\]) in DDO68 in $V$ and $B$ bands since 1988 is based on the new and archive data and the data from [@DDO68LBV Pustilnik et al. (2017)]. All magnitudes are for the aperture with r=2.5$''$. The dotted lines at $V = 20.20$ and $B = 20.25$ correspond to the minimal observed light of the entire Knot 3. These minimal levels were slightly reduced due to a more advanced background determination with respect of that adopted in paper by [@PerepelPustilnik2017 Perepelitsyna & Pustilnik (2017)]. These magnitudes are consistent, in particular, with Knot 3 light on the night 2005.01.12, when the LBV was too faint and did not show up in the spectrum of Knot 3.
With except of one direct photometry (the HST image), all other magnitudes are derived as the ‘residual light’ via subtraction of the constant luminosity of the underlying HII region ($V=20.20$) from the lightcurve on the Fig. \[fig3\]. We observe a very rare case of LBV ‘giant eruption’ ([@Smith2006 Smith & Owocki (2006)]) during the years 2008-2011, with the total amplitude of the LBV optical variability $\delta$V$\sim$4.5$^m$, reaching $M_{\rm V}$ = -10.5. Series of ‘giant eruptions’ in LBVs which form several expanding shells, can precede their SN explosions at rather short time scale. Observations of light variations of DDO68-V1 after the ‘giant eruption’, since Year 2015 reveal the behaviour resembling the phenomenon of S Doradus ([@SDor Sterken, 2003]). In the right panel of Figure 3, the photometric variability is observed up to 2.5$^m$ over the periods of 0.5–2 years.
Implications and conclusions
============================
1. We extend the recently published lightcurve for the period of 2005 – 2015 for DDO68-V1 ([@DDO68LBV Pustilnik et al. 2017]), adding our fresh (years 2016–2018) Zeiss-1000 and BTA telescopes photometry of the HII region Knot 3 (containing the LBV = V1) and the photometry from the archive images at ten epochs with ten different telescopes over the period of 1988 – 2013.
2. The data allow us at the first time to determine the reliable amplitude of this LBV lightcurve. All available data suggest that the LBV $V$-band light varied during the last decade in the range of $\sim$20.0$^m$ to fainter than 24.5$^m$. This corresponds to the absolute magnitude $M_{\rm V}$ range of -6.0$^m$ to -10.5$^m$.
3. If the photometric behavior of the most metal-poor LBV is similar to that of more typical LBVs, the DDO68-V1 light variations during the last 28 years suggest that it underwent a ‘giant eruption’ during the years 2008 – 2011.
4. We call to the community for the campaign of DDO68-V1 multiwavelength monitoring that can give the new insights in the lowest metallicity LBV properties and prove the substantial increase of its bolometric luminosity.
5. Having in mind other known examples of extragalactic SN impostors, one can occasionally catch this unique object in the SN impostor phase. Moreover, in the case of the great luck, we can catch even the unique case of a nearby SNII explosion related to the extremely low-Z massive star.
The full-format paper presenting all details of observational data and their analysis as well as a wider discussion of all available data is prepared for publication in MNRAS.
Acknowledgements {#acknowledgements .unnumbered}
================
The work was supported by the grant of Russian Science Fund No. 14-12-00965. The authors thank O. Spiridonova, V. Goransky and A. Moskvitin for their help with DDO68 observations at the SAO 1m telescope. The authors are grateful to L. van Zee, D. Hunter, B. Elmegreen, U. Hopp, L. Makarova, R. Swaters, B. Mendez, V. Taylor, R. Jansen, R.A. Windhorst, S.C. Odewan, J.E. Hibbard for providing archival CCD images of DDO68 obtained for their observational programs. We are pleased to thank P. Kaigorodov and D. Kolomeitsev for their kind help in extracting the data from archive tapes.
2001, , 349, 125
2008, *MNRAS*, 391, 881
2007, *ApJ*, 665, 1115 2009, *ApJ*, 690, 1797
2017, *MNRAS*, 466, 556
2017, *ASP Conference Series*, V.510, Yu.Yu. Balega, D.O. Kudryavtsev, I.I. Romanyuk, and I.A. Yakunin, eds. p.484 2005, *A&A*, 443, 91
2008, *MNRAS*, 388, L24
2017, *MNRAS*, 465, 4985
2016, *ApJ*, 830, 3
2006, *ApJ*, 645, L45
2003, *ASP Conference Series*, V.292, C.Sterken, ed. p.437
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that the low lying spin states of two electrons in a semiconductor quantum dot can be strongly mixed by electron-electron asymmetric exchange. This mixing is generated by the coupling of electron spin to its orbital motion and to the relative orbital motion of the two electrons. The asymmetric exchange can be as large as $50\%$ of the isotropic exchange, even for cylindrical quantum dots. The resulting spin mixing contributes to understanding spin dynamics in quantum dots, including light polarization reversal.'
address: '*[Naval Research Laboratory, Washington DC 20375]{}*'
author:
- 'Ş. C. Bădescu'
- 'T. L. Reinecke'
title: 'Mixing of two-electron spin states in a semiconductor quantum dot'
---
An electron spin in a semiconductor quantum dot (QD) is an attractive qubit for quantum computing [@Loss98]: the spin in the ground orbital state can have long coherence time [@Kroutvar04]; a single qubit can be initialized or read optically by transient electron-hole pair excitation giving a negative trion $X^-$ [@Cortez02; @Ware05]; and the manipulation of the spin exchange between neighboring spins can be the basis for two-qubit gates [@Loss98]. A detailed picture of correlations between spins in QDs is essential for understanding the spin dynamics. The dominant interaction between two electrons ([*e-e*]{}) is the Heisenberg-like spin-symmetric $J\hat{\bm
s}_1$$\cdot$$\hat{\bm s}_2$ (symmetric exchange), which conserves the total spin $\hat{\bm S}$$=$$\hat{\bm s}_1$$+$$\hat{\bm s}_2$. Additional spin-asymmetric [*e-e*]{} interactions (asymmetric exchange) do not conserve $\hat{\bm S}$ and thus decrease the fidelity of gate operations.
Among a number of recent experiments giving information about spin dynamics are those involving an optical polarization reversal [@Cortez02; @Ware05]. For them, it has been suggested that this effect results from spin flipping due to electron-hole ([*e-h*]{}) exchange in QDs with lateral asymmetry. However, these experiments require strong spin mixing, inconsistent with [*e-h*]{} exchange alone [@Ware05].
Spin-orbit (s-o) interactions play a key role in understanding mixing of spin states. They arise from effective magnetic fields created by the orbital motion of electrons [@B-L-P]. Electrons in QD ground states with dominant $s$ components have small orbital angular momentum and thus small s-o coupling. A number of experiments of interest involve excited electrons in excited states of the QD. Linear combinations of nearly degenerate excited states in a plane ([*e.g.*]{} $p_x$- and $p_y$-like) can give rise to 2D orbital motion with an effective magnetic field perpendicular to the plane, and thus to large s-o coupling. This is analogous to the $\hat{\bm L}$$\cdot$$\hat{\bm S}$ coupling in atoms [@L-L]. Thus, relatively symmetric QDs ([*e.g.*]{} cylindrical) can have significant s-o effects, as we show here.
There are three sources of s-o coupling that lead to the mixing of spin states. The largest two contributions arise from the ${\bm
k}$$\cdot$$\hat{\bm p}$ mixing between the conduction and valence bands near the zone center, as described in the effective mass approach [@BirPikus]. We derive them by treating the potentials from the structure and from the [*e-e*]{} Coulomb repulsion on the same footing with ${\bm k}$$\cdot$$\hat{\bm p}$ terms, using the Kane model [@Kane]. We have in mind QDs with a strong confinement to a single state $\xi(z)$ along the growth axis ${\bm e}_z$, and a weaker confinement in the transverse directions, which give the electron states $\phi_i({\bm
r})$$=$$\xi(z)$$\varphi_i({\bm \rho})$.
The first contribution to the s-o coupling, $\hat{\bm h}^V$, arises from the [*structure potential*]{} $V({\bm r})$ of the QD. It gives a single-electron s-o coupling of the form [@ConstantV]: $$\hat{\bm h}^V\cdot\hat{\bm
s}=\gamma_s^V\left[\partial_zV\left(\hat{\bm p}^\bot\times\hat{\bm
s}^\bot\right)+\left(\partial_{\bm \rho}V\times\hat{\bm
p}^\bot\right)\hat{s}^z\right]{\bm e}_z\,,\label{h_R}$$ where $p^z$ is not present due to the strong vertical confinement (for a single state $\xi(z)$, $\langle \xi|p^z|\xi
\rangle$$=$$0$). The first term in Eq.(\[h\_R\]) is the usual Rashba coupling $\gamma^V({\bm e}_z$$\times$$\hat{\bm p}^\bot)$, where $\gamma^V$$=$$\gamma_s^V\langle\xi|\partial_z V|\xi\rangle$, associated to asymmetry in the growth direction [@Rashba84]. The second term is important for excited states whose main components are inversion-asymmetric ($p$-like), where it gives the dominant s-o coupling, independent of structure or bulk inversion asymmetries. This term vanishes in the QD ground state, whose main component is inversion-symmetric ($s$-like).
The second contribution, $\hat{\bm h}^C$, arises from the interaction of each spin with the orbital motion of the other. We have obtained it within a two-particle [*${\bm
k}$$\cdot$$\hat{\bm p}$*]{} approach for electrons interacting through the [*Coulomb potential*]{} $U_C({\bm r}_r)$ [@Badescu05; @ConstantC]. For $k=$$1,2$ and ${\bm r}_r$$=$${\bm
r}_1$$-$${\bm r}_2$, we have: $$\hat{\bm h}_k^C\cdot\hat{\bm s}_k=(-1)^k\gamma_s({\bm
\nabla}_{{\bm r}_r}U_C\times \hat{\bm p}_k)\cdot\hat{\bm
s}_k\,.\label{h_rels-o}$$
The coupling $\hat{\bm h}^V$ from Eq.(\[h\_R\]) is analogous to the Pauli s-o interaction, while $\hat{\bm h}^C$ from Eq.(\[h\_rels-o\]) is analogous to the Breit-Pauli spin-relative orbit coupling [@B-L-P]. The Pauli and Breit-Pauli couplings in vacuum or in atoms are relativistically small, due to the large energy gap $2m_0c^2$ between electron and positron bands, whereas the present gap $E_g$ is smaller, giving larger s-o couplings.
There is also a smaller contribution, $\hat{\bm h}^\mathrm{B}$, from the [*Dresselhaus coupling*]{} due to the lack of bulk inversion symmetry [@Dresselhaus]. It arises from the mixing of the conduction band with the remote upper bands and it gives a single-particle s-o coupling in the form $\hat{\bm
h}^\mathrm{B}$$\cdot$$\hat{\bm
s}$=$\gamma^B_b\epsilon^{\alpha\beta\delta}\hat{p}_\alpha(\hat{p}_\beta^2$$-$$\hat{p}_\delta^2)\hat{s}_\alpha$, with indices denoting crystal symmetry axis. In the QDs with strong vertical confinement consider here $\langle
p^{z2}\rangle$$\gg$$\langle p^{\bot2}\rangle$, thus the effective Dresselhaus coupling contains only the transverse components ${\bm
h}^{B,\bot}$$\simeq$$\gamma^B(\hat{p}_x,-\hat{p}_y)$, with $\gamma^B$$=$$\gamma^B_{b}\langle\xi|p_z^2|\xi\rangle$.
We use a model of QDs [@ModelPotential] resembling those from self-assembled growth [@SK] along crystal axis $[001]$. The lateral potential $\mathcal{V}({\bm \rho})$ contains a part $\mathcal{V}_s$ symmetric for the inversion ${\bm
\rho}$$\rightarrow$$-{\bm \rho}$, and it may also contain an inversion-asymmetric part $\mathcal{V}_a$ [@AsymDots]. We take the principal axes ${\bm e}_{x,y}$ of the QD to be along the crystal axes $[110]$ and $[1\overline{1}0]$. To construct accurate states we use a large basis set of harmonic oscillator wavefunctions. The lateral sizes $a_x$,$a_y$ are given by the curvature at the potential minimum [@Single-particle-Phi_i], which is determined entirely by the symmetric part $\mathcal{V}_s$. $\mathcal{V}_a$ contains $\mathcal{V}_{ax}$ ($\mathcal{V}_{ay}$), odd in $x$ ($y$), and is parametrized by $E_x$ ($E_y$) [@ModelPotential]. The effective lateral electric field in the ground state $\langle\varphi_1|$$-\partial_{x,y}\mathcal{V}_a$$|\varphi_1\rangle\propto
E_{x,y}$ and vanishes for lateral inversion symmetry.
First we consider the two-electron wavefunctions without s-o coupling. They are obtained by diagonalizing hamiltonian $H_0$ that contains the Coulomb interaction $U_C$$=$$\frac{e^2}{\kappa
r_r}$ ($\kappa$ is the dielectric constant) with a band-mixing correction $\gamma_c\delta({\bm r}_r)$ [@ConstantC] and the QD potential $V({\bm r})$, in the basis of harmonic oscillator wavefunctions. These basis functions separate into the symmetric and antisymmetric sets $\{S^{(0)}_n\}$, $\{T^{(0)}_m\}$ by their permutation symmetry [@Two-particle-basis]. In general, each eigenstate of $H_0$ can be written in terms of functions having definite $s$, $x$, $y$, $d$ symmetry, [*e.g.*]{}: $$\begin{aligned}
T_1&=&T_1^x+E_xT_1^s+E_xE_yT_1^y+E_yT_1^d\,,\nonumber\\
T_2&=&T_2^y+E_yT_2^s+E_xE_yT_2^x+E_xT_2^d\,,\label{WF-comp}\\
S_2&=&S_2^x+E_xS_2^s+E_xE_yS_2^y+E_yS_2^d\,,\nonumber\\
S_3&=&S_3^y+E_yS_3^s+E_xE_yS_3^x+E_xS_3^d\,,\nonumber\end{aligned}$$ where $T_1$ ($T_2$) labels the lowest triplet with a larger $x$ ($y$) component, and $S_2$ ($S_3$) labels the lowest singlet with a larger $x$ ($y$) component. The lowest lying states are shown in Fig.\[Asymmetry-Cases\](a), where $E_y$$=$$0$. Higher lying states not shown are $T_3$ ($T_4$), which are the lowest $d$ ($s$) symmetry triplets, and $S_4$ (the lowest $d$-symmetry singlet). The isotropic part of the exchange for a triplet $T_i$ (with energy $\epsilon_i^t$) and a singlet $S_j$ (with energy $\epsilon_j^s$) is given by $J_{ij}$$=$$2($$\epsilon_i^t$$-$$\epsilon_j^s)/\hbar^2$. Here we have chosen the energy splitting between the electron ground and excited states to be in the range $20$$-$$45$ meV; this gives an exchange splitting ($J_{13}$ between $T_1$ and $S_3$) of the order $5$$-$$10$ meV, in the range of the experiments.
Next, the triplet-singlet mixing is generated by adding the s-o terms ${\bm h}^V$, ${\bm h}^C$, ${\bm h}^B$ to $H_0$. These give a hamiltonian composed of a spin-symmetric part $H_s$ that conserves the total spin, and a spin-antisymmetric part $H_a$: $$\begin{aligned}
H_s&=&H_{0}+\frac{1}{2}\left(\hat{\bm h}_1+\hat{\bm
h}_2+\gamma_s\partial_{{\bm \rho}_r}U_C\times \hat{\bm
p}_r^\bot\right)\cdot\hat{\bm S}\,,\label{Two_Particles-Diag}\\
H_a&=&\frac{1}{2}\left(\hat{\bm h}_1-\hat{{\bm
h}}_2+2\gamma_s\partial_{{\bm \rho}_r}U_C\times \hat{\bm
p}_c^\bot\right)\cdot\left(\hat{\bm s}_1-\hat{\bm
s}_2\right)\,,\nonumber\end{aligned}$$ where $\hat{\bm h}_k$$=$$\hat{\bm h}_k^V$$+$$\hat{\bm h}_k^B$, $\hat{\bm p}_r$$=$$\hat{\bm p}_1$$-$$\hat{\bm p}_2$, and $\hat{\bm
p}_{c}$$=$$(\hat{\bm p}_1$$+$$\hat{\bm p}_2)/2$. $H_a$ can be written as: $$\begin{aligned}
H_a&=&\sum_{i,j}{\bm \beta}_{ij}\cdot(\hat{\bm s}_1-\hat{\bm
s}_2)|T_i\rangle\langle S_j|+h.c.\label{H_st}\,,\\
{\bm \beta}_{ij}&=&\langle T_i|\hat{\bm
h}_1+\gamma_s\partial_{{\bm \rho}_r}U_C\times\hat{\bm
p}_c^\bot|S_j\rangle\nonumber\,,\end{aligned}$$ where ${\bm \beta}_{ij}$ gives the asymmetric exchange. States of different total spin $|\bm S|$ are coupled via the operator $\hat{\bm s}_1$$-$$\hat{\bm s}_2$, which is equivalent to the Dzyaloshinskii-Morya form $\frac{2i}{\hbar}$$(\hat{\bm
s}_1$$\times$$\hat{\bm s}_2)$ [@D-M]. The asymmetric exchange can be written: $${\bm \beta}\cdot\left(\hat{\bm s}_1-\hat{\bm
s}_2\right)=\beta^z\left(\hat{s}^z_1-\hat{s}^z_2\right)+{\bm
\beta}^\bot\cdot\left(\hat{\bm s}^\bot_1-\hat{\bm
s}^\bot_2\right)\,.\label{IntroEq}$$ The longitudinal component $\beta^z$ conserves the total spin projection $S^z$, [*i.e.*]{}, it mixes singlets with triplets that have $S^z$$=$$0$ (“longitudinal mixing”). This is equivalent to a precession of the total spin around axis ${\bm e}_z$ ($\Delta
S^z$$=$$0$). The transverse components ${\bm \beta}^\bot$ mix states with different total-spin projection ($|\Delta
S^z|$$\neq$$0$), equivalent with a total-spin precession around in-plane axis (“transverse mixing”).
(8.0,9.0)
It is convenient to group the operators from the matrix element giving ${\bm \beta}_{ij}$ in Eq.(\[H\_st\]) into an axial vector operator $\hat{\bm A}$$\equiv$$\hat{A}{\bm e}_z$ and two polar vector operators $\hat{\bm P}$$\equiv$$\hat{P}{\bm e}_z$, $\hat{\bm R}$$\equiv$$\hat{\bm R}^\bot$: $$\begin{aligned}
\hat{\bm A}&=&2\gamma_s^V\left(\hat{\partial}_{{\bm\rho}_1}\mathcal{V}_{s}\times\hat{\partial}_{{\bm\rho}_1}\right)-2\gamma_s\left(\partial_{{\bm\rho}_r}U_C\times\hat{\partial}_{{\bm\rho}_c}\right)\,,\nonumber\\
\hat{\bm P}&=&2\gamma_s^V\left(\hat{\partial}_{{\bm\rho}_1}\mathcal{V}_{a}\times\hat{\partial}_{{\bm\rho}_1}\right)\label{A-P_components}\\
\hat{\bm R}&=&-2\gamma^V\left({\bm
e}_z\times\hat{\partial}_{{\bm\rho}_1}\right)+2\gamma^B\left({\bm
e}_x\hat{\partial}_{x_1}-{\bm e}_y\hat{\partial}_{y_1}
\right)\nonumber\,.\end{aligned}$$ $\hat{\bm A}$ and $\hat{\bm P}$ include the vertical magnetic field from the $2D$ motion in the nearly degenerate excited states, and they generate $\beta^z$ in Eq.(\[IntroEq\]). $\hat{\bm R}$ arises from the Rashba and Dresselhaus terms, and it generates ${\bm \beta}^\bot$.
[c|c|c|c|c]{} &$S_1$ ($\approx$ $s$-symmetry)&$S_2$ ($\approx$ $x$-symmetry) &$S_3$ ($\approx$ $y$-symmetry) & $S_4$ ($\approx$ $d$-symmetry)\
$T_4$&&&& $+$\
($\approx$ $s$) & $E_xE_y[A_{41}^{xy}$$+$$A_{41}^{yx}$$+$$A_{41}^{sd}$$+$$A_{41}^{ds}]$ & $E_y[A_{42}^{sd}$$+$$A_{42}^{yx}$$+$$E_x^2(A_{42}^{xy}$$+$$A_{42}^{ds})]$ & $E_x[A_{43}^{sd}$$+$$A_{43}^{xy}$$+$$E_y^2(A_{43}^{xy}$$+$$A_{43}^{sd})]$ & $E_x^2A_{44}^{xy}$$+$$E_y^2(A_{44}^{yx}$$+$$E_x^2A_{44}^{ds})$\
$T_1$ && & $~+$ &\
($\approx$ $x$) & $E_y[A_{11}^{xy}$$+$$A_{11}^{ds}$$+$$E_x^2(A_{11}^{sd}$$+$$A_{11}^{yx})]$ & $E_xE_y[A_{12}^{xy}$$+$$A_{12}^{yx}$$+$$A_{12}^{sd}$$+$$A_{12}^{ds}]$ &$E_x^2A_{13}^{sd}$$+$$E_y^2(A_{13}^{ds}$$+$$E_x^2A_{13}^{yx})$ & $E_x[A_{14}^{xy}$$+$$A_{14}^{sd}$$+$$E_y^2(A_{14}^{ds}$$+$$A_{14}^{yx})]$\
$T_2$ && $+$ & &\
($\approx$ $y$) & $E_x[A_{21}^{yx}$$+$$A_{21}^{ds}$$+$$E_y^2(A_{21}^{sd}$$+$$A_{21}^{xy})]$ & $E_x^2A_{22}^{ds}$$+$$E_y^2(A_{22}^{sd}$$+$$E_x^2A_{22}^{xy})$ &$E_xE_y[A_{23}^{xy}$$+$$A_{23}^{yx}$$+$$A_{23}^{sd}$$+$$A_{23}^{ds}]$ & $E_y[A_{24}^{yx}$$+$$A_{24}^{sd}$$+$$E_x^2(A_{24}^{ds}$$+$$A_{24}^{xy})]$\
$T_3$& $+$&&&\
($\approx$ $d$) & $E_x^2A_{31}^{yx}$$+$$E_y^2(A_{31}^{xy}$$+$$E_x^2A_{31}^{sd})$ & $E_x[A_{32}^{ds}$$+$$A_{32}^{yx}$$+$$E_y^2(A_{32}^{xy}$$+$$A_{32}^{sd})]$ & $E_y[A_{33}^{ds}$$+$$A_{33}^{xy}$$+$$E_x^2(A_{33}^{yx}$$+$$A_{33}^{sd})]$ & $E_xE_y[A_{34}^{xy}$$+$$A_{34}^{yx}$$+$$A_{34}^{sd}$$+$$A_{34}^{ds}]$\
\[Table-A\]
[c|c|c|c|c]{} &$S_1$ &$S_2$ &$S_3$ & $S_4$\
$T_4$ & $E_x[P^{sx}_{41}$$+$$P^{xs}_{41}$$+$$E_y^2(P^{yd}_{41}$$+$$P^{dy}_{41})]+$ & $P^{sx}_{42}$$+$$E_x^2P^{xs}_{42}$$+$$E_y^2(P^{yd}_{42}$$+$$E_x^2P^{dy}_{42})+$ & $P^{sy}_{43}$$+$$E_x^2P^{xd}_{43}$$+$$E_y^2(P^{zys}_{43}$$+$$E_x^2P^{dx}_{43})+$ & $E_x[P^{sy}_{44}$$+$$P^{xd}_{44}$$+$$E_y^2(P^{ys}_{44}$$+$$P^{dx}_{44})]+$\
& $E_y[P^{sy}_{41}$$+$$P^{ys}_{41}$$+$$E_x^2(P^{xd}_{41}$$+$$P^{ds}_{41})]$ & $E_xE_y(P^{sy}_{42}$$+$$P^{xd}_{42}$$+$$P^{ys}_{42}$$+$$P^{dx}_{42})$ & $E_xE_y(P^{sx}_{43}$$+$$P^{xs}_{43}$$+$$P^{yd}_{z43}$$+$$P^{dy}_{43})$ & $E_y[P^{sx}_{44}$$+$$P^{yd}_{44}$$+$$E_y^2(P^{xs}_{44}$$+$$P^{dy}_{44})]$\
$T_1$ & $P^{xs}_{11}$$+$$E_x^2P^{sx}_{11}$$+$$E_y^2(P^{dy}_{11}$$+$$E_x^2P^{yd}_{11})+$ & $E_x[P^{xs}_{12}$$+$$P^{sx}_{12}$$+$$E_y^2(P^{sy}_{12}$$+$$P^{ys}_{12})]+$ & $E_x[P^{xd}_{13}$$+$$P^{sy}_{13}$$+$$E_y^2(P^{sy}_{13}$$+$$P^{dx}_{13})]+$ & $P^{xd}_{14}$$+$$E_x^2P^{sy}_{14}$$+$$E_y^2(P^{dx}_{14}$$+$$E_x^2P^{ys}_{14})+$\
& $E_xE_y(P^{xd}_{11}$$+$$P^{sy}_{11}$$+$$P^{dx}_{11}$$+$$P^{ys}_{11})$ & $E_y[P^{xd}_{12}$$+$$P^{dx}_{12}$$+$$E_x^2(P^{sy}_{12}$$+$$P^{ys}_{12})]$ & $E_y[P^{xs}_{13}$$+$$P^{dy}_{13}$$+$$E_x^2(P^{sx}_{13}$$+$$P^{yd}_{13})]$ & $E_xE_y(P^{xs}_{14}$$+$$P^{sx}_{14}$$+$$P^{dy}_{14}$$+$$P^{yd}_{14})$\
$T_2$ & $P^{ys}_{21}$$+$$E_x^2P^{dx}_{21}$$+$$E_y^2(P^{sy}_{21}$$+$$E_x^2P^{ys}_{xd})+$ & $E_x[P^{ys}_{22}$$+$$P^{dx}_{22}$$+$$E_y^2(P^{sy}_{22}$$+$$P^{xd}_{22})]+$ & $E_x[P^{yd}_{23}$$+$$P^{dy}_{23}$$+$$E_y^2(P^{sx}_{23}$$+$$P^{xs}_{23})]+$ & $P^{yd}_{24}$$+$$E_x^2P^{dy}_{24}$$+$$E_y^2(P^{sx}_{24}$$+$$E_x^2P^{xs}_{24})+$\
& $E_xE_y(P^{yd}_{21}$$+$$P^{dy}_{21}$$+$$P^{sx}_{21}$$+$$P^{xs}_{21})$ & $E_y[P^{yd}_{22}$$+$$P^{sx}_{22}$$+$$E_x^2(P^{dy}_{22}$$+$$P^{xs}_{22})]$ & $E_y[P^{ys}_{23}$$+$$P^{sy}_{23}$$+$$E_x^2(P^{dx}_{23}$$+$$P^{xd}_{23})]$ & $E_xE_y(P^{ys}_{24}$$+$$P^{dx}_{24}$$+$$P^{sy}_{24}$$+$$P^{xd}_{24})$\
$T_3$ & $E_x[P^{dx}_{31}$$+$$P^{ys}_{31}$$+$$E_y^2(P^{xd}_{31}$$+$$P^{sy}_{31})]+$ & $P^{dx}_{32}$$+$$E_x^2P^{ys}_{32}$$+$$E_y^2(P^{xd}_{32}$$+$$E_x^2P^{sy}_{32})+$ & $P^{dy}_{33}$$+$$E_x^2P^{yd}_{33}$$+$$E_y^2(P^{xs}_{33}$$+$$E_x^2P^{sx}_{33})+$ & $E_x[P^{dy}_{34}$$+$$P^{yd}_{34}$$+$$E_y^2(P^{xs}_{34}$$+$$P^{sx}_{34})]+$\
& $E_y[P^{dy}_{31}$$+$$P^{xs}_{31}$$+$$E_x^2(P^{yd}_{31}$$+$$P^{sx}_{31})]$ & $E_xE_y(P^{dy}_{32}$$+$$P^{yd}_{32}$$+$$P^{xs}_{32}$$+$$P^{sx}_{32})$ & $E_xE_y(P^{dx}_{33}$$+$$P^{ys}_{33}$$+$$P^{xd}_{33}$$+$$P^{sy}_{33})$ & $E_y[P^{dx}_{34}$$+$$P^{xd}_{34}$$+$$E_x^2(P^{ys}_{34}$$+$$P^{sy}_{34})]$\
\[Table-Pd\]
Table \[Table-A\] gives the matrix elements between $T_i$ ($i$$=$$1,4$), and $S_j$ ($j$$=$$1,4$), from the spin mixing operator $\hat{\bm A}$ in Eq.(\[A-P\_components\]). The states are characterized by the symmetry of their dominant wavefunction components, [*e.g.*]{} $S_2$ $\approx x$-symmetry. Table \[Table-Pd\] gives corresponding results from $P$. The terms in small boxes in Table \[Table-A\] are dominant and are independent of lateral asymmetries. All the other terms in Tables I and II are non-zero only for cases of lateral asymmetry. The central $2\times 2$ block highlighted is of interest for the dynamics of $X^-$ in the “$p$” shell [@Cortez02; @Ware05].
The matrix elements in Tables I and II can be understood by writing the operators in the basis $\{T^{(0)}_m,{S^{(0)}_n\}}$: $\hat{\bm A}$$+$$\hat{\bm
P}$$=$$\sum_{m,n}\left(A_{mn}+P_{mn}\right){\bm
e}_z|T^{(0)}_m\rangle\langle S^{(0)}_n|+h.c.$. The matrix elements $A_{ij}^{\alpha\beta}$ and $P_{ij}^{\alpha\beta}$ in the tables are sums of the matrix elements $A_{mn}$ and $P_{mn}$ with the same symmetry. From Eq.(\[A-P\_components\]), it is seen that $A_{mn}$ is nonzero only for $|T^{(0)}_m\rangle\langle S^{(0)}_n|$ odd both in $x$ and in $y$, thus $\hat{\bm A}$ can produce longitudinal mixing $\beta^z_{ij}$ between two-electron eigenstates $T_i$ and $S_j$ if one of these contains a $x$ $(s)$-symmetry component, and the other has a $y$ ($d$)-symmetry part \[Table \[Table-A\]\]. $P_{mn}$ is non-zero only for QD asymmetries ($\mathcal{V}_a$$\neq$$0$) and for $|T^{(0)}_m\rangle\langle S^{(0)}_n|$ odd either only in $x$ or only in $y$. Thus, $\hat{\bm P}$ contributes to the longitudinal spin mixing $\beta^z_{ij}$ between $T_i$ and $S_j$ if one of them has a $s$ or $d$ component and the other has a $x$ or $y$ component \[Table \[Table-Pd\]\]. $\hat{\bm R}$ can be written as $\hat{\bm R}$$=$$\sum_{m,n}{\bm
R}^\bot_{mn}|T^{(0)}_m\rangle\langle S^{(0)}_n|+h.c.$ Results for the matrix elements of ${\bm R}^\bot_{mn}$ are not given explicitly here. They require QD lateral asymmetry and are nonzero for $|T^{(0)}_m\rangle\langle S^{(0)}_n|$ odd in one of $x$ or $y$. They can give transverse spin mixing ${\bm \beta}^\bot_{ij}$ of states with different $z$ spin projection. The degree of triplet-singlet mixing is given by the ratio of the asymmetric to the symmetric exchange: $$\mu_{ij}^{z}=\hbar^{-1}\beta_{ij}^z/J_{ij}\,\,\,,\,\,\,{\bm
\mu}_{ij}^\bot\,=\,\hbar^{-1}{\bm
\beta}_{ij}^\bot/J_{ij}\,.\label{ratios}$$
We now consider QDs with different asymmetries and consider the longitudinal spin mixing $\mu^z$ from them. This mixing does not have contributions from the Dresselhaus and Rashba couplings.
[*i. QDs with lateral inversion symmetry*]{}. For them $E_x$$=$$E_y$$=$$0$. Examples are shown in Fig.\[Asymmetry-Cases\](b) and by the $E_x$$=$$0$ points in Figs.\[Asymmetry-Cases\](a,c) and Fig.\[EllipticQD\]. In such QDs, the two-electron states $T_i$, $S_j$ have well-defined symmetries. The spin-mixing is due only to $\hat{\bm A}$, on the second diagonal (in small boxes) in Table \[Table-A\]. “Pure” states of $x$ ($y$)-symmetry such as $T_1$ ($T_2$) couple only to “pure” states of $y$ ($x$)-symmetry such as $S_3$ ($S_2$). The first order longitudinal spin mixing of $T_1$ ($T_2$) is from $S_3$ ($S_2$), which is the closest in energy. $T_3$ (the lowest $d$-symmetry triplet) couples by $\hat{\bm A}$ to $s$-symmetry singlets like $S_1$. $T_4$ (the lowest $s$-symmetry triplet) couples by $\hat{\bm A}$ to $d$-symmetry singlets such as $S_4$.
From Fig.\[Asymmetry-Cases\](b) and Fig.\[EllipticQD\] (at $E_x$$=$$0$) we can see that the asymmetric exchange can be a substantial fraction of the symmetric exchange (up to $\approx50\%$). Fig.\[Asymmetry-Cases\](b) shows that the asymmetric exchange is smaller for larger QDs, which results from larger orbits giving smaller effective magnetic fields in the s-o coupling. In this case $\beta_{22}^z$$=$$-\beta_{13}^z$ because of degeneracy. The orbital momentum $\hat{L}_z$ eigenstates $\frac{1}{\sqrt{2}}(S_2$$\pm$$iS_3)$ are strongly coupled to $\frac{1}{\sqrt{2}}(T_1$$\pm$$iT_2)$ and obey $\Delta L_z$$=$$0$. From Fig. \[EllipticQD\] we see that the asymmetric exchange decreases as the degeneracy of the first two excited states is removed by different $a_x$ and $a_y$. In this case $L_z$ is not conserved. The stronger confinement along ${\bm e}_y$ ($a_x$$>$$a_y$) leads to $J_{13}$$>$$J_{22}$, and thus to $|\mu_{13}^z|$$<$$|\mu_{22}^z|$.
(8.0,3.0)
(8.5,2.8)
[*ii. QDs with a single vertical plane of reflection*]{}. For this case $E_x$$\neq$$0$ and $E_y$$=$$0$. This gives more non-zero matrix elements in Tables \[Table-A\] and \[Table-Pd\], [*e.g.*]{} now $T_4$ is mixed with $S_3$ as well as with $S_4$. This case is illustrated in Fig.\[Asymmetry-Cases\](a,c) and in Fig.\[EllipticQD\]. $\mu^z$ for the lowest triplet is seen to decrease with increasing $E_x$. For these cases, the terms proportional to $E_x$ and $E_x^2$ in Table \[Table-A\] and also the terms from $P$$=$$2\widetilde{\gamma}^V$$(\partial_{{\bm
\rho}_1}$$\mathcal{V}_{ax}$$\times$$\hat{\partial}_{{\bm
\rho}_1})_{mn}^z\propto E_x$ from Table \[Table-Pd\] are nonzero, and they tend to cancel partially the larger terms in the boxes in Table \[Table-A\]. For some triplet-singlet pairs, such as $S_3$ and $T_1$, the symmetric exchange becomes larger and thus their mixing decreases. Other singlet-triplet pairs can be degenerate, such as $T_2$ and $S_2$ in Fig.\[Asymmetry-Cases\](a) at $E_x\approx1.5$; then nonzero $\beta^z_{22}$ leads to strong singlet-triplet mixing \[Fig.\[EllipticQD\]\]. For this case, $L_z$ is not conserved. Triplets with $\langle\hat{L}_z\rangle$$\approx$$\pm\hbar$ can be coupled to singlets that have $\langle\hat{L}_z\rangle$$\approx$$\mp\hbar$.
[*iii. QDs with no vertical plane of reflection*]{}. For this case $E_x$$\neq$$0$ and $E_y$$\neq$$0$. Then all states in Tables \[Table-A\] and \[Table-Pd\] are mixed, and the degree of longitudinal spin mixing can be larger than in the previous cases.
In addition to the longitudinal spin-mixing described above, there is also mixing that changes the spin projection $S^z$ (transverse mixing ${\bm \mu}^\bot$). This arises exclusively from the Dresselhaus and Rashba couplings, which give ${\bm R}$ in Eq.(\[A-P\_components\]). For QDs with lateral inversion symmetry, ${\bm R}$ mixes states which typically differ by the single-particle energy splitting, [*e.g.*]{} $T_1$ with $S_1$ and $S_4$ [*etc.*]{}. For them the mixing from $\hat{\bm R}$ is small, due to large $J_{11}$ and $J_{14}$. For QDs with only one vertical plane of reflection, $\hat{\bm R}$ mixes $T_1$ with $S_2$ or $S_3$, which are closer in energy and therefore give larger mixing. We show in Fig.\[Figure-Relaxation\] this transverse spin-mixing for $T_1$ and $S_2$ and for $T_1$ and $S_3$. This mixing occurs only for non-zero asymmetric potential ($E_x$$\neq$$0$). It is generally smaller than the longitudinal spin mixing discussed earlier, but it can become appreciable for large asymmetries, and it is larger in smaller QDs.
In recent experiments on light polarization reversal in QDs after excited state pumping [@Cortez02; @Ware05], the separation between excited $p_x$ and $p_y$-like states is small and cannot be resolved. One picture of this effect [@Cortez02; @Ware05] is that it involves mixing of electron triplets and singlets with simultaneous spin flips of an electron and of the hole caused by axially-asymmetric [*e-h*]{} exchange. For this mechanism, however, the experimental results require large QD asymmetries [@Ware05]. The additional mixing of triplets and singlets by asymmetric [*e-e*]{} exchange here gives consistency with experiments using more realistic QD potentials.
In addition, the present results for triplet-singlet mixing provide an alternate process for the light polarization reversal. The angular momentum from the light can be given to the orbital motion of excited-state electrons. Then the triplet-singlet coupling given here can mix two-electron states that differ in their [*orbital*]{} angular momentum by $2\hbar$, leading to reversed light polarization. This is given for example in QDs with unequal lateral sizes even in the inversion-symmetric case ([*i*]{}) above.
We are grateful for discussions with Y. Lyanda-Geller, B. V. Shanabrook, D. Gammon and M. E. Ware. This work was supported in part by the ONR and by DARPA.
[200]{}
D. Loss, D. P. DiVincenzo, Phys. Rev. A [**57**]{}, 120 (1998)
M. Kroutvar [*et al.*]{}, Nature [**432**]{} (7013), 81 (2004)
S. Cortez [*et al.*]{}, Phys. Rev. Lett. [**89**]{}(20), 207401 (2002)
M. E. Ware [*et al.*]{}, Phys. Rev. Lett. [**95**]{}(17), 177403 (2005)
V. B. Berestetskii, E. M. Lifsitz, L. P. Pitaevskii, [*Quantum Electrodynamics*]{}, (Butterworth-Heinemann, 1998)
L. D. Landau and E. M. Lifshitz, [*Quantum Mechanics*]{}, (Butterworth-Heinemann, 1998)
G. L. Bir and G. E. Pikus, [*Symmetry and Strain-induced Effects in Semiconductors*]{}, (John Wiley $\&$ Sons, 1974)
The Kane model [@BirPikus] uses parameters: band gap $E_g$, energy of the split-off band $\Delta$, and matrix element $P$ of operator $\hbar\hat{p}_x/m_0$ between the Bloch states of the conduction and valence bands.
$\gamma_s^V$ comes from the combination of the ${\bm k}$$\cdot$$\hat{\bm p}$ band mixing with the potential for holes $V_h$$=$$-c_{h} V$ ($0$$<$$c_{h}$$<$$1$), which gives $\gamma_s^V$$=$$-\frac{c_h}{\hbar^2}\frac{2P^2}{3
E_g^2}\frac{\Delta (2 E_g +\Delta)}{(E_g +\Delta)^2}$. The procedure is the same as that for treatment of the Coulomb interaction [@Badescu05; @ConstantC]
Yu. L. Bychkov, E. I. Rashba, JETP Lett. [**39**]{}, 78 (1984)
Ş. C. Bădescu [*et al.*]{}, Phys. Rev. B [**72**]{}, 161304(R) (2005)
$\gamma_s$ comes from the combination of the ${\bm k}$$\cdot$$\hat{\bm p}$ terms with the [*e-e*]{} Coulomb potential, which gives $\gamma_s$$=$$\frac{1}{\hbar^2}\frac{2P^2}{3 E_g^2}\frac{\Delta (2
E_g +\Delta)}{(E_g +\Delta)^2}$ [@Badescu05]. This also gives to a spin-independent correction to the energy, $\gamma_c$$\delta({\bm r})$, where $\gamma_c$$=$$2\pi
\frac{e^2}{\varepsilon}\frac{2P^2}{3 E_g^2}\frac{(E_g
+\Delta)^2+E_g^2}{(E_g +\Delta)^2}$.
G. Dresselhaus, Phys. Rev. [**100**]{}, 580 (1955).
We consider dots of height $W$, with confining potential $V({\bm
r})$$=$$-$$U_0\theta(|z$$-$$W/2|)(1$$+$$E_z
z)\widetilde{\mathcal{V}}({\bm \rho})$, where $U_0$ is the conduction band offset and $\theta$ is the step function. The vertical function $\xi (z)$ is the solution of $V_z(z)$$=$$-$$U_0\theta(|z$$-$$W/2|)(1$$+$$E_z z)$, and $\varphi_i$ are solutions of the lateral potential $\mathcal{V}({\bm \rho})$$=$$-$$U_0 \widetilde{\mathcal{V}}({\bm
\rho})$. For the symmetric part of the potential we use the Gaussian form $\mathcal{V}_s({\bm \rho})$$=$$-U_0
e^{-\left(x/A_x\right)^2-\left(y/A_y\right)^2}$, and for the antisymmetric part we use $\mathcal{V}_a({\bm \rho})$$=$$
\mathcal{V}_s({\bm
\rho})\left(E_x\left(\frac{x}{A_x}\right)^3+E_y\left(\frac{y}{A_y}\right)^3\right)$. $E_{x,y}$ are perturbations to $\mathcal{V}_s$ and characterize the lateral inversion asymmetry of the system. The energy scale of the lowest electron states is controlled by $A_{x,y}$ and is chosen comparable to the experiment. Here we take $W$$=$$4$ nm and use InAs/GaAs parameters with a band offset $U_0$$=$$0.6$ eV.
W. Seifert [*et al.*]{}, Prog. Cryst. Growth Ch. [**33**]{}, 423 (1996)
R. Krebs [*et al.*]{}, J. Cryst. Growth [**251**]{}, 742 (2003)
The single-particle basis is reducible to four subspaces: $\{|\mathrm{s}\rangle\}$ even in $x$ and $y$, $\{|\mathrm{x}\rangle\}$ odd in $x$, $\{|\mathrm{y}\rangle\}$ odd in $y$, and $\{|\mathrm{d}\rangle\}$ odd in $x$ and in $y$. If the system has inversion symmetry, these provide four independent subspaces for the eigenstates.
The two-particle basis is reducible to four subspaces, respectivelly of $s$, $x$, $y$, and $d$-symmetry. Symmetric combinations $S^{(0)}_n$ form the singlet basis $\{\sigma_{\mathrm{ss'}},\sigma_{\mathrm{xx'}},\sigma_{\mathrm{yy'}},\sigma_{\mathrm{dd'}}\}$$\oplus$$\{\sigma_{\mathrm{sx}},\sigma_{\mathrm{yd}}\}$$\oplus$$\{\sigma_{\mathrm{sy}},\sigma_{\mathrm{xd}}\}$$\oplus$$\{\sigma_{\mathrm{xy}},\sigma_{\mathrm{sd}}\}$; antisymmetric combinations $T^{(0)}_m$ form the triplet basis $\{\tau_{\mathrm{ss'}},\tau_{\mathrm{xx'}},\tau_{\mathrm{yy'}},\tau_{\mathrm{dd'}}\}$$\oplus$$\{\tau_{\mathrm{sx}},\tau_{\mathrm{yd}}\}$$\oplus$$\{\tau_{\mathrm{sy}},\tau_{\mathrm{xd}}\}$$\oplus$$\{\tau_{\mathrm{xy}},\tau_{\mathrm{sd}}\}$.
I. E. Dzyaloshinskii, Phys. Chem. Solids 4, 241 (1958); T. Morya, Phys. Rev. 120, 91 (1960)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The online data reduction service *reductus* transforms measurements in experimental science from laboratory coordinates into physically meaningful quantities with accurate estimation of uncertainties based on instrumental settings and properties. This reduction process is based on a few well-known transformations, but flexibility in the application of the transforms and algorithms supports flexibility in experiment design, supporting a broader range of measurements than a rigid reduction scheme for data. The user interface allows easy construction of arbitrary pipelines from well-known data transforms using a visual dataflow diagram. Source data is drawn from a networked, open data repository. The Python backend uses intelligent caching to store intermediate results of calculations for a highly responsive user experience. The reference implementation allows immediate reduction of measurements as they are recorded for the three neutron reflectometry instruments at the NIST Center for Neutron Research (NCNR), without the need for visiting scientists to install additional software on their own computers.'
author:
- William
- Paul
bibliography:
- 'webreduce.bib'
title: '***reductus*: a stateless Python data-reduction service with a browser frontend**'
---
[Ratcliff II]{}
[Kienzle]{}
A web-based, flexible scientific data reduction system is presented, which accesses published data stores and transforms raw measurements into interpretable data through data-flow diagrams which are converted to advanced calculations on a Python backend; the results are returned in real time through the web interface. The application was developed for handling neutron and X-ray reflectometry results at a user facility.
Motivation
==========
The transformation of raw measurement output into meaningful, interpretable data with attached uncertainties (data reduction) is a ubiquitous task in the experimental sciences. In the case where the workflow is well-established and the community is small the most direct way to accomplish this is to develop a custom application to be installed on a limited number of dedicated computers. However, at a scientific user facility a large number of visiting researchers are using the measurement tools on a part-time basis. As such, it is necessary to make reduction capabilities widely available and flexible. In this case a web-based application is an attractive alternative to distributing dedicated installable executables.
The main benefit of a web application is the almost universal accessibility to potential users. On top of this, a centralized reduction server also benefits from the ability to update the calculation code at any time without requiring all users to update their software, as well as largely eliminating the non-trivial time cost of maintaining an installable application for more than one target platform. Users can also install the service locally much as they would if reduction were a desktop application. This allows them to access their own data and develop reduction applications for their own instrumentation. The specific implementation described herein was developed to provide reduction for reflectometry instruments, but the system was designed to be extensible and usable for other data processing problems and work is underway to support reduction for off-specular reflectometry, small-angle neutron scattering (SANS), and triple axis spectrometry (TAS). The approach can be adapted to any type of computational problem, but the level of flexibility and computational complexity supported is driven by the needs of a data reduction engine.
Data reduction for reflectometry
--------------------------------
The measurements for reflectometry are performed at a particular configuration in laboratory coordinates (primarily incident and detector angles) as recorded in the raw data files, with some uncertainty in the precise values.[@dura2006; @salah2007] In counting experiments with a beam of particles incident on the sample there are many independent events for each configuration, each with a slightly different wavelength and flight path. When accumulated, this variation manifests as instrument resolution, which can be convolved with the theoretical model of the experiment to produce an expected count rate for that configuration. Usually the experiment will be set up so that configuration uncertainty is much less than instrument resolution in order to maximize the information gained from the measurement, and so configuration uncertainty is largely ignored. In order to determine the relative portion of scattered counts for each instrument configuration, the total incident beam rate must be measured independently, then simple division can normalize to a relative count rate. With the incident beam generated according to a Poisson distribution, the uncertainty in counts grows as $\surd n$, which leads to count rate estimates which grow as $1/\surd n$; counting longer leads to smaller error bars.
There will be some background count rate, either from alternative paths to the detector such as scattering through air, or from completely distinct processes such as electronic noise, which can be subtracted from the measured scattering signal giving an estimate of the true signal. Background subtraction can lead to negative values, which is an unphysical estimate of the true count rate. A more complete analysis of the data would model both the signal and background, and determine parameter estimates consistent with both the background measurement alone and with the combined signal and background measurement. By exposing the complete reduction chain the user is free to choose the style of analysis.
Dataflow diagram as a template for computation
==============================================
The reduction process can be described as a series of data transformations, with data flowing from one transformation module into the next. Corrections such as background subtraction and incident rate normalization can be chained together in a dataflow diagram,[@sutherland1966] with the final data set saved as the result of the reduction. This yields the estimated count per incident count for each configuration. The incident angles and wavelengths converted to more meaningful reciprocal space $q$ (inverse distance) coordinates (see Figure \[fig:simple\_diagram\]). The individual modules may have control parameters, such as a scale factor for a scaling module (needed if there is an attenuator in one measurement but not in the others).
\[fig:simple\_diagram\]
{width="\linewidth"}
Data types
----------
The data flowing between modules has a type associated with it. In order to connect the output of one module to the input of the next module, the type of the output parameter on the first module must match the type of the input parameter on the subsequent module. For each data type, there are methods to store and load the data, to convert the data to display form, and to export the data. The stored format should be complete, so that reloading saved data returns an equivalent data set. The displayed and exported forms might be a subset of the total data.
Operations
----------
By implementing the transformation modules in Python, the instrument scientist has access to a large library of numerical processing facilities based on NumPy and SciPy,[@oliphant2007python] and in particular, libraries which support propagation of uncertainties.[@lebigot2013uncertainties] The *reductus* library includes additional facilities for unit conversion, data rebinning, interpolation and weighted least squares solving as well as its own simplified uncertainties package for handling large data sets.
Bundles of inputs
-----------------
It is often the case that many measurements need to be combined, with the same computation steps applied to each data file. Rather than defining a separate chain for each file, *reductus* instead sends bundles of files between nodes. To interpret the bundles, the module parameters are defined as either single or multiple. If the primary input is single, then the module action operates on each of the inputs separately; if multiple, then all inputs are passed to the module action as a single list. For example, if several measurements must be scaled independently then joined together into a single data set, both the scale module input would be tagged as single, but the join module input would be tagged multiple. The scale factor would be tagged multiple, indicating a separate scale factor for each input. Outputs can also be single or multiple. Unlike the join module, which produces a single output from multiple inputs, a split module would produce multiple outputs from a single input. A module which rebalances data amongst inputs ([*e.g.,*]{} to correct for leakage between polarization states in a polarized beam experiment), takes multiple inputs and produces multiple outputs.
Instrument and module definition
--------------------------------
An instrument is a set of data types and the computation modules for working with them. A computation module has a number of properties, including name, description, version, module action and parameters. Each parameter has an id, a label, a type, a description, and some flags indicating whether the parameter is optional or required, and if it is single or multiple. Input and output parameters use one of the data types defined for the instrument. Control parameters can have a variety of types, including simple integers, floats or strings, or more complicated types such as indices into the data set or coordinates on the graph, allowing some parameter values to be set with mouse pointer interaction in the user interface.
Module interface language
-------------------------
To encourage accurate and complete module documentation, the module interface can be extracted from the stylized module documentation and function definition. The module interface language starts with an overview of the module action. For each input, control and output parameter it gives the data type and units and provides a short description of the parameter which can be displayed as a tool tip in the user interface. Inputs and control parameters are distinguished by examining the action declaration; the positional parameters are inputs that can be wired to another node’s output and the keyword parameters are control parameters. After the parameters, the module documentation should define the author and version. The module name is set to the name of the action function.
The module interface language is valid ReStructured Text (RST), which means that the standard `docutils` toolset for Python can be used to convert the documentation string to hypertext markup (HTML) or portable document format (PDF). The conversion to HTML is performed with `Sphinx`, allowing for the creation of an independent user manual for each instrument; it is also done dynamically for each module for display in the user interface. Embedded equations are rendered in HTML using `mathjax`, a TeX equation interpreter for javascript.
Serialization of the diagram
----------------------------
A dataflow diagram is represented as a list of nodes numbered from $0$ to $n$, with each node having a computation module, a label, an $(x,y)$ position, and values for the control parameters of the computation module. The connections are defined as a list of links, with each link having a source (node number and output parameter name), and a target (node number and input parameter name).
Every diagram can be used as a template, with the configuration values for the nodes packaged separately from the diagram. The computation engine looks first in the configuration for control parameter values for the node, using the value given in the diagram if a specific configuration value is not found. If no value is provided in the configuration or in the diagram then the default parameter value for the module is used.
Backend
=======
The backend is composed of two main pieces: a traditional web (HTTP) server providing static resources including HTML, javascript, and cascading style sheet (CSS) source code for the client application in the user’s browser, and a computation engine that handles requests for reduction calculations as remote procedure calls (RPC), as diagrammed in Fig. \[fig:architecture\]. A pool of shared calculation engines is shown but is only recommended for a production server, as a single engine is sufficient for a single-user test environment. A shared, disk-backed computation cache is not required but is strongly recommended for a responsive server, even in a single-user environment (a per-instance, non-persistent in-memory cache is the fallback option.)
The `Data Store` in Fig. \[fig:architecture\] is not part of the server but is an HTTP-accessible source of raw data[^1] which is loaded as the first step of a reduction. This arrangement makes it possible to do data reduction without handling the raw data files on the client computer—the user can download just the reduced data if they wish.
\[fig:architecture\]
{width="\linewidth"}
Computation server
------------------
The point of contact for the client is the web server, which serves the static resources (HTML, JS, CSS) as well as being a proxy gateway to the calculation engines through the Python web services gateway interface (WSGI).
### Converting the diagram to computations {#sec:computing_templates}
A dataflow diagram is a directed acyclic graph (DAG), with the modules as the nodes in the graph and the connections from outputs to inputs as the links between the nodes. No cycles are allowed, which means that the output of a module cannot be used to compute its own input. Every DAG has topological order, so the nodes can be arranged linearly such that the independent nodes appear first, and each dependent node appears after all of its input nodes. By computing nodes in topological order all inputs are guaranteed to be computed before they are needed. Although there are linear time algorithms for sorting a DAG, the diagram sizes for data reduction are small enough that a naïve $O(n^2)$ algorithm can be used.
### Results and source caching
The results of every calculation request are cached on the server. There are several choices in the configuration of the server but the default is to use a `Redis`[@sanfilippo2012redis] key-value store with a Least-Recently-Used (LRU) expiry algorithm and a maximum cache size. This can be started automatically as a system service at startup on a Linux server, allowing worry-free operation between server reboots.
The server cache is very important to the performance of the service: the slowest part of many computations is retrieving the source data files from the data repository to the server. With caching, this typically will only happen once per data file per experiment, and after that the server-local cached version will be used. For data files, it is assumed that they can be uniquely identified by their resource path (URL) and last-modified time, (accurate to a second due to the file system `mtime` limitations). It is therefore fallible if the file is rewritten in a time span of less than a second, but the data files we are using are updated much more slowly than that.
Each calculation step is identified by a unique id, created by hashing all of the input values along with the version number of the code for the step itself. For inputs which are the result of a previous calculation step, the hash of that step is used to identify the input. Since the calculations on the same data should give the same results each time, the results can be cached using this key and retrieved for future calculation steps or returned to the web client for display. If the calculation parameters change (for example, the scale factor for a scaling step), then the hash will change, and the step will be recalculated and cached with the new key. This will change the input values for the subsequent step, which will change its hash, which will cause it to be recalculated as well. In this way, if the source data (timestamp) changes, all reduction steps which depend upon it will be updated. By including the version number of each step in the cache, changes to the server will automatically cause all future reductions to be recomputed, even if they are already cached.
### Data Provenance and Reproducibility
In a highly interactive environment, where parameters, files and even work flows can be modified and the results saved at any time, it is important to have a record of the inputs as well as the results.[@simmhan2005] Therefore, the reduction template and all its parameters are stored along with the reduced data in each saved file. By loading a file into *reductus* the precise steps required to reproduce the data become visible. The NCNR data source is referenced through a Digital Object Identifier (DOI), with the implicit promise of a stable reference even if the data is moved to a new URL, thus providing long term reproducibility of the reduction. The server uses the current version of each reduction step to evaluate the outputs. This effectively acts as a behind-the-scenes update to all steps in the reduction process; any steps that are newer will be recomputed, and the updated results can be re-exported. This is particularly useful for reductions performed during data acquisition. As newer measurements are added the updated timestamp will force recomputation of all subsequent steps.
In order to reproduce an existing reduction, the server version at the time of the reduction must be used. The server source is managed with the `git` source control system,[@torvalds2010git] available on GitHub at <https://github.com/reductus/reductus>. Git creates a unique hash of the entire source tree for each commit which is stored as part of the template. To reproduce the data from a specific reduction this hash must be used to retrieve the source code and run it as a local server. The specific versions of the dependencies (`scipy`, `numpy`, `uncertainties`, $\ldots$) can be recorded in the source tree as well to protect against changing interfaces. Because users can easily revert to older versions of the software, developers are free to modify the code at will and maintain a clean code base.
### Statelessness
The computation engine maintains no state. The user interface manages the interactions of the end-user with the engine, and keeps a working copy of the active dataflow template(s); the browser session is the only real context. This has distinct operational advantages for the compute engine—it can be restarted at any time with close to zero impact on the availability and continuity of the service. The cache is persistent between engine restarts, but can be completely wiped if necessary without destroying user data (performance will suffer temporarily as the calculations are re-cached).
Server configurations
---------------------
The system is designed to be modular, allowing a number of possible configurations of the needed resources.
### Simple single-computer configuration
The simplest configuration is to run the web server, calculation nodes and cache on the same computer. A server implementation using the Python `flask` package is provided, which can simultaneously serve the static client resources as well as the RPC calculation requests. This server is suggested for use when running a private reduction service.
A private server is required for processing data sets stored locally; since the service is stateless, providing neither data upload nor storage of reduction diagrams and reduced data, there is no other way to reduce data that is not present in the public data repositories.
Local file access is only enabled for servers running on the private “localhost” network interface. Such servers should not allow external connections since access to local files is a security risk.
Similarly, a private server is required for custom reduction steps that are not part of the standard service since the standard service does not allow the execution of arbitrary Python code. Users at facilities that do not allow external access to the web will need to copy all the data files to a local machine and reduce it with a private server.
### Container-based configuration
The service can be run using a linux container environment. A recipe for Docker is provided in the source code for running the application as three coordinated containers: one for the web server, one for the Python calculation engine and one for the Redis cache. The current snapshot of the source code in the users’ directory is copied into the containers as part of the build step, so this is a useful setup for development and testing. The user need only install Docker and Docker-Compose. The supporting tools, including Python, Redis and all dependent libraries are pulled in by the Docker-Compose recipe. This greatly eases ability of users to extend the project and to test new features.
### Scalable production server configuration
For production (public-facing) servers, the static files can be copied from the `/static` subfolder of the source repository to a web server such as Apache or nginx, where requests to the `/RPC2` path are forwarded to a pool of Python calculation engines ([*e.g.,*]{} using uwsgi to run the calculations with Apache mod\_uwsgi\_proxy acting as the load balancer), sharing a Redis instance for caching.
An elastic on-demand service could be built from these components as well, with multiple (replicated) Redis caches and no limit on the number of worker Python calculation engines that can be employed by a High-Availabilty web server setup. The statelessness of the server means that no complicated synchronization is required between server components.
Web interface
=============
The user interface for *reductus* is a javascript application that runs in a browser. The application relies on a number of advanced javascript libraries, and as such it is supported only by browsers that have reasonably standards-compliant implementations of javascript (`ECMAScript` version $\geq 5$). The application is made up of these key components (see Fig. \[fig:user\_interface\]):
- visual, interactive dataflow diagram
- panel for setting parameters for individual computation modules
- plotting panel to show results
- data source file browser
\[fig:user\_interface\]
{width="\linewidth"}
The dataflow display and editor, plotting routines and parameter-setting panels were implemented using the `D3.js` visualization library.[@bostock2011-d3]
Once connected to the server, the web client can query the computation service for the public data stores, the available reduction steps, and a set of predefined template diagrams representing the usual reduction procedures for the data. The menus and default options are populated based on this information.
Dataflow diagram
----------------
The user interacts with the dataflow diagram in order to navigate the reduction chain: by clicking on a module within the diagram to bring up the parameters panel for that module, or by clicking on the input (left) or output (right) terminals to display calculation results. Changes made in the parameters panel are by default immediately applied to the active template, though by un-checking the `auto-accept parameters` box in the `data` menu an additional confirmation step (pressing the button in the panel) can be imposed to avoid accidental changing of the active template. In that context, if a user makes parameter change but then selects a different module without first pressing , that change is lost.
The button removes all values from the active parameters; the client then repopulates the panel with default values for each parameter.
The web client creates a javascript object notation (JSON) representation of the dataflow diagram along with an indicator of the input or output of interest, and sends the request to the server via HTTP POST. The response is a plottable representation of the data for that connector encoded either as JSON or MSGPACK, which is then displayed in the plotting panel.
Clicking on any output within the dataflow diagram (including the rightmost, “final” output) will trigger a calculation of all ancestor results for that result in the diagram on the server, as described in section \[sec:computing\_templates\].
So, while the user has the option to inspect intermediate calculation steps, for a routine reduction dataflow one can simply populate the file inputs and click on the final output for export the completely reduced data.
Parameters panel
----------------
At the beginning of a reduction, a user chooses an instrument definition for working with the data. As described in \[sec:computing\_templates\], this includes a list of data types and a list of reduction steps (modules) to act on those data types. The parameters panel is rendered based on the definition for the chosen module type, using the predefined types for each input field to the module function, mapping the simple known types (`int, float, bool,` [*etc.*]{}) to HTML form elements. Some field types have renderers with enhanced interaction with the plot panel, such as an `index` type which allows clicking on data points to add them to an index list in the parameters panel. Another example of enhanced interactions is the `scale` type which enables dragging a whole dataset on the plot to set the scaling factor in the parameters panel.
When parameters are changed in this panel and committed with , they will be used in any calculation of data flowing through that module.
Browser caching of calculations
-------------------------------
In addition to the caching provided on the server for avoiding recalculation of identical results, a local browser cache of calculations is provided. This is particularly useful for the initial source data load, in which metadata from all of the files in a source directory are passed to the client for inspection and sorting in the source file browser. Naturally in a data reduction scheme, the amount of data on the input side (loaders) is much greater than the output result, so caching of the inputs helps tremendously when making small adjustments interactively to the dataflow algorithm or parameters.
Sessions and persistence
------------------------
The *reductus* server is stateless; the reduction diagrams created by the user are not stored (a unique hash of the template representation may be associated with cached calculations on the server, but no user template is ever stored there). The only state associated with a session is stored in the browser or on the filesystem of the user’s computer.
### Stashing in-browser
Results of calculations can be “stashed” in the local persistent memory of the browser. A list of these results can be recalled in the client and reused in two ways: by reloading the entire calculation into the active dataflow panel, or by selecting multiple stashed results to directly compare their outputs.
### Saving to and loading from filesystem {#subsec:saving}
In addition to the local browser store, the user may download a text version of the dataflow diagram in JSON format with `Menu\rightarrowTemplate\rightarrowDownload`, which can be reloaded with `Menu\rightarrowTemplate\rightarrowUpload`. The file contains the diagram along with any field values, including the names of the input files. The actual data is not included.
The data for the currently selected node can be saved with `Menu\rightarrowData\rightarrowExport`. This prompts the user for a filename, then produces a tab-delimited column-format text file with the dataflow diagram prepended as a comment header. The stored diagram allows a full reduction to be reloaded into the client with `Menu\rightarrowData\rightarrowReload Exported` (note that this may trigger recalculation if the raw data has been updated since the reduction was exported).
Due to security limitations built into all current browsers the data may only be saved to the user’s “Downloads” folder, while uploads can of course originate from any user-readable folder.
### Sharing data among collaborators
The dataflow diagram is self contained. The reduced-data text files produced by the *reductus* system can be shared with others easily by email or portable media, and provide both useful data and a recipe in the header for recreating the data from known sources with known transformations; the chain of logic can be inspected and verified by reloading the dataflow into the web client at any time, thus assuring data provenance. This allows for easy collaboration amongst users without the need for accounts or passwords.
Conclusions
===========
The *reductus* system is an interesting experiment in providing stateful web services with a stateless server. Although users lose the convenience of cloud services for managing their data, they are free from the inconvenience of maintaining yet another user ID and password. Files can be stored and shared using familiar tools such as email. The server is easy to adapt and install locally for the rare user that needs more than the rigid set of functions and data sources provided in the remote web service; this is no more complex than adapting and installing the equivalent desktop application would be. For the developer, the stateless server needs very little maintenance. There are no database migrations needed, no backups required, and moving the service to a different computer is as simple as installing the software and redirecting the domain name service (DNS) to an alternate IP. Javascript provides a flexible environment for interactive applications, with a rich and growing ecosystem of libraries that work across most browser platforms. The Python backend provides an ecosystem for rapid development of numerical code. The web services middleware gives us scalability with no additional effort.
Making the dataflow graph visible and modifiable increases flexibility without increasing complexity. Users with simple reduction problems can enter their data on the left of the graph and retrieve the results on the right, ignoring all steps in between. If there are problems, they can examine inputs and outputs at each step and identify the cause. Although 85% of reduction is performed on-site during the experiment, there were over 150 external users access the system from across the United States and Europe in 2017. Without the need to install a local version of the software we have far fewer support requests; now they only occur for unusual datasets.
Feedback from the users has been overwhelmingly positive, and the new system has completely supplanted our old reduction software for the three neutron reflectometry instruments at the NCNR, with work underway to adapt the system to several new instrument classes at the facility.
contributions to the early code were made by Brendan Rowan, Alex Yee, Ophir Lifshitz and Elakian Kanakaraj as part of the NIST Summer High school Internship Program (SHIP) and Joseph Redmon as part of the Summer Undergraduate Research Fellowship (SURF), supported by NSF CHRNS grant DMR-1508249. We are grateful to Prof. Robert Briber of the Dept. of Materials Science and Engineering at the University of Maryland for providing significant technical support and operating the servers for early deployments of the *reductus* service.
The identification of commercial products or services in this paper does not imply recommendation or endorsement by the National Institute of Standards and Technology, nor does it imply that the equipment or service used is necessarily the best available for the purpose.
[^1]: for example, the NIST Center for Neutron Research data store is located at <https://dx.doi.org/10.18434/T4201B>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'At small momenta, the Girvin-MacDonald-Platzman (GMP) mode in the fractional quantum Hall (FQH) effect can be identified with gapped nematic fluctuations in the isotropic FQH liquid. This correspondence would be exact as the GMP mode softens upon approach to the putative point of a quantum phase transition to a FQH nematic. Motivated by these considerations as well as by suggestive evidence of an FQH nematic in tilted field experiments, we have sought evidence of such a nematic FQHE in a microscopic model of interacting electrons in the lowest Landau level at filling factor 1/3. Using a family of anisotropic Laughlin states as trial wave functions, we find a continuous quantum phase transition between the isotropic Laughlin liquid and the FQH nematic. Results of numerical exact diagonalization also suggest that rotational symmetry is spontaneously broken, and that the phase diagram of the model contains both a nematic and a stripe phase.'
author:
- 'N. Regnault'
- 'J. Maciejko'
- 'S. A. Kivelson'
- 'S. L. Sondhi'
bibliography:
- 'v1v3v5.bib'
title: Evidence of a fractional quantum Hall nematic phase in a microscopic model
---
10000
Introduction
============
The FQH nematic is a hypothesized state of matter simultaneously characterized by a broken symmetry and topological order. In this state, the defining characteristics of topological order (quasiparticles with fractional charge and statistics and topological ground-state degeneracy [@QHE]) coexist with those of nematic order (spontaneous breaking of spatial rotational symmetry [@DeGennes]). Early approaches to broken rotational symmetry in FQH states involved the construction of effective field theories [@balents1996] and trial wave functions [@musaelian1996]. Observations of anisotropic transport in a two-dimensional electron gas (2DEG) at filling fraction $\nu=7/3$ [@xia2011] (and, more recently, $\nu=5/2$ [@liu2013]) motivated the construction of effective field theories [@mulligan2010; @mulligan2011; @maciejko2013; @you2014] of a putative FQH isotropic-nematic quantum phase transition. An independent motivation for the study of such physics comes from the work of Haldane [@haldane2011; @haldane2009; @haldane2011b], who proposed a geometrical description of FQH states in which the Girvin-MacDonald-Platzman (GMP) mode [@girvin1985; @girvin1986]—a gapped neutral collective mode similar to the phonon-roton mode of superfluids—is interpreted as a fluctuating but unimodular guiding-center spatial metric $g_{ab}({\boldsymbol{r}},t)$ [@luo2016] that describes the shape of the correlation hole in the FQH fluid. In particular, the long-wavelength limit of the GMP mode was identified as a spin-2 (quadrupolar) excitation, analogous to the graviton [@lee1991; @dev1992; @yang2012b; @yang2013b]. Because a unimodular metric $g_{ab}$ is equivalent via matrix exponentiation $g=\exp Q$ to a traceless symmetric nematic order parameter $Q_{ab}$, the GMP mode of the Laughlin liquid can also be identified as the gapped fluctuations of a nematic order parameter in a disordered (isotropic) phase [@maciejko2013].
Deep in the FQH phase, the GMP mode (and all other excitations) occur at high energies, and so are sensitive to microscopic considerations. However, if a transition from an isotropic FQH liquid to a FQH nematic can be induced by varying some parameters in the problem, the GMP should soften upon approach to the associated quantum critical point (QCP). Proximate to the QCP, the physics should be universal and accurately describable by an effective quantum field theory. This scenario and the resulting effective field theory of the FQH nematic state [@maciejko2013] can be realized in a model of 2DEG in a magnetic field with attractive quadrupolar interactions, within the composite-fermion mean-field theory [@you2014]. Ref. [@you2014] also clarified the relation between the Berry phase for nematic fluctuations [@maciejko2013] and the Hall viscosity of the isotropic FQH liquid [@read2009; @read2011].
There are two additional routes to arriving at a FQH nematic in addition to the transition from a gapped FQH state described above. In the first the effective field theory approach was extended to the problem of the half-filled Landau level, in order to describe a transition from an isotropic to a nematic composite Fermi liquid [@you2016]. This is relevant in the context of the observation of compressible anisotropic phases in half-filled Landau levels [@lilly1999a; @du1999; @lilly1999b; @pan1999; @xia2010; @samkharadze2016]. In the composite fermion approach, the FQH nematic is obtained as a Pomeranchuk instability [@pomeranchuk1958; @oganesyan2001] of composite fermions [@doan2007], and the broken rotational symmetry corresponds to the condensation of a fermion bilinear. This is a Fermi-liquid-like or weak-coupling mechanism for the formation of a nematic. In the second, which is a strong-coupling perspective, a nematic electron fluid is viewed as a partially melted solid [@fradkin2010]. In this picture, inspired from the theory of classical liquid crystals, a nematic state is proximate to various phases with broken translational symmetry, such as stripe or smectic phases. If under the influence of some external parameter, topological defects in the latter proliferate in such a way that translational order is melted but orientational order is preserved, a nematic state (or in general, a $\ell$-atic with $\ell\geq 2$) results. This leads, for example, to an alternate description of the compressible nematic phase observed in half-filled Landau levels as a quantum melted stripe phase rather than as a nematic composite Fermi liquid [@fradkin1999; @fradkin2000; @wexler2001; @radzihovsky2002]. According to this perspective, one might expect an incompressible FQH nematic to be proximate not only to an isotropic FQH liquid but also to phases with translational order.
More generally, there is increasingly strong evidence that a host of highly correlated electronic systems support one or another form of electronic liquid crystalline phases [@kivelsonfradkinemery; @fradkin2010; @vojta2009; @fernandes2014; @ianfisherscience]. In many ways, the FQH nematic studied here is the paradigmatic example, as the electronic Hamiltonian is simpler, with none of the complexity associated with the solid-state chemistry of typical highly correlated materials, and more symmetric (to good approximation, the system is fully rotationally invariant). Thus, the results of the present study may be conceptually useful more broadly.
In this paper we report the results of a microscopic study of spatially ordered phases proximate to the incompressible, isotropic FQH liquid. To do so, we explore possible quantum phase transitions out of the isotropic FQH liquid in a model of interacting electrons in the lowest Landau level (LLL) at filling factor $\nu=1/3$. Our model contains only the first three Haldane pseudopotentials [@haldane1983] $V_1$, $V_3$, and $V_5$, with the ratios $V_3/V_1$ and $V_5/V_1$ as tuning parameters. We study the model with numerical exact diagonalization (ED) for up to $N=13$ electrons as well as with a variational approach using trial wave-functions of a sort that have been used successfully in studies of explicitly anisotropic Hamiltonians [@qiu2012; @yang2012; @wang2012; @papic2013; @qiu2013; @yang2013; @johri2016; @apalkov2014; @ghazaryan2015; @ghazaryan2014]. We find that the isotropic $\nu=1/3$ Laughlin liquid is obtained at small values of the tuning parameters but gives way to other phases at larger values (Fig. \[fig:fig1\]). Our results show (although not entirely unambiguously) the existence of three phases separated by direct transitions: (1) An isotropic (Laughlin) FQH liquid; (2) A nematic FQH liquid crystal; (3) A striped FQH liquid crystal phase (which breaks both rotational and translational symmetry).
![Conjectured schematic phase diagram of the $V_1$-$V_3$-$V_5$ model: A = isotropic Laughlin liquid, B = stripe/smectic phase, C = FQH nematic.[]{data-label="fig:fig1"}](fig1.pdf){width="0.5\columnwidth"}
Microscopic model
=================
The general Hamiltonian for a 2DEG in a uniform perpendicular magnetic field and interacting via translationally invariant two-body interactions is $$H= \sum_{{\boldsymbol{q}}} V({{\boldsymbol{q}}}) :\bar{\rho}_{{\boldsymbol{q}}} \bar{\rho}_{-{{\boldsymbol{q}}}}:,\label{GenericHamiltonian}$$ when projected to the LLL, which corresponds to ignoring Landau-level mixing effects. In Eq. (\[GenericHamiltonian\]), $\bar{\rho}_{{\boldsymbol{q}}}$ is the projected density operator in momentum space and the colons denote normal ordering. The projected two-body interaction $V({{\boldsymbol{q}}})$ can be decomposed as $$V({{\boldsymbol{q}}})=\sum_{n=1,3,5,\ldots} V_n {\cal L}_n({{\boldsymbol{q}}}^2)\label{Pseudopotentials},$$ where $V_n$ are the Haldane pseudopotentials and ${\cal L}_n$ is the $n$th Laguerre polynomial. The model we consider is one in which $V_1,V_3,V_5$ are nonzero and all the other pseudopotentials are set to zero. Thus our Hamiltonian depends on two tuning parameters $V_3/V_1$ and $V_5/V_1$, $$\begin{aligned}
H(V_3/V_1,V_5/V_1)&=&V_1\sum_{{\boldsymbol{q}}}\biggl({\mathcal{L}}_1({{\boldsymbol{q}}}^2)+\frac{V_3}{V_1}{\mathcal{L}}_3({{\boldsymbol{q}}}^2){\nonumber}\\
&&+\frac{V_5}{V_1}{\mathcal{L}}_5({{\boldsymbol{q}}}^2)\biggr):\bar{\rho}_{{\boldsymbol{q}}} \bar{\rho}_{-{{\boldsymbol{q}}}}:,\label{v1v3v5}\end{aligned}$$ with $V_1$ simply setting the overall energy scale. For $V_3/V_1=V_5/V_1=0$, one recovers the model Hamiltonian for which the Laughlin state is the exact zero-energy ground state at $\nu=1/3$. For our calculations, we consider $N$ interacting electrons in the LLL on a torus geometry pierced by $N_\Phi$ flux quanta. The torus is spanned by the two vectors $L_x {{\boldsymbol{e}}}_x$ and $L_y {{\boldsymbol{e}}}_y$, with two orthogonal unit vectors ${{\boldsymbol{e}}}_x$ and ${{\boldsymbol{e}}}_y$. The aspect ratio of the torus is defined as $L_y/L_x = 1 + \delta$. For $\delta=0$, the system has a square aspect ratio and exhibits an additional $C_4$ symmetry. We fix the filling factor $\nu=N/N_\Phi$ to be $\nu=1/3$, and neglect spin effects.
Variational approach {#sec:variational}
====================
To search for a possible isotropic-to-nematic transition in the microscopic model (\[v1v3v5\]), a natural first line of attack is a variational approach. This requires (1) a family of trial wave functions parametrized by an appropriate set of variational parameters, and (2) a microscopic definition of an order parameter that can be calculated from these wave functions.
In 2D systems such as the FQH liquids of interest to us, nematic order is described in the long-wavelength limit by a real traceless symmetric rank-two tensor $Q_{ab}$ with two independent parameters $Q_{11}=-Q_{22}$ and $Q_{12}=Q_{21}$ [@DeGennes]. It is convenient to use a complex representation $Q\equiv Q_{11}+iQ_{12}=|Q|e^{i\phi}$ where $|Q|$ and $\phi$ are the amplitude and phase of the nematic order parameter, respectively [@fradkin2010b]. In this representation, phase rotations are spatial rotations; here we define $\phi$ such that a rotation of it by $2\pi$ corresponds to a physical spatial rotation by $\pi$, which leaves the nematic order parameter invariant. In an isotropic phase, $|Q|$ vanishes and $\phi$ is free to fluctuate, while in a nematic phase, $|Q|$ is nonzero and $\phi$ selects a specific direction, breaking rotational symmetry spontaneously.
Building on the equivalence between Haldane’s unimodular metric $g_{ab}$ and a nematic order parameter [@maciejko2013], $$\begin{aligned}
\label{gab}
g_{ab}=\left(\begin{array}{cc}
\mathop{\mathrm{ch}}|Q|+\mathop{\mathrm{sh}}|Q|\cos\varphi & \mathop{\mathrm{sh}}|Q|\sin\varphi \\
\mathop{\mathrm{sh}}|Q|\sin\varphi & \mathop{\mathrm{ch}}|Q|-\mathop{\mathrm{sh}}|Q|\cos\varphi
\end{array}\right),\end{aligned}$$ we use the anisotropic LLL wave functions of Ref. [@qiu2012] as a continuous family of trial ground-state wave functions $|\Psi(|Q|,\phi)\rangle$ parametrized by $|Q|$ and $\phi$ (see Appendix \[app:trialWF\] for the explicit form of these wave functions). To determine the possibility of an FQH nematic phase in the Hamiltonian (\[v1v3v5\]), we minimize the variational energy $$\begin{aligned}
\label{VarE}
&E(|Q|;V_3/V_1,V_5/V_1)=\nonumber\\
&\hspace{10mm}\frac{\langle\Psi(|Q|,\phi)|H(V_3/V_1,V_5/V_1)|\Psi(|Q|,\phi)\rangle}
{\langle\Psi(|Q|,\phi)|\Psi(|Q|,\phi)\rangle},\end{aligned}$$ with respect to the variational parameter $|Q|$. Unlike in previous studies [@qiu2012; @yang2012; @wang2012; @papic2013; @qiu2013; @yang2013; @johri2016] where the optimal metric $g_{ab}$ is determined variationally for a system with an anisotropic band effective mass or an anisotropic dielectric tensor, i.e., for a microscopic Hamiltonian that *explicitly* breaks rotation symmetry, here our microscopic Hamiltonian (\[v1v3v5\]) is rotationally invariant and the variational energy (\[VarE\]) is independent of $\phi$. However, rotation symmetry can be broken *spontaneously* if Eq. (\[VarE\]) is minimized by a nonzero nematic amplitude $|Q|$ for certain values of $V_3/V_1$ and $V_5/V_1$.
Finally, we must find a definition of the nematic order parameter in terms of the electronic degrees of freedom of our microscopic model. In principle, any microscopic observable with quadrupolar symmetry is a suitable candidate. Consider for example the quantity $$\begin{aligned}
\label{fd}
f({{\boldsymbol{d}}})=\langle \Psi|\bar{\rho}({{\boldsymbol{r}}}+{{\boldsymbol{d}}}/2)\bar{\rho}({{\boldsymbol{r}}}-{{\boldsymbol{d}}}/2)|\Psi\rangle,\end{aligned}$$ where ${{\boldsymbol{r}}}$ is the position vector, $\bar{\rho}({{\boldsymbol{r}}})$ is the projected density operator in real space, $|\Psi\rangle$ is the ground state, and ${{\boldsymbol{d}}}$ is a vector with a microscopic length (so the order parameter remains local) and angle $\theta$ in the plane. The length of ${{\boldsymbol{d}}}$ only affects the overall scale of the order parameter amplitude and can be chosen for simplicity to be the magnetic length $\ell_B$, which is the only length scale in the problem. For a translationally invariant ground state, Eq. (\[fd\]) is independent of ${{\boldsymbol{r}}}$. For a rotationally invariant ground state, $f({{\boldsymbol{d}}})=0$. Furthermore, $f({{\boldsymbol{d}}})=f(-{{\boldsymbol{d}}})$, hence $f({{\boldsymbol{d}}})$ measures the breaking of rotational symmetry up to the equivalence ${{\boldsymbol{d}}}\sim-{{\boldsymbol{d}}}$, as does a nematic order parameter (which can also be thought of as a headless vector [@DeGennes]). Expanding $f({{\boldsymbol{d}}})$ in angular momentum components, we find that the first nontrivial term has angular momentum $l=2$ and thus its complex coefficient $f_2=\int_0^{2\pi}\frac{d\theta}{2\pi}e^{-2i\theta}f({{\boldsymbol{d}}})$ can be used as a microscopic definition of our complex nematic order parameter. Because our numerical simulations are performed on a torus, the full $SO(2)$ rotation symmetry is in fact explicitly broken to a discrete $C_4$ rotation symmetry by the periodic boundary conditions (see Appendix \[app:finitesize\] for a detailed discussion). The resulting Ising nematic [@abanin2010; @metlitski2010] is thus more appropriately described by an order parameter $N_{x^2-y^2}\propto f(\hat{x})-f(\hat{y})$ with $d_{x^2-y^2}$ symmetry, $$\begin{aligned}
\label{NematicOP}
N_{x^2-y^2}\equiv\sum_{{\boldsymbol{q}}}(\cos q_x-\cos q_y)\langle \Psi|\bar{\rho}_{{{\boldsymbol{q}}}}\bar{\rho}_{-{{\boldsymbol{q}}}}|\Psi\rangle,\end{aligned}$$ working in units such that $\ell_B=1$, and assuming nematic order along the $x$ axis without loss of generality.
![(a) Value of the variational parameter $|Q|$ that minimizes the ground-state energy (\[VarE\]) for a system with $N=13$ electrons. A continuous quantum phase transition from an isotropic to a nematic FQH state is seen at $V_3/V_1\approx 0.6$. (b) Amplitude of the microscopic nematic order parameter (\[NematicOP\]) as a function of $|Q|$.[]{data-label="fig:fig6"}](fig2.pdf){width="\columnwidth"}
In Fig. \[fig:fig6\](a) we plot the value of the variational parameter $|Q|$ that minimizes the variational energy (\[VarE\]), while in Fig. \[fig:fig6\](b) we give the correspondence between $|Q|$ and the microscopic nematic order parameter $N_{x^2-y^2}$, i.e., Eq. (\[NematicOP\]) evaluated in the optimal trial state. The isotropic FQH liquid is stable up to a value $V_3/V_1\approx 0.6$, beyond which rotation symmetry is broken spontaneously. Given that the trial states $|\Psi(|Q|,\phi)\rangle$ have the same topological order as the isotropic $\nu=1/3$ Laughlin state for all $|Q|$ and $\phi$ [@qiu2012], this broken-symmetry state is a FQH nematic. In this variational approach, the isotropic-nematic transition is continuous; however, in other approaches such as composite fermion mean-field theory [@you2014] the transition is sometimes found to be first order. Figure \[fig:fig6\] was calculated numerically and is shown for $N=13$, but finite-size effects are negligible given the variational nature of this approach. The variational approach thus suggests a first phase boundary in our schematic phase diagram (Fig. \[fig:fig1\]), i.e., the nearly vertical phase boundary that signals the destruction of the isotropic Laughlin liquid (phase A).
Exact diagonalization {#sec:ED}
=====================
![Exact diagonalization study of the $V_1$-$V_3$-$V_5$ model with $N$ electrons: (a)-(c) gap between ground and first excited state in the zero-momentum sector; (d)-(f) overlap of the ground state with the $\nu=1/3$ Laughlin state; (g)-(i) overlap of the ground state with the $V_3/V_1\rightarrow\infty$ ground state.[]{data-label="fig:fig2"}](fig3.pdf){width="\columnwidth"}
To complement our variational approach, we study the model (\[v1v3v5\]) with ED for up to $N=13$ electrons. Our main results are shown in Fig. \[fig:fig2\]. In order to compare gaps between different values of $V_3/V_1$, we have set the energy scale to be one for the two-particle problem, irrespective of the value of $V_3/V_1$. While finite-size effects are still noticeable from a quantitative standpoint, from a qualitative standpoint the combined data of the gap in the zero-momentum sector \[Fig. \[fig:fig2\](a)-(c)\] and the overlap of the ground state with the Laughlin \[Fig. \[fig:fig2\](d)-(f)\] and $V_3/V_1\rightarrow\infty$ \[Fig. \[fig:fig2\](g)-(i)\] ground states allow us to further refine our schematic phase diagram (Fig. \[fig:fig1\]). Three phases are noticeable, separated by gap-closing transitions. In agreement with the variational results, phase A occurs for $V_3/V_1$ less than $\sim 0.5$ and appears to persist for relatively large $V_5/V_1$. It is gapped and adiabatically connected to the isotropic $\nu=1/3$ Laughlin liquid. The second phase (denoted by B) occurs for $V_3/V_1\gtrsim 0.5$ and small $V_5/V_1$. The gap in the zero-momentum sector is much smaller than that of the Laughlin state and appears to be strongly size-dependent. The ground state in this part of the phase diagram has significant overlap with the $V_3/V_1\rightarrow\infty$ ground state (obtained by considering Eq. (\[Pseudopotentials\]) with $V_3$ nonzero only). Since this overlap grows with increasing $V_3/V_1$, we suspect the ground state is adiabatically connected to the $V_3/V_1\rightarrow\infty$ ground state. This picture holds when looking at the gap to the first excited state irrespective of its momentum (such a calculation can only be performed for $N=10$ and $N=11$). The $V_3$-only model has only been the object of a few studies [@Wojs-PhysRevB.69.205322; @Wojs-PhysRevB.71.245331; @Mukherjee-PhysRevLett.112.016801], and we discuss it further in Sec. \[sec:largeV3\]. Finally, the third phase (denoted by C) occurs for large $V_3/V_1$ and $V_5/V_1$. It has negligible overlap with both the Laughlin state and the $V_3/V_1\rightarrow\infty$ ground state; we discuss it further in Sec. \[sec:largeV3V5\]. Note that we did not consider a system with $N=12$ electrons. Indeed, the Laughlin liquid has an enhanced stability due to commensuration effects (see Appendix \[app:numerical\]).
As already mentioned, the gap-closing phase boundary obtained in ED between phase A and phases B and C roughly matches the isotropic-nematic phase boundary obtained in the variational approach. Since phases B and C are on the nematic side of the transition, this suggests they are characterized by some form of spatial order in the thermodynamic limit. As we discuss in Sec. \[sec:largeV3V5\], phase C exhibits a large nematic susceptibility, and we conjecture it is an FQH nematic in the thermodynamic limit. Given that the variational approach with a single nematic variational parameter predicts an FQH nematic also in the region corresponding to phase B, while our ED results suggest this is a distinct phase but also with spatial order (as we will argue in Sec. \[sec:largeV3\]), we conjecture that phase B is a spatially ordered phase proximate to a nematic, that has a lower energy than the nematic in this region of parameter space. The most natural possibility is a phase with spontaneously broken translation symmetry, i.e., a smectic/stripe phase, from which the FQH nematic descends by quantum melting at the B-C phase boundary. Such phases have been considered in the quantum Hall context before [@fradkin1999; @fradkin2000; @wexler2001; @radzihovsky2002].
Large $V_3$: possible stripe phase {#sec:largeV3}
----------------------------------
![Dependence of gap to first excited state on system size: comparison between the Laughlin liquid (blue curve, $V_1=1$ and $V_3=V_5=0$) and the pure $V_3$ phase (red curve, $V_3=1$ and $V_1=V_5=0$ i.e. $V_3/V_1\rightarrow\infty$). The gap was computed up to $N=14$ (resp. $N=13$) electrons for the pure $V_3$ phase (resp. Laughlin liquid).[]{data-label="fig:fig3"}](fig4.pdf){width="0.9\columnwidth"}
Compared to the $V_1$-only Hamiltonian whose exact ground state at filling factor $\nu=1/3$ is the celebrated Laughlin state, the $V_3$-only Hamiltonian at the same filling factor has been studied much less (the exact ground state of the $V_3$-only Hamiltonian is the $\nu=1/5$ Laughlin state, but at filling factor $1/5$). A series of numerical ED studies [@Wojs-PhysRevB.69.205322; @Wojs-PhysRevB.71.245331; @Mukherjee-PhysRevLett.112.016801] have argued that the ground state in this limit is an incompressible state with shift 7 on the sphere and topologically distinct from the Laughlin state, the so-called Wójs-Yi-Quinn (WYQ) state. Interpreting the $V_3$-only Hamiltonian as describing the interaction of composite fermions in the second LL at filling factor $\nu_\text{CF}=1+1/3$, the resulting WYQ state of composite fermions was proposed as a candidate state for a recently discovered FQHE at filling factor $\nu=4/11$ [@pan2015; @samkharadze2015].
Here we propose the alternate possibility that the $V_3/V_1\rightarrow\infty$ ground state, and by adiabatic continuity, phase B in Fig. \[fig:fig1\], might be gapless in the thermodynamic limit. First, as mentioned before, the gap in phase B appears to be strongly size-dependent. In Fig. \[fig:fig3\] we compare the dependence on system size $N$ of the gap between ground and first excited state for the Laughlin liquid ($V_1$-only model) and the $V_3$-only model. While the gap in the Laughlin liquid is essentially system-size independent, the gap in the pure $V_3$ model exhibits a strong dependence on $N$ (and almost closes at $N=14$ electrons, the largest size we can reach), which suggests the existence of gapless degrees of freedom in the thermodynamic limit. It is interesting to note in this context that in a recent study [@mukherjee2015], the neutral collective mode spectrum of the composite fermion WYQ state mentioned above was found to be nearly gapless, with the wave vector and energy of the magnetoroton minimum an order of magnitude smaller than in the $\nu=1/3$ Laughlin state.
Second, in Fig. \[fig:fig4\] we plot the dependence of the gap on the aspect ratio $\delta$ of the torus. Here we plot both the direct gap (gap in the zero-momentum sector) and the absolute gap, which in these systems is indirect (i.e., occurs at finite momentum). In the Laughlin liquid there is essentially no dependence of the gaps on the aspect ratio, as one expects for an isotropic incompressible phase. By contrast, in the pure-$V_3$ case both direct and indirect gaps are strongly dependent on the aspect ratio, which is at odds with the scenario proposed previously of an isotropic, incompressible topological phase. Within the limitations of numerical ED on small systems, this result nonetheless suggests that the ground state of the $V_3$-model is strongly susceptible to spatial perturbations that break the spatial $C_4$ rotational symmetry of the torus. In the thermodynamic limit, this is suggestive of a phase with a spontaneously broken spatial symmetry. As we discuss in Sec. \[sec:largeV3V5\], phase C is characterized by a large nematic susceptibility, leading us to speculate that it becomes an FQH nematic in the thermodynamic limit. Since phase B is a distinct phase, but proximate to a nematic, we conjecture it is a smectic/stripe phase.
![Dependence of gap to first excited state on aspect ratio $\delta=L_y/L_x-1$ of the torus in both the zero-momentum sector (blue lines) and at finite momentum (red lines). We have considered both $N=12$ electrons (solid lines) and $N=11$ electrons (dotted lines): (a) Laughlin liquid ($V_3=V_5=0$), (b) pure $V_3$ phase ($V_3/V_1\rightarrow\infty$).[]{data-label="fig:fig4"}](fig5.pdf){width="\columnwidth"}
Large $V_3$ and $V_5$: possible FQH nematic {#sec:largeV3V5}
-------------------------------------------
In the large $V_3,V_5$ regime, the ground state has essentially zero overlap with both the Laughlin state and the $V_3/V_1\rightarrow\infty$ ground state; we surmise it corresponds to a distinct phase of matter in the thermodynamic limit (phase C in Fig. \[fig:fig1\]). As in the case of phase B, the gap in phase C appears to be strongly size-dependent \[Fig. \[fig:fig2\](a)-(c)\], suggestive of a gapless phase in the thermodynamic limit. Combined with the variational results of Sec. \[sec:variational\], this strongly suggests phase C is a FQH nematic.
To further investigate this possibility and go beyond the variational approach, which is necessarily biased, we would like to show directly that the ground state obtained in ED in the large $V_3,V_5$ regime exhibits nematic order. However, since there is no spontaneous symmetry breaking in a finite system, we can only calculate a nematic susceptibility, defined as the rate of change of the nematic order parameter with respect to a suitable symmetry-breaking field. Since a nonzero aspect ratio parameter $\delta$ explicitly breaks the $C_4$ rotation symmetry of the torus, we adopt the following microscopic definition of nematic susceptibility, $$\begin{aligned}
\label{chiQ}
\chi_Q\equiv\lim_{\delta\rightarrow 0}\frac{\partial N_{x^2-y^2}(\delta)}{\partial\delta},\end{aligned}$$ where the nematic order parameter (\[NematicOP\]) is calculated numerically with the ED ground state $|\Psi\rangle$ (by opposition to the trial ground state) and as a function of $\delta$. In practice, we obtain $\chi_Q$ by numerically determining the slope of the curve of $N_{x^2-y^2}(\delta)$ vs $\delta$ for small values of $\delta$.
In Fig. \[fig:fig5\] we plot the nematic susceptibility $\chi_Q$ as a function of the tuning parameters $V_3/V_1$ and $V_5/V_1$, for different system sizes. While we are clearly in a regime where finite-size effects are still important, at least for $N=10$ and $N=11$ electrons there is a strong nematic susceptibility in this phase. This is evidence of strong nematic correlations, and further supports our conjecture that phase C is an FQH nematic in the thermodynamic limit.
![Ground-state nematic susceptibility (\[chiQ\]), capped for readability at $\chi_\textrm{max}=10$.[]{data-label="fig:fig5"}](fig6.pdf){width="\columnwidth"}
Implications for experiments
============================
Experimental evidence of nematic $\nu=7/3$ and $\nu=8/3$ FQH states was reported in the tilted field experiments of Xia [*et al.*]{} [@xia2011]. They argued convincingly that the tilted field has two important effects in the particular devices they studied: Of course, it produces a symmetry breaking field in the 2DEG which defines a preferred nematic “x” axis parallel to the in-plane component of the magnetic field. It also affects the effective interactions, and it is apparently this latter effect that tunes the system through an isotropic-to-nematic transition. In the relevant range of tilt angles, the Hall conductance is appropriately quantized at low temperatures. However, the longitudinal resistivity exhibits an astonishing $T$ dependence: While at elevated $T$, $\rho_{xx}$ and $\rho_{yy}$ differ by a tilt-angle dependent factor, both decrease with decreasing $T$ in a similar fashion. However, below a critical temperature $T_\text{nem} \sim 50$ mK, $\rho_{xx}$ exhibits a sharp change such that it is an increasing function of decreasing $T$, while $\rho_{yy}$ continues its smooth decrease, tending to 0 as $T\to 0$.
The existence of a finite-$T$ transition is expected [@fradkin2000] for a nematic phase, and this and the large enhancement of the resistance anisotropy are the strongest pieces of evidence in favor of this identification [@fradkin2000; @mulligan2010; @mulligan2011]. However, it is far from obvious how to understand the particulars of the results. Most obviously, the quasiparticle creation energy and character should not change dramatically across the nematic transition, and consequently well inside the nematic state, either in the regime of thermally activated transport or variable range hopping, $\rho_{xx}/\rho_{yy}$ should approach a $T$ independent constant value, which is not readily reconciled with the strongly $T$ dependent anisotropy that is observed.
To resolve this issue, we offer a speculative proposal, which could be readily tested in future experiments. One of the most important characteristics of classical liquid crystals is that in appropriate circumstances they form macroscopic textures — textures whose length scales are set by the size of system. The ubiquitous existence of such textures reflects the fact that the broken symmetries involved are spatial as opposed to internal symmetries. The nature of the textures depends sensitively on the nature of the broken symmetries. For theory at the level of the effective Hamiltonian studied in the present paper, there is full $U(1)$ rotational symmetry, so the nematic is an XY order parameter. However, in the GaAs heterostructures in which the 2DEG is realized, corrections to the effective mass approximation break that symmetry down to a $C_4$ discrete subgroup. These corrections are small in proportion to $(a/\ell_B)^4$ (see Appendix \[app:anisotropy\]), so they are unimportant for most purposes, but at long distances in the broken symmetry phase they necessarily play an essential role, and correspondingly any macroscopic textures should be characteristic of an Ising order parameter.
In the present context, we note that generically one particular order parameter orientation is favored at any edge of the system. So long as this preference is sufficiently strong, it must be satisfied in the equilibrium state, which forces the order parameter to be spatially varying in the bulk. In the case of an Ising order parameter and a square sample in the presence of a weak symmetry breaking field, or more generally in any rectangular sample, this forces the existence of two domain walls which split the sample into three macroscopic domains running along one direction. The nature of the states associated with the domain wall is, itself, an interesting problem whose solution depends on a number of microscopic considerations; however, typically such domain walls are associated with a dramatically reduced gap to current carrying excitations. Where this is the case, the existence of such textures below $T_\text{nem}$ could account for the opposite $T$ dependence of $\rho_{xx}$ and the suddenness of its onset, even in the presence of a symmetry breaking field. Clearly, further theoretical and experimental work is necessary to test this proposal.
Assuming that what is observed in these experiments is, indeed, an FQH nematic, there are a number of other interesting quantities that could be measured. Most importantly from the present viewpoint, the GMP mode, which can be observed as a sharp mode in Raman [@golkar2016] or acoustic wave [@yang2016] spectroscopy, should soften as the transition is approached from the isotropic side. Observing this mode would confirm its identification as the soft-mode associated with a proximate nematic state; more generally, it would be interesting to observe as the condensed matter version of a graviton.
Conclusion
==========
In conclusion, we have explored the possibility of engineering quantum phase transitions out of the isotropic FQH liquid towards spatially ordered phases in a microscopic model of an interacting 2DEG in a perpendicular magnetic field. Combining a variational study with ED results for up to $N=13$ electrons, we arrived at the schematic phase diagram illustrated in Fig. \[fig:fig1\]. Phase A was seen to be adiabatically connected to the isotropic $\nu=1/3$ Laughlin state, and thus represents the region of stability of the isotropic FQH liquid. In phase B, numerical evidence suggests that this phase is adiabatically connected to the $V_3/V_1\rightarrow\infty$ ground state. The many-body gap of the latter was seen to be strongly dependent on system size and aspect ratio of the torus on which our simulations were defined, which suggests that phase B would have spontaneously broken spatial symmetries in the thermodynamic limit. Finally, a variational analysis with anisotropic Laughlin trial wave functions combined with the study of a suitably defined nematic susceptibility led us to propose that phase C could be a FQH nematic in the thermodynamic limit. General arguments suggest that a nematic should be proximate to a phase with translational order: we were thus led to conjecture that phase B could be a smectic/stripe phase, which melts into a nematic at the B-C phase boundary.
Of course, the main limitation of our study is the neglect of quantum fluctuations in our variational analysis as well as the presence of finite-size effects (especially in phases B and C). We hope nonetheless that our work will stimulate further numerical studies of the model Hamiltonian considered here. In particular, it would be interesting to study this model and test the conjectures presented here using algorithms recently developed to apply the density-matrix renormalization group to FQH systems [@zaletel2012; @zaletel2013; @zaletel2015], which might reduce finite-size effects.
Finally we note that weak disorder will destroy both the nematic and the stripe phase in $d=2$ but the nematic will have a quadratically longer correlation length in this limit and exponentially longer if the spatial anisotropy is taken into account. So the practical effect of weak disorder is to stabilize the nematic relative to the stripe phase [^1].
We wish to acknowledge helpful discussions with E. Fradkin and Z. Papić. N. R. was supported by the Packard Foundation, MURI-130-6082, a Keck grant and the Princeton Global Scholarship. J. M. was supported by NSERC grant \#RGPIN-2014-4608, the CRC Program, CIFAR, and the University of Alberta. S. A. K. was supported by NSF-DMR via Grant No. 1265593, and S. L. S. was supported by NSF-DMR via Grant No. 1311781.
Variational wave functions for the FQH isotropic-to-nematic transition {#app:trialWF}
======================================================================
In Sec. \[sec:variational\] we use the anisotropic Laughlin states $\Psi_L^{1/3}(\gamma)$ at filling factor $\nu=1/3$ constructed in Ref. [@qiu2012] as a continuous family of trial ground-state wave functions for the FQH isotropic-to-nematic transition. These anisotropic Laughlin states are many-body wave functions at filling factor $1/q$ in the LLL, $$\begin{aligned}
\label{AnisLaughlin}
\Psi_L^{1/q}(Q,\bar{Q})=f_{Q,\bar{Q}}^{1/q}(\{z_i\}) e^{-\sum_i|z_i|^2/2},\end{aligned}$$ where $f_{Q,\bar{Q}}^{1/q}(\{z_i\})$ is a holomorphic function of the dimensionless complex electron coordinates $z_i=(x_i+iy_i)/\sqrt{2}\ell_B$ given by [@qiu2012] $$\begin{aligned}
f_{Q,\bar{Q}}^{1/q}(\{z_i\})=&\prod_i\sqrt{\frac{\lambda}{2\pi}}e^{-\gamma z_i^2/2}\nonumber\\
&\times\prod_{i<j}\left[z_i-z_j+z_0^2\left(\frac{\partial}{\partial z_i}-\frac{\partial}{\partial z_j}\right)\right]^q 1,\end{aligned}$$ where $\lambda$, $\gamma$, and $z_0^2$ are dimensionless parameters. Interpreting Haldane’s intrinsic metric as a nematic order parameter via Eq. (\[gab\]), we can relate these dimensionless parameters to our complex nematic variational parameter $|Q|e^{i\phi}$ in the following way, $$\begin{aligned}
\lambda&={\mathop{\mathrm{sech}}}(|Q|/2),\\
\gamma&=e^{-i\phi}\tanh (|Q|/2),\\
z_0^2&=\frac{1}{2}e^{i\phi}\sinh |Q|.\end{aligned}$$ When the amplitude of nematic order vanishes $|Q|=0$, we have $\lambda=1$, $\gamma=0$, and $z_0^2=0$, and $\Psi_L^{1/q}(Q,\bar{Q})$ reduces to the usual isotropic Laughlin wave function [@laughlin1983].
Ising anisotropy from finite-size effects {#app:finitesize}
=========================================
An important consequence of studying an isotropic-to-nematic transition on a torus of finite dimensions $L_x\times L_y$ is the explicit breaking of the continuous $SO(2)$ rotation symmetry to a discrete $C_4$ rotation symmetry. This has the effect of reducing the continuous degeneracy of broken-symmetry nematic states to a discrete two-fold degeneracy. In this Appendix, we derive an effective Hamiltonian \[Eq. (\[Heff\])\] for this 2D ground state manifold. For $L_x=L_y$ and in the absence of tunneling effects the system is effectively an Ising nematic with two degenerate ground states related by a 90$^\circ$ rotation of the nematic director. As we will see, a change of aspect ratio $L_x\neq L_y$ splits the degenerate ground states by a “Zeeman” energy $\propto V^{-1/2}\delta$ for small $\delta$ where $L_y/L_x=1+\delta$ and $V=L_xL_y$ is the system volume, and tunneling effects also induce a splitting $\propto\exp(-\textrm{const.}\times V^{1/4})$.
Landau-Ginzburg theory of the isotropic-to-nematic transition
-------------------------------------------------------------
Assuming we are close enough to the transition, we can ignore the coupling to the topological degrees of freedom of the FQH liquid, i.e., the Chern-Simons gauge fields and Laughlin quasiparticles [@maciejko2013], and we focus on the nematic sector alone. As mentioned previously, the nematic order parameter can be expressed as a complex scalar $Q=|Q|e^{i\phi}$, where $0\leq\phi<2\pi$ corresponds to physically distinct configurations. We denote the complex conjugate by $\bar{Q}=Q^*$, and use a complex notation for spatial derivatives, $\partial=\partial_x+i\partial_y$ and $\bar{\partial}=\partial_x-i\partial_y$. Under an $SO(2)$ spatial rotation by an angle $\theta$, the fields transform as $$\begin{aligned}
\label{rotation}
Q\rightarrow e^{2i\theta}Q,\,
\bar{Q}\rightarrow e^{-2i\theta}\bar{Q},\,
\partial\rightarrow e^{i\theta}\partial,\,
\bar{\partial}\rightarrow e^{-i\theta}\bar{\partial},\end{aligned}$$ so an $SO(2)$ invariant Lagrangian should be invariant under this $U(1)$ transformation. In our earlier work [@maciejko2013], the nematic sector was described by a Lagrangian of the form $$\begin{aligned}
\label{L0}
\mathcal{L}_0=i\lambda \bar{Q}\partial_tQ-\mathcal{H}_0,\end{aligned}$$ with a Hamiltonian $$\begin{aligned}
\label{H_0}
\mathcal{H}_0=\kappa\partial\bar{Q}\bar{\partial}Q+r\bar{Q}Q+u\bar{Q}\bar{Q}QQ,\end{aligned}$$ which is manifestly invariant under the transformation (\[rotation\]). This Landau-Ginzburg theory is meant to be an expansion in powers of $Q,\bar{Q}$ and its derivatives. At this order in the expansion, the Lagrangian (\[L0\]) has the same form as that for an internal degree of freedom (e.g., an XY model), which is reflected in the fact that the Goldstone mode has an isotropic dispersion $\varepsilon({{\boldsymbol{q}}})\propto|{{\boldsymbol{q}}}|$. In the nematic language, this means the Frank free energy [@ChaikinLubensky] has equal splay and bend coefficients, $$\begin{aligned}
\label{FrankFreeEnergy}
\partial\bar{Q}\bar{\partial}Q\propto\left\{(\nabla\cdot{{\boldsymbol{n}}})^2+[{{\boldsymbol{n}}}\times(\nabla\times{{\boldsymbol{n}}})]^2
\right\},\end{aligned}$$ where ${{\boldsymbol{n}}}=(\cos(\phi/2),\sin(\phi/2))$ is the nematic director, which is related to the nematic order parameter by $Q_{ab}=|Q|(2n_an_b-\delta_{ab})$. Note that there is no twist in 2D, ${{\boldsymbol{n}}}\cdot(\nabla\times{{\boldsymbol{n}}})=0$. In writing Eq. (\[FrankFreeEnergy\]) we assumed a uniform magnitude $|Q|$ of the nematic order parameter, which is appropriate for a nonlinear sigma model description of the ordered phase.
To see that we are breaking a spatial symmetry and not just an internal symmetry, we need to go to higher orders in the Landau-Ginzburg expansion. The simplest terms that do this (and give different splay and bend coefficients) are cubic terms of the form $Q\nabla Q\nabla Q$ [@lubensky1970]. In our complex notation, four such terms are allowed, $$\begin{aligned}
\mathcal{L}_\textrm{cubic}&=&
c_1\left[Q(\partial\bar{Q})^2+\mathrm{c.c.}\right]
+ic_2\left[Q(\partial\bar{Q})^2-\mathrm{c.c.}\right]\nonumber\\
&&+c_3(Q\bar{\partial}Q\bar{\partial}\bar{Q}+\mathrm{c.c.})
+ic_4(Q\bar{\partial}Q\bar{\partial}\bar{Q}-\mathrm{c.c.}),\nonumber\\\end{aligned}$$ where $c_1,c_2,c_3,c_4$ are real coefficients. We now imagine that we are in the nematic phase $\langle Q\rangle=Q_0e^{i\phi_0}\neq 0$ and wish to examine the effects of fluctuations. We write $Q({{\boldsymbol{r}}},t)=(Q_0+\chi({{\boldsymbol{r}}},t))e^{i(\phi_0+\phi({{\boldsymbol{r}}},t))}$ and expand in powers of the fluctuations $\chi({{\boldsymbol{r}}},t)\ll Q_0$, $\phi({{\boldsymbol{r}}},t)\ll 2\pi$ and their derivatives. The amplitude fluctuations $\chi$ are massive and can be integrated out. The imaginary-time Lagrangian for the Goldstone mode $\phi$ is of the form $$\begin{aligned}
\label{Lag}
\mathcal{L}(\phi)&=&\frac{1}{2}\bigl\{(\partial_\tau\phi)^2+v^2(\nabla\phi)^2+\lambda(\phi_0)[(\partial_x\phi)^2-(\partial_y\phi)^2]\nonumber\\
&&+2\eta(\phi_0)\partial_x\phi\partial_y\phi\bigr\},\end{aligned}$$ where $v^2\propto Q_0^2$, and $$\begin{aligned}
\lambda(\phi_0)&=A\cos\phi_0-B\sin\phi_0,\\
\eta(\phi_0)&=B\cos\phi_0+A\sin\phi_0,\end{aligned}$$ with $A,B\propto Q_0^3$, hence close to the transition we have $Q_0\ll 1$ and thus $\lambda,\eta\ll v^2$. The Goldstone mode dispersion is $$\begin{aligned}
\label{GoldstoneEq}
\varepsilon({{\boldsymbol{q}}})=\sqrt{v^2{{\boldsymbol{q}}}^2+\lambda(\phi_0)(q_x^2-q_y^2)+2\eta(\phi_0)q_xq_y},\end{aligned}$$ which is clearly anisotropic.
Finite-size corrections to the ground-state energy
--------------------------------------------------
We will now use the approach of Fisher [@fisher1989] to calculate the finite-size corrections to the ground-state energy. We consider putting the system on a torus of dimensions $L_x\times L_y$. We define $$\begin{aligned}
\label{DeltaE0}
\Delta E_0(L_x,L_y)=E_0(L_x,L_y)-E_0(L_x=\infty,L_y=\infty),\end{aligned}$$ where $E_0$ is the ground-state energy. From Eq. (\[Lag\]), we have $$\begin{aligned}
E_0(L_x,L_y)=\frac{1}{2}\int_{-\infty}^\infty\frac{d\omega}{2\pi}\sum_{q_xq_y}\ln\left(\omega^2+v^2{{\boldsymbol{q}}}^2+\lambda(\phi_0)(q_x^2-q_y^2)+2\eta(\phi_0)q_xq_y\right).\end{aligned}$$ In the limit $L_x,L_y\rightarrow\infty$ we have $\sum_{q_xq_y}\rightarrow L_xL_y\int\frac{d^2q}{(2\pi)^2}$ where the integral ranges over the entire momentum-space plane $\mathbb{R}^2$, hence Eq. (\[DeltaE0\]) becomes $$\begin{aligned}
\Delta E_0(L_x,L_y)&=\frac{1}{2}\int_{-\infty}^\infty\frac{d\omega}{2\pi}\left(\sum_{q_xq_y}-L_xL_y\int\frac{d^2q}{(2\pi)^2}\right)\ln\left(\omega^2+v^2{{\boldsymbol{q}}}^2+\lambda(\phi_0)(q_x^2-q_y^2)+2\eta(\phi_0)q_xq_y\right)\\
&=\frac{1}{2}\left(\sum_{q_xq_y}-L_xL_y\int\frac{d^2q}{(2\pi)^2}\right)\varepsilon({{\boldsymbol{q}}}),\end{aligned}$$ with $\varepsilon({{\boldsymbol{q}}})$ given in Eq. (\[GoldstoneEq\]). Using the Poisson summation formula, we can show that [@fisher1989] $$\begin{aligned}
\sum_{q_xq_y}\varepsilon({{\boldsymbol{q}}})=\sum_{q_x\in\frac{2\pi\mathbb{Z}}{L_x}}\sum_{q_y\in\frac{2\pi\mathbb{Z}}{L_y}}\varepsilon(q_x,q_y)
=\sum_{\ell_x,\ell_y\in\mathbb{Z}}\int_{-\infty}^\infty d\theta_x\int_{-\infty}^\infty d\theta_y\,e^{2\pi i(\ell_x\theta_x+\ell_y\theta_y)}
\varepsilon\left(\frac{2\pi\theta_x}{L_x},\frac{2\pi\theta_y}{L_y}\right),\end{aligned}$$ whereas $$\begin{aligned}
L_xL_y\int\frac{d^2q}{(2\pi)^2}\varepsilon({{\boldsymbol{q}}})=\int_{-\infty}^\infty d\theta_x\int_{-\infty}^\infty d\theta_y\,
\varepsilon\left(\frac{2\pi\theta_x}{L_x},\frac{2\pi\theta_y}{L_y}\right),\end{aligned}$$ via a simple rescaling $q_x=\frac{2\pi\theta_x}{L_x}$, $q_y=\frac{2\pi\theta_x}{L_y}$. Therefore $$\begin{aligned}
\label{DeltaE0sum}
\Delta E_0(L_x,L_y)&=\frac{1}{2}\sum_{\ell_x\neq 0}\sum_{\ell_y\neq 0}\int_{-\infty}^\infty d\theta_x\int_{-\infty}^\infty d\theta_y\,e^{2\pi i(\ell_x\theta_x+\ell_y\theta_y)}
\varepsilon\left(\frac{2\pi\theta_x}{L_x},\frac{2\pi\theta_y}{L_y}\right)\nonumber\\
&=\frac{1}{2}L_xL_y\sum_{\ell_x\neq 0}\sum_{\ell_y\neq 0}\tilde{\varepsilon}(\ell_xL_x,\ell_yL_y),\end{aligned}$$ where $\tilde{\varepsilon}(x,y)$ is the Fourier transform of $\varepsilon({{\boldsymbol{q}}})$ defined in Eq. (\[GoldstoneEq\]). The Fourier integral is difficult to evaluate exactly, but we can expand in powers of the anisotropy $\frac{\lambda}{v^2},\frac{\eta}{v^2}\ll 1$ if we are close enough to the transition. To first order in these quantities, we have $$\begin{aligned}
\label{EpsilonFirstOrder}
\tilde{\varepsilon}(r,\varphi)=v\int_0^\infty\frac{dq\,q^2}{2\pi}\int_0^{2\pi}\frac{d\theta}{2\pi}\,e^{iqr\cos(\theta-\varphi)}\left(1+\frac{\lambda(\phi_0)}{2v^2}\cos 2\theta+\frac{\eta(\phi_0)}{2v^2}\sin 2\theta\right)+\mathcal{O}(\lambda^2,\eta^2,\lambda\eta),\end{aligned}$$ where ${{\boldsymbol{q}}}=q(\cos\theta,\sin\theta)$ and ${{\boldsymbol{r}}}=r(\cos\varphi,\sin\varphi)$. Performing the integration over $\theta$ and $q$, we obtain $$\begin{aligned}
\label{epsrphi}
\tilde{\varepsilon}(r,\varphi)=-\frac{v}{2\pi r^3}\left(1+\frac{3\lambda(\phi_0)}{2v^2}\cos 2\varphi+\frac{3\eta(\phi_0)}{2v^2}\sin 2\varphi\right),\end{aligned}$$ hence, from Eq. (\[DeltaE0sum\]), we have $$\begin{aligned}
\label{DE0IsAnis}
\Delta E_0(L_x,L_y)=\Delta E_0^\textrm{is}+\Delta E_0^\textrm{anis}(\phi_0),\end{aligned}$$ where the isotropic contribution $\Delta E_0^\textrm{is}$ comes from the $\phi_0$-independent term in Eq. (\[epsrphi\]), and the anisotropic contribution $\Delta E_0^\textrm{anis}(\phi_0)$ is $$\begin{aligned}
\Delta E_0^\textrm{anis}(\phi_0)=-\frac{3L_xL_y}{8\pi v}\sum_{\ell_x\neq 0}\sum_{\ell_y\neq 0}\frac{\lambda(\phi_0)(\ell_x^2L_x^2-\ell_y^2L_y^2)+2\eta(\phi_0)\ell_x\ell_yL_xL_y}{(\ell_x^2L_x^2+\ell_y^2L_y^2)^{5/2}}.\end{aligned}$$ We change variables from $L_x$ and $L_y$ to the total volume $V=L_xL_y$ and aspect ratio $R=L_y/L_x$. The term proportional to $\eta(\phi_0)$ vanishes because the summand is odd in $\ell_x$ or $\ell_y$. We have $$\begin{aligned}
\Delta E_0^\textrm{anis}(\phi_0,V,R)=-\frac{3\lambda(\phi_0)}{8\pi v\sqrt{V}}f(R),\end{aligned}$$ where $$\begin{aligned}
f(R)=R^{3/2}\sum_{\ell_x\neq 0}\sum_{\ell_y\neq 0}\frac{\ell_x^2-R^2\ell_y^2}{(\ell_x^2+R^2\ell_y^2)^{5/2}}.\end{aligned}$$ This function can be evaluated numerically. Alternatively, for aspect ratios close to one, $R=1+\delta$ with $\delta\ll 1$, we can expand for small $\delta$. We find $$\begin{aligned}
f(R)=\frac{\delta}{2}\sum_{\ell_x\neq 0}\sum_{\ell_y\neq 0}\frac{3\ell_x^4-14\ell_x^2\ell_y^2+3\ell_y^4}{(\ell_x^2+\ell_y^2)^{7/2}}+\mathcal{O}(\delta^2)
\simeq -1.37\delta+\mathcal{O}(\delta^2),\end{aligned}$$ hence we obtain $$\begin{aligned}
\Delta E_0^\textrm{anis}(\phi_0,V,\delta)\simeq\textrm{const.}\times\frac{\lambda(\phi_0)\delta}{\sqrt{V}},\end{aligned}$$ for $\delta\ll 1$. However, there should be a four-fold anisotropy for a finite-size system when $\delta=0$. The expansion to first order in $\lambda,\eta$ in Eq. (\[EpsilonFirstOrder\]) was insufficient. We go back and expand to second order. The correction to $\tilde{\varepsilon}({{\boldsymbol{r}}})$ at this order is $$\begin{aligned}
\delta\tilde{\varepsilon}(r,\varphi)&=-\frac{1}{8v^3}\int_0^\infty\frac{dq\,q^2}{2\pi}\int_0^{2\pi}\frac{d\theta}{2\pi}\,e^{iqr\cos(\theta-\varphi)}\left(
\lambda^2(\phi_0)\cos^22\theta+\eta^2(\phi_0)\sin^22\theta
+2\lambda(\phi_0)\eta(\phi_0)\sin 2\theta\cos 2\theta\right)
\nonumber\\
&=\frac{1}{32\pi v^3r^3}\left\{\eta^2(\phi_0)+\lambda^2(\phi_0)
+15\left[(\eta^2(\phi_0)-\lambda^2(\phi_0)]\cos 4\varphi
-2\eta(\phi_0)\lambda(\phi_0)\sin 4\varphi\right]\right\}.\end{aligned}$$ The term proportional to $\eta(\phi_0)\lambda(\phi_0)$ is odd in $\ell_x$ or $\ell_y$ and vanishes upon summation over $\ell_x,\ell_y$. We find that the dependence on the aspect ratio parameter $\delta$ of the contribution of $\delta\tilde{\varepsilon}({{\boldsymbol{r}}})$ to $\Delta E_0(L_x,L_y)$ begins at order $\delta^2$ only, so we can set $\delta=0$ in the resulting expressions. To leading (zeroth) order in $\delta$ therefore, we find $$\begin{aligned}
\delta\Delta E_0(\phi_0,V,\delta)&=\frac{1}{64\pi v^3\sqrt{V}}\left\{\left[\eta^2(\phi_0)+\lambda^2(\phi_0)\right]
\sum_{\ell_x\neq 0}\sum_{\ell_y\neq 0}\frac{1}{(\ell_x^2+\ell_y^2)^{3/2}}
+15\left[\eta^2(\phi_0)-\lambda^2(\phi_0)\right]
\sum_{\ell_x\neq 0}\sum_{\ell_y\neq 0}\frac{\ell_x^4-6\ell_x^2\ell_y^2+\ell_y^4}{(\ell_x^2+\ell_y^2)^{7/2}}\right\}\nonumber\\
&\simeq\frac{1}{64\pi v^3\sqrt{V}}\left\{4.22
\left[\eta^2(\phi_0)+\lambda^2(\phi_0)\right]
-15\times 1.94\left[\eta^2(\phi_0)-\lambda^2(\phi_0)\right]\right\}\nonumber\\
&=\frac{Q_0^3}{\sqrt{V}}\left(c+a\cos 2\phi_0+b\sin 2\phi_0\right),\end{aligned}$$ where $a,b,c$ are independent of $\phi_0,V,\delta$. The term proportional to $c$ contributes to the uninteresting isotropic part $\Delta E_0^\textrm{is}$. Ignoring the isotropic terms, the finite-size correction to the ground-state energy in the nematic phase is therefore $$\begin{aligned}
\label{DeltaE0final}
\Delta E_0^\textrm{anis}(\phi_0,V,\delta)=\frac{Q_0^2}{\sqrt{V}}\left[Q_0(a\cos 2\phi_0+b\sin 2\phi_0)+\delta(a'\cos\phi_0+b'\sin\phi_0)\right]+\mathcal{O}(\delta^2).\end{aligned}$$ The $a,b$ terms describe the finite-size four-fold anisotropy due to the toroidal geometry, while the $a',b'$ terms (proportional to $\delta$) describe the perturbation due to the aspect ratio.
What about terms of cubic and higher order in $\lambda,\eta$ in Eq. (\[EpsilonFirstOrder\])? At order $n$ in the expansion, we will have terms of the form $Q_0^{n+1}\times\{\cos 2n\varphi,\sin 2n\varphi\}\times\{\cos n\phi_0,\sin n\phi_0\}$ summed over the square lattice $(\ell_x,\ell_y)\in\mathbb{Z}^2\backslash\{(0,0)\}$. Due to the symmetries of the lattice certain terms will vanish identically when summed over the entire lattice. The mirror plane at $\ell_y=0$ sends $\varphi\rightarrow-\varphi$, hence $\sin 2n\varphi$ vanishes when summed over for all $n$. The mirror plane at $\ell_x=\ell_y$ sends $\varphi\rightarrow\frac{\pi}{2}-\varphi$, under which $\cos 2n\varphi$ changes sign for odd $n$ and stays unchanged for even $n$. Therefore the terms $\cos 2n\varphi$ with even $n$ survive the sum. In general therefore, we will have $$\begin{aligned}
\Delta E_0^\textrm{anis}(\phi_0,V,\delta)&=\frac{1}{\sqrt{V}}\left[\sum_{n=1}^\infty Q_0^{2n+1}\left(a_n\cos 2n\phi_0+b_n\sin 2n\phi_0\right)+\delta\sum_{n=1}^\infty Q_0^{2n}\left(a'_n\cos(2n-1)\phi_0
+b'_n\sin(2n-1)\phi_0\right)\right]\nonumber\\
&\hspace{5mm}+\mathcal{O}(\delta^2),\end{aligned}$$ where $a_n,b_n,a'_n,b'_n$ are constants. We note that $\Delta E_0^\textrm{anis}(\phi_0,V,0)$ is invariant under $\phi_0\rightarrow\phi_0+\pi$ while $\phi_0$ has periodicity $2\pi$, which implies two degenerate minima for square aspect ratio. Sufficiently close to the critical point where $Q_0\ll 1$, we can neglect terms of higher order in $Q_0$ and work with the simpler potential (\[DeltaE0final\]).
Zero-mode dynamics and effective Hamiltonian
--------------------------------------------
The finite-size correction to the ground-state energy (\[DeltaE0final\]) corresponds to the zero-point energy of the ${{\boldsymbol{q}}}\neq 0$ Goldstone bosons. However, when discussing symmetry breaking in finite systems it is important to consider the effect of the uniform (${{\boldsymbol{q}}}=0$) mode [@anderson1952]. The lowest-energy ${{\boldsymbol{q}}}\neq 0$ fluctuation has a finite-size gap $\propto\frac{1}{\sqrt{V}}$. On the other hand, as will be seen shortly, the uniform fluctuation ${{\boldsymbol{q}}}=0$ has a finite-size gap $\propto\frac{1}{V}$, which is much smaller than $\frac{1}{\sqrt{V}}$ in the thermodynamic limit $V\rightarrow\infty$. Therefore the ${{\boldsymbol{q}}}\neq 0$ fluctuations are very fast compared to the uniform fluctuation, and it is appropriate to use an approximation à la Born-Oppenheimer and think of $\phi_0$ in Eq. (\[DeltaE0final\]) as a dynamical field $\phi_0=\phi_0(\tau)$. The kinetic term in the action for $\phi_0$ is obtained from Eq. (\[Lag\]) with $\phi({{\boldsymbol{r}}},\tau)=\phi_0(\tau)$, $$\begin{aligned}
S_0[\phi_0]=\frac{1}{2}\int_0^\beta d\tau\int d^2r\,(\partial_\tau\phi_0)^2=\frac{V}{2}\int_0^\beta d\tau\,(\partial_\tau\phi_0)^2,\end{aligned}$$ which corresponds to the first-quantized Hamiltonian of an $SO(2)$ rotor, $$\begin{aligned}
H_0(\phi_0)=-\frac{1}{2V}\frac{\partial^2}{\partial\phi_0^2},\end{aligned}$$ where the wave functions $\Psi(\phi_0)$ must obey the periodic boundary conditions $\Psi(\phi_0)=\Psi(\phi_0+2\pi)$. Using the ansatz $\Psi(\phi_0)\propto e^{iN\phi_0}$ where $N\in\mathbb{Z}$, we obtain the “tower-of-states” spectrum $E(N)=\frac{N^2}{2V}$, hence the uniform fluctuation has a finite-size gap $\propto\frac{1}{V}$. The finite-size energy (\[DeltaE0final\]) is then essentially a potential for $\phi_0$, and we consider the Hamiltonian $$\begin{aligned}
\label{Hphi0}
H(\phi_0)=-\frac{1}{2V}\frac{\partial^2}{\partial\phi_0^2}+U(\phi_0),\end{aligned}$$ where $$\begin{aligned}
U(\phi_0)=\frac{Q_0^2}{\sqrt{V}}\bigl[&Q_0(a\cos 2\phi_0+b\sin 2\phi_0)\nonumber\\
&+\delta(a'\cos\phi_0+b'\sin\phi_0)\bigr].\end{aligned}$$ Since $\frac{1}{\sqrt{V}}\gg\frac{1}{V}$ in the thermodynamic limit, we can minimize the potential energy $U(\phi_0)$ first and consider the kinetic term as a perturbation. Since $\delta\ll 1$, we can in fact minimize the $a,b$ terms first and consider the second term in $U(\phi_0)$ as a perturbation. Because $0\leq\phi_0<2\pi$, the $a,b$ terms have two inequivalent minima. Writing $$\begin{aligned}
a\cos 2\phi_0+b\sin 2\phi_0=A\cos 2(\phi_0-f),\end{aligned}$$ with $A^2=a^2+b^2$ (not the same $A$ as before) and $\tan 2f=b/a$, and assuming $A>0$, there are two degenerate minima at $$\begin{aligned}
\phi_0=\frac{\pi}{2}+f,\hspace{5mm}
\phi_0=\frac{3\pi}{2}+f.\end{aligned}$$ These are the two states of our effective Ising nematic, which we can denote by a pseudospin variable $\uparrow,\downarrow$, $$\begin{aligned}
|\uparrow\rangle=\left|\phi_0=\frac{\pi}{2}+f\right\rangle,\hspace{5mm}
|\downarrow\rangle=\left|\phi_0=\frac{3\pi}{2}+f\right\rangle.\end{aligned}$$ Because these are $\phi_0$-eigenstates, the term proportional to $\delta$ in $U(\phi_0)$ is diagonal in the pseudospin basis. Writing $$\begin{aligned}
a'\cos\phi_0+b'\sin\phi_0=B\cos(\phi_0-f'),\end{aligned}$$ where $B^2=(a')^2+(b')^2$ and $\tan 2f'=b'/a'$, the term proportional to $\delta$ becomes a “Zeeman” term for the Ising pseudospin, $$\begin{aligned}
\frac{BQ_0^2\delta}{\sqrt{V}}\cos(\phi_0-f')\rightarrow E_Z\tau_z,\end{aligned}$$ where $\tau_z$ is the third Pauli matrix acting in pseudospin space, and $$\begin{aligned}
E_Z=\frac{BQ_0^2\delta}{\sqrt{V}}\sin(f'-f)=\textrm{const.}\times\frac{Q_0^2\delta}{\sqrt{V}},\end{aligned}$$ is the “Zeeman” energy. Finally, besides the $\tau_z$ term, even at $\delta=0$ the kinetic term will lift the degeneracy between the two pseudospin states because of tunneling effects, which corresponds to a $\tau_x$ term. We can study these effects via a simple instanton calculation that is equivalent to the WKB approximation. Ignoring the Zeeman term, the action corresponding to the Hamiltonian (\[Hphi0\]) is $$\begin{aligned}
S[\phi_0]=\int_0^\infty d\tau\left[\frac{V}{2}\left(\frac{d\phi_0}{d\tau}\right)^2+\frac{AQ_0^3}{\sqrt{V}}\cos 2(\phi_0-f)\right],\end{aligned}$$ at zero temperature. The Euler-Lagrange equation is the sine-Gordon equation, $$\begin{aligned}
V\frac{d^2\phi_0}{d\tau^2}=\frac{dU(\phi_0)}{d\phi_0}.\end{aligned}$$ We search for a kink solution with $\phi_0(\tau=-\infty)=\frac{\pi}{2}+f$ and $\phi_0(\tau=+\infty)=\frac{3\pi}{2}+f$. The solution $$\begin{aligned}
\tilde{\phi}_0(\tau)=\pi+f+2\arctan\left(\tanh\sqrt{\frac{AQ_0^3}{V^{3/2}}}(\tau-\tau_0)\right),\end{aligned}$$ where $\tau_0$ is the (arbitrary) position of the kink, satisfies the equation as well as the boundary conditions at $\tau=\pm\infty$. The action for this instanton is $$\begin{aligned}
S_\textrm{inst}=S[\tilde{\phi}_0]&\simeq\frac{4AQ_0^3}{\sqrt{V}}
\int_{-\infty}^\infty d\tau\,\mathrm{sech}^2\left(2\sqrt{\frac{AQ_0^3}{V^{3/2}}}(\tau-\tau_0)\right)\nonumber\\
&=4\sqrt{A}Q_0^{3/2}V^{1/4},\end{aligned}$$ where we have made the approximation of extending the range of integration from $(0,\infty)$ to $(-\infty,\infty)$. Therefore the final effective Hamiltonian for the 2D ground-state manifold is $$\begin{aligned}
\label{Heff}
H_\textrm{eff}=\Delta\tau_x+E_Z\tau_z,\end{aligned}$$ where $$\begin{aligned}
E_Z\propto\frac{Q_0^2\delta}{\sqrt{V}},\hspace{5mm}
\Delta\propto e^{-\textrm{const.}\times Q_0^{3/2}V^{1/4}}.\end{aligned}$$
Ising anisotropy from corrections to effective mass theory {#app:anisotropy}
==========================================================
Even for an infinite system, due to the underlying crystal structure of the material hosting the 2DEG continuous rotation symmetry will be broken down to a discrete point group symmetry, with the consequence that the Goldstone mode in the FQH nematic phase will acquire a gap. In this Appendix we show that for a lattice with $C_4$ symmetry, the gap $E_\text{gap}$ of the quasi-Goldstone mode scales like $E_\text{gap}\sim(a/\ell_B)^4$, where $a$ is the lattice constant and $\ell_B$ is the magnetic length.
Corrections to effective mass theory
------------------------------------
As discussed in Appendix \[app:finitesize\], nematic order is described by the Landau-Ginzburg Lagrangian (\[L0\]). This description assumes the 2DEG has a perfect continuous $SO(2)$ rotation symmetry (\[rotation\]), which is valid if we consider the microscopic first-quantized Hamiltonian $$\begin{aligned}
H=\sum_{i=1}^N\frac{\boldsymbol{\pi}_i^2}{2m^*}+\sum_{i<j}V({{\boldsymbol{r}}}_i-{{\boldsymbol{r}}}_j),\end{aligned}$$ for a system of $N$ electrons, with $\boldsymbol{\pi}_i={{\boldsymbol{p}}}_i-e{{\boldsymbol{A}}}({{\boldsymbol{r}}}_i)$ and $V({{\boldsymbol{r}}})$ the Coulomb interaction. In general this symmetry will be broken explicitly by lattice effects. At long wavelengths this is captured by corrections to effective mass theory, which for a lattice with $C_4$ symmetry begins at quartic order in the momentum $\boldsymbol{\pi}_i$. (One could also consider anisotropic corrections to the dielectric tensor implicit in $V({{\boldsymbol{r}}})$.)
For a lattice with $C_4$ symmetry the leading corrections to effective mass theory near the $\Gamma$ point are $$\begin{aligned}
\label{DeltaH}
\Delta H=\sum_{i=1}^N\left(\frac{\alpha}{2}\{\pi_{ix}^2,\pi_{iy}^2\}+\beta\left(\pi_{ix}^4+\pi_{iy}^4\right)\right).\end{aligned}$$ The second and first terms could come from nearest-neighbor $t$ and next-nearest-neighbor $t'$ hopping on the square lattice, respectively. The first term has been symmetrized because $\pi_{ix}$ and $\pi_{iy}$ do not commute. Let us assume that $t'\sim t$ for simplicity, and neglect all factors of order one. Then $ta^2\sim \hbar^2/m^*$ and $\alpha,\beta\sim ta^4/\hbar^4$, where $a$ is the lattice constant. Therefore $\alpha,\beta\sim a^2/m^*\hbar^2$.
First-order perturbation theory
-------------------------------
Using the variational wave functions presented in Appendix \[app:trialWF\], we wish to compute perturbatively the form of the symmetry-breaking corrections to the Hamiltonian (\[H\_0\]) that originate from (\[DeltaH\]): $$\begin{aligned}
\label{AnisCorr}
\Delta\mathcal{H}(Q,\bar{Q})=\langle\Psi_L^{1/q}(Q,\bar{Q})|\Delta H|\Psi_L^{1/q}(Q,\bar{Q})\rangle+\ldots,\end{aligned}$$ where the first term corresponds to first-order perturbation theory in $\Delta H$ and $\ldots$ to higher orders in perturbation theory. We assume the trial wave functions $|\Psi_L^{1/q}(Q,\bar{Q})\rangle$ are normalized, $\langle\Psi_L^{1/q}(Q,\bar{Q})|\Psi_L^{1/q}(Q,\bar{Q})\rangle=1$. $\Delta\mathcal{H}(Q,\bar{Q})$ should contain terms that are only invariant under rotations by $\theta=n\pi/2$, $n\in\mathbb{Z}$, corresponding to $Q\rightarrow-Q$, $\bar{Q}\rightarrow-\bar{Q}$ for $n$ odd. These terms break the $U(1)$ symmetry of the nematic to $\mathbb{Z}_2$. Listing the most relevant terms first, we expect to have $$\begin{aligned}
\label{DH_mass}
\Delta\mathcal{H}(Q,\bar{Q})=c_1(Q^2+\bar{Q}^2)+ic_2(Q^2-\bar{Q}^2)+\ldots,\end{aligned}$$ where $c_1,c_2$ are real constants. In the nematic phase, these terms will correspond to a mass term for the transverse Goldstone mode.
Defining the inter-Landau level (LL) annihilation operator $b_i=\ell_B(\pi_{ix}+i\pi_{iy})/\sqrt{2}\hbar$ and its Hermitian conjugate $b_i^\dag$, where $\ell_B=\sqrt{\hbar/eB}$ is the magnetic length, we obtain $$\begin{aligned}
\label{DeltaH2}
\Delta H&=\frac{\hbar^4}{\ell_B^4}\sum_{i=1}^N\biggl[(\alpha+6\beta)b_i^\dag b_i+\left(\frac{\alpha}{2}+3\beta\right)b_i^\dag b_i^\dag b_ib_i\nonumber\\
&\hspace{5mm}+\left(-\frac{\alpha}{4}+\frac{\beta}{2}\right)(b_ib_ib_ib_i+b_i^\dag b_i^\dag b_i^\dag b_i^\dag)\biggr],\end{aligned}$$ plus a constant that we neglect. Note that $b_i,b_i^\dag$ are dimensionless. The expectation value in $\Psi_L^{1/q}(Q,\bar{Q})$ of the first two terms in Eq. (\[DeltaH2\]) vanishes, because $b_i$ acts on a LLL wave function. Very explicitly, consider the action of this operator on the trial wave function. Using the fact that $b_j=\frac{1}{2}z_j+\partial/\partial\bar{z}_j$, we have $$\begin{aligned}
b_j\Psi_L^{1/q}(Q,\bar{Q})=&\left(\frac{1}{2}z_j+\frac{\partial}{\partial\bar{z}_j}\right)f_{Q,\bar{Q}}^{1/q}(\{z_i\})e^{-|z_j|^2}\nonumber\\
&\times\prod_{i\neq j}e^{-|z_i|^2/2}\nonumber\\
&=\frac{1}{2}z_j\Psi_L^{1/q}(Q,\bar{Q})+f_{Q,\bar{Q}}^{1/q}(\{z_i\})e^{-|z_j|^2/2}\nonumber\\
&\times\left(-\frac{1}{2}z_j\right)\prod_{i\neq j}e^{-|z_i|^2/2}\nonumber\\
&=0,\end{aligned}$$ using Eq. (\[AnisLaughlin\]). For the same reason, we have $$\begin{aligned}
b_ib_ib_ib_i\Psi_L^{1/q}(Q,\bar{Q})=0,\end{aligned}$$ and the term with $b_i^\dag b_i^\dag b_i^\dag b_i^\dag$ will involve terms such as $$\begin{aligned}
&\int\prod_i d^2z_i \Psi_L^{1/q}(Q,\bar{Q})^* b_j^\dag b_j^\dag b_j^\dag b_j^\dag \Psi_L^{1/q}(Q,\bar{Q})\nonumber\\
&=\int\prod_i d^2z_i \left(b_jb_jb_jb_j\Psi_L^{1/q}(Q,\bar{Q})\right)^* \Psi_L^{1/q}(Q,\bar{Q})\nonumber\\
&=0.\end{aligned}$$ Thus $\Delta\mathcal{H}(Q,\bar{Q})$ in Eq. (\[AnisCorr\]) vanishes to first order in perturbation theory, and we must go to second order.
 is slightly different but the gross features of the two plots are mostly identical. (b) Gap as a function of the ratio $V_3/V_1$ for $N=13$ electrons and a square aspect ratio. We show both the gap in the zero-momentum sector (red line) and the gap to the first excited with a non zero-momentum (blue line). Note that in order to compare gaps between different values of $V_3/V_1$, we have set the energy scale to be one for the two-particle problem, irrespective of the value of $V_3/V_1$. Inset: Close-up of the gap in the zero-momentum sector around the point $V_3/V_1 \simeq 0.500$ where the gap almost closes ($1.5 \times 10^{-4}$).[]{data-label="fig:IndirectGap"}](v1v3v5_fullgap.pdf){width="\columnwidth"}
Second-order perturbation theory
--------------------------------
The second-order correction would look like $$\begin{aligned}
\Delta\mathcal{H}(Q,\bar{Q})=\sum_{n\neq 0}\frac{\langle\Psi_L^{1/q}(Q,\bar{Q})|\Delta H|n\rangle\langle n|\Delta H|\Psi_L^{1/q}(Q,\bar{Q})\rangle}
{E_n-E_0},\end{aligned}$$ where $|n\rangle$, $n\neq 0$ are many-body excited states and $E_0$ is the energy of the trial wave function. We first observe that the terms $b_i^\dag b_i$, $b_i^\dag b_i^\dag b_i b_i$, and $b_ib_ib_ib_i$ will again give a vanishing contribution to $\Delta\mathcal{H}$ since these terms annihilate the trial wave function and thus make the matrix element $\langle n|\Delta H|\Psi_L^{1/q}(Q,\bar{Q})\rangle$ vanish for all $n$. The only nonvanishing contribution must come from the $b_ib_ib_ib_i$ and $b_i^\dag b_i^\dag b_i^\dag b_i^\dag$ terms, which cause LL mixing. Schematically, $$\begin{aligned}
&\Delta\mathcal{H}(Q,\bar{Q})\sim\left(\frac{\hbar^4}{\ell_B^4}\right)^2\left(\frac{a^2}{m^*\hbar^2}\right)^2\nonumber\\
&\times\sum_{n\neq 0}\frac{\langle\Psi_L^{1/q}(Q,\bar{Q})|bbbb|n\rangle\langle n|b^\dag b^\dag b^\dag b^\dag|\Psi_L^{1/q}(Q,\bar{Q})\rangle}{E_n-E_0},\end{aligned}$$ where we used the fact that $\alpha,\beta\sim a^2/m^*\hbar^2$. Now, $|n\rangle$ cannot be in the LLL because this would again lead to the vanishing of the matrix element $\langle\Psi_L^{1/q}(Q,\bar{Q})|bbbb|n\rangle$. The largest contribution to $\Delta\mathcal{H}$ comes from virtual transitions to excited states in the first LL, such that $E_n-E_0\sim\hbar\omega_c$ where $\omega_c=eB/m^*$ is the cyclotron frequency. Since $b,b^\dag$ are dimensionless, the matrix elements are order one. While determining exactly the dependence on $Q,\bar{Q}$ of $\Delta\mathcal{H}(Q,\bar{Q})$ would require a more detailed analysis, we see no reason why the leading terms in Eq. (\[DH\_mass\]) would not appear. Therefore we expect the magnitude of the coefficients $c_1,c_2$ to be given by $$\begin{aligned}
c_1,c_2\sim\left(\frac{\hbar^4}{\ell_B^4}\right)^2\left(\frac{a^2}{m^*\hbar^2}\right)^2\frac{1}{\hbar\omega_c}\sim\left(\frac{a}{\ell_B}\right)^4\hbar\omega_c,\end{aligned}$$ where we used $\omega_c=eB/m^*$. We thus estimate the gap of the quasi-Goldstone mode in the nematic phase to be $$\begin{aligned}
E_\textrm{gap}\sim Q_0^2\left(\frac{a}{\ell_B}\right)^4\hbar\omega_c,\end{aligned}$$ where $Q_0$ is the (dimensionless) amplitude of the nematic order parameter.
![Gap as a function of the ratio $V_3/V_1$ for $N=12$ electrons and a square aspect ratio. We show both the gap in the zero-momentum sector (red line) and the gap to the first excited with a non zero-momentum (blue line). In order to compare gaps between different values of $V_3/V_1$, we have set the energy scale to be one for the two-particle problem, irrespective of the value of $V_3/V_1$. Inset: Overlap of the ground state with the $\nu=1/3$ Laughlin state (red dashed line) and the $V_3/V_1\rightarrow\infty$ ground state (blue bashed line) as function of the ratio $V_3/V_1$ for $N=12$ electrons and a square aspect ratio.[]{data-label="fig:V1V3N12"}](v1v3_n_12.pdf){width="0.8\columnwidth"}
In Ref. [@xia2011], the $\nu=7/3$ plateau is found at $B\approx 2.8$ T, which using the rule of thumb $\ell_B\approx 25\text{ nm}/\sqrt{B[\text{T}]}$ corresponds to a magnetic length $\ell_B\approx 14.9$ nm. The lattice constant of GaAs is $a\approx 5.65$ Å, which gives $$\begin{aligned}
\left(\frac{a}{\ell_B}\right)^4\approx 2\times 10^{-6},\end{aligned}$$ i.e., a millionth of the cyclotron gap.
Additional numerical results {#app:numerical}
============================
In this Appendix we will provide a more complete numerical survey that might be relevant for the expert readers. In Sec. \[sec:ED\], we have focused on the gap in the zero-momentum sector. We mentioned that looking at the gap to the first excited state (irrespective of its momentum) does not qualitatively change the phase diagram, at least for the system sizes we can reach, i.e., $N=10$ and $N=11$. Indeed, we show the gap to the first excited state for $N=11$ in Fig. \[fig:IndirectGap\](a). It should be compared to Fig. \[fig:fig2\](b) that provides the gap in the zero-momentum sector for this number of fermions. (Performing a similar calculation for $N=13$, including the full $(V_3/V_1, V_5/V_1)$ diagram, would be computationally too demanding.) We have however calculated both the gap in the zero-momentum sector and the indirect gap along the $V_5/V_1=0$ line for this system size \[Fig. \[fig:IndirectGap\](b)\]. Once again, we see that the gap in the zero-momentum sector is sufficient to characterize the different regions.
![Exact diagonalization study of the $V_1$-$V_3$-$V_5$ model with $N=12$ electrons: (a) overlap of the ground state with the $\nu=1/3$ Laughlin state; (b) overlap of the ground state with the $V_3/V_1\rightarrow\infty$ ground state; (c) gap between ground and first excited state in the zero-momentum sector; (d) ground-state nematic susceptibility (\[chiQ\]), capped for readability at $\chi_\textrm{max}=10$.[]{data-label="fig:DataN12"}](v1v3v5_n_12.pdf){width="0.92\columnwidth"}
We have mentioned that the system with $N=12$ electrons was not considered in Sec. \[sec:ED\] due to an enhanced stability of the Laughlin phase most probably related to commensuration effects. Indeed, on the sphere geometry at the shift of the Laughlin $\nu=1/3$ state, the gap never closes when increasing $V_3/V_1$ for $N=12$ electrons. The torus geometry is not biased by the shift choice. Thus we still observe the gap closing but at a larger value of $V_3/V_1$ \[see Fig. \[fig:V1V3N12\], that should be compared to Fig. \[fig:IndirectGap\](b)\]. Despite this greater robustness of the Laughlin phase, the full $(V_3/V_1, V_5/V_1)$ phase diagram for this system size does not exhibit any major qualitative change. Indeed the various quantities (overlaps, gap, and nematic susceptibility) shown in Fig. \[fig:DataN12\] for $N=12$ have large similarities with those of Figs. \[fig:fig6\] and \[fig:fig2\].
[^1]: For example, see Ref. [@nie2014].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This survey summarizes the results discussed in a talk at “Bielefeld Geometry & Topology Days” held at Bielefeld University in July 2015. We are interested in quantitative and qualitative properties of Bott-Chern cohomology. We announce new results obtained in [@angella-tardini-1] jointly with Nicoletta Tardini. In particular, we prove an upper bound of the dimensions of Bott-Chern cohomology in terms of Hodge numbers. We also introduce a notion, called Schweitzer qualitative property, which encodes the existence of a non-degenerate pairing in Bott-Chern cohomology, like the Poincaré duality for the de Rham cohomology. We show that this property characterizes the $\partial\overline\partial$-Lemma.'
address: |
Centro di Ricerca Matematica “Ennio de Giorgi”\
Collegio Puteano, Scuola Normale Superiore\
Piazza dei Cavalieri 3\
56126 Pisa, Italy
author:
- Daniele Angella
title: 'On the Bott-Chern and Aeppli cohomology'
---
[^1]
Introduction {#introduction .unnumbered}
============
We are aimed at decoding some of the information on the complex geometry of a compact complex manifold from the cohomologies associated to its double complex of differential forms. In particular, we focus on the information contained in the Bott-Chern and Aeppli cohomologies. They provide, in a sense, a bridge between the holomorphic contents of the Dolbeault cohomology, and the topological contents of the de Rham cohomology. In this sense, it is expected that they both provide a better control on the holomorphic structure, (see, e.g., [@angella-tomassini-3] and Theorem \[thm:main-thm\],) and furnish natural tools for treating geometric aspects, (see, e.g., [@tosatti-weinkove]). A possible link between these two settings would be provided by a proof of the conjecture that [*compact complex manifolds satisfying the $\partial\overline\partial$-Lemma admit balanced metrics in the sense of Michelsohn*]{}, see [@popovici §6].
In this survey, we focus on quantitative properties of Bott-Chern and Aeppli cohomologies towards the study of their qualitative properties.
More precisely, as for the “quantitative” aspects, we recall a result proven by the author and A. Tomassini in [@angella-tomassini-3]. It states that, once fixed the topological structure, there is a lower bound on the dimension of the Bott-Chern cohomology in terms of the Betti numbers. See [@angella-tomassini-3 Theorem A] or Theorem \[thm:frolicher-bc\] for a precise statement. Moreover, the lower bound is attained if and only if the complex manifold satisfies the $\partial\overline\partial$-Lemma, (that is, the identity induces natural isomorphisms between Bott-Chern, Aeppli, Dolbeault cohomologies). We announce here an upper bound for the Bott-Chern cohomology in terms of Hodge numbers, obtained in joint work with Nicoletta Tardini: further details and complete proofs can be found in [@angella-tardini-1]. More precisely, in Theorem \[thm:upper-bound\], we prove that, on a compact complex manifold $X$ of complex dimension $n$, for any $k\in\mathbb{Z}$, there holds $$\sum_{p+q=k} \dim_\mathbb{C} H^{p,q}_{A}(X) \;\leq\; (n+1) \, \left( \sum_{p+q=k} \dim_\mathbb{C} H^{p,q}_{\overline\partial}(X) + \sum_{p+q=k+1} \dim_\mathbb{C} H^{p,q}_{\overline\partial}(X) \right) \;.$$ Note that a topological upper bound is not possible.
As a consequence, we get that the difference $\sum_{p+q=k} \left( \dim_\mathbb{C} H^{p,q}_{BC}(X) - \dim_\mathbb{C} H^{p,q}_{A}(X) \right)$ is bounded from both above and below by Hodge numbers. Such a quantity yields another characterization of the $\partial\overline\partial$-Lemma: in Theorem \[thm:char-deldelbar-minus\], in joint work with Nicoletta Tardini [@angella-tardini-1], we show that a compact complex manifold $X$ satisfies the $\partial\overline\partial$-Lemma if and only if $$\sum_{k\in\mathbb{Z}} \left| \sum_{p+q=k} \left( \dim_\mathbb{C} H^{p,q}_{BC}(X) - \dim_\mathbb{C} H^{p,q}_{A}(X) \right) \right| \;=\; 0 \;.$$ This has to be compared with the characterization in [@angella-tomassini-3 Theorem B], which uses instead the vanishing of the non-negative degrees $\Delta^k$ in : $$\sum_{k\in\mathbb{Z}} \left( \sum_{p+q=k} \left( \dim_\mathbb{C} H^{p,q}_{BC}(X) + \dim_\mathbb{C} H^{p,q}_{A}(X) \right) - 2\, b_k \right) \;=\; 0 \;,$$ and with [@angella-tomassini-5 Corollary 4.14], which is deduced from interpretation of complex structures as generalized-complex structures in the sense of N. Hitchin. Note also that, by the Schweitzer duality between Bott-Chern cohomology and Aeppli cohomology [@schweitzer §2.c], the condition in Theorem \[thm:char-deldelbar-minus\] can be written just in terms of Bott-Chern cohomology as $$\sum_{k\in\mathbb{Z}} \left| \sum_{p+q=k} \dim_\mathbb{C} H^{p,q}_{BC}(X) - \sum_{p+q=2n-k} \dim_\mathbb{C} H^{p,q}_{BC}(X) \right| \;=\; 0 \;;$$ compare with the Poincaré duality and the Serre duality.
Our final aim would be to attempt the study of “qualitative” aspects, namely, of the concrete realization of the algebra structure in Bott-Chern cohomology. Some previous attempts were done in [@angella-tomassini-6; @tardini-tomassini], with the final motivation of understanding whether a possible notion of formality [*à la*]{} Sullivan for Bott-Chern cohomology makes sense. Roughly speaking, a manifold would be “formal with respect to Bott-Chern cohomology” when the Bott-Chern cohomology functor can be made “concrete” by means of a zigzag of morphisms of bi-differential bi-graded algebras being quasi-isomorphisms with respect to Bott-Chern cohomology. This is the case when there is a suitable choice of the representatives having by themselves a structure of algebra. (A stronger request would be to ask for harmonic representatives with respect to some Hermitian metric.) We introduce here a “qualitative” notion, motivated by the following observation. From the geometric point of view, the algebra structure is mainly important in connection also with the Poincaré duality. But Hermitian duality does not preserve Bott-Chern cohomology. We will introduce a property, called [*Schweitzer qualitative property*]{}, from the work in [@schweitzer], that implies in fact the validity of the $\partial\overline\partial$-Lemma, thanks to the above quantitative results, see [@angella-tardini-1].
[*Acknowledgments.*]{} This note has been written for the Workshop “Bielefeld Geometry & Topology Days” held at Bielefeld University on July 2nd–3rd, 2015. The author warmly thanks the organizer Giovanni Bazzoni for the kind invitation and hospitality. Thanks to him and to all the participants for the enthusiastic and fruitful environment they contributed to. We acknowledge many important discussions with Nicoletta Tardini and Adriano Tomassini on the subject: we thank also Michela Zedda, Giovanni Bazzoni, Oliver Goertsches for interesting conversations and useful suggestions. Thanks to Valentino Tosatti and Jim Stasheff for useful comments that improved the presentation of the paper. The new results that we announce here are obtained in joint work with Nicoletta Tardini, [@angella-tardini-1].
Bott-Chern and Aeppli cohomology
================================
We study the geometry of compact complex manifolds $X$ as encoded in their double complex of forms, namely, $\left(\wedge^{\bullet,\bullet}X,\partial,\overline\partial\right)$. This is an object of $\mathbf{bba}$, that is, the category of bi-differential $\mathbb{Z}^2$-graded algebras. Note that any compact complex manifold admits two natural structures: a real structure; and a non-degenerate pairing structure. These are encoded also in symmetries of the double complex: conjugation yields a symmetry around the bottom-left/top-right diagonal; the duality yields a symmetry around the bottom-right/top-left diagonal.
We keep in mind the case that the double complex is a direct sum of the following two indecomposable objects[^2]:
- [*zigzags*]{} of length $\ell+1$, where $\ell\in\mathbb{N}$ counts the number of arrows; $$\xymatrix{
\cdots & \bullet &&& \\
& \bullet \ar[u]^{\overline\partial} \ar[r]_{\partial} & \bullet & & \\
&& \ddots \ar[u]^{\overline\partial} \ar[r]_{\partial} & \bullet & \\
&& & \bullet \ar[u]^{\overline\partial} & \cdots \\
}$$
- [*squares*]{} of isomorphisms: $$\xymatrix{
\bullet \ar[r]^{\partial}_{\simeq} & \bullet \\
\bullet \ar[u]^{\overline\partial}_{\simeq} \ar[r]_{\partial}^{\simeq} & \bullet \ar[u]_{\overline\partial}^{\simeq}
}$$
In particular, zigzags of length one are called [*dots*]{}. By Hodge theory and elliptic PDE theory, the cohomologies have finite dimension: whence the number of zigzag is finite, and the number of squares is infinite.
For example, the double complex associated to a hypothetical complex structure on the $6$-dimensional sphere $\mathbb{S}^6$ should be as in Figure \[fig:S6\]. This example is constructed by using the results in [@ugarte-sphere] on the Frölicher spectral sequence of a hypothetical complex structure on the six-sphere.
(0,0) grid (4\*[2]{},4\*[2]{});
in [0,...,3]{} at ([2]{}\*.5+[2]{}\*,-.3) ; in [0,...,3]{} at (-.3,[2]{}\*.5+[2]{}\*) ;
\(A) at (0\*[2]{}+1/2\*[2]{}, 0\*[2]{}+1/2\*[2]{}); (B) at (0\*[2]{}+1/4\*[2]{}, 1\*[2]{}+1/4\*[2]{}); (C) at (0\*[2]{}+1/2\*[2]{}, 1\*[2]{}+3/4\*[2]{}); (D) at (0\*[2]{}+1/4\*[2]{}, 2\*[2]{}+1/2\*[2]{}); (E) at (1\*[2]{}+1/2\*[2]{}, 0\*[2]{}+3/4\*[2]{}); (F) at (1\*[2]{}+1/4\*[2]{}, 0\*[2]{}+1/4\*[2]{}); (G) at (1\*[2]{}+1/4\*[2]{}, 1\*[2]{}+1/4\*[2]{}); (H) at (1\*[2]{}+3/4\*[2]{}, 1\*[2]{}+1/4\*[2]{}); (I) at (1\*[2]{}+1/2\*[2]{}, 1\*[2]{}+3/4\*[2]{}); (L) at (1\*[2]{}+1/4\*[2]{}, 2\*[2]{}+1/2\*[2]{}); (M) at (1\*[2]{}+3/4\*[2]{}, 2\*[2]{}+1/2\*[2]{}); (N) at (2\*[2]{}+1/2\*[2]{}, 0\*[2]{}+3/4\*[2]{}); (O) at (2\*[2]{}+1/4\*[2]{}, 0\*[2]{}+1/4\*[2]{}); (P) at (2\*[2]{}+3/4\*[2]{}, 1\*[2]{}+1/4\*[2]{}); (Q) at (2\*[2]{}+3/4\*[2]{}, 1\*[2]{}+3/4\*[2]{}); (R) at (2\*[2]{}+3/4\*[2]{}, 2\*[2]{}+1/2\*[2]{}); (S) at (3\*[2]{}+3/4\*[2]{}, 1\*[2]{}+3/4\*[2]{});
\(A) circle ([1\*[2]{}pt]{}); (B) circle ([1\*[2]{}pt]{}); (C) circle ([1\*[2]{}pt]{}); (D) circle ([1\*[2]{}pt]{}); (E) circle ([1\*[2]{}pt]{}); (H) circle ([1\*[2]{}pt]{}); (I) circle ([1\*[2]{}pt]{}); (L) circle ([1\*[2]{}pt]{}); (M) circle ([1\*[2]{}pt]{}); (N) circle ([1\*[2]{}pt]{}); (O) circle ([1\*[2]{}pt]{}); (P) circle ([1\*[2]{}pt]{}); (S) circle ([1\*[2]{}pt]{});
\(B) – (G) – (F) – (O); (C) – (I); (D) – (L); (E) – (N); (H) – (P); (M) – (R) – (Q) – (S);
@size[6]{}@mathfonts at (0\*[2]{}+1/2\*[2]{}, 0\*[2]{}+1/2\*[2]{}-.3) [$1$]{}; at (0\*[2]{}+1/4\*[2]{}, 1\*[2]{}+1/4\*[2]{}-.3) [$\alpha$]{}; at (0\*[2]{}+1/2\*[2]{}+.2, 1\*[2]{}+3/4\*[2]{}-.3) [$h^{0,2}+1-\alpha$]{}; at (0\*[2]{}+1/4\*[2]{}+.1, 2\*[2]{}+1/2\*[2]{}-.3) [$h^{0,2}-\beta$]{}; at (1\*[2]{}+1/2\*[2]{}+.1, 0\*[2]{}+3/4\*[2]{}-.3) [$h^{1,0}$]{}; at (1\*[2]{}+3/4\*[2]{}+.6, 1\*[2]{}+1/4\*[2]{}-.3) [$h^{1,1}-h^{0,2}+\alpha-1$]{}; at (1\*[2]{}+3/4\*[2]{}, 2\*[2]{}+1/2\*[2]{}-.3) [$\beta$]{};
\[fig:S6\]
Several cohomologies can be defined associated to a double complex. The first ones that appear are the de Rham cohomology, $$H_{dR}(X;\mathbb{C}) \;=\; \frac{\ker (\partial+\overline\partial)}{\mathrm{imm} (\partial+\overline\partial)} \circ \mathrm{Tot} \;,$$ (here, $\mathrm{Tot}\colon \mathbf{bba}\to\mathbf{dga}$ denotes the totalization functor, $$\mathrm{Tot}(A^{\bullet,\bullet},\partial,\overline\partial) \;:=\; \left(\bigoplus_{p+q=\bullet}A^{p,q},\partial+\overline\partial\right) \;,$$ the category $\mathbf{dga}$ having differential $\mathbb{Z}$-graded algebras as objects,) and the Dolbeault cohomology and its conjugate, $$H_{\overline\partial}(X) \;=\; \frac{\ker \overline\partial}{\mathrm{imm} \overline\partial} \qquad \text{ and } \qquad H_{\partial}(X) \;=\; \frac{\ker \partial}{\mathrm{imm} \partial} \;.$$ In the model we keep in mind, constituted by squares and zigzags, the Dolbeault cohomology is easily computed by erasing vertical arrows with their ends from the diagram, and counting the remaining points.
Notice that Dolbeault and de Rham cohomology does not suffice, in general, for detecting the complete structure of the double complex. For example, zigzags of odd length do not contribute to the difference between Dolbeault and de Rham cohomology; in other words, to the higher terms in the Frölicher spectral sequence. This means that symmetric zigzags of odd length cannot be detected. For example, the following diagrams have the same de Rham and Dolbeault cohomologies[^3]:
(0,0) grid (4,4);
in [0,...,3]{} at (.5+,-.3) ; in [0,...,3]{} at (-.3,.5+) ;
\(a) at (1.5,2.5); (b) at (2.5,1.5);
\(a) circle (2.5pt); (b) circle (2.5pt);
(0,0) grid (4,4);
in [0,...,3]{} at (.5+,-.3) ; in [0,...,3]{} at (-.3,.5+) ;
(a1) at (1.5+.2,2.5+.2); (b1) at (2.5+.2,1.5+.2); (a2) at (1.5-.2,2.5-.2); (b2) at (2.5-.2,1.5-.2); (c) at (2.5+.2,2.5+.2); (d) at (1.5-.2,1.5-.2);
(a1) circle (2.5pt); (b1) circle (2.5pt); (a2) circle (2.5pt); (b2) circle (2.5pt); (c) circle (2.5pt); (d) circle (2.5pt);
(a1) – (c); (b1) – (c); (a2) – (d); (b2) – (d);
The last discussion serves us as a motivation for introducing Bott-Chern and Aeppli cohomologies. Indeed, the above diagrams differ as to the number of corners; whence we get the need for having an invariant that counts the corners. In [@bott-chern; @aeppli], [*Bott-Chern cohomology*]{} and [*Aeppli cohomology*]{} are defined as $$H_{BC}^{\bullet,\bullet}(X) \;=\; \frac{\ker\partial\cap\ker\overline\partial}{\mathrm{imm}\partial\overline\partial}
\qquad \text{ and } \qquad
H_{A}^{\bullet,\bullet}(X) \;=\; \frac{\ker\partial\overline\partial}{\mathrm{imm}\partial+\mathrm{imm}\overline\partial} \;.$$ As for the Bott-Chern cohomology, it is easy to recognize that it counts the corners possibly having ingoing arrows, except for the squares[^4]: $$\xymatrix{
\circ \ar@{-->}[r] & \bullet \\
&\circ \ar@{-->}[u]
}$$ Dually, the Aeppli cohomology counts the corners possibly having outgoing arrows, except for the squares: $$\xymatrix{
\circ & \\
\bullet \ar@{-->}[r] \ar@{-->}[u] & \circ
}$$
Finally, we have the following diagram, where the maps are morphisms of either $\mathbb{Z}$-graded or $\mathbb{Z}^2$-graded algebras naturally induced by the identity: $$\xymatrix{
& H^{\bullet,\bullet}_{BC}(X) \ar[d]\ar[ld]\ar[rd] & \\
H^{\bullet,\bullet}_{\partial}(X) \ar[rd] \ar@{=>}[r] & H^{\bullet}_{dR}(X;\mathbb{C}) \ar[d] & H^{\bullet,\bullet}_{\overline\partial}(X) \ar[ld] \ar@{=>}[l] \\
& {\phantom{\;.}} H^{\bullet,\bullet}_{A}(X) \;. &
}$$ Here, the link between de Rham and (conjugate) Dolbeault cohomology is the Frölicher spectral sequence, naturally arising from the structure of double-complex.
A compact complex manifold $X$ is said to satisfy the [*$\partial\overline\partial$-Lemma*]{} if the natural map $H^{\bullet,\bullet}_{BC}(X)\to H^{\bullet,\bullet}_{A}(X)$ induced by the identity is injective. This is equivalent to all the maps in the above diagram being isomorphisms, see [@deligne-griffiths-morgan-sullivan Lemma 5.15]. This is also equivalent to asking that the double complex associated to $X$ be a direct sum of squares and dots, [@deligne-griffiths-morgan-sullivan Proposition 5.17].
Quantitative properties of Bott-Chern cohomology
================================================
In this section, we are interested in determining [*quantitative*]{} relations between the dimensions of the above cohomologies. This is ultimately related to what sequences of integers can appear as dimensions of cohomologies of double complexes of compact complex manifolds. The known restrictions on such sequences are essentially:
- restrictions arising from dimension, compactness, and connectedness;
- symmetries arising from the real structure and the non-degenerate pairing structure;
- inequalities of algebraic type as in [@frolicher Theorem 2], [@angella-tomassini-3 Theorem A], and Theorem \[thm:upper-bound\];
- inequalities of analytic type, as the ones that hold for compact complex surfaces, see [@angella-dloussky-tomassini] and subsequent work.
A first result in this direction is the [*Frölicher inequality*]{} in [@frolicher Theorem 2]. It states that the structure of double complex yields a spectral sequence of the form $$H^{\bullet,\bullet}_{\overline\partial} \Rightarrow H^\bullet_{dR}(X;\mathbb{C}) \;,$$ whence the inequality $$\text{for any }k\in\mathbb{Z}\;, \qquad h^{k}_{\overline\partial} - b_k \;\geq\; 0 \;.$$ (As a matter of notation, we write, e.g., $h^k_{\overline\partial}:=\sum_{p+q=k}\dim_{\mathbb{C}} H^{p,q}_{\overline\partial}(X)$; moreover, $b_k$ denotes the $k$th Betti number of $X$.)
The discussion in the previous paragraph motivates a similar inequality for the Bott-Chern cohomology. Indeed, recall that Dolbeault cohomology does not care horizontal arrows, conjugate Dolbeault cohomology does not care vertical arrows, Bott-Chern cohomology counts possibly incoming corners, Aeppli cohomology counts possibly outgoing corners, with the exception, in any case, of squares. Hence, just by combinatorial arguments, one recognizes that the sum of the dimension of the Bott-Chern and Aeppli cohomologies is greater than or equal than the sum of the dimension of Dolbeault and conjugate Dolbeault cohomologies, which is greater than or equal than twice the Betti number. Moreover, both equalities hold if and only if the double complex is direct sum of squares and dots, that is, if the manifold satisfies the $\partial\overline\partial$-Lemma. This heuristically explains the results in [@angella-tomassini-3], to which we refer for details. The proof there uses the Varouchas exact sequences in [@varouchas].
\[thm:frolicher-bc\] Let $X$ be a compact complex manifold. Then, for any $k\in\mathbb{Z}$, the $k$th [*non-$\partial\overline\partial$-degree*]{} $$\label{eq:non-kahler-degree}
\Delta^k(X) \;:=\; h^{k}_{BC} + h^{k}_{A} - 2\, b_k$$ is a non-negative integer. Moreover, $X$ satisfies the $\partial\overline\partial$-Lemma if and only if $\Delta^k(X)=0$ for any $k\in\mathbb{Z}$; equivalently, $$\sum_{k\in\mathbb{Z}}\Delta^k(X) \;=\; 0 \;.$$
In a sense, the previous result states that an un-natural isomorphism (the one given by equality of the dimensions — a quantitative property) ensures a natural isomorphism (the one induced by identity — a qualitative property).
As a special case, consider compact complex surfaces. In [@angella-dloussky-tomassini] and in further discussions with A. Tomassini and M. Verbitsky, we showed that the non-$\partial\overline\partial$-degree of compact complex surfaces are topological invariants. More precisely, $\Delta^1=0$ and $\Delta^2\in\{0,2\}$ according to the complex surface admitting Kähler metrics. (Notice that, for compact complex surfaces, the $\partial\overline\partial$-Lemma is in fact equivalent to the existence of a Kähler metric, by the Lamari and the Buchdahl criterion.)
We announce now an upper bound for the dimension of Bott-Chern cohomology in terms of Hodge numbers, obtained in joint work with Nicoletta Tardini [@angella-tardini-1]. Note that we cannot get a just topological upper bound. That is, an opposite Frölicher inequality cannot be obtained. This is because of the even-length zigzags, which contribute to the Dolbeault cohomology but not to the de Rham cohomology.
\[thm:upper-bound\] Let $X$ be a compact complex manifold of complex dimension $n$. Then, for any $k\in\mathbb{Z}$, there holds $$\begin{aligned}
h^k_{A} &\leq& \min\{k+1, (2n-k)+1\} \, \left( h^k_{\overline\partial} + h^{k+1}_{\overline\partial} \right) \\[5pt]
&\leq& (n+1) \, \left( h^k_{\overline\partial} + h^{k+1}_{\overline\partial} \right) \;,
\end{aligned}$$ whence also $$\begin{aligned}
h^k_{BC} &\leq& \min\{k+1, (2n-k)+1\} \, \left( h^k_{\overline\partial} + h^{k-1}_{\overline\partial} \right) \\[5pt]
&\leq& (n+1) \, \left( h^k_{\overline\partial} + h^{k-1}_{\overline\partial} \right) \;.
\end{aligned}$$
We give just the idea behind the proof. We refer to [@angella-tardini-1] for details. The point is that a contribution to Aeppli cohomology arises from zigzags of positive length $\ell+1$. Any such zigzag, when placed between total degrees $k$ and $k+1$, creates exactly two non trivial classes in either Dolbeault or conjugate Dolbeault cohomology at either degree $k$ or degree $k+1$, and at most $\lfloor\ell/2\rfloor+1$ classes in Aeppli cohomology at degree $k$. (In particular, $\lfloor\ell/2\rfloor+1\leq \min\{k+1,(2n-k)+1\}\leq n+1$.) The inequality for Bott-Chern cohomology follows from the Schweitzer duality [@schweitzer §2.c] and by the Serre duality.
In particular, for any $k\in\mathbb{Z}$, there holds $$-2(n+1) \, \left( h^{k-1}_{\overline\partial} + h^{k}_{\overline\partial} \right) \;\leq\; h^k_{A}-h^{k}_{BC} \;\leq\; 2(n+1) \, \left( h^{k}_{\overline\partial} + h^{k+1}_{\overline\partial} \right) \;.$$ In the following result, from [@angella-tardini-1], we give a characterization of the $\partial\overline\partial$-Lemma in terms of the above inequality.
\[thm:char-deldelbar-minus\] A compact complex manifold $X$ satisfies the $\partial\overline\partial$-Lemma if and only if $$\sum_{k\in\mathbb{Z}} \left| h^{k}_{BC} - h^{k}_{A} \right| \;=\; 0 \;.$$
We give the idea of the proof, referring to [@angella-tardini-1] for details. Recall that the Bott-Chern cohomology counts the corners with possible incoming arrows, and the Aeppli cohomology counts the corners with possible outcoming arrows, with the exceptions of squares. Therefore the hypothesis can be restated as: for any anti-diagonal, the number of ingoing arrows equals the number of outgoing arrows, except for squares. Since no ingoing arrow can enter the anti-diagonal of total degree $0$, it follows that there is no zigzag of positive length in the whole diagram. That is, the $\partial\overline\partial$-Lemma holds.
Qualitative properties of Bott-Chern cohomology
===============================================
By investigating the [*qualitative*]{} properties of some cohomology, we mean the study of what and how algebraic structures are induced in cohomology from the space of forms.
For example, let us focus on the differential graded algebra structure on the space of forms given by the wedge product and the exterior differential, and on the de Rham cohomology. By the Leibniz rule, it induces a structure of algebra in cohomology. We look at $H_{dR}$ as a functor inside the category $\mathbf{dga}$ of differential $\mathbb{Z}$-graded algebras: $$H_{dR} \colon \mathbf{dga} \rightsquigarrow \mathbf{dga} \;.$$ We ask for what objects $X$ this functor can be made “concrete”, that is, when it can be realized as a composition of quasi-isomorphisms and formal inverses of quasi-isomorphisms in $\mathbf{dga}$: e.g., $$\xymatrix{
X \ar@{~>}[rrrrrrr] \ar[dr]_{\text{qis}} & & & & & & & H_{dR}(X) \\
& C_1 & C_2 \ar[l]^{\text{qis}} & C_3 \ar[l]^{\text{qis}}\ar[r]_{\text{qis}} & \cdots \ar[r]_{\text{qis}} & C_{h-1} & C_h \ar[l]^{\text{qis}} \ar[ur]_{\text{qis}} &
}$$ By the existence of minimal models, see, e.g., [@wu Theorem in §II.3 at page 29], [@bousfield-gugenheim Proposition 7.7], this corresponds to the dga of forms and the dga of de Rham cohomology sharing the same model. A compact complex manifold whose double complex of forms satisfies such a property is called [*formal*]{} in the sense of Sullivan [@sullivan]. Note that the minimal model contains information on the rational homotopy groups of the manifold [@sullivan]: hence the rational homotopy type of formal manifolds is a formal consequence of their de Rham cohomology. Compact complex manifolds satisfying the $\partial\overline\partial$-Lemma (e.g., compact Kähler manifolds) are formal in the sense of Sullivan, [@deligne-griffiths-morgan-sullivan Main Theorem]. A theory of Dolbeault formality for complex manifolds has been developed in [@neisendorfer-taylor].
Notice that every compact complex manifold is formal [*in the category of $A_\infty$-algebras*]{}[^5]. This follows from the Homotopy Transfer Principle by T. V. Kadeishvili [@kadeishvili]. See [@merkulov; @zhou; @lu-palmieri-wu-zhang] for the explicit construction of the Merkulov model. By [@lu-palmieri-wu-zhang], the induced $A_\infty$-structure in cohomology yields the Massey products, up to sign. See also [@positselski].
With these notations, a compact complex manifold is formal in the sense of Sullivan if there exists a system of representatives $H^\bullet$ for the cohomology such that the induced $A_\infty$-structure on $H^\bullet$ is actually an algebra structure. A particular case is when the chosen representatives are actually the harmonic representatives with respect to some Hermitian metric. This last situation is referred as [*geometric formality*]{} in the sense of Kotschick [@kotschick]. In this case, a possible category to which one can restrict is the category whose objects are dgas with a non-degenerate pairing.
Note that the wedge products on forms induces an algebra structure in Bott-Chern cohomology, and just an $H_{BC}$-module structure in Aeppli cohomology. The duality pairing given by any fixed Hermitian metric is internal in de Rham and Dolbeault cohomologies by Poincaré and Serre dualities, and it yields an isomorphism between Bott-Chern and Aeppli cohomologies [@schweitzer §2.c].
This is an important issue, for example, in defining Massey products for Bott-Chern cohomology. In [@angella-tomassini-6], triple Aeppli-Bott-Chern Massey products are defined, starting from Bott-Chern classes, and yielding a class in Aeppli cohomology, up to indeterminacy.
Hence, in [@angella-tardini-1], we propose the following definition, which takes into consideration the structure of non-degenerate pairing.
A compact complex manifold $X$ of complex dimension $n$ is said to satisfy the [*Schweitzer qualitative property*]{} if the natural pairing $$H^{\bullet,\bullet}_{BC}(X) \times H^{\bullet,\bullet}_{BC}(X) \to \mathbb{C} \;, \qquad \left( [\alpha], [\beta] \right) \mapsto \int_X \alpha\wedge\beta$$ induced by the wedge product and by the pairing with the fundamental class $[X]$ is non-degenerate.
The above qualitative properties implies a quantitative property which, in turn, characterizes the qualitative property of the $\partial\overline\partial$-Lemma, thanks to Theorem \[thm:char-deldelbar-minus\]. This is stated in [@angella-tardini-1].
\[thm:main-thm\] Let $X$ be a compact complex manifold. If it satisfies the Schweitzer qualitative property, then it satisfies the $\partial\overline\partial$-Lemma.
[48]{}
<span style="font-variant:small-caps;">A. Aeppli</span>, *On the cohomology structure of Stein manifolds*, in [*Proc. Conf. Complex Analysis (Minneapolis, Minn., 1964)*]{}, Springer, Berlin, 1965, 58–70.
<span style="font-variant:small-caps;">D. Angella</span>, <span style="font-variant:small-caps;">G. Dloussky</span>, <span style="font-variant:small-caps;">A. Tomassini</span>, *On Bott-Chern cohomology of compact complex surfaces*, to appear in [Ann. Mat. Pura Appl.]{}, DOI: 10.1007/s10231-014-0458-7.
<span style="font-variant:small-caps;">D. Angella</span>, <span style="font-variant:small-caps;">N. Tardini</span>, *Quantitative and qualitative cohomological properties for non-Kähler manifolds*, `arXiv:1507.07108`.
<span style="font-variant:small-caps;">D. Angella</span>, <span style="font-variant:small-caps;">A. Tomassini</span>, *On the $\partial\overline{\partial}$-lemma and Bott-Chern cohomology*, [Invent. Math.]{} **192** (2013), no. 1, 71–81.
<span style="font-variant:small-caps;">D. Angella</span>, <span style="font-variant:small-caps;">A. Tomassini</span>, *Inequalities à la Frölicher and cohomological decompositions*, [J. Noncommut. Geom.]{} **9** (2015), no. 2, 505–542.
<span style="font-variant:small-caps;">D. Angella</span>, <span style="font-variant:small-caps;">A. Tomassini</span>, *On Bott-Chern cohomology and formality*, [J. Geom. Phys.]{} **93** (2015), 52–61.
<span style="font-variant:small-caps;">R. Bott</span>, <span style="font-variant:small-caps;">S. S. Chern</span>, *Hermitian vector bundles and the equidistribution of the zeroes of their holomorphic sections*, [Acta Math.]{} **114** (1965), no. 1, 71–112.
<span style="font-variant:small-caps;">A. K. Bousfield</span>, <span style="font-variant:small-caps;">V. K. A. M. Gugenheim</span>, *On PL de Rham theory and rational homotopy type*, [Mem. Amer. Math. Soc.]{} **8** (1976), no. 179.
<span style="font-variant:small-caps;">P. Deligne</span>, <span style="font-variant:small-caps;">Ph. A. Griffiths</span>, <span style="font-variant:small-caps;">J. Morgan</span>, <span style="font-variant:small-caps;">D. P. Sullivan</span>, *Real homotopy theory of Kähler manifolds*, [Invent. Math.]{} **29** (1975), no. 3, 245–274.
<span style="font-variant:small-caps;">V. Dotsenko</span>, <span style="font-variant:small-caps;">S. Shadrin</span>, <span style="font-variant:small-caps;">B. Vallette</span>, *De Rham cohomology and homotopy Frobenius manifolds*, [J. Eur. Math. Soc. (JEMS)]{} **17** (2015), no. 3, 535–547.
<span style="font-variant:small-caps;">A. Frölicher</span>, *Relations between the cohomology groups of Dolbeault and topological invariants*, [Proc. Natl. Acad. Sci. USA]{} **41** (1955), no. 9, 641–644.
<span style="font-variant:small-caps;">E. Getzler</span>, <span style="font-variant:small-caps;">J. D. S. Jones</span>, *$A_\infty$-algebras and the cyclic bar complex*, [Illinois J. Math.]{} **34** (1990), no. 2, 256–283.
<span style="font-variant:small-caps;">T. V. Kadeishvili</span>, *The algebraic structure in the homology of an A($\infty$)-algebra*, [Soobshch. Akad. Nauk Gruzin. SSR]{} **108** (1982), no. 2, 249–252.
<span style="font-variant:small-caps;">B. Keller</span>, *Introduction to $A$-infinity algebras and modules*, [Homology Homotopy Appl.]{} **3** (2001), no. 1, 1–35.
<span style="font-variant:small-caps;">D. Kotschick</span>, *On products of harmonic forms*, [Duke Math. J.]{} **107** (2001), no. 3, 521–531.
<span style="font-variant:small-caps;">D.-M. Lu</span>, <span style="font-variant:small-caps;">J. H. Palmieri</span>, <span style="font-variant:small-caps;">Q.-S. Wu</span>, <span style="font-variant:small-caps;">J. J. Zhang</span>, *$A$-infinity structure on Ext-algebras*, [J. Pure Appl. Algebra]{} **213** (2009), no. 11, 2017–2037.
<span style="font-variant:small-caps;">S. A. Merkulov</span>, *Strong homotopy algebras of a Kähler manifold*, [Internat. Math. Res. Notices]{} **1999** (1999), no. 3, 153–164.
<span style="font-variant:small-caps;">J. Neisendorfer</span>, <span style="font-variant:small-caps;">L. Taylor</span>, *Dolbeault homotopy theory*, [Trans. Amer. Math. Soc.]{} **245** (1978), 183–210.
<span style="font-variant:small-caps;">D. Popovici</span>, *Volume and Self-Intersection of Differences of Two Nef Classes*, `arXiv:1505.03457`.
<span style="font-variant:small-caps;">L. Positselski</span>, *Koszulity of cohomology = $K(\pi,1)$-ness + quasi-formality*, `arXiv:1507.04691`.
<span style="font-variant:small-caps;">M. Schweitzer</span>, *Autour de la cohomologie de Bott-Chern*, Prépublication de l’Institut Fourier no. 703 (2007), `arXiv:0709.3528`.
<span style="font-variant:small-caps;">J. D. Stasheff</span>, *Homotopy associativity of H-spaces. I, II*, [Trans. Amer. Math. Soc.]{} **108** (1963), 275–292; ibid. **108** (1963) 293–312.
<span style="font-variant:small-caps;">D. Sullivan</span>, *Infinitesimal computations in topology*, [Inst. Hautes Études Sci. Publ. Math.]{} **47** (1977), no. 1, 269–331.
<span style="font-variant:small-caps;">N. Tardini</span>, <span style="font-variant:small-caps;">A. Tomassini</span>, *On geometric Bott–Chern formality and deformations*, `arXiv:1502.03706`.
<span style="font-variant:small-caps;">V. Tosatti</span>, <span style="font-variant:small-caps;">B. Weinkove</span>, *The complex Monge-Ampère equation on compact Hermitian manifolds*, [J. Amer. Math. Soc.]{} **23** (2010), no. 4, 1187–1195.
<span style="font-variant:small-caps;">L. Ugarte</span>, *Hodge numbers of a hypothetical complex structure on the six sphere*, [Geom. Dedicata]{} **81** (2000), no. 1–3, 173–179.
<span style="font-variant:small-caps;">J. Varouchas</span>, *Proprietés cohomologiques d’une classe de variétés analytiques complexes compactes*, in [*Séminaire d’analyse P. Lelong-P. Dolbeault-H. Skoda, années 1983/1984*]{}, Lecture Notes in Math., vol. **1198**, 233–243, Springer, Berlin (1986).
<span style="font-variant:small-caps;">W.-t. Wu</span>, [*Rational homotopy type. A constructive study via the theory of the $I^*$-measure*]{}, Lecture Notes in Mathematics, **1264**, Springer-Verlag, Berlin, 1987.
<span style="font-variant:small-caps;">J. Zhou</span>, *Hodge theory and $A_\infty$-structures on cohomology*, [Internat. Math. Res. Notices]{} **2000** (2000), no. 2, 71–78.
[^1]: The author is supported by the Project PRIN “Varietà reali e complesse: geometria, topologia e analisi armonica”, by the Project FIRB “Geometria Differenziale e Teoria Geometrica delle Funzioni”, by SNS GR14 grant “Geometry of non-Kähler manifolds”, and by GNSAGA of INdAM
[^2]: it seems that this assumption can be assumed without loss of generality, see the answer by Greg Kuperberg in the MathOverflow discussion at <http://mathoverflow.net/questions/25723/>, where he quotes Mikhail Khovanov
[^3]: this example was suggested by Michela Zedda
[^4]: in the following diagram, the filled dot at the top-right is a generator of the Bott-Chern cohomology. Indeed, it is $\partial$-closed and $\overline\partial$-closed: this is pictured by the fact that there is no outgoing arrow. It is not $\partial\overline\partial$-exact, that is, it is not the top-right corner of a square. But it can be $\partial$-exact and/or $\overline\partial$-exact, whence the dotted arrows in the notation. Analogous considerations hold for the diagram for the Aeppli cohomology.
[^5]: that is, strongly homotopy associative algebras: the category of $A_\infty$-algebras is equivalent to the category of differential graded co-algebras, by means of the bar construction. We refer to [@stasheff; @keller; @getzler-jones; @dotsenko-shadrin-vallette] for definitions and details.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A noncommutative Grassmannian $A ={\mathop{\mathrm{NGr}}\nolimits}(m,n)$ is introduced by Efimov, Luntz, and Orlov in [*Deformation theory of objects in homotopy and derived categories III: Abelian categories*]{} as a noncommutative algebra associated to an exceptional collection of $n-m+1$ coherent sheaves on ${\ensuremath{\mathbb P}}^n$. It is a graded Calabi–Yau ${\ensuremath{\mathbb Z}}$-algebra of dimension $n-m+1$. We show that this algebra is coherent provided that the codimension $d=n-m$ of the Grassmannian is two. According to [*op. cit.*]{}, this gives a $t$-structure on the derived category of the coherent sheaves on the noncommutative Grassmannian. The proof is quite different from the recent proofs of the coherence of some graded 3-dimensional Calabi–Yau algebras and is based on properties of a PBW-basis of the algebra $A$.'
address: 'Department of Mathematics for Economics, Myasnitskaya str. 20, State University ‘Higher School of Economics’, Moscow 101990, Russia '
author:
- Dmitri Piontkovski
title: Noncommutative Grassmannian of codimension two has coherent coordinate ring
---
[^1]
Introduction
============
Recall that a connected [$\mathbb N$]{}-graded algebra of the form $A = A_0 \oplus
A_1 \oplus \dots $ such that $A_0$ is a copy of the basic field $k$ is called [*regular*]{} if it has finite global dimension (say, $d$) and satisfies the following Gorenstein property: $$\ext^i_A(k,k) \cong\left\{ \begin{array}{ll}
k^* [l] \mbox{ for some } l\in {\ensuremath{\mathbb Z}},& i =d \\
0, & i\ne d .
\end{array} \right.$$ The same notion of regularity is extend (following Bondal and Polishchuk [@bp]) to a slightly more general case of a ${\ensuremath{\mathbb Z}}$-algebra $A$, see Subsection \[subs:z\] below.
Regular algebras play the roles of coordinate rings of noncommutative projective spaces in a version of noncommutative projective geometry [@pol; @BvdB] which generalizes the well-known approach of Artin and Zhang [@AZ]. Namely, suppose that a regular algebra of global dimension $d$ is graded coherent. Consider the quotient category ${\mathop{\mathrm{qgr}}\nolimits}A = {\mathop{\mathrm{cmod}}\nolimits}A / {\mathop{\mathrm{tors}}\nolimits}A$ of the category ${\mathop{\mathrm{cmod}}\nolimits}A$ of finitely presented (=graded coherent) right graded $A$-modules by its subcategory ${\mathop{\mathrm{tors}}\nolimits}A$ of finite-dimensional modules. This category ${\mathop{\mathrm{qgr}}\nolimits}A $ plays the role of the category of coherent sheaves on a noncommutative $(d-1)$-dimensional projective space. Here we are interested in the case $d=3$, that is, in the case of noncommutative planes. The famous classification of 3-dimensional regular algebras $A$ of polynomial growth is obtained Artin and Shelter [@AS]. Particularly, they have shown that these algebras are Noetherian (hence, coherent). It is not hard to construct also non-Noetherian 3-dimensional regular algebras (e.g., one may follow the approach of [@AS Sections 2 and 3]). In contrast, it is often not easy to prove that such an algebra is coherent. However, there are important examples for which the coherence is established. These are the octonion algebra of P. Smith [@smi], the Yang–Mills algebra introduced by Movshev and Swartz [@ms] which is coherent by a theorem of Herscovich [@hers] (the Yang–Mils algebra introduced by Connes and Dobios–Violette [@cdv] is a particular case of it), and 3-Calabi–Yau algebras which are Ore extensions of 2-Calaby–Yau ones by He, Oystaeyen, and Zhang [@hoz].
In all these cases, the coherence property is proved using the same lemma [@Pi1 Prop. 3.2]. It states that if a non-trivial two-sided ideal $I$ in a graded algebra $A$ is free as a left module and the quotient algebra $A/I$ is right Noetherian, then $A$ is graded coherent. So, the known examples of coherent regular algebras are, in a sense, extensions of Noetherian algebras along free modules.
In this paper, we prove the coherence property of another 3-dimensional regular algebra. In contrast to the previous cases, it seems that the approach based on the above lemma fails for this algebra.[^2] This algebra is introduced by Efimov, Luntz, and Orlov [@elo] under the name [*Noncommutative Grassmannian*]{} (see Subsection \[subs:gras\] below for an explicit definition). According to [@elo Section 7], a noncommutative Grassmannian ${\mathop{\mathrm{NGr}}\nolimits}(m,n)$ ‘is a true noncommutative moduli space of the structure sheaves ${\mathcal O_{{\ensuremath{\mathbb P}}(W)} \in D^b_{coh}({\ensuremath{\mathbb P}}^n)}$’, where $W \subset k^n$ runs the vector subspaces of dimension $m$. The connection with structure sheaves is based on the theory of helices in derived categories by A.Bondal and A.Polishchuk [@bondal; @bp]. The construction of [@elo] gives a description (by constructing a $t$-structure) of the derived category of ${\mathop{\mathrm{qgr}}\nolimits}{\mathop{\mathrm{NGr}}\nolimits}(m,n)$ provided that the algebra ${\mathop{\mathrm{NGr}}\nolimits}(m,n)$ of coherent. It is pointed out in [@elo Remark 8.23] that ${\mathop{\mathrm{NGr}}\nolimits}(m,n)$ is coherent if the codimension $d = n-m$ of the Grassmannian is equal to 1 (since the algebra ${\mathop{\mathrm{NGr}}\nolimits}(m,n)$ has global dimension 2 in this case). The first non-trivial case is the Grassmannian of codimension $d= 2$, when the algebra has global dimension three. The main result of this paper is the following.
\[th:main\_intro\] The noncommutative Grassmannian algebra ${\mathop{\mathrm{NGr}}\nolimits}(m,n)$ is coherent provided that $n-m=2$.
The paper is organized as follows. In Section \[sec:defs\] we briefly remind necessary facts about ${\ensuremath{\mathbb Z}}$-algebras and, in particular, about the noncommutative Grassmannian algebra ${\mathop{\mathrm{NGr}}\nolimits}(m,n)$. In Subsection \[subs: hat A\], we note that the ${\ensuremath{\mathbb Z}}$–algebra $A ={\mathop{\mathrm{NGr}}\nolimits}(m,n)$ with $n-m=2$ is 3-periodic, so, its properties are essentially the same as the properties of the corresponding algebra $\hat A$ over a triangle quiver. We immediately calculate here the Hilbert series of $\hat A$. This obviously gives also the Hilbert series of $A$. In Subsection \[subs:PBW\], we show that $A$ is a PBW algebra as a 6-periodic ${\ensuremath{\mathbb Z}}$-algebra (in particular, it admits a quadratic Gröbner basis of relations). Note that we do not know if $\hat A$ is PBW or not. In Proposition \[prop:3-proc\], we prove that $A$ satisfies a property of bounded processing [@pi01], that is, the structure of the multiplication of paths in this quiver algebra is essentially depend only on bounded segments of the multipliers. We recall necessary definitions in Subsection \[subs:proc\]. Using a result of [@pi01], we finally deduce in Corollary \[cor:main\] that the algebra $A$ is coherent. A stronger consequence of bounded processing is briefly discussed in Remark \[rem:uni-coh\].
Acknowledgement
---------------
I am grateful to Alexei Bondal, Alexander Efimov, and Dmitri Orlov for stimulating questions and fruitful discussions.
Background and notations {#sec:defs}
========================
[$\mathbb Z$]{}-algebras {#subs:z}
------------------------
Recall that a ${\ensuremath{\mathbb Z}}$-algebra is a path algebra (with relations) over the infinite line quiver $\dots \to
-1 \to 0 \to 1
\to 2\to \dots$ with multiple arrows. We refer the reader to [@bp] or [@PP Ch. 4, Sect. 9–10] for the basic definitions of quadratic, Koszul and PBW ${\ensuremath{\mathbb Z}}$–algebras. We say that a ${\ensuremath{\mathbb Z}}$-algebra $A = \oplus_{i,j \in {\ensuremath{\mathbb Z}}}$ (where $A_{i,j} = 0$ for $i<j$ and $A_{i,i}=k$) is [*regular*]{} of dimension $d$ if each irreducible module $k_i = P_i / \oplus_{j> i}A_{i,j}$ has global dimension $d$ (where $P_i = \oplus_{j\ge i}A_{i,j}$ is the corresponding projective module) and the Exts of these modules satisfy an analogous Gorenstein condition [@bp Sect. 4] $$\ext^i (k_s, P_t)\cong\left\{ \begin{array}{ll} k^* \mbox{ if } t = s+l ,& i =d,
\\
0, & \mbox{ otherwise}
\end{array} \right.$$ for some $l\in {\ensuremath{\mathbb Z}}$. Note that regular algebras are sometimes also called [*AS-regular*]{}.
Noncommutative Grassmannian {#subs:gras}
---------------------------
A noncommutative Grassmannian is defined by Efimov, Lunts, and Orlov [@elo Part 3] as a noncommutative scheme associated to the following algebra. Given two positive integers $m<n$ and an $n$-dimensional vector space $V$, let $A = A^{m,V}$ be a quadratic ${\ensuremath{\mathbb Z}}$-algebra with $A_{ij} = k$ (a basic field) and generators $$A_{i,i+1} = \left\{\begin{array}{ll} {{\Lambda}}^d V, & i|(d+1),\\
V^*, &\mbox{ otherwise, }
\end{array} \right.$$ where $d=n-m$. The quadratic relations of $A$ are defined via the natural exact sequences $$0\to \Lambda^{d-1} V \to A_{i+1,i+2}\otimes A_{i,i+1} \to
A_{i,i+2} \to 0 \mbox{ for } (d+1)|i,i+1$$ and $$0\to \Lambda^2 V^* \to A_{i+1,i+2}\otimes A_{i,i+1} \to A_{i,i+2}
\to 0, \mbox{ otherwise}.$$
Obviously, $A$ is $(d+1)$-periodic (that is, $A_{i,j}$ is naturally isomorphic to $A_{i+d+1,j+d+1}$). By [@elo Prop. 8.18], $A$ isomorphic to the automorphism ${\ensuremath{\mathbb Z}}$-algebra of the helix generated by the exceptional collection of $(d+1)$ coherent sheaves on ${\ensuremath{\mathbb P}}(k^n)$ $$E = ({\mathcal O}_{{\ensuremath{\mathbb P}}(k^n)}(m-n), \dots, {\mathcal O}_{{\ensuremath{\mathbb P}}(k^n)}(-1), {\mathcal O}_{{\ensuremath{\mathbb P}}(k^n)} ).$$ It follows from [@bp] that $A$ is Koszul and Gorenstein of global dimension $d+1$.
Note that $A$ is a so-called graded Calabi–Yau algebra of dimension 3 (in the sense of [@bock]). This follows from the same property of the corresponding algebra $\hat A$, see Subsection \[subs: hat A\] below.
It is pointed out in [@elo Remark 8.23] that the description of the derived category of ${\mathop{\mathrm{QMod}}\nolimits}A$ (“quasicoherent sheaves” on the noncommutative Grassmannian) can the transferred to $D^b({\mathop{\mathrm{qmod}}\nolimits}A)$ (derived category of the “coherent sheaves”) provided that the category ${\mathop{\mathrm{qmod}}\nolimits}A$ is Abelian, that is, $A$ is coherent. Namely, in this case a $t$-structure on the derived category of the finite-dimensional modules over a finite-dimensional algebra $\oplus_{1\le i,j \le n} A_{i,j}$ constructed in [@elo Section 8] induces a $t$-structure on $D^b({\mathop{\mathrm{qmod}}\nolimits}A)$.
The results
===========
Algebra $\hat A$ and its Hilbert series {#subs: hat A}
---------------------------------------
Since $A$ is $(d+1)$-periodic (in terms of [@bp Section 4], it is geometric of period $d+1$), its categories of graded modules (all , finitely generated, finitely presented, finite dimensional…) are equivalent to the ones of the algebra $\hat
A = \oplus_{i=0}^d \oplus_{j\in {\ensuremath{\mathbb Z}}} A_{ij}$ considered as a path algebra over a cyclic quiver of length $d+1$ with multiple arrows. It follows that $\hat A$ is Koszul and Gorenstein of global dimension $d+1$ as well as $A$ is. The components of $\hat A$ are indexed by the pairs $(i,j)$ with $i \in {\ensuremath{\mathbb Z}}/ (d+1){\ensuremath{\mathbb Z}}$ and $j\in
{\ensuremath{\mathbb Z}}$, where elements of $\hat A_{i,j}$ are considered as path from $i$ to $(j-i) \mod (d+1)$. The surjection $A\to \hat A$ is induced by the natural surjection of quiver algebras.
In our case $d=2$, the algebra $\hat A$ is a path algebra (=quiver algebra with relations) over the quiver $Q:0\to 1\to 2\to 0$ where the arrows are multiple, namely, with $n(n-1)/2$ arrows $0\to 1$, $n$ arrows $1\to 2$ and $n$ arrows $2\to 0$. Note that it follows from [@bock Theorem 3.1] that $\hat A$ is a graded Calabi–Yau algebra of dimension 3. In the notations of Subsection \[subs:PBW\] below, the corresponding superpotential is equal to a cyclic element represented by $$\sum_{t\in 3{\ensuremath{\mathbb Z}}} \sum_{i,j\in [1..n], i\ne j} x^{t-1}_i e^t_{ij} x^{t+1}_j .$$
Recall that the Hilbert series $H_{\hat A}$ of the graded path algebra $\hat A$ over the above quiver with 3 vertices is a $3 \times 3$ matrix $H_{\hat A}= (h_{ij})$ over ${\ensuremath{\mathbb Z}}[[t]]$ defined by $$h_{ij} = \sum_{m \in {\ensuremath{\mathbb Z}}} (\mbox{the number of paths $i\to j$ of
length } m) t^m = \sum_{n\in Z} t^{3n+j-i} \dim \hat A_{i,3n+j}.$$
\[prop:Hilb\_ser\] The Hilbert series $H_{\hat A}$ of the path algebra $\hat A$ is $$H_{\hat A} = \left(
\begin{array}{ccc}
1 -t^3& \binom{n}{2} t & n t^2 \\
n t^2 & 1-t^3 & -n t \\
-n t & \binom{n}{2} t^2 & 1-t^3\\
\end{array}
\right)^{-1}$$ $$=
\left( \begin {array}{ccc}
1+ \frac{4-5{n}^{2}+{n}^{4}}{4}
{t}^{3}
& \frac{n(n-1)}{2} t
&n \left( {n}^{2}-n-2 \right)
{t}^{ 2}/2\\
\noalign{\medskip} \frac{n (n+1)}{2}
{t}^{2}
& 1+\frac{4-5{n}^{2}+{n}^{4}}{4}
{t}^{3}
&nt\\
\noalign{\medskip} nt&
n \frac{{n}^{2}-n-2}{2}
{t}^{2}&
1 + \frac{2+{n}^{4}-{n}^{3}-4{n}^{2}}{2}
{t}^{3}
\end {array}
\right) + O(t^4).$$
Note that the Koszul dual algebra of $\hat A$ has the following components: $$\hat A_{ii}^! = k \mbox{ for all } i =0, 1,2,$$ $$\hat A_{i,i+1}^! = (\hat A_{i,i+1})^* \approx
\left\{
\begin{array}{ll}
\left( {\Lambda}^2 V \right)^*,
& i=0,\\
V, & i=1,2,
\end{array}
\right.$$ $$\hat A_{i,i+2}^! = (\mbox{relations of $A$ in} A_{i+1,i+2}\otimes
A_{i,i+1})^* \approx
\left\{
\begin{array}{ll}
\left( \Lambda^{d-1} V \right)^* = V^*, & i=0,2\\
\left( \Lambda^2 V^* \right)^*, & i=1,
\end{array}
\right.$$ and, by the Gorenstein property, $$\hat A_{i,i+3}^! \approx k.$$ All other components of $\hat A^!$ vanish.
It follows that the Hilbert series of $\hat A^!$ is the following matrix of order $d+1 = 3$ $$H_{\hat A^!} (t) = \left(
\begin{array}{ccc}
1 +t^3& \binom{n}{2} t & n t^2 \\
\binom{n}{2} t^2 & 1+t^3 & n t \\
n t & n t^2 & 1+t^3\\
\end{array}
\right)$$ Since $\hat A$ is Koszul, we have $H_{\hat A}(t) H_{\hat A}(-t) = \Id$, thus, $$H_{\hat A}(t) =H_{\hat A}(-t)^{-1}.$$
PBW property {#subs:PW}
------------
\[subs:PBW\]
Let us fix a pair of dual bases $x = \{x_1,\dots, x_n \}$ in $V^*$ and $e = \{e_1,\dots, e_n \}$ in $V$. Given a finite sequence $\alpha = (\alpha_1, \dots, \alpha_s)$, we denote by $e_\alpha$ the product $e_{\alpha_1} \wedge \dots \wedge e_{\alpha_s} \in
\Lambda^s V$, so that the set $e=\{ e_{ij} | i<j \}$ is a basis of $\Lambda^2 V$. For $t\in {\ensuremath{\mathbb Z}}$, we denote by $x^t=\{x_1^t,\dots,
x_n^t \}$ and $e^t = \{e_{ij}^t | i<j \}$ the corresponding bases of $A_{t,t+1} $ in the cases $3 |t$ and $3
\not| t$, respectively. We will sometimes omit the upper indexes for elements of these bases.
In this notations, the relations of the algebra $A$ of the grading component $(t,t+2)$ has the following form (where for $i>j$ we use the sign $e_{ij}:=-e_{ji}$): $$\label{eq:rel_A}
\begin{array}{lll}
f_i = f_i^t:= \sum\limits_{j \ne i} e_{ij}^t x_j^{t+1}, & i = 1..n, & t\in 3{\ensuremath{\mathbb Z}},\\
c_{ij} = c_{ij}^t:= x_i^t x_j^{t+1} - x_j^t x_i^{t+1}, & 1\le i <
j \le n, & t\in 3{\ensuremath{\mathbb Z}}+1 ,\\
g_i = g_i^t:= \sum\limits_{j \ne i} x_j^{t} e_{ji}^{t+1} , & i = 1..n, & t\in 3{\ensuremath{\mathbb Z}}+2.\\
\end{array}$$
Let us fix orderings of the elements of the bases $x^t$ and $e^t$ in the following (6-periodic) way: $$\begin{array}{lll}
e_{ij}>e_{kl} & \mbox{ if } i<k \mbox{ or } i=k,j<l & \mbox{ for } t \in
6{\ensuremath{\mathbb Z}},\\
e_{ij} > e_{kl} & \mbox{ if } i>k \mbox{ or } i=k,j>l & \mbox{ for } t \in
6{\ensuremath{\mathbb Z}}+3,\\
x_{1}<\dots <x_n & &\mbox{ for } t \in
6{\ensuremath{\mathbb Z}}\pm 1,\\
x_{1}>\dots >x_n & &\mbox{ for } t \in
6{\ensuremath{\mathbb Z}}\pm 2.
\end{array}$$
Let us introduce the reverse lexicographical order on the paths of the quiver $Q$, that is, for two paths $v=v_1\dots v_t$ and $w=w_1\dots w_t$ of length $t$ with the same head and the same tail we set $v<w$ iff $v_j<w_j , v_{j+1} = w_{j+1}, \dots $ and $ v_{t} = w_{t}$ for some $j$.
\[prop: A is PBW\] The above quadratic relations of the algebra $\hat A$ form a Gröbner basis of the ideal of relations with respect to the above order, that is, the algebras $\hat A$ and $A$ are PBW algebras.
The leading monomials of the relations (\[eq:rel\_A\]) w. r. t. the above reverse lexicographical order are $$\label{eq:lead_terms}
\begin{array}{lll}
e_{in} x_n \mbox{ for } 1\le i <n, e_{n-1,n}x_{n-1}, & t \in 6 {\ensuremath{\mathbb Z}},\\
x_i x_j \mbox{ for } n\ge i > j \ge 1, & t \in 6 {\ensuremath{\mathbb Z}}+1, \\
x_1 e_{1i} \mbox{ for } n\ge i>1, x_{2} e_{12}, & t \in 6 {\ensuremath{\mathbb Z}}+2,\\
e_{1i} x_1 \mbox{ for } n\ge i>1, e_{12} x_{2} , & t \in 6 {\ensuremath{\mathbb Z}}+3,\\
x_i x_j \mbox{ for } 1 \le i < j \le n, & t \in 6 {\ensuremath{\mathbb Z}}+4, \\
x_n e_{in} \mbox{ for } 1\le i <n, x_{n-1} e_{n-1,n}, & t \in 6 {\ensuremath{\mathbb Z}}+5,\\
\end{array}$$
Let $ B $ be the $Z$-algebra defined by the same generators as $A$ and the above monomial relations. We have $\dim B_{ij} = \dim
A_{ij}$ for $j=i, j=i+1, j=i+2$ and $\dim B_{ij} \ge \dim A_{ij}$ for $j\ge i+3$. By [@PP Prop. 10.1 of Ch. 4], to show that $A$ is PBW it is sufficient to check the equalities $\dim B_{i,
i+3} = \dim A_{i, i+3}$ for all $i$.
Because of the symmetry of the monomial relations above, we have $\dim B_{i, i+3} = \dim B_{i+3s, i+3s+3} $ for all $s\in {\ensuremath{\mathbb Z}}$. Since the algebra $A$ is 3-periodic, it is enough to show that the equalities $\dim B_{i,
i+3} = \dim A_{i, i+3}$ hold for $i=0,1,2$. Here $\dim A_{i,i+3}
= \dim\hat A_{i \mod 3,i+3}$ is the coefficient of $t^3$ in the $(i \mod 3 ,i \mod 3)$-th entry of the Hilbert series given in Proposition \[prop:Hilb\_ser\], that is, $$\dim \hat A_{0,3} = \dim \hat A_{1,4} =
1-5/4\,{n}^{2}+1/4\,{n}^{4}$$ or $$\dim \hat A_{2,5} = 1+1/2\,{n}^{4}-1/2\,{n}^{3}-2\,{n}^{2}.$$
To find $\dim B_{t, t+3}$, let us calculate the nonzero paths of length 3 in the algebra $B$. The integers $i,j,k,l$ below belong to the interval $[1,\dots,n]$. We have $$\dim B_{0, 3}$$ $$= {\mathop{\mathrm{Card}}}\{ e_{ij}x_k x_l|i<j, k\le l\} - {\mathop{\mathrm{Card}}}\{ e_{in}x_n x_l|i<n \} -{\mathop{\mathrm{Card}}}\{ e_{n-1,n}x_{n-1} x_l|l \le
n-1\}$$ $$= \frac{n(n-1)}{2}\frac{n(n+1)}{2} - n(n-1)-(n-1) =
\frac{n^4}{4} -\frac{5n^2}{4}+1,$$ $$\dim B_{1, 4} = \dim B_{0, 3} = \frac{n^4}{4} -\frac{5n^2}{4}+1
\mbox{ by symmetry}$$ and $$\dim B_{2, 5} = {\mathop{\mathrm{Card}}}\{ x_i e_{jk} x_l |j<k \}
- {\mathop{\mathrm{Card}}}\{ x_n e_{in} x_l|i<n \}
- {\mathop{\mathrm{Card}}}\{ x_{n-1} e_{n-1,n} x_l \}$$ $$- {\mathop{\mathrm{Card}}}\{ x_i e_{1k} x_1|k>1\} - {\mathop{\mathrm{Card}}}\{ x_i e_{12} x_2 \}
+{\mathop{\mathrm{Card}}}\{ x_n e_{1n} x_1 \}$$ $$=
n^2\frac{n(n-1)}{2} - \frac{n(n-1)}{2} - n
- \frac{n(n-1)}{2} - n +1 = \frac{n^4}2 -\frac{n^3}2-2n^2+1.$$ We obtain the equality $\dim B_{t, t+3} = \dim \hat
A_{t,t+3}$ for each $t=0,1,2$, thus, the algebra $A$ is PBW.
There is another proof of Proposition \[prop: A is PBW\] based on the Diamond Lemma and the Buchberger criterion for Gröbner bases.
Note that while the algebra $A$ is 3-periodic, we have shown only that $A$ is PBW as a 6-periodic algebra. We do not know whether $A$ is PBW as a 3-periodic algebra, that is, whether $\hat A$ is PBW or not.
Bounded processing {#subs:proc}
------------------
Given an algebra $R$ defined by a set of generators $X$ and a Gröbner basis of relations $G$ (given an admissible order of monomials on $X$), one can identify each element of the algebra with a linear combinations of the words on $X$ which are normal (=irreducible) with respect to $G$. Recall that the multiplication of normal words after this identification is defined as follows. We may assume that the Gröbner basis $G$ is reduced, that is, each its element $g\in G$ has the form $g = \hat g - \bar g$, where $\hat g$ is its leading monomial and $-\bar g$ is a linear combination of lower monomials. Given two normal words $u$ and $v$, one applies (if possible) to the concatenation $uv$ a [*reduction*]{} by some element of the Gröbner $g\in G$, that is, one replaces a subword $\hat g$ in $uv$ by the noncommutative polynomial $(-\bar g)$. In the resulted noncommutative polynomial, one applies to all its nonzero terms additional reductions, etc. After a finite number of steps, all nonzero terms of the resulted noncommutative polynomial $u*v$ became irreducible. By the definition of Gröbner basis, the linear combination of the normal words $u*v$ is defined uniquely and is identified with the product of $u$ and $v$ in the algebra $R$.
Thus, one can consider the calculation of the product $u*v$ as a processing of some “machine” (like a Turing machine, if $G$ is finite) which takes a concatenation $uv$ as an input, finds the first occurrence of a leading monomial $\hat g$ of some element of the Gröbner basis, replaces it by $-\bar g$, etc. Since the words $u$ and $v$ are normal, the first replaced subword $\hat g$ should overlap the both parts $u$ and $v$ of $uv$. In the next steps, the region of processing in each subword should overlap one of the words from some $\bar g$ which appeared in a previous step. An algebra $R$ is said to be an [*algebra of $r$-processing*]{} [@pi01] for some $r>0$ if the region of processing do not spread beyond $r$ letters to the right from the beginning of the right part $v$ of the initial word $uv$, that is, for each pair of normal words $u$ and $v = ws$, where the word $w$ has length at least $r$, we have $$u*v = (u*w)s.$$
Note that the above definition is compatible with the standard assumptions of the Gröbner basis theory for ideals in path algebras [@ffg]. In this theory, it is assumed that the above set of generators $X$ of a path algebra $kQ$ of a quiver $Q$ consists of two parts, $X = V \cup E$, where $V$ is the set of vertices and $E$ is the set of arrows of the quiver $Q$. By definition, the normal words are the paths of the corresponding quiver, that is, the paths of length 0 which are the vertices and the paths of positive length which are sequences of arrows.
We see that the definition of length of a normal word in the quiver algebra is slightly different from the general definition of length as the number of letters in word used above. However, if we try to check the property of $r$-processing (with $r\ge 1$) for a quotient of the path algebra by an ideal $I$ generated by linear combination of paths of positive length (quiver algebra $R = kQ /I$), then we can assume that the words $u,v$, and $w$ are paths of positive length. For these words $u,v$, and $w$, the both definitions of lengths coincide, so, one can use the second one.
In particular, the path algebra $kQ$ of any quiver $Q$ is an algebra of 1-processing. Another example of a quiver algebra with $r$-processing is our algebra $A$.
\[prop:3-proc\] The algebra $A$ is an algebra of 3-processing with respect to the generators and the Gröbner basis introduced in Subsection \[subs:PBW\].
Let $uv = \dots e_{ab}x_ix_je_{kl}x_sx_t\dots$ be a product of two normal words $u$ and $v$. Suppose that the right subword $x_sx_t\dots$ belongs to the second part $v$. It is sufficient to show that in each stage of the processing, the subword which begins with $x_t$ is stable. Since in each stage of the processing the region of processing is extended by at most one letter (because all leading monomials of the elements of the Gröbner basis have length 2), it is sufficient to show that this region does not rich $x_t$.
In the reductions w. r. t. the elements of the Gröbner basis $G$ from Subsection \[subs:PBW\], any two-letter word $c$ is reduced to a linear combinations of another two-letter words $c'$ with the following properties:
\(1) each letter $e$ is replaced by some $e$, and each $x$ is replaced by some $x$;
\(2) if $c = yz$ and $c' = y'z'$ for some letters $y,z,y',z'$, then $z\ge z'$.
In particular, we have
(3’) if $c= e_{\alpha \beta} x_{s}$, then $c'= e_{\alpha' \beta'} x_{s'}$ with $x_{s'} \le x_s$.
It follows from observation of the list of the leading monomials of elements of $G$ given in (\[eq:lead\_terms\]) that
\(4) if the monomial $x_sx_t$ is irreducuble (i. e., normal), then for each $x_{s'} \le x_s$ the monomial $x_{s'}x_t$ is irreducible too.
Since the monomial $x_sx_t$ is a subword of the normal word $v$, it is irreducible. By (1), in each stage of the processing this subword will be replaced by a linear combination of the words of the form $x_{s'}x_{t'}$, where $x_{s'} \le x_s$ by (3’). By (4), it follows that $x_{t'} = x_t$, that is, the region of processing does not reach $x_t$. It follows that $x_t$ and all letters to the right of it will be stable in each stage of the processing.
Due to the right–left symmetry in list of the leading monomials of the Gröbner basis, $A$ is an algebra of left $3$-processing as well.
Note that the sufficient condition for $r$–processing given in [@pi01 Prop. 2] does not hold for the algebra $A$.
Coherence
---------
\[cor:main\] The algebra $A$ is right and left coherent.
The right coherence of $A$ follows from Proposition \[prop:3-proc\] and [@pi01 Th. 8]. The left coherence follows from symmetry.
Note that it follows that $A$ is coherent both in graded and non-graded sense.
\[rem:uni-coh\] The property of $3$–processing implies also the following estimate for the degrees of relations of ideals [@pi01 Prop. 7]: If a right sided ideal $I$ in $A$ is generated in degreed $\le d$ for some $d$, then its relations are concentrated in degrees $\le d+6$. It follows that $A$ is universally coherent in terms of [@pi4], see also [@pi4 Prop. 4.10]. Similar linear estimates for the generators of the entries of the minimal projective resolution for each finitely presented $A$–module follow from [@pi4 Prop. 4.3].
[Biblio]{}
M. Artin and W.F. Schelter, [*Graded algebras of global dimension 3,*]{} Adv. Math., [**66**]{} (1987), pp. 171–216
M. Artin, J. J. Zhang, [*Noncommutative projective schemes,*]{} Adv. Math., [**109**]{} (1994), 2, p. 228–287
Bocklandt R., [*Graded Calabi Yau algebras of dimension 3*]{}, Journal of pure and applied algebra, [**212**]{} (2008), 1, pp. 14–32 (with Appendix by M. Van den Bergh)
Bondal A. I., [*Representation of associative algebras and coherent sheaves*]{}, Izv. Akad. Nauk SSSR Ser. Mat. [**53**]{} (1989), 1, pp. 25–44 \[Russian\]; translation in [**34**]{} (1990), 1, pp. 23–42
A. I. Bondal, A. Polishchuk, [*Homological properties of associative algebras: the method of helices*]{} (Russian. Russian summary), Izv. Ross. Akad. Nauk Ser. Mat. 57 (1993), no. 2, 3–50; translation in Russian Acad. Sci. Izv. Math. 42 (1994), 2, pp. 219–260
A. Bondal, M. Van den Bergh, [*Generators and representability of functors in commutative and noncommutative geometry,*]{} Mosc. Math. J., [**3**]{} (2003), 1, pp. 1–36, 258
A. Connes, M. Dubois-Violette, [*Yang-Mills algebra*]{}, Lett. Math. Phys., [**61**]{} (2002), 2, pp. 149–158
A. I. Efimov, V. A. Lunts, D. O. Orlov, [*Deformation theory of objects in homotopy and derived categories III: Abelian categories*]{}, Adv. Math., [**226**]{} (2011), 5, pp. 3857–3911
D. R. Farkas, C. D. Feustel, E. L. Green, [*Synergy in the theories of Gröbner bases and path algebras*]{}, Canadian Journal of Mathematics, [**45**]{} (1993), 4, pp. 727–739
Herscovich E., [*Representations of Super Yang-Mills Algebras*]{}, Communications in Mathematical Physics, [**320**]{} (2012), 3, pp. 783–820
He J., Van Oystaeyen F., Zhang Y., [*Graded 3-Calabi-Yau algebras as Ore extensions of 2-Calabi-Yau algebras*]{}, arXiv:1303.5293 (2013)
Movshev, M., Schwarz, A., [*Algebraic structure of Yang-Mills theory*]{}, The unity of mathematics, Progr. Math., Vol. 244, Birkhauser Boston, 2006, pp. 473–523 D. Piontkovski, [*Noncommutative Groebner bases, coherence of associative algebras, and divisibility in semigroups,*]{} Fundam. Prikl. Mat., [**7**]{} (2001), 2, pp. 495–513 \[Russian\] D. Piontkovski, [*Linear equations over noncommutative graded rings,*]{} J. Algebra, [**294**]{} (2005), 2, pp. 346–372
D. Piontkovski, [*Coherent algebras and noncommutative projective lines*]{}, Journal of Algebra, [**319**]{} (2008), 8, pp. 3280–3290
Polishchuk, A. Positselski, L. [*Quadratic algebras.*]{} University Lecture Series, 37. American Mathematical Society, Providence, RI, 2005. xii+159 pp.
A. Polishchuk, [*Noncommutative proj and coherent algebras*]{}, Math. Res. Lett., [**12**]{} (2005), 1, pp. 63–74
S. P. Smith, [*A 3-Calabi-Yau algebra with $G_2$ symmetry constructed from the octonions*]{}, arXiv:1104.3824 (2011)
[^1]: This study was supported by “The National Research University Higher School of Economics Academic Fund Program” in 2013–2014 (research grant 12-01-0134). It is also supported by RFBR project 14-01-00416.
[^2]: At least, for this algebra $A$ calculations of Hilbert series for several ideals $I$ such that the quotient algebra $A/I$ is Noetherian by natural reasons show that $I$ cannot be projective as a left module.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The complex N44 in the Large Magellanic Cloud (LMC) provides an excellent site to perform a detailed study of star formation in a mild starburst, as it hosts three regions of star formation at different evolutionary stages and it is not as complicated and confusing as the 30 Doradus giant region. We have obtained [*Spitzer Space Telescope*]{} observations and complementary ground-based 4m $uBVIJK$ observations of N44 to identify candidate massive young stellar objects (YSOs). We further classify the YSOs into Types I, II, and III, according to their spectral energy distributions (SEDs). In our sample of 60 YSO candidates, $\sim 65$% of them are resolved into multiple components or extended sources in high-resolution ground-based images. We have modeled the SEDs of 36 YSOs that appear single or dominant within a group. We find good fits for Types I and I/II YSOs, but Types II and II/III YSOs show deviations between their observed SEDs and models that do not include PAH emission. We have also found that some Type III YSOs have central holes in their disk components. YSO counterparts are found in four ultracompact regions and their stellar masses determined from SED model fits agree well with those estimated from the ionization requirements of the regions. The distribution of YSOs is compared with those of the underlying stellar population and interstellar gas conditions to illustrate a correlation between the current formation of O-type stars and previous formation of massive stars. Evidence of triggered star formation is also presented.'
author:
- 'C.-H. Rosie Chen, You-Hua Chu and Robert A. Gruendl'
- 'Karl D. Gordon'
- Fabian Heitsch
title: '[*Spitzer*]{} View of Young Massive Stars in the LMC HII Complex N44'
---
Introduction
============
Star formation frequently takes place in high concentrations of massive stars at intense levels, a phenomenon referred to as “starbursts.” The massive stars in a starburst can inject energies into the interstellar medium (ISM) to photoionize the ambient medium and to dynamically sweep up the medium into expanding shells. The expanding shells can trigger subsequent star formation either by compressing ambient dense clouds or by collect-and-collapse within the shell [@El98]. Starbursts can easily spread over areas 10$^2$-10$^3$ pc across, and play a vital role in determining the large-scale structures of their host galaxies.
While starbursts are the most prominent features in a galaxy, their detailed properties cannot be easily studied: in distant galaxies the stellar content is not resolved and in the Milky Way the distances and association among stars in a starburst are uncertain. The Large and Small Magellanic Clouds (LMC & SMC; MCs) are the only galaxies in which stars are at common, known distances and can be individually resolved down to $\sim$1 $M_\odot$ at optical wavelengths [@GBH06; @Noetal06]. Recent [*Spitzer Space Telescope*]{} observations in the mid-infrared (mid-IR) enabled the detection of individual young stellar objects (YSOs) in the MCs, revealing the on-going star formation [@CYetal05; @JTetal05; @CGC08]. It is now possible to use the resolved stellar population to map out star formation as a function of space and time in the MCs. As the ISM of the MCs has been surveyed in great detail, it is further possible to determine the relationship between star formation and the physical properties of the ISM. Consequently, studies of starbursts in the MCs allow us to investigate fundamental issues about star formation, such as whether and how the stellar energy feedback from massive stars triggers the formation of next-generation stars, whether the initial mass function (IMF) depends on the interstellar conditions for star formation, and whether massive stars are formed only under specific conditions [@ZY07].
We have chosen the complex LH$\alpha$ 120-N44 [N44; @He56] in the LMC to carry out a comprehensive study of star formation in a mild starburst situation, as it is one of the top ranking complexes in the LMC but is not as complicated and confusing as the giant region 30 Doradus. N44 contains three OB associations, LH47, 48, and 49 [@LH70], that are in different evolutionary stages and interstellar structures: LH47 in the central superbubble, LH48 in one contiguous region at the northeast rim of the superbubble, and LH49 in a group of regions to the southeast exterior of the superbubble (Figures. \[fig:n44opt\]a,b). Along the western rim of the superbubble exist a number of dense regions where star formation may have been triggered by the expansion of the superbubble [@OM95]. Surveys of CO in the N44 complex [Fig. \[fig:n44opt\]d; @FYetal01] show a high concentration of molecular gas at the western rim of the central superbubble, another high concentration at the regions around LH49, and a weaker concentration to the north of the superbubble where a couple of small regions are visible but no OB associations are identified.
The effects of stellar energy feedback are evident in N44. The [H$\alpha$]{} image (Fig. \[fig:n44opt\]a) shows the reach of stellar UV radiation. The X-ray image (Fig. \[fig:n44opt\]c) reveals 10$^6$ K gas heated by fast stellar winds and supernova explosions. While the expansion of the central superbubble may have triggered the intense star formation along its western rim, the hot gas outflow from the superbubble may be responsible for the onset of star formation at the northeastern extension of the south molecular peak.
To study the current star formation in N44, we have obtained [*Spitzer*]{} mid-IR and complementary ground-based optical and near-IR observations. These observations have been analyzed and the results are reported in this paper. Section 2 describes the observations and data reduction. Section 3 reports the initial identification of YSO candidates and Section 4 describes a further pruning of the YSO candidate list. In Section 5 we classify the YSOs, and in Section 6 we determine their physical properties by modeling their spectral shapes. In Section 7 we discuss the massive star formation properties. A summary is given in Section 8.
Observations and Data Reduction
===============================
We have obtained [*Spitzer*]{} mid-IR observations to diagnose YSOs. To extend the spectral energy distribution (SED) and to improve the angular resolution, we have further obtained ground-based optical and near-IR imaging observations of N44. We have also retrieved available images in the [*Hubble Space Telescope*]{} ([*HST*]{}) archive to examine the optical counterparts and environments of the YSOs.
[*Spitzer*]{} IRAC and MIPS Observations
----------------------------------------
Our [*Spitzer*]{} observations of N44 were made with the InfraRed Array Camera [IRAC; @FGetal04] on 2005 March 27 and the Multiband Imaging Photometer for [*Spitzer*]{} [MIPS; @RGetal04] on 2005 April 7. The IRAC observations were obtained in the 3.6, 4.5, 5.8, and 8.0 [$\mu$m]{} bands using the mapping mode to cover a $\sim30' \times 30'$ area in each band. The pixel size was 12 pixel$^{-1}$, and the common area covered by all four bands is $\sim30' \times 23'$. At each pointing, exposures were made with a five-point cyclic dithering pattern in the 30 s high dynamic range mode. The total integration time at each pointing is $\sim$ 150 s. The mosaicked maps were processed with the Spitzer Science Center (SSC) pipeline (ver. S11.4.0) and provided to us as part of the Post Basic Calibrated Data (PBCD) products.
The MIPS observations were made in the 24, 70, and 160 [$\mu$m]{} bands using the scan map mode at the medium scan rate. The mapping consists of sixteen $0\fdg5$ scan legs with a cross-scan step of 148$''$ to cover a region of $20' \times 30'$ in all three MIPS bands. The MIPS DAT version 3.00 [@GKetal05] was used for the basic processing and final mosaicking of the individual images. In addition, the 24 [$\mu$m]{} image has been corrected for a readout offset and divided by a scan-mirror-independent flat field, and the 70 and 160 [$\mu$m]{} images have been corrected for a pixel-dependent background using a low-order polynomial fit to the source-free regions. The final mosaicked maps have pixel scales of 125, 985, and 160 pixel$^{-1}$, and exposure times of roughly 160, 80, and 16 s at 24, 70, and 160 [$\mu$m]{}, respectively.
The common area covered by all IRAC and MIPS bands is $20' \times 23'$. To achieve a more complete view of the molecular cloud to the north of N44 superbubble, we have used the [*Spitzer*]{} survey of the LMC [SAGE, @MMetal06] to extend our IRAC 3.6 and 5.8 [$\mu$m]{} data in the declination direction. The final IRAC and MIPS field we have analyzed is $20' \times 25'$.
To carry out aperture photometry for point sources in the IRAC images we use the IRAF package `apphot`. First, the sources were identified with the automated source finding routine `daofind`, using parameters optimized to find the majority of point-like sources while minimizing the inclusion of peaks of extended dust emission. In regions containing multiple sources superposed on extended emission, `daofind` does not always identify the same sources at the same locations in the four IRAC bands. Therefore, we identified sources in each of the four IRAC bands and merged the four source lists into a master source list using 12 (= 1 pixel) as the criterion for coincidence. The resulting master list is then used for photometric measurements in all four bands. The photometric measurements were made with a source aperture of 36 (3-pixels) radius and an annular background aperture at radii of 36–84 (3–7 pixels). Finally, an aperture correction was applied, and the fluxes were converted into magnitudes using the correction factors and zero-magnitude fluxes provided in the IRAC Data Handbook and listed in Table \[photpar\].
The photometric measurements from long- and short-exposure IRAC observations were averaged with weights proportional to the inverse square of their errors. For sources that were saturated in the long exposures, their measurements from the short exposures were adopted. The results were used to produce a photometric catalog of 17,002 IRAC sources in the $20' \times 25'$ field of N44. This catalog is presented as an ASCII table and an example is shown in Table \[photcat\].
MIPS images have lower angular resolution. Source identification using `daofind` is feasible only for the 24 [$\mu$m]{} images. In the 70 and 160 [$\mu$m]{} images, point sources cannot be easily resolved from one another or from a bright diffuse background; therefore, the few apparent point sources were identified by visual inspection. The photometric measurements were made with parameters appropriate for the point spread functions (PSFs), as recommended in the MIPS Data Handbook and given in Table \[photpar\]. Note that the aperture corrections adopted for 70 and 160 [$\mu$m]{} measurements are those for sources of temperatures 15 and 10 K, respectively. For sources of higher temperatures, such as 1000–3000 K, the adopted aperture corrections will result in fluxes 8% too high in 70 [$\mu$m]{} and 2% too high in 160 [$\mu$m]{} because the source emission peaks at shorter wavelengths; however, these errors do not significantly affect the analysis and conclusions of our study of YSOs. In the 160 [$\mu$m]{} image, only one point source can be clearly identified; therefore, the 160 [$\mu$m]{} photometry will be discussed in the text but not included in the photometric catalog. The catalogs of the MIPS 24 and 70 [$\mu$m]{} bands are merged with the IRAC catalog and included in Table \[photcat\].
The [*Spitzer*]{} images of N44 in the 3.6, 8.0, and 24 [$\mu$m]{} bands are shown in Figure \[fig:n44img\]. The 3.6 [$\mu$m]{} image is dominated by stellar emission, the 8.0 [$\mu$m]{} image shows the polycyclic aromatic hydrocarbon (PAH) emission, while the 24 [$\mu$m]{} image is dominated by dust continuum emission [@LD01; @LD02; @DL07]. To better illustrate the relative distribution of emission in the different bands, we have produced a color composite with 3.6, 8.0, and 24 [$\mu$m]{} images mapped in blue, green, and red, respectively. In this color composite, shown in Fig. \[fig:n44img\]d, dust emission appears red and diffuse, stars appear as blue point sources, red supergiants appear yellow, and dust-shrouded YSOs and AGB stars appear red.
CTIO 4 m ISPI and MOSAIC Observations
-------------------------------------
We obtained near-IR images in the $J$ and $K_s$ bands with the Infrared Side Port Imager (ISPI) on the Blanco 4 m telescope at Cerro Tololo Inter-American Observatory (CTIO) on 2005 November 14–15. The images were obtained with the $2{\rm K} \times 2$K HgCdTe HAWAII-2 array, which had a pixel scale of 03 pixel$^{-1}$ and a field-of-view of $10\farcm25 \times 10\farcm25$. Six fields were observed to map N44. Each field was observed with ten 30 s exposures in the $J$ band and twenty 30 s exposures in the $K_s$ band (each of the latter was coadded from two 15 s frames to avoid background saturation). The observations were dithered to aid in the removal of transients and chip defects. Owing to the diffuse emission in N44, we obtained sky frames at $\sim20'$ south of N44 to aid in sky subtractions. The sky observations were made before and after each set of ten on-source exposures. All images were processed using the IRAF package `cirred` for dark and sky subtraction and flat-fielding. The astrometry of individual processed images was solved with the routine `imwcs` in the package `wcstools`. The astrometrically calibrated images are then coadded to produce a total exposure map for each filter. The flux calibration was carried out using 2MASS photometry of isolated sources.
We obtained the SDSS $u$ and Johnson-Cousins $BVI$ broadband images of N44 with the MOSAIC II CCD Imager on the CTIO Blanco 4 m telescope on 2006 February 2. The MOSAIC Imager consists of eight $2{\rm K} \times 4$K SITe CCDs. The CCDs have a pixel scale of 027 pixel$^{-1}$, yielding a total field-of-view of $36'\times36'$. The entire N44 was imaged in a single field. All $uBVI$ images have been processed with the standard reduction procedure: bias and dark were subtracted, flat-fielding was applied, and multiple frames in each filter were combined to remove cosmic rays and improve the S/N. The astrometry for the processed images were performed by referencing to stars in the USNO B1.0 catalog. The flux calibration was carried out using photometric measurements of isolated sources in @Ch07 and @ZDetal04.
Archival [*HST*]{} Images
-------------------------
We have searched the [*HST*]{} archive for Wide Field Planetary Camera 2 (WFPC2) images in the field of N44. The available observations are listed in Table \[wfpc2obs\], in which the coordinates, program ID, PI, filter, and exposure time are given. Most of the observations contain multiple exposures for the same pointing and filter, and these images are combined using the IRAF routine `crrej` to remove cosmic rays and produce a total exposure map. The astrometry was refined for each resultant image by referencing to stars in the USNO B1.0 catalog.
Seven fields have observations, but not all are useful; for example, the wide $U$ band (F300W) images have very low S/N. The most useful images are those taken with the [H$\alpha$]{} (F656N) or Strömgren $y$ (F547M) filter. The former shows ionized gas and the latter shows stars at high resolution.
Other Useful Archival Datasets
------------------------------
To construct SEDs for the sources in our [*Spitzer*]{} photometric catalog of N44, we have expanded the catalog by adding $UBV$ photometric measurements obtained from CCD images taken with the CTIO 0.9 m telescope [@Ch07]. In regions not covered by these 0.9 m $UBV$ observations, we use the $UBVI$ photometry from the Magellanic Cloud Photometric Survey [MCPS, @ZDetal04]. We have also added near-IR data from the Point Source Catalog of the Two Micron All Sky Survey [2MASS, @SMetal06].
When merging the datasets, we allow a 1$''$ error margin for matching [*Spitzer*]{} sources with optical or near-IR sources. The final catalog lists each source’s right ascension, declination, and magnitudes in the order of increasing wavelengths, i.e., $U$, $B$, $V$, $I$, $J$, $H$, $K_s$, 3.6, 4.5, 5.8, 8.0, 24, and 70 [$\mu$m]{}. The entire catalog is presented as an ASCII table and an example is shown in Table \[photcat\]. These magnitudes can be converted to flux densities using the corresponding zero-magnitude flux listed in Tables \[photpar\] and \[flux2mag\] and then used to construct SEDs for individual sources.
Finally, we have used [H$\alpha$]{} images of N44 from the Magellanic Cloud Emission Line Survey [MCELS, @SRetal99] to examine the large-scale distribution of dense ionized gas and to compare with images at other wavelengths. As the angular resolution of this survey is $\sim2''$, we have used additional [H$\alpha$]{} images taken with the CTIO 0.9 m telescope by M. A. Guerrero [@NYetal02]. These images, with a pixel size of 04 pixel$^{-1}$, are used to show the immediate environments of YSOs.
Initial Identification of Massive YSO Candidates
================================================
Expected Spectral Properties of YSOs
------------------------------------
YSOs have SEDs that differ from stars because they are shrouded in dust, which absorbs the stellar radiation and irradiates at IR wavelengths; therefore, it is possible to identify YSOs from their IR excesses. It is, however, difficult to decipher the IR excess of massive YSOs because the distribution of their circumstellar dust is not well known. As their formation mechanism is still uncertain [e.g., @SPH00; @ZY07], it is not even known whether massive YSOs ubiquitously possess accretion disks [e.g., @CRetal07].
We will use the commonly accepted physical structure and evolution of low-mass YSOs as a starting point to decipher the SEDs of higher-mass YSOs. Low-mass YSOs are believed to be initially surrounded by a small accretion disk and a large infalling envelope with bipolar cavities, and as they evolve, the envelope and disk dissipate. Different evolutionary stages result in different SEDs that have been used to classify low-mass YSOs, i.e., the Class I/II/III system [@La87]. In this classification system, a Class I YSO has a compact accretion disk and a large infalling envelope with bipolar cavities; its SED is dominated by emission from the envelope and rises longward of 2 [$\mu$m]{}. A Class II YSO has dispersed most of its envelope and is surrounded by a flared disk; its SED, dominated by emissions from the central source and disk, is flat or falls longward of 2 [$\mu$m]{}. A Class III YSO has cleared most of the disk so its SED shows stellar photospheric emission with little or no excess in near-IR.
These geometries of dust disk and envelope applicable to low-mass YSOs have been adopted in radiative transfer models of high-mass YSOs by a number of investigators [e.g., @WBetal04b; @RTetal06]. In their models, at an early evolutionary stage, a YSO has a small disk and a large envelope with narrow bipolar conical cavities; its SED is dominated by the envelope emission and shows a generally rising trend from the shortest detectable wavelength to beyond 24 [$\mu$m]{}. At an intermediate evolutionary stage, the opening angle of the bipolar cavity increases and the star and disk may be exposed; the YSO’s SED thus shows double peaks: one below 1 [$\mu$m]{} (stellar emission) with its intensity dependent on the viewing angle, and the other rising longward of 1 [$\mu$m]{} but turning flat or falling shortward of 10–20 [$\mu$m]{}. At a late evolutionary stage when most of the envelope and disk have been dispersed, the SED shows a bright stellar emission with a modest mid-IR excess. These models provide useful links between circumstellar dust structures and SEDs; thus, we will use the general trends of the SEDs to identify YSO candidates in N44.
Selection of Massive YSO Candidates
-----------------------------------
Owing to their excess IR emission, YSOs are positioned in redder parts of the color-color and color-magnitude diagrams (CMDs) than normal stars. However, background galaxies and asymptotic giant branch (AGB) stars can also be red sources, and these contaminants exist in non-negligible numbers. To separate YSOs from these contaminants, we have examined several color-color diagrams and CMDs with solely IRAC bands as well as combinations of IRAC with $JHK_s$ or MIPS bands [@GC08]. We find that for many sources the 2MASS catalog is too shallow to detect their counterparts in $JHK_s$, and the MIPS observations cannot resolve them from nearby sources or bright diffuse background. To include most of the sources, we decide to concentrate on diagnostic diagrams using only the IRAC bands. The IRAC \[3.6\]$-$\[4.5\] vs. \[4.5\]$-$\[8.0\] color-color diagram has been suggested by @Setal07 and the \[8.0\] vs. \[4.5\]$-$\[8.0\] CMD has been suggested by @HPetal06 to offer the best separation of YSOs from contaminants. We have adopted the latter approach for the initial selection of massive YSOs because galaxies and evolved stars have different distributions in brightness and can be effectively excluded using simple criteria.
Figure \[fig:cmds\] displays the \[8.0\] vs. (\[4.5\]$-$\[8.0\]) CMD of all sources detected in N44. The sources in the prominent vertical branch centered at $([4.5]-[8.0]) \sim 0.0$ are mostly main-sequence, giant, and supergiant stars. The contaminating background galaxies are concentrated in the lower part of the CMD, and the AGB and evolved stars are distributed mostly in the upper part of the CMD. Below we discuss the criteria used to exclude these contaminating sources from the high-mass YSO candidates.
### Excluding Normal and AGB Stars
Normal stars, i.e., main-sequence, giant, and supergiant stars, do not have excess IR emission and thus it is relatively easy to distinguish them from YSOs with a color criterion $([4.5]-[8.0]) < 0.5$ [@WBetal04a; @HPetal06]. On the other hand, evolved stars, e.g., AGB and post-AGB (pAGB) stars, can have circumstellar dust and show excess IR emission and red colors. To examine the locations of such evolved stars in the CMD, we first use known objects in the field of N44. Using the SIMBAD database, we have found 10 confirmed AGB stars at various evolutionary stages, such as carbon stars, M-type variables, IR carbon stars, and OH/IR stars. These 10 objects are marked by open squares in the \[8.0\] vs. \[4.5\]$-$\[8.0\] CMD (Fig. \[fig:cmds\]). These sources have a color range of 0.2 $<$ (\[4.5\]$-$\[8.0\]) $<$ 1.3 and a magnitude range of 6.4 $<$ \[8.0\] $<$ 11.4.
We have also used models for Galactic C- and O-rich AGB stars [@Gr06] to illustrate their expected locations in the CMD (Fig. \[fig:cmds\]). To avoid crowding, we plot only models for a stellar luminosity of 3000 $L_\odot$ to illustrate the range of colors. For a luminosity range $1\times10^3$ to $6\times10^4 L_\odot$ [@Po93], the expected loci of AGB stars in the CMD can move vertically by 1.2 to $-3.3$ mag. As the chemistry of AGB atmospheres is dominated by nucleosynthesized material, these Galactic models are good approximations for LMC objects although the LMC metallicity is only $1/3$ solar.
We adopt the (\[4.5\]$-$\[8.0\]) $\ge 2.0$ criterion to exclude normal and AGB/pAGB stars. As shown in Fig. \[fig:cmds\], this criterion does exclude all known AGB/pAGB stars and a great majority of AGB models. In our N44 field, 60 luminous sources are found with 0.5 $<$ (\[4.5\]$-$\[8.0\]) $<$ 2.0 and \[8.0\] $\le$ 12.0, and indeed almost all of them have SEDs consistent with those of AGB/pAGB stars, with the remaining few appearing to be normal stars contaminated by nebular background.
### Avoiding Background Galaxies \[bggal\]
Background galaxies are concentrated in the lower part of the CMD in Fig. 3, bounded by (\[4.5\]$-$\[8.0\]) $> 0.5$ and \[8.0\] $\ge 14.0-$(\[4.5\]$-$\[8.0\]), as suggested by @HPetal06 using the [*Spitzer*]{} Wide-Area Infrared Extragalactic Survey [SWIRE, @LCetal03]. We compare the surface density of sources bounded by \[8.0\] $\ge 14.0-$(\[4.5\]$-$\[8.0\]) and \[8.0\] $\le$ 13.0 in the CMD of N44 to that of the SWIRE Survey. For an area of $20'\times25'$, 284 sources are detected in N44 within this wedge of CMD. This surface density, 0.57 sources arcmin$^{-2}$, is much higher than that of SWIRE Survey, 0.06 sources arcmin$^{-2}$, although the SWIRE Survey had longer exposure times and thus higher sensitivities [@GC08]. This higher surface density of sources in N44 is most likely attributed to a population of low-mass YSOs, as such YSOs are expected to occupy this part of CMD for the LMC’s distance modulus of $\sim$18.5 [@Fe99; @WBetal04a; @RTetal06]. Therefore, as we adopt the criterion \[8.0\] $< 14.0-$(\[4.5\]$-$\[8.0\]) to exclude background galaxies, we have also excluded YSOs with masses $\lesssim4$ $M_\odot$.
### Massive YSO Candidates
After applying the two criteria $([4.5]-[8.0]) \ge 2.0$ and $[8.0] < 14 - ([4.5]-[8.0])$ to exclude most of normal and AGB stars and background galaxies, we obtain a list of 99 YSO candidates. To search for additional candidates that are more embedded and hence not detected in the 4.5 [$\mu$m]{} band, we resort to the 24 [$\mu$m]{}sources. Only one 24 [$\mu$m]{} source does not have a corresponding 4.5 [$\mu$m]{} source. This source appears extended in the 8.0 [$\mu$m]{} image, indicating that it is likely a small interstellar dust feature. Therefore, no new objects are added to the list of 99 YSO candidates.
The 99 YSO candidates selected from the above CMD criteria still include a significant number of small dust features, obscured evolved stars, and bright background galaxies. These contaminants need to be examined closely to assess their nature and need to be excluded from the YSO list. In the next section we discuss how we use SEDs and multi-wavelength images to confirm and classify YSOs.
Further Pruning of YSO Candidates
=================================
To differentiate between YSOs and contaminants, we examine each YSO candidate’s morphology, environment, brightness, and SED to utilize as much information as possible in our consideration. We have prepared the following multi-wavelength images with identical field-of-view for each YSO candidate: MCELS [H$\alpha$]{}; MOSAIC $B$, $V$, and $I$; ISPI $J$ and $K_s$; IRAC 3.6, 4.5, 5.8, and 8.0 [$\mu$m]{}; and MIPS 24 and 70 [$\mu$m]{} observations. For some YSO candidates that have [*HST*]{} WFPC2 [H$\alpha$]{} and Strömgren $y$ images available, these [*HST*]{} images replace the ground-based [H$\alpha$]{} and $V$ images. For YSO candidates that are not covered by our ISPI observations, 2MASS images are used. In addition, we have constructed an SED for each source from optical to 70 [$\mu$m]{} using the extended photometric catalog described in §2.4. For each YSO candidate, we display and examine the multi-wavelength images, its SED, and its location in the \[8.0\] vs. (\[4.5\]$-$\[8.0\]) CMD simultaneously. The nature of each YSO candidate is assessed by three of us (Chen, Chu, and Gruendl) independently multiple times. We gain experience from each round of examination and use our new knowledge to aid in the next round of examination. For most sources our classifications converge, but a few sources have ambiguous properties and their classifications are thus uncertain.
Identification of Background Galaxies
-------------------------------------
Background galaxies can be identified from their morphologies if they are resolved. To diagnose unresolved background galaxies, we have to resort to SEDs. Galaxies co-located with YSO candidates in the \[8.0\] vs. (\[4.5\]$-$\[8.0\]) CMD are abundant in gas and dust, such as late-type galaxies or active galactic nuclei (AGN). The SEDs of late-type galaxies are characterized by two broad humps, one from stellar emission over optical and near-IR wavelength range and the other from dust emission over mid- to far-IR range. The observed SED of an AGN depends on the viewing angle with respect to its dust torus; it can be flat from optical to far-IR or obscured in optical, showing only mid- to far-IR emission [@FAetal05; @HEetal05; @RRetal05]. In the latter case, the SED can be falling or rising at 24 [$\mu$m]{}. We further use the interstellar environment as a secondary diagnostic for background galaxies, since galaxies are statistically less likely to occupy preferred positions in prominent dust filaments. Based on these considerations, two of our CMD-selected YSO candidates, sources 052042.0$-$674307.7 and 052106.8$-$675715.9, are reclassified as background galaxies.
Identification of Evolved Stars
-------------------------------
The known AGB stars in N44 have SEDs peaking between 1 and 8 [$\mu$m]{}, as shown in Figure \[fig:agbseds\]. These SEDs are similar to those of Galactic AGB stars, corresponding to dust temperatures of $\sim 400-1,000$ K [@RRetal86]. These known AGB stars in N44 are bluer than our selection criteria for YSOs; however, the more obscured AGBs may have lower dust temperatures, exhibit redder colors, and occupy the same regions as the YSO candidates in the CMD [e.g., @Betal06]. We identify such obscured AGB or evolved stars based on their SEDs, whose shapes are similar to those shown in Fig. \[fig:agbseds\] but peaking at longer wavelengths. We have also used the interstellar environment as a secondary criterion to identify AGB and evolved stars, since evolved stars are not expected to be located at preferred positions in diffuse dust emission. From these considerations, two of our YSO candidates, sources 052221.0$-$680515.3 and 052351.1$-$675326.6, are reclassified as AGB/evolved stars.
Identification of Dust Clumps
-----------------------------
Warm interstellar dust may show SEDs similar to those of circumstellar dust in YSOs. As the angular resolution of IRAC images is $\sim 2''$, corresponding to 0.5 pc for a LMC distance of 50 kpc, some small dust clumps may be identified as point sources and included in the YSO candidate list. Owing to our conservative compilation of master source list for IRAC photometry (§2.1), a star projected near a dust clump may also make its way into our YSO candidate list. To identify these two types of YSO imposters, we use the ISPI $JK_s$ images that have higher angular resolution. In the ISPI $K_s$ images, stars or YSOs appear unresolved, while dust clumps may appear as extended emission. The $J - K_s$ color further differentiates between stars and YSOs, as stars are brighter in $J$ and YSOs brighter in $K_s$. Aided by the ISPI images, we find that 23 of our 99 YSO candidates are interstellar dust clumps, and another 12 are stars projected near dust clumps.
Final Massive YSO Sample
------------------------
The results of our examination of 99 YSO candidates are given in Table \[ysoclass\], which lists source name, ranking of the brightness at 8 [$\mu$m]{}, magnitudes in the $U$, $B$, $V$, $I$, $J$, $H$, $K_s$, 3.6, 4.5, 5.8, 8.0, 24, and 70 [$\mu$m]{} bands, source classification, and remarks. Note that some of the YSOs appear as single sources in IRAC images but are resolved into multiple sources in ISPI and MOSAIC images; for these sources, the photometric measurements in Table \[ysoclass\] were made for the dominant YSO sources and these entries are remarked in Column “Flag”.
After excluding background galaxies, evolved/AGB stars, and dust clump imposters, 60 YSO candidates remain. As these are most likely bona fide YSOs, we will simply call them YSOs in the rest of the paper.
Comparison with YSO Samples Identified by Others
------------------------------------------------
YSOs in N44 have been identified using other methodologies by, e.g., @GC08 and @WBetal08. @GC08 used the same [*Spitzer*]{} data of N44 and similar selection criteria, but identified only 49 YSOs. This is understandable because their automated search for point sources was unable to find faint sources near a bright neighbor or over a bright background. Such faint sources can be found and measured more easily with human intervention, as we did in this paper for the small area of N44 only.
The comparison with the YSO sample of @WBetal08 is not as straightforward because they used a very complex set of criteria. Within the area of N44, they found only 19 YSO candidates, of which one was identified as a planetary nebula (PN). In Figure \[SAGE\_comp\]a, we mark the locations of these YSO candidates and our YSOs in the 8 [$\mu$m]{} image of N44, and in Fig. \[SAGE\_comp\]b, we compare our sources with @WBetal08 YSO candidates in a \[8.0\] vs. (\[4.5\]$-$\[8.0\]) CMD. The differences can be summarized as the following:
1. Within the YSO wedge bounded by \[4.5\]$-$\[8.0\] $\ge$ 2.0 and \[8.0\] $\ge 14.0-$(\[4.5\]$-$\[8.0\]), we find 60 YSOs, while @WBetal08 find only 13 YSO candidates, a subset of our sample. To investigate why the majority of the YSOs in N44, including bright sources with \[8.0\] $<$ 7.0, are missed in Whitney et al., we further compare our sources with their original catalog, i.e., the SAGE point source catalog (kindly provided by Remy Indebetouw and Marta Sewilo). We find that the missed YSOs do not result from different selection criteria, but from defining point sources and applying signal-to-noise (S/N) thresholds when making final lists in these two studies. For a survey of the entire LMC, the SAGE catalog used stringent parameters for the automated search for point sources and Whitney et al. applied a high S/N threshold. The former discards any irregular point sources compared to IRAC PSFs. The latter tends to exclude sources near bright neighbors or over bright background, as shown in Fig. \[SAGE\_comp\]a, since varying backgrounds make photometry more difficult, and the SAGE pipeline increases the photometric uncertainties in such areas, reporting conservative S/N ratios.
2. Six YSO candidates of the Whitney et al. sample were rejected by our stringent color-magnitude criteria for YSO selection. The SEDs of these six sources are shown in Figure \[SAGE\_comp\]c. With the addition of optical fluxes in the SEDs, it can be more clearly seen that at least YSO candidate W546 is likely an evolved star with circumstellar dust and that W530 and W574 may be background galaxies.
We conclude that our sample of massive YSOs in N44 is the most complete among the three compared.
Classification of Massive YSOs
==============================
As mentioned earlier, there is not yet a well-defined classification system for massive YSOs. In Robitaille et al.’s (2006) study of YSO models, they suggested a physical classification scheme for YSOs of all masses. This classification is analogous to the Class scheme for low-mass YSOs, but uses physical quantities, instead of the slope of the SED, to define the evolutionary stage of the models. The three stages are defined by the ratio of envelope accretion rate ($\dot{M}_{\rm env}$) to central stellar mass ($M_\star$) and the ratio of disk mass ($M_{disk}$) to $M_\star$: Stage I sources have $\dot{M}_{\rm env}/M_\star > 10^{-6}$ yr$^{-1}$, Stage II sources have $\dot{M}_{\rm env}/M_\star < 10^{-6}$ yr$^{-1}$ and $M_{\rm disk}/M_\star > 10^{-6}$, and Stage III sources have $\dot{M}_{\rm env}/M_\star < 10^{-6}$ yr$^{-1}$ and $M_{\rm disk}/M_\star < 10^{-6}$. These physical quantities are not directly observable; they have to be inferred from model fits to the observed SEDs. This classification scheme, while physical, is model-dependent and the comparison with observations requires data taken with sufficient angular resolutions to separate among individual sources and from backgrounds, which is not always the case for LMC objects. Thus, it is not an appropriate description of observed properties of YSOs.
Here we suggest an empirical classification of the observed SEDs of massive YSOs, and use a “Type” nomenclature for distinction from the “Classes” for low-mass YSOs and the “Stages” for model YSOs. In our scheme:
- [**Type I**]{} has an SED rising steeply from near-IR to 24 [$\mu$m]{} and beyond, indicating that the radiation is predominantly from a circumstellar envelope. Examined in images, a Type I YSO is not visible at optical or even in the $J$ band, but emerges in the $K_s$ band and continuously brightens up all the way to 24 and 70 [$\mu$m]{}. Type I YSOs are usually in or behind dark clouds; an example is given in Figure \[fig:ysoimg\]a, where the YSO’s optical and IR images, SED, and location in the \[8.0\] vs. (\[4.5\]$-$\[8.0\]) CMD are shown.
- [**Type II**]{} shows an SED with a low peak at optical wavelengths and a high peak at 8–24 [$\mu$m]{}, corresponding respectively to the central source and the inner circumstellar material (e.g., a disk) and indicating that the envelope has dissipated. In images, a Type II YSO would appear faint at optical, brighten up from $J$ to 8 [$\mu$m]{}, and appear faint again at 24 [$\mu$m]{}. An example of a Type II YSO is given in Fig. \[fig:ysoimg\]b.
- [**Type III**]{} shows an SED with bright optical stellar emission and modest dust emission at near- and mid-IR, indicating that the young star is largely exposed but still surrounded by remnant circumstellar material. In direct images, a Type III YSO is bright at optical, but fades toward longer wavelengths. It may even be surrounded by an region; as described later in §7.3, one of the Type III YSOs has an [*HST*]{} [H$\alpha$]{} image available and it shows a compact region. An example of a Type III YSO is given in Fig. \[fig:ysoimg\]c.
As our classification is based on mainly the SEDs and as the observed SED of a somewhat evolved YSO with bipolar cavities is dependent on its viewing angle, an unknown fraction of our Type I YSOs may have the same physical structure of the Type II YSOs but viewed outside the cavity. Moreover, not all YSOs can be classified unambiguously into these three types as some YSOs may be transitional between two evolutionary stages. A number of sources show SEDs resembling two adjacent types; for these we simply classify them as I/II or II/III.
Additional complications arise because of the limited angular resolution of [*Spitzer*]{} and complex surroundings of YSOs. First, IRAC’s $\sim 2''$ angular resolution translates into 0.5 pc in the LMC, which can hide a small group of YSOs. Indeed, 20 of the YSOs are resolved by our MOSAIC or ISPI images into multiple sources or even small clusters within the IRAC PSF, and another 19 YSOs appear more extended than the PSF of MOSAIC and ISPI. While the ISPI $JK_s$ images allow us to identify the dominant YSO among the multiple sources for accurate photometric measurements, the IRAC measurements correspond to the integrated light of all sources, causing uncertainty in the classification, especially when two or more YSOs are present within the IRAC PSF. Fig. \[fig:ysoimg\]d shows an example of a YSO resolved into multiple sources in MOSAIC and ISPI images. Second, YSOs are often found in dark clouds and dust columns. These interstellar features can be identified in the optical images as dark patches against dense stellar background or photoionized bright-rimmed dust pillars. The emission from these dust features can be blended with that of the YSOs, especially at the 24 [$\mu$m]{} band, causing large uncertainties in the YSO classification. Fig. \[fig:ysoimg\]e shows an example of a YSO at the tip of a dust column in N44F.
Our classification of the 60 YSOs and remarks on their multiplicity and association with dark cloud and dust column are given in Table \[ysoclass\]. Although for multiple systems the IRAC and MIPS fluxes are contaminated by other sources, these fluxes are most likely dominated by the most massive (and hence the brightest) YSO since the luminosity of a YSO is proportional to its mass cubed [@BM96]. Thus, for a YSO that appears single or is clearly the dominant source within the IRAC PSF, its SED can be modeled to assess its physical properties, as discussed in the next section.
Determining YSO Properties from Model Fits of SEDs
==================================================
Modeling the SEDs
-----------------
The observed SED of a YSO can be compared with model SEDs, and the best-fit models selected by the $\chi^2$ statistics can be used to infer the probable ranges of physical parameters for the YSO. This can be carried out with the fitting tool Online SED Model Fitter[^1] [@RTetal07]. The database of this fitting tool includes model SEDs pre-calculated for 20,000 YSO models each with 10 different viewing angles. A typical YSO model assumes a stellar core, a flared accretion disk, and a rotationally flattened infalling envelope with bipolar conical cavities. To fully describe this complex structure, each model is defined by 14 parameters [@RTetal06]. The stellar parameters include the star’s mass ($M_\star$), radius ($R_\star$), and temperature ($T_\star$); these are sets of parameters at different ages ($t_\star$) along the pre-main-sequence stellar evolutionary tracks modeled by @BM96 and @SLetal00. For a given set of $M_\star$, $R_\star$, and $T_\star$, the stellar radiation determined from stellar atmosphere models of @Ku93 or @BH05 is adopted and fed to the radiative transfer code to produce the YSO’s model SED. The disk parameters include the disk’s accretion rate ($\dot{M}_{\rm disk}$), mass ($M_{\rm disk}$), inner radius ($R^{\rm min}_{\rm disk}$), outer radius ($R^{\rm max}_{\rm disk}$), scale height factor ($z_{\rm factor}$), and flaring angle ($\beta$). The envelope parameters include envelope accretion rate ($\dot{M}_{\rm env}$), envelope outer radius ($R^{\rm max}_{\rm env}$), cavity density ($\rho_{\rm cavity}$), cavity opening angle ($\theta_{\rm cavity}$), and the ambient density ($\rho_{\rm ambient}$) at which the density of the envelope reaches the lowest value as the density of the ambient ISM. Outside of the envelope is the foreground extinction ($A_V$). The 20,000 YSO models are generated to sample each of the disk and envelope parameters and the ambient density for a variety of pre-main-sequence stellar model [@RTetal06].
We have used the above SED fitting tool to analyze 36 YSOs in our sample that appear single or are clearly the dominant source within the IRAC PSF. The input parameters of the SED model fitter include the fluxes of a YSO and uncertainties of the fluxes. The uncertainty of a flux has two origins: the measurement itself and the absolute flux calibration. The fluxes and their measurement errors are given in Table \[ysoclass\]. The calibration errors are 5% in $U$, 3% in $B$, $V$, 10% in $I$, $J$, $K_s$, 3.6, 4.5, 5.8, 8.0, and 24 [$\mu$m]{}, and 20% in 70 [$\mu$m]{} (Chen 2007; Gruendl & Chu 2008; IRAC Data Handbook; MIPS Data Handbook). The total uncertainty of a flux is thus the quadratic sum of the measurement error and the calibration error. The resulting best-fit and acceptable SED models are shown in Figure \[fig:fit\], in which the YSOs are arranged in the order of Types I, I/II, II, II/III, and III (from our empirical classification in §5), and within each type in order of increasing \[8.0\] magnitude. In each panel, data points are plotted in filled circles and upper limits are plotted in filled triangles; error bars are plotted but they are usually smaller than the plot symbols. The best-fit model, with minimum $\chi^2$, is shown in solid black line, and the radiation from the stellar core reddened by the best-fit $A_V$ is shown in dashed black line. For most YSOs, their SEDs can be fitted similarly well by a range of models, and these “acceptable” models are plotted in grey lines (see Robitaille et al. 2007 for definitions of $\chi^2$ and acceptable fits).
The results of the SED model fits are given in Table \[sedfits\], where the YSOs are listed in the same order as in Fig. \[fig:fit\]. The source name, \[8.0\] magnitude, and type from our empirical classification are listed to the left of the table, and physical parameters of the best-fit models are listed to the right. Among the 14 parameters of each model, we have only listed the stellar parameters ($M_\star$, $T_\star$, $R_\star$, $t_\star$), accretion rates ($\dot{M}_{\rm env}$, $\dot{M}_{\rm disk}$), disk mass ($M_{\rm disk}$), and foreground extinction ($A_V$). We have also listed the viewing angle ($i$, the angle between the sightline and the polar axis), and the total luminosity ($L_{\rm tot}$). In addition to the parameters of the best-fit model, we have also used the acceptable models to show a possible range of stellar mass ($M_\star$ Range), and used $\dot{M}_{\rm env}/M_\star$ and $M_{\rm disk}/M_\star$ from the acceptable models to estimate a possible range of its evolutionary stage, the Stage Range, as defined by @RTetal06. The possible ranges of stellar mass and evolutionary stage are also given in Table \[sedfits\].
The SED fits of the 36 YSOs, as displayed in Fig. \[fig:fit\], show different degrees of goodness-of-fit among our empirically defined YSO types. The seven Type I and I/II YSOs have model SEDs agree well with the observed SEDs, although the best and acceptable models span a large mass range. The 21 Type II and II/III YSOs show good agreement between model and observed SEDs, except at 4.5 [$\mu$m]{}, where the observed fluxes are systematically lower than the modeled. This discrepancy will be discussed below in §6.2.1. Among the eight Type III YSOs analyzed, only three of them show good agreement in the SED fits, while the other five exhibit significant discrepancies between model and observed SEDs. These discrepancies and possible causes will be discussed in §6.2.2.
Significant Discrepancies Between Model and Observed SEDs
---------------------------------------------------------
### PAH Emission in Massive YSOs
A great majority of our Type II and III YSOs show a brightness dip at 4.5 [$\mu$m]{} in their SEDs (Fig. \[fig:fit\]). This dip is not an absorption feature as it appears. Instead, it is caused by PAH emission features at 3.3, 6.2, 7.7, and 8.6 [$\mu$m]{} in the other three IRAC bands [@LD01; @LD02; @DL07]. The emission mechanism of these features is ultraviolet (UV) excitation of PAHs followed by IR fluorescence [@ATB89]; thus these emission features have been observed in regions with UV radiation and dust, such as disks around Herbig Ae/Be stars [e.g., @RB03; @VBetal04] and photodissociation regions (PDRs) surrounding regions [e.g., @HT97]. Our Type II and III YSOs have similar physical conditions, and the resultant PAH emission cause the apparent dip at 4.5 [$\mu$m]{} in the SEDs.
The PAH emission of the YSOs is unlikely to originate from a large-scale, diffuse interstellar structure, e.g., a superbubble rim, since the contamination from such interstellar emission is minimized by background subtraction in the photometric measurement. Furthermore, there is a correlation between PAH emission and our YSO types: PAH emission is not observed in Type I, appears in some of the Type I/II YSOs, and is almost ubiquitous in Type II and III YSOs. This correlation argues against the large-scale interstellar origin. Therefore, the PAH emission in our YSOs is most likely of a circumstellar origin, such as a disk or a compact region, or both. A Type I YSO is deeply embedded in a massive envelope, of which only the dust continuum is observable, and thus no PAH emission is expected. A Type II YSO may show prominent PAH emission from the disk, if viewed through its bipolar cavities. A Type III massive YSO will ionize its surrounding gas and show PAH emission from its PDRs. While the SEDs reveal signatures of PAH emission, the geometry of the emitting region (whether a disk or a PDR of an region) cannot be determined without more detailed spectral information.
The SEDs of massive YSOs show evidence of PAH emission, but the SED models of @RTetal06 do not include PAHs or small grains. This fundamental difference causes the dust continuum determined from SED fitting to be biased by the three IRAC bands that contain PAH emission. Consequently, the observed 4.5 [$\mu$m]{} fluxes are below the model fits. Moreover, the error in the best-fit continuum level is propagated to the disk parameters, which are predominantly determined from the fluxes in the IRAC bands. Thus, SED model fits for YSOs with PAH emission may have large errors in the viewing angle or disk mass. These errors may be further propagated into the simultaneously fitted envelope and stellar parameters, but the effects are unlikely to be large because the stellar and envelope parameters are determined mainly by the SED at optical and mid-IR wavelengths, respectively.
### Discrepancies in Type III YSOs
Among the eight Type III YSOs, five show significant discrepancies between the best-fit models and the observed SEDs: 052129.7$-$675106.9, 052315.1$-$680017.0, 052157.0$-$675700.1, 052159.6$-$675721.7, and 052340.6$-$680528.5 (Fig. \[fig:fit\]). The brightest object, 052129.7$-$675106.9, is peculiar and will be discussed later in §7.2. The other four YSOs show deep V- or U-shape SEDs with the flex point in the IRAC bands, where the best-fit model fluxes are much higher than the observed fluxes, exhibiting the most prominent discrepancies.
To understand these discrepancies, we can divide the SED of a YSO into three segments, optical/near-IR ($UBVIJK_s$), near/mid-IR (IRAC bands), and mid/far-IR (MIPS bands and beyond), and look into the dominant origin of emission for each segment. At the shortest wavelengths, the optical/near-IR segment, the radiation is dominated by stellar photospheric emission; for example, Figure \[fig:sfit\] shows the optical/near-IR segment of the SEDs of four Type III YSOs can be well fitted by stellar atmosphere models [@Ku93]. At longer wavelengths, the stellar emission diminishes and dust emission rises. The near/mid-IR segment consists of emission from PAHs and warm dust, while the mid/far-IR segment consists of emission mostly from colder dust. Qualitatively, warm dust that emits at near/mid-IR wavelengths needs to be near the star in order to reach the required emitting temperatures; therefore, the near/mid-IR segment of a YSO’s SED is dominated by disk emission. Colder dust that emits at mid/far-IR wavelengths is farther away from the star, and thus the mid/far-IR segment of a YSO’s SED is dominated by envelope emission.
The aforementioned discrepancies between observations and best-fit models of SEDs of Type III YSOs in the IRAC bands suggest that the YSOs do not have as much disk emission as in the models. Below we analyze the SEDs quantitatively to understand their physical implications. The mid- to far-IR parts of the V- or U-shaped SEDs of these Type III YSOs (Fig. \[fig:fit\]) suggest that the dust continuum peaks at $\ge$24 [$\mu$m]{}. The corresponding blackbody temperature of the dust is thus $\le$120 K. The distance of such dust temperature to the central star can be estimated if the stellar spectral type and the implied stellar effective temperature are known. Among these four YSOs, 052157.0$-$675700.1, 052159.6$-$675721.7, and 052340.6$-$680528.5 have been spectroscopically classified as O8.5V$+$neb, O7.5V$+$neb, and B\[e\] stars, respectively [@OM95], and 052315.1$-$680017.0 has $(U-B)=-0.79$, $(B-V)=0.09$, and $V=14.10$ (Table \[ysoclass\]) that are consistent with the colors and magnitudes of a B0-O5 V star [@SK82] reddened by $A_V \sim 1.2$. For a dust temperature of 120 K and an albedo of 0.5, the closest distance of such dust to a B0 V star with a stellar effective temperature of 30,000 K is 860 AU; this distance is larger for stars with higher effective temperatures. This result implies a low dust content, or a hole, within the central $\sim$1000 AU.
While our assessment of the observed SEDs suggests $R^{\rm min}_{\rm disk}$ $>$ 860 AU, no models with $R^{\rm min}_{\rm disk}$ $>$ 100 AU are available in the database of pre-calculated SEDs for model fitting [@RTetal07]. To illustrate how $R^{\rm min}_{\rm disk}$ affects the SEDs, we have used the @WBetal03 radiative transfer code[^2] to calculate SEDs for a YSO model with stellar parameters appropriate for the observed spectral types but with customized disk parameters. The YSO model \#3019532 is chosen because its $T_\star=$ 34718 K and $R_\star=$ 5.628 $R_\odot$ are representative of late-type O stars. In this Stage III YSO model, the envelope has dissipated completely. We have calculated two SEDs with disk radii of 50–1000 AU and 1000–2000 AU, respectively; these model SEDs are overplotted on the observed SED of 052157.0$-$675700.1 in Figure \[fig:fit\_hole\]. The model SED with $R^{\rm min}_{\rm disk}$ = 50 AU show prominent excess emission in the near/mid-IR bands, while the model SED with $R^{\rm min}_{\rm disk}$ = 1000 AU adequately reproduce the V-shaped SED without excess near/mid-IR emission. This comparison suggests that dust has been cleared out to nearly 1000 AU. This clearing of dust is consistent with the expectation from YSOs at late evolutionary stages. We thus conclude that the large discrepancies between the best-fit models and the observed SEDs are caused by the absence of appropriate model SEDs in the database of the Online SED Model Fitter.
Similar V-shaped SEDs have been observed in Herbig Ae/Be stars RCW34 and TY Cra, and the SED modeling suggested an analogous evolutionary stage [@HLetal92]. Similar near-IR dips have also been observed in SEDs of the low-mass YSOs DM Tau and GM Aur; these dips have been interpreted as the clearing of dust from the inner regions of their disks [@RWetal03; @CNetal05].
Evolutionary Stage of YSOs
--------------------------
### Type vs. Stage
We use our analysis of 36 YSOs to compare our empirical classification, Type, with the theoretical classification, Stage [@RTetal06]. For each YSO, the $\dot{M}_{\rm env}/M_\star$ and $M_{disk}/M_\star$ ratios from the best-fit and acceptable models are used to determine its Stages, and the possible range of Stage is listed in Table \[sedfits\] under the column “Stage Range”. At first glance, there does not seem to be any systematic correspondence between Types and Stages.
The lack of overwhelming correspondence may be attributed to three reasons. First, the definitions of Stages are somewhat arbitrary, and it is not even known whether the Stages can be universally applied to YSOs of all masses. Furthermore, a Stage II YSO viewed along directions near the disk plane will mimic a Stage I YSO in the SED, causing ambiguity. Second, the 24 [$\mu$m]{} fluxes are vital in constraining the model fits, but they are the most uncertain measurements, especially for the faint YSOs. For example, the uncertain 24 [$\mu$m]{} fluxes of Type I YSOs 052216.9$-$680403.6 and 052211.9$-$675818.1 cause their model fits to be ill constrained (see Fig. \[fig:fit\]). Third, the limited angular resolution of [*Spitzer*]{} may cause inclusion of extraneous dust emission from unresolved regions. This is particularly relevant to YSOs at late evolutionary stages. For example, the Type III YSO 052207.3$-$675819.9, as shown in §7.3, has a small region that is resolved only in [*HST*]{} WFPC2 images; without these WFPC2 images, circumstellar disk or envelope would have been invoked to explain the YSO’s far-IR emission.
Still, we expect the extreme Stage I and III YSOs to correspond to our Type I and III, respectively, since these extreme types are either deeply embedded in massive envelopes or have little circumstellar material so that their SEDs are less dependent on the viewing angles. Examined closely, it can be seen that Type I YSOs with good flux measurements at 24 [$\mu$m]{} are indeed well fitted by models at Stage I, and Type III YSOs without unresolved regions indeed correspond to Stage III, if they have good 24 [$\mu$m]{} flux measurements and the model fits do not have near/mid-IR excesses (see §6.2.2). In summary, we conclude that meaningful comparisons between Types and Stages can be made only if good mid/far-IR flux measurements and high angular resolution images are available.
### Evolutionary Stages of YSOs in Ultra-compact Regions
Some young massive stars are known to be associated with small (diameter $\le10^{17}$ cm), dense ($\ge10^4$ cm$^{-3}$) regions of ionized gas, called ultracompact regions (UCHIIs). Five UCHIIs have been identified in N44: B0522$-$5800, B0523$-$6806(NE), B0523$-$6806(SE), B0523$-$6806(SW), and B0523$-$6806 [@IJC04]. All five UCHIIs are coincident with [*Spitzer*]{} sources within 1$''$. The [*Spitzer*]{} counterpart of B0523$-$6806(SE), 052324.8$-$680641.6, is faint and falls below the cutoff line of \[8.0\] $\ge 14.0-$(\[4.5\]$-$\[8.0\]) in the \[8.0\] vs. \[4.5\]$-$\[8.0\] CMD, where background galaxies and YSOs with masses $\le4~M_\odot$ are located. As low-mass stars cannot produce UCHIIs, this source is most likely a background star-forming galaxy. We have examined the MOSAIC $I$, ISPI $JK_s$, and [*Spitzer*]{} IRAC and MIPS images of this source. The $I$ and $J$ images show a brighter source and a faint source separated by less than 1$''$. The faint source becomes the brighter of the two in the $K_s$ band, and is the dominant source in the [*Spitzer*]{} bands. This source is not in a giant molecular cloud or surrounded by bright diffuse PAH emission. These results support the identification of B0523$-$6806(SE), or 052324.8$-$680641.6, as a background star-forming galaxy.
The [*Spitzer*]{} counterparts of the other four UCHIIs are bright and have been identified as Type I-II YSOs (see Table \[uch2\]). As these sources are among the top 10 most luminous YSOs in N44 at 8.0 [$\mu$m]{}, they have good 24 [$\mu$m]{} flux measurements and consequently models of their SEDs are well constrained. The stellar masses determined from the best-fit models to their SEDs can be translated into spectral types, assuming the relationship for main sequence stars. These spectral types can be compared with those implied by the ionizing fluxes determined from radio continuum observations [@IJC04]. As seen in Table \[uch2\], excellent agreement exists between these two independently determined spectral types.
The development of a UCHII depends on not only the ionizing flux provided by the central star, but also the opacity of the circumstellar medium. For infalling rates higher than some critical value, $\dot{M}_{\rm crit}$, the circumstellar medium will have such high opacities that the ionized region will be too small and too optically thick to be detectable [@CE02]. For the respective spectral types of the central stars, the $\dot{M}_{\rm crit}$ of the four UCHIIs in N44 are computed to be $\sim 1-4\times10^{-5} M_\odot$ yr$^{-1}$ (Table \[uch2\]). Compared to the $\dot{M}_{\rm env}$ determined from the best-fit models for these four YSOs (Table \[uch2\]), it is seen that only the Type II YSO 052255.2$-$680409.5 has $\dot{M}_{\rm env}~< ~\dot{M}_{\rm crit}$, and the other three YSOs have $\dot{M}_{\rm env}~\gg ~\dot{M}_{\rm crit}$. We have further examined whether any acceptable models of the latter three YSOs yield smaller envelope accretion rates. We find that only 052343.6$-$680034.2 has a few models with $\dot{M}_{\rm env} \le \dot{M}_{\rm crit}$, but these models have $T_\star =$ 38000–40000 K, higher that that for an O9 V star ($T_\star \sim$ 33000 K) estimated from the ionization requirement of the UCHII. All the other acceptable models have $\dot{M}_{\rm env} \gg
\dot{M}_{\rm crit}$.
Our limited sample shows that the Type II YSO in UCHII has $\dot{M}_{\rm env} < \dot{M}_{\rm crit}$, and the Type I and I/II YSOs in UCHIIs have $\dot{M}_{\rm env} \gg \dot{M}_{\rm crit}$. As Types I and I/II YSOs are still dominated by envelopes, it is possible that most of the infalling envelope material is used in forming an accretion disk, instead of the stellar core, as modeled by @YS02. Therefore, $\dot{M}_{\rm env}$ should not be interpreted as the accretion rate of mass onto the stellar core.
Masses of YSOs
--------------
In our sample of 60 YSOs in N44, 24 do not have mass estimates from SED model fits as their IRAC fluxes are contaminated by neighboring stars; nevertheless, it can be judged from their lower brightnesses that they are probably on the low-mass end in the sample. The other 36 YSOs have reliable SEDs that can be modeled to determine their masses; their mass estimates from the best and acceptable fits are listed in Table \[sedfits\]. Although many of these YSOs have large uncertainties in their mass estimates, i.e., large $M_\star$ Range, 30 of them have $M_\star$ Range all greater than $8 M_\odot$ and are thus most likely bona fide massive YSOs. The remaining six YSOs have $M_\star$ Range extending from intermediate- to high-mass; these may also be massive YSOs, particularly for the four with best-fit masses $\ge 8 M_\odot$.
It has been suggested that the criterion \[8.0\] $\le 8.0$ may be used to select massive YSOs in the LMC [@GC08]. We find that indeed the $M_\star$ Ranges of the brightest YSOs, with \[8.0\] $\le 8.0$, are all $\ge 8 M_\odot$. This criterion may be too conservative, as almost all (25 out of 27) YSOs with \[8.0\] $\le 9.0$ still have masses $\ge 8 M_\odot$
At the high-mass end, nine YSOs have $M_\star$ Ranges all $\ge 17 M_\odot$; these are most likely O-type YSOs. Five YSOs that show $M_\star$ Range with an upper mass limit $\ge 17 M_\odot$ and a lower mass limit $\ll 17 M_\odot$; these may or may not be O-type stars. To improve the census of O-type YSOs, we have checked the optical spectral classifications of the Type III YSOs to search for O stars that are missed by our estimates of masses from SED model fits. We find that two Type III YSOs with optical spectral types of O7.5V and O8.5V [@OM95] have masses determined from the SED model fits to correspond to B0–3V spectral types (052157.0$-$675700.1 and 052159.6$-$675721.7). Therefore, there exist at least 11 O-type YSOs in N44. These most massive YSOs will be discussed further in §7.1.
Massive Star Formation in N44
=============================
It is difficult to study the relationship between interstellar conditions and the formation of massive stars because massive stars’ UV radiation fluxes and fast stellar winds quickly ionize and disperse the ambient ISM. Massive YSOs, on the other hand, have not significantly altered the physical conditions of their surrounding medium on a large scale, and thus can be used to probe massive star formation. The large number of massive YSOs found in N44 provides an excellent opportunity to investigate issues such as relationship between star formation properties and interstellar conditions, progression of star formation, and evidence of triggered star formation.
Interstellar Environments and Star Formation Properties
-------------------------------------------------------
We examine the star formation properties of the molecular clouds in N44 as these clouds contain the bulk material to form stars. The NANTEN CO survey of the LMC [@FYetal01] shows three large concentrations of molecular material in N44, i.e., the central, southern, and northern peaks (Figure \[fig:yso\_pos\]). These three concentrations exhibit different numbers of massive stars formed in the last few Myr, as evidenced by their different amounts of ionized gas. The central molecular peak is associated with prominent [H$\alpha$]{}emission from a supershell and bright regions along the shell rim. Star formation has been occurring at this site for an extended period of time, with the supershell encompassing 10-Myr old massive stars and the bright regions containing $\sim$5-Myr old massive stars [@OM95]. The southern molecular peak shows one bright region and several smaller, disjoint regions. The massive stars of these regions have not been studied spectroscopically, but the absence of shell structures produced by fast stellar winds and supernovae implies that massive stars are most likely also $\sim$5-Myr old or that star formation has started only in the last few Myr. The northern molecular peak has a couple small regions, indicating that only modest massive star formation has taken place.
Fig. \[fig:yso\_pos\] shows that almost all of the YSOs in N44 are found in molecular clouds and about 75% of the YSOs are congregated toward the three molecular peaks. The YSOs of the three molecular peaks show different characteristics in their spatial distributions and interstellar environments. The central molecular peak has the highest concentration of YSOs, as 21 of them aggregate in the prominent regions along the southwest rim of the supershell. The southern molecular peak has loosely distributed YSOs, and most of these 11 YSOs are associated with the disjoint regions. The northern molecular peak has 12 YSOs that are also loosely distributed, but the majority of the YSOs are not associated with any ionized gas.
The YSOs of the three molecular peaks in N44 also show differences in their mass distributions. In Fig. \[fig:yso\_pos\], we have marked the YSOs with circles in three sizes that represent O-type stars with $M_\star \ge$ 17 $M_\odot$, B-type stars with $M_\star \ge$ 8 $M_\odot$, and intermediate-mass stars with $M_\star <$ 8 $M_\odot$, respectively. It is striking that $\sim$80% of the O-type and B-type YSOs are in or adjacent to regions. The YSOs with intermediate masses or without mass estimates do not show such strong bias in spatial distribution, although this may be partly due to an observational bias, as these YSOs are fainter and it is more difficult to detect faint YSOs over the bright background dust emission in regions. The central molecular peak, having the most prominent regions, possesses the highest concentration of O- and B-type YSOs. The southern molecular peak has a respectable number of O- and B-type YSOs, while the northern molecular peak has no O-type YSOs at all.
The characteristics of the current star formation in the three molecular peaks of N44 appear to be dependent on the massive star formation that occurred in the recent past. It is possible that the pattern of star formation is controlled by properties of the molecular clouds. The central, southern, and northern peaks correspond to the giant molecular clouds (GMCs) LMC$/$M5221$-$6802, LMC$/$M5239$-$6802, and LMC$/$M5221$-$6750 cataloged by @MNetal01. Their FWHM line-widths ($\Delta V$) at the peak position are 7.2, 15.8, and 3.8 km s$^{-1}$, respectively. It is interesting to note that the $\Delta V$ of the southern molecular peak of N44 is the highest and that of the northern molecular peak is nearly the lowest among all GMCs in the LMC. While the small $\Delta V$ of the northern molecular peak reflects the low level of stellar energy feedback in the last few Myr, the larger $\Delta V$ of the other two peaks do not seem to scale with star formation activity. Detailed mapping of these three molecular clouds is needed to search for fundamental differences among these three clouds that might be responsible for their different star formation characteristics.
[*HST*]{} Images: a Closer Look at the YSOs
-------------------------------------------
To examine the immediate surroundings of the YSOs in N44, we have searched the [*HST*]{} archive and found useful WFPC2 images of three fields that contain YSOs. These three fields encompass the regions N44C, N44F, and N44H, respectively, and their locations in N44 are shown in Fig. \[fig:n44opt\]a. These regions and their associated YSOs and ionizing stars are individually discussed below.
N44C is a bright region located at the southwest rim of the N44 superbubble. The [*HST*]{} [H$\alpha$]{} image of N44C shows an overall shell morphology: the northern part is bright and centered on the ionizing O7 V star [@OM95], while the southern part is faint and consists of multiple circular filaments superposed by radial filaments streaming to the southwest (Figure \[fig:n44c\_img\]a). It is not clear whether these southern radial and circular filaments are physically associated with each other. N44C abuts against a dark cloud to the north, which coincides with the core of the GMC M5221$-$6802 revealed by high-resolution ESO-SEST observations [@CYetal97]. The observed density variations suggest that N44C is a blister region on the surface of a molecular cloud.
Seven YSOs are within the field-of-view of the [*HST*]{} [H$\alpha$]{}image of N44C. One is projected at the base of a bright-rimmed dust pillar in the northwest part of N44C. Four are in the dark cloud 2–3 pc exterior to the northwest edge of N44C; all four are superposed on diffuse 8 [$\mu$m]{} PAH emission (Fig. \[fig:n44c\_img\]b), indicating that they may be associated with PDRs. One is located further northwest and projected in the N44 supershell rim; it is also superposed on large-scale diffuse 8 [$\mu$m]{} PAH emission and may be associated with PDRs. The last one is located in the southwest outskirts of N44C; it does not have prominent PDRs around it. It is remarkable that the majority of the YSOs associated with N44C are in a molecular core and superposed on prominent PDRs. Apparently they were formed in an interstellar environment that has been affected by energy feedback from stars that were formed a few Myr earlier.
N44F is a bright, ring-shaped region located at the northwest outskirts of the superbubble. The region is ionized by an O8III star [@WBD97]. Two prominent bright-rimmed dust pillars can be readily recognized in the WFPC2 [H$\alpha$]{} image (Figure \[fig:n44f\_img\]a). One of these dust pillars has a YSO, 052136.0$-$675443.4, emerging at its tip, reminiscent of those seen in the Eagle Nebula [M16, @HJetal96]. The mass estimated from SED fits for this YSO is $\sim 6-11~M_\odot$, more massive than those with $\sim 3-4~M_\odot$ seen in the Eagle Nebula [@TSH02].
N44H is located to the southeast of the N44 superbubble. This region contains the luminous blue variable (LBV) HD269445 [@SOetal84; @HD94] and four blue stars whose photometric measurements suggest spectral types of early-B [@Ch07]; thus, the LBV is most likely responsible for photoionizing the region. The YSO 052249.2$-$680129.0 is projected within N44H toward its bright northwest rim, $\sim 60''$, or 15 pc, from the LBV. The WFPC2 H$\alpha$ image of this YSO shows a close pair of sources and another source at $\sim$3$''$ south (Figure \[fig:n44h\_img\]a). The 8 [$\mu$m]{} image (Figure \[fig:n44h\_img\]b) shows a bright source and a faint extension to the south that are coincident with the optical pair and southern source, respectively. It is possible that all three sources are YSOs.
Triggered Star Formation
------------------------
We use the YSOs and their interstellar environment to investigate whether some of the current star formation in N44 is triggered. Fig. \[fig:yso\_pos\] shows that on a large scale the central supershell of N44 exhibits the most prominent association between stellar energy feedback and star formation. The alignment of massive YSOs in the southwest rim of the supershell suggests that the expansion of the supershell into the molecular cloud has triggered the star formation. This triggering mechanism may have been going on for $\sim$ 5 Myr and caused the formation of the massive stars in the bright regions along the shell rim.
Examined closely, the majority of massive YSOs are located at the edges of regions or in PDRs of dark clouds. These are suggestive examples of star formation triggered by external thermal pressure raised by photoionization or photodissociation. YSOs are also found in bright-rimmed dust pillars. Among them, the YSO in N44F, as aforementioned in §7.2, is in a simple interstellar structure and its immediate surroundings can be examined with WFPC2 images (Fig. \[fig:n44f\_img\]a), providing an opportunity to investigate whether the star formation is triggered. For this YSO to be formed from triggering instead of merely being exposed by the ionization front of the region, the formation time scale of the YSO should not be longer than the time for the ionization front to traverse from the tip to the bottom of the pillar. For the pillar’s projected length of $\sim$ 1.25 pc and a shell expansion velocity of 12 km s$^{-1}$ [@NYetal02], the traverse time is $\sim$ 0.1 Myr. This time scale is comparable to the formation time scale for a 10$M_\odot$ YSO, $\sim$ 0.1 Myr [@BM94], making it a plausible case of triggered star formation. To further determine statistically the importance of triggering for star formation, a systematic survey of the immediate surroundings of YSOs with high-resolution images, such as [*HST*]{} images, is needed.
The supershell of N44 shows a breakout at the south rim, and hot gas has been escaping and deflected to the east. The juxtaposition of this hot gas flow and the star forming regions associated with LH49 in the southern GMC is intriguing and may suggest triggered star formation. However, there is no direct physical evidence that the hot gas outflow compresses the molecular cloud to form stars. In fact, the thermal energy of the hot gas outflow, $\sim 1\times10^{50}$ ergs [@MEetal96], is much smaller than the kinetic energy of the southern GMC, $\sim 3\times10^{51}$ ergs estimated from a mass of 2.1$\times10^{6}$ $M_\odot$ and a velocity dispersion of 12.6 km s$^{-1}$ [@MNetal01]. Furthermore, the star formation in the southern GMC is mostly concentrated in regions not in contact with the hot gas flow. We conclude that the breakout of N44’s supershell is not responsible for triggering the star formation in the southern molecular concentration.
A Small Region around YSO 052207.3$-$675819.9
----------------------------------------------
It is interesting to note that a small nebula is detected around YSO 052207.3$-$675819.9 projected in the dark cloud northwest of N44C. A closeup of this nebula in Figure \[fig:n44c\_hii\] shows that it is detected in the [H$\alpha$]{} emission-line image but not in the $y$ continuum image, indicating that it is an region, instead of a reflection nebula. This small region has a size of 22$\times$15, or 0.53pc$\times0.36$pc, and an average H$\alpha$ surface brightness[^3] of $1.9\times10^{-14}$ ergs cm$^{-2}$ s$^{-1}$ arcsec$^{-2}$, corresponding to an emission measure of $9.4\times10^3$ cm$^{-6}$ pc. Adopting an average diameter of 0.45 pc as the path length of H$\alpha$ emission, the rms electron density of the region is then $\sim 145$ cm$^{-3}$. This region has a size comparable to those of compact regions, i.e., $\lesssim 0.5$ pc, but its emission measure and rms electron density are much lower than the typical values for compact regions, $\gtrsim 10^7$ cm$^{-6}$ pc and $\gtrsim 5\times10^3$ cm$^{-3}$ [@Fr00]. The lower densities of this small region suggest that it has expanded and is more evolved than a compact region.
This small region provides an opportunity for us to use its ionization requirement to assess the spectral type of the ionizing star and compare its implied mass with that determined from the YSO SED fitting. The $UBV$ photometry of YSO 052207.3$-$675819.9, given in Table \[ysoclass\], suggests that its central star is a $\sim$ B2 V star reddened by $A_V \sim 1.3$. Applying this extinction correction to the observed [H$\alpha$]{} flux, we obtained an [H$\alpha$]{} luminosity of $L_{{\rm H}\alpha}$ = 3.7$\times$10$^{34}$ ergs s$^{-1}$, which requires an ionizing power of $Q({\rm H}^0) = 7.4\times10^{11}~L_{{\rm H}\alpha}$ photons s$^{-1}$ = $2.7\times10^{46}$ photons s$^{-1}$. This ionizing power can be provided by a main-sequence star of effective temperature of 26,200 K [@Pa73], corresponding to a spectral type of B1 V [@SK82][^4]. The mass estimated from SED model fits for this YSO is 9–15 $M_\odot$, corresponding to spectral type B1-2 V. It is satisfactory that the three independent methods have produced consistent mass estimates.
This YSO in a small region illustrates a caveat in analyzing objects that are not fully resolved by [*Spitzer*]{} images. The SED model fits of YSO 052207.3$-$675819.9 infers that it is at evolutionary Stage II. The presence of a visible region and the visibility of the YSO in $UBV$ bands indicate that the star has little circumstellar dust and that the YSO is likely at Stage III. This discrepancy in evolutionary stage is most likely caused by the inclusion of dust emission from the region in the flux measurements, since it is not resolved in the ISPI and [*Spitzer*]{} images. It is thus important to use high-resolution images to examine the immediate surroundings of massive YSOs in order not to be mis-led by the SEDs.
The Source in N44A: Young or Evolved?
-------------------------------------
The source 052129.7$-$675106.9 is coincident with a small compact region, identified as an [H$\alpha$]{} knot and cataloged as N44A by @He56. Previously, this IR source was identified as IRAS 05216$-$6753 and classified as an region [@RPetal87] or an obscured supergiant or AGB [@WPetal92; @LCetal97]. Figure \[fig:agb2yso\] shows that the central star of this small region is bright in optical wavelengths. @Ch07 measured $V = 14.17 \pm 0.02$, ($B-V$) $= 0.79 \pm 0.03$ and ($U-B$) $= -0.64 \pm 0.06$. These magnitudes and colors cannot be reproduced by any combination of normal spectral type and luminosity class. The observed ($U-B$) indicates a blue star, so the intrinsic color ($B-V$)$_0$ must be in the range of 0 to $-$0.3; therefore, E($B-V$) = 0.8–1.1, or $A_v$ = 2.5–3.4, and $M_V$ = $-$6.8 to $-$8, suggesting that this star is a blue supergiant of luminosity class I. It cannot be an AGB star.
We can make another independent estimate of the spectral type based on the [H$\alpha$]{} luminosity of N44A. Using the flux-calibrated MCELS images of N44, we subtracted the red continuum from the [H$\alpha$]{} image and measured a continuum-subtracted [H$\alpha$]{} flux of 4.7$\times$10$^{-12}$ ergs s$^{-1}$ cm$^{-2}$ from N44A. To achieve feasible ionizing power, the star has to be blue with ($B-V$)$_0$ $\sim$ $-$0.3, thus we adopt an extinction of $A_v \sim 3.4$ and find an [H$\alpha$]{} luminosity of $1.4\times10^{37}$ ergs s$^{-1}$. This [H$\alpha$]{} luminosity requires the ionizing power of an O9I star.
We thus suggest that the central star of N44A is an O9I star. The strong mid-IR excess of this star indicates the existence of abundant dust around the star. The anomalous combination of ($U-B$) and ($B-V$) colors may be caused by the inclusion of extra scattered light in the $U$ band. If the star is a young star, the dust and region would both be interstellar. If the star is an evolved star, the dust and region would both be ejected stellar material and show enhanced elemental abundances. The current data do not allow us to distinguish whether the central star of N44A is a YSO or an evolved massive star, such as an LBV. High-resolution optical spectra of the [H$\alpha$]{} and \[\] lines are needed to search for N-enriched ionized gas. If the ionized gas is enriched, the central star of N44A is an evolved star. If the ionized gas is not enriched, the central star of N44A must be young, otherwise its stellar wind would have blown away the small region.
Summary
=======
We have observed the starburst region N44 in the LMC with the [*Spitzer*]{} IRAC and MIPS at 3.6, 4.5, 5.8, 8.0, 24, 70, and 160 [$\mu$m]{}, CTIO Blanco 4 m telescope ISPI in $J$ and $K_s$ and MOSAIC in SDSS $u$ and Johnson-Cousins $BVI$ bands. Point sources were identified and their photometric measurements were made. To identify YSOs, we first constructed an \[8.0\] vs. \[4.5\]$-$\[8.0\] CMD and used the criteria \[4.5\]$-$\[8.0\] $\ge$ 2.0 to exclude normal and evolved stars and \[8.0\] $<$ 14.0$-$(\[4.5\]$-$\[8.0\]) to exclude background galaxies. A total of $\sim$100 YSO candidates were identified. We then inspected the SED and close-up images of each YSO candidate in H$\alpha$, $BVIJK_s$, IRAC, and MIPS bands simultaneously to further identify and exclude evolved stars, galaxies, and dust clumps, resulting a final sample of 60 YSO candidates that are most likely bona fide YSOs of high or intermediate masses.
Following the classification of Classes I, II, and III for low-mass YSOs, we suggest the classification of Types I, II, and III for higher-mass YSOs, according to their SEDs. Type I YSOs are not detectable in optical to $J$ bands, and have SEDs rising beyond 24 [$\mu$m]{}. Type II YSOs become visible in optical to $J$ bands, and have SEDs peaking in the mid-MIR but falling beyond 24 [$\mu$m]{}. Type III YSOs are bright in optical, but still show excess IR emission.
To assess the physical properties of the YSOs, we use the Online SED Model Fitter [@RTetal07] for 36 YSOs whose SEDs are reliably determined. We find the SEDs of Type I and Type I/II YSOs can be modeled well, but the SEDs of Type II and Type II/III YSOs show prominent discrepancies at 4.5 [$\mu$m]{} because the SED models of @RTetal06 do not include the PAH emission. Some Type III YSOs show deep V- or U-shaped SEDs that require a low dust content within the central $\sim$1,000 AU, indicating a central hole in the disk. Using the SED model fits of YSOs and published spectroscopic observations of two Type III YSOs, we find at least 11 O-type YSOs in N44.
We have examined the five UCHIIs in N44 reported by @IJC04. One is probably a mis-identified background galaxy, and the remaining four have Type I-II YSO counterparts. The stellar masses of these YSOs determined from their SED model fits agree well with the masses implied by the ionizing fluxes required by the UCHIIs. However, the SED model fits suggest envelope accretion rates much higher than the critical mass accretion rate for the formation of UCHIIs.
N44 encompasses three molecular concentrations with different star formation histories. It is remarkable that the current formation of O-type stars is well correlated with the formation of such massive stars in the last few Myrs. The alignment of YSOs along the southwest rim of N44’s supershell suggests that their formation may have been triggered by the expansion of the supershell. [*HST*]{} images show that some YSOs in N44 are in bright-rimmed dust pillars and PDRs with PAH emission, indicating that the thermal pressure raised by photo-ionization or photo-dissociation may have triggered the star formation.
We caution that small regions may be associated with YSOs and thus their dust emission may contaminate the SEDs, causing uncertainties in the model fits. [*HST*]{} images are needed to resolve such small regions. Finally, we have analyzed the photometric measurements and ionizing power of the central star of N44A, and we find the star to be O9I star, instead of an AGB star as suggested previously.
We thank Leslie Looney for discussion on star formation, Remy Indebetouw on comparisons with YSOs in the SAGE catalog, and Barbara Whitney on SED fits. We also thank Armin Rest’s help on taking CTIO 4m MOSAIC images. This work was supported through NASA grants JPL 1264494 ([*Spitzer Space Telescope*]{}) and HST-AR-10942.01-A ([*Hubble Space Telescope*]{}). CHRC acknowledges support in part from JPL grants 1282653 and 1288328 (University of Virginia; PI: Indebetouw). This study made use of data products of the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and the Space Administration and the National Science Foundation.
[79]{} natexlab\#1[\#1]{}
, L. J., [Tielens]{}, G. G. M., & [Barker]{}, J. R. 1989, , 71, 733
, M., & [Mitalas]{}, R. 1994, , 95, 517
, P. A., & [Maeder]{}, A. 1996, , 307, 829
, M., & [Murdin]{}, P. 2000, Encyclopedia of Astronomy and Astrophysics
, V. M., & [McCarthy]{}, M. F. 1990, , 100, 674
, I., & [Hauschildt]{}, P. H. 2005, in ESA Special Publication, Vol. 576, The Three-Dimensional Universe with Gaia, ed. C. [Turon]{}, K. S. [O’Flaherty]{}, & M. A. C. [Perryman]{}, 565
, C. L., [et al.]{} 2006, , 132, 1890
, N., [et al.]{} 2005, , 630, L185
, A., [Gruendl]{}, R. A., & [Chu]{}, Y.-H. 2008, , 678, 200
, R., [Galli]{}, D., [Lodato]{}, G., [Walmsley]{}, C. M., & [Zhang]{}, Q. 2007, in Protostars and Planets V, ed. B. [Reipurth]{}, D. [Jewitt]{}, & K. [Keil]{}, 197
, C.-H. R. 2007, PhD thesis
, Y.-N., [Henkel]{}, C., [Whiteoak]{}, J. B., [Millar]{}, T. J., [Hunt]{}, M. R., & [Lemme]{}, C. 1997, , 317, 548
, Y.-H., [Mac Low]{}, M.-M., [Garcia-Segura]{}, G., [Wakker]{}, B., & [Kennicutt]{}, Jr., R. C. 1993, , 414, 213
, Y.-H., [et al.]{} 2005, , 634, L189
, E. 2002, , 40, 27
, M., [Wheaton]{}, W. A., & [Megeath]{}, S. T. 2003, , 126, 1090
, B. T., & [Li]{}, A. 2007, , 657, 810
, M. P., [Van Dyk]{}, S. D., & [Price]{}, S. D. 2001, , 122, 1844
, B. G. 1998, in ASP Conf. Ser. 148: Origins, ed. C. E. [Woodward]{}, J. M. [Shull]{}, & H. A. [Thronson]{}, Jr., 150
, G. G., [et al.]{} 2004, , 154, 10
, M. 1999, in IAU Symp. 190: New Views of the Magellanic Clouds, ed. Y.-H. [Chu]{}, N. [Suntzeff]{}, J. [Hesser]{}, & D. [Bohlender]{}, 542
, A., [et al.]{} 2005, , 129, 2074
, J., [Kurtz]{}, S. E., [Garc[í]{}a-Segura]{}, G., & [Hofner]{}, P. 2000, , 272, 169
, Y., [Mizuno]{}, N., [Yamaguchi]{}, R., [Mizuno]{}, A., & [Onishi]{}, T. 2001, , 53, L41
, K. D., [et al.]{} 2005, , 117, 503
, D., [Brandner]{}, W., & [Henning]{}, T. 2006, , 636, L133
, M. A. T. 2006, , 448, 181
, R. A., & [Chu]{}, Y.-H. 2008, , submitted
, P. M., [et al.]{} 2006, , 644, 307
, E., [et al.]{} 2005, , 129, 1198
, K. G. 1956, , 2, 315
, J. J., [et al.]{} 1996, , 111, 2349
, L. A., [Strom]{}, S. E., [Vrba]{}, F. J., & [Keene]{}, J. 1992, , 397, 613
, D. J., & [Tielens]{}, A. G. G. M. 1997, , 35, 179
, R. M., & [Davidson]{}, K. 1994, , 106, 1025
, R., [Johnson]{}, K. E., & [Conti]{}, P. 2004, , 128, 2206
, T. J., [Woodward]{}, C. E., [Boyer]{}, M. L., [Gehrz]{}, R. D., & [Polomski]{}, E. 2005, , 620, 731
, R. 1993, ATLAS9 Stellar Atmosphere Programs and 2 km/s grid. Kurucz CD-ROM No. 13. Cambridge, Mass.: Smithsonian Astrophysical Observatory, 1993., 13
, C. J. 1987, in IAU Symp. 115: Star Forming Regions, ed. M. [Peimbert]{} & J. [Jugaku]{}, 1
, A., & [Draine]{}, B. T. 2001, , 554, 778
—. 2002, , 576, 762
, C. J., [et al.]{} 2003, , 115, 897
, C., [Zijlstra]{}, A. A., [Waters]{}, L. B. F. M., & [Groenewegen]{}, M. A. T. 1997, , 125, 419
, P. B., & [Hodge]{}, P. W. 1970, , 75, 171
, E. A., [Chu]{}, Y.-H., [Points]{}, S. D., [Hwang]{}, U., & [Smith]{}, R. C. 1996, , 464, 829
, M., [et al.]{} 2006, , 132, 2268
, N., [et al.]{} 2001, , 53, 971
, Y., [Chu]{}, Y.-H., [Guerrero]{}, M. A., [Oey]{}, M. S., [Gruendl]{}, R. A., & [Smith]{}, R. C. 2002, , 124, 3325
, A., [et al.]{} 2006, , 640, L29
, M. S., & [Massey]{}, P. 1995, , 452, 210
, N. 1973, , 78, 929
, S. R. 1993, in Infrared Astronomy, ed. A. [Mampaso]{}, M. [Prieto]{}, & F. [Sanchez]{}, 63
, N., [Glass]{}, I. S., & [Catchpole]{}, R. M. 1988, , 232, 53
, M. E., & [Barsony]{}, M. 2003, , 584, 832
, W. K. M., [Wood]{}, K., [Armitage]{}, P. J., [Whitney]{}, B. A., & [Bjorkman]{}, J. E. 2003, , 342, 79
, G. H., [et al.]{} 2004, , 154, 25
, T. P., [Whitney]{}, B. A., [Indebetouw]{}, R., & [Wood]{}, K. 2007, , 169, 328
, T. P., [Whitney]{}, B. A., [Indebetouw]{}, R., [Wood]{}, K., & [Denzmore]{}, P. 2006, , 167, 256
, P. F., [Aitken]{}, D. K., & [Smith]{}, C. H. 1987, , 228, 269
, M., [Lock]{}, T. D., [Walker]{}, D. W., & [Harris]{}, S. 1986, , 222, 273
, M., [et al.]{} 2005, , 129, 1183
, T. 1982, [in Landolt-B[ö]{}rnstein New Series, Group 6, Volume 2b, Stars and Star Clusters]{}, ed. K. [Schaifers]{} & H. H. [Voigt]{} (Berlin: Springer)
, L., [Dufour]{}, E., & [Forestini]{}, M. 2000, , 358, 593
, J. D., [et al.]{} 2007, , 669, 327
, M. F., [et al.]{} 2006, , 131, 1163
, R. C., & [The MCELS Team]{}. 1999, in IAU Symp. 190: New Views of the Magellanic Clouds, ed. Y.-H. [Chu]{}, N. [Suntzeff]{}, J. [Hesser]{}, & D. [Bohlender]{}, 28
, O., [Wolf]{}, B., [Leitherer]{}, C., [Zickgraf]{}, F.-J., [Krautter]{}, J., & [de Groot]{}, M. 1984, , 140, 459
, S. W., [Palla]{}, F., & [Ho]{}, P. T. P. 2000, Protostars and Planets IV, 327
, R. I., [Smith]{}, B. A., & [Hester]{}, J. J. 2002, , 570, 749
, R., [Waters]{}, L. B. F. M., [Dominik]{}, C., [Dullemond]{}, C. P., [Tielens]{}, A. G. G. M., & [de Koter]{}, A. 2004, , 418, 177
, B. A., [Indebetouw]{}, R., [Bjorkman]{}, J. E., & [Wood]{}, K. 2004, , 617, 1177
, B. A., [Wood]{}, K., [Bjorkman]{}, J. E., & [Cohen]{}, M. 2003, , 598, 1079
, B. A., [et al.]{} 2004, , 154, 315
—. 2008, , 136, 18
, J.-M., [Bomans]{}, D. J., & [Dieball]{}, A. 1997, , 123, 455
, P. R., [Whiteoak]{}, J. B., [Hughes]{}, S. M. G., [Bessell]{}, M. S., [Gardner]{}, F. F., & [Hyland]{}, A. R. 1992, , 397, 552
, H. W., & [Sonnhalter]{}, C. 2002, , 569, 846
, D., [Harris]{}, J., [Thompson]{}, I. B., & [Grebel]{}, E. K. 2004, , 128, 1606
, H., & [Yorke]{}, H. W. 2007, , 45, 481
Fig. \[fig:ysoimg\] — Continued.
Fig. \[fig:ysoimg\] — Continued.
Fig. \[fig:ysoimg\] — Continued.
Fig. \[fig:ysoimg\] — Continued.
.
\
Fig. \[fig:fit\] — Continued.
\
Fig. \[fig:fit\] — Continued.
\
Fig. \[fig:fit\] — Continued.
\
Fig. \[fig:fit\] — Continued.
.
.
[rcccc]{}
IRAC 3.6 [$\mu$m]{} & 3.6 & 3.6-8.4 & 1.124 & 277.5\
4.5 [$\mu$m]{} & 3.6 & 3.6-8.4 & 1.127 & 179.5\
5.8 [$\mu$m]{} & 3.6 & 3.6-8.4 & 1.143 & 116.6\
8.0 [$\mu$m]{} & 3.6 & 3.6-8.4 & 1.234 & 63.1\
MIPS 24 [$\mu$m]{} & 6 & 20-32 & 1.699 & 7.14\
70 [$\mu$m]{}& 16 & 39-65 & 2.087 & 0.775\
160 [$\mu$m]{}& 40 & 75-125 & 1.884 & 0.159\
[ccccccccccccccccr]{}
052030.6-680316.7 & 80.12759 & -68.05464 & & 19.68 0.03 & 17.86 0.04 & 15.95 0.04 & 14.63 0.04 & 13.82 0.05 & 13.52 0.05 & 13.18 0.09 & 13.39 0.15 & 13.24 0.03 & 12.83 0.06 & & & 10\
052030.7-674817.8 & 80.12783 & -67.80495 & & & & & & & & & & & 14.46 0.21 & & & 10\
052030.7-680353.3 & 80.12783 & -68.06480 & & 21.44 0.12 & 20.61 0.12 & 19.30 0.07 & & & & 17.08 0.07 & & 16.41 0.47 & & & & 10\
052030.7-674946.3 & 80.12787 & -67.82953 & & 20.57 0.05 & 19.69 0.04 & & & & & 16.83 0.06 & 16.78 0.09 & & & & & 0\
052030.7-674436.7 & 80.12789 & -67.74352 & 21.49 0.45 & 20.57 0.06 & 19.46 0.05 & 18.44 0.06 & & & & & 17.20 0.10 & & & & & 10\
052030.7-675920.8 & 80.12793 & -67.98911 & 19.46 0.05 & 18.31 0.01 & 17.05 0.02 & & 14.89 0.04 & 14.20 0.07 & 14.10 0.07 & 13.90 0.17 & 13.94 0.01 & 13.81 0.04 & 13.50 0.08 & & & 0\
[rrllc]{}
05 21 45.3 & -68 06 13.5 & 7553/MacKenty & F300W & $1\times2900$\
05 21 49.8 & -67 54 59.1 & 6698/Chu & F656N & $2\times500$\
& & & F673N & $2\times500$\
05 21 58.4 & -68 06 36.2 & 7553/MacKenty & F300W & $1\times2900$\
05 22 07.9 & -67 59 13.0 & 6623/Garnett & F502N & $3\times700$\
& & & F547M & $2\times80$\
& & & F656N & $3\times700$\
& & & F675W & $2\times80$\
05 22 35.7 & -67 58 12.8 & 6698/Chu & F656N & $2\times500$\
& & & F673N & $2\times500$\
05 22 54.9 & -68 01 41.9 & 6540/Schulte-Ladbeck & F656N & $2\times5$, $2\times30$, $2\times500$\
05 23 40.2 & -67 59 46.8 & 6253/MacKenty & F300W & $1\times160$, $1\times260$\
[cccccccc]{}
Effective & & & & & & &\
Wavelength ([$\mu$m]{})& 0.367 & 0.436 & 0.545 & 0.797 & 1.235 & 1.662 & 2.159\
0$^{th}$ mag & & & & & & &\
$F_\nu$ (Jy) & 1780 & 4000 & 3600 & 2420 & 1594 & 1024 & 666.8\
[rrrrrrrrrll]{}
052042.0-674307.7 & 96 & ... ... & ... ... & ... ... & ... ... & 16.06 0.11 & 15.04 0.11 & 14.71 0.12 &&\
&& 13.85 0.01 & 13.81 0.01 & 13.61 0.04 & 11.64 0.02 & 9.30 0.25 & ... ... & 0 & G &\
052046.6-675255.1 & 15 & ... ... & ... ... & ... ... & 19.23 0.01 & 17.78 0.12 & ... ... & 15.95 0.06 &&\
&& 12.45 0.01 & 12.46 0.08 & 9.53 0.02 & 7.81 0.01 & 3.88 0.01 & -0.79 0.10 & 11 & T2 &\
052106.8-675715.9 & 99 & 19.76 0.17 & 20.29 0.06 & 18.85 0.16 & ... ... & 16.61 0.18 & 15.44 0.16 & 15.18 0.20 &&\
&& 14.23 0.02 & 13.99 0.01 & ... ... & 11.79 0.03 & ... ... & ... ... & 0 & G &\
052117.5-680204.6 & 41 & ... ... & ... ... & ... ... & 19.66 0.02 & 17.85 0.15 & ... ... & 15.98 0.07 &&\
&& 13.36 0.01 & 13.25 0.01 & 10.52 0.04 & 8.73 0.02 & 5.40 0.01 & -0.26 99.9 & 11 & G? YSC? &\
052120.7-674706.6 & 31 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.03 0.02 & 12.96 0.03 & 10.21 0.02 & 8.51 0.02 & 5.34 0.01 & ... ... & 0 & D &\
052120.7-674725.5 & 74 & 17.39 0.07 & 18.50 0.06 & 17.81 0.08 & 17.25 0.08 & ... ... & ... ... & ... ... &&\
&& 13.65 0.09 & 13.18 0.07 & 11.30 0.12 & 9.54 0.13 & ... ... & ... ... & 0 & S & on IR D\
052120.9-674716.7 & 82 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.96 0.07 & 13.20 0.04 & ... ... & 10.40 0.21 & ... ... & ... ... & 0 & D &\
052120.9-680217.6 & 98 & ... ... & 22.50 0.50 & 21.91 0.18 & 19.94 0.11 & ... ... & ... ... & ... ... &&\
&& 14.95 0.04 & 13.94 0.02 & 12.70 0.23 & 11.68 0.22 & 7.80 99.9 & ... ... & 0 & ES? &\
052122.0-674729.0 & 5 & ... ... & ... ... & ... ... & 18.77 0.01 & 17.12 0.12 & ... ... & 14.88 0.04 &&\
&& 12.08 0.06 & 11.76 0.06 & 9.27 0.03 & 7.44 0.01 & 1.86 0.01 & -2.17 0.06 & 11 & YSO+RN? &\
052126.2-674742.1 & 71 & 17.62 0.03 & 18.03 0.02 & 17.91 0.02 & 19.43 0.01 & 17.76 0.15 & ... ... & 16.39 0.07 &&\
&& 13.54 0.03 & 13.38 0.02 & ... ... & 9.47 0.05 & 5.92 99.9 & ... ... & 11 & T2, mul & in MC\
052127.2-675915.1 & 92 & 15.37 0.01 & 15.43 0.01 & 14.86 0.01 & ... ... & 13.71 0.02 & 13.38 0.04 & 13.35 0.04 &&\
&& 13.08 0.09 & 13.38 0.01 & 12.50 0.07 & 11.24 0.09 & ... ... & ... ... & 0 & blue S &\
052129.7-675106.9 & 1 & 14.32 0.01 & 14.96 0.01 & 14.17 0.01 & 14.07 0.01 & 13.00 0.03 & 12.02 0.03 & 10.46 0.02 &&\
&& 8.03 0.01 & 6.96 0.01 & 5.83 0.01 & 4.12 0.01 & sat. ... & -2.95 0.04 & 10 & ES &\
052133.3-674420.8 & 97 & 16.36 0.07 & 17.11 0.02 & 16.85 0.12 & 17.93 0.01 & 16.52 0.06 & ... ... & 15.21 0.05 &&\
&& ... ... & 13.82 0.01 & ... ... & 11.65 0.05 & 7.98 0.01 & ... ... & 11 & mul YSOs? &\
052135.5-675500.2 & 86 & ... ... & 18.49 0.02 & 17.29 0.01 & ... ... & 15.09 0.06 & 14.34 0.06 & 14.20 0.08 &&\
&& 14.04 0.26 & 13.94 0.10 & 12.75 0.15 & 10.89 0.13 & ... ... & ... ... & 0 & S & in DR\
052136.0-675443.4 & 65 & 17.16 0.05 & 17.75 0.02 & 17.28 0.04 & 17.31 0.01 & 16.93 0.06 & ... ... & 16.10 0.03 &&\
&& 13.65 0.27 & 13.01 0.20 & 11.48 0.19 & 9.32 0.07 & ... ... & ... ... & 11 & T2/3 & tip of DC\
052136.3-674643.2 & 89 & ... ... & ... ... & ... ... & 21.49 0.12 & 20.48 0.67 & ... ... & 17.54 0.08 &&\
&& 15.30 0.14 & 13.65 0.05 & 12.49 0.09 & 11.06 0.16 & 5.36 0.01 & ... ... & 11 & mul YSOs &\
052136.6-675449.4 & 75 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.85 0.10 & 13.87 0.17 & 11.44 0.15 & 9.54 0.06 & 4.04 0.01 & ... ... & 0 & D &\
052138.0-674630.3 & 72 & ... ... & ... ... & ... ... & 19.30 0.24 & 17.95 0.11 & ... ... & 16.75 0.03 &&\
&& 13.72 0.24 & 13.91 0.07 & 11.20 0.09 & 9.52 0.04 & 4.80 0.01 & ... ... & 11 & T2 &\
052141.9-675324.1 & 63 & 20.40 0.25 & 20.42 0.10 & 19.68 0.13 & 19.62 0.02 & 18.13 0.17 & ... ... & 16.47 0.05 &&\
&& 13.90 0.05 & 13.81 0.04 & 11.00 0.07 & 9.26 0.04 & 5.32 0.02 & ... ... & 11 & T2, mul? &\
052144.5-674541.5 & 45 & ... ... & 21.01 0.10 & 19.88 0.13 & 20.58 0.03 & 19.15 0.20 & ... ... & 15.41 0.04 &&\
&& 13.38 0.01 & 13.36 0.01 & 10.65 0.01 & 8.91 0.01 & 5.87 0.03 & 0.37 99.9 & 11 & YSC? &\
052147.1-675656.7 & 17 & 14.05 0.01 & 14.91 0.01 & 15.30 0.01 & 14.86 0.02 & 14.94 0.03 & ... ... & 14.31 0.03 &&\
&& 12.07 0.04 & 11.89 0.06 & 9.63 0.02 & 7.84 0.01 & 2.54 0.01 & ... ... & 11 & T3 &\
052152.8-675449.5 & 91 & 13.98 0.01 & 14.94 0.01 & 14.97 0.01 & ... ... & 14.69 0.06 & 14.43 0.07 & 14.25 0.10 &&\
&& 13.68 0.18 & 13.39 0.19 & 12.62 0.08 & 11.18 0.11 & ... ... & ... ... & 0 & S &\
052154.2-674737.1 & 90 & ... ... & ... ... & ... ... & 21.54 0.04 & 19.73 0.33 & ... ... & 16.58 0.06 &&\
&& 14.41 0.03 & 13.50 0.23 & 12.54 0.21 & 11.10 0.10 & ... ... & ... ... & 11 & T2, mul? & in DC\
052155.3-674730.2 & 61 & ... ... & ... ... & ... ... & 20.36 0.02 & 18.48 0.18 & ... ... & 16.14 0.01 &&\
&& 12.90 0.09 & 11.77 0.04 & 10.49 0.04 & 9.20 0.02 & 5.00 0.01 & ... ... & 11 & T2, mul &\
052155.3-675634.9 & 46 & 18.54 0.07 & 19.47 0.03 & 19.01 0.04 & 18.49 0.01 & 17.47 0.12 & ... ... & 15.36 0.01 &&\
&& 12.90 0.10 & 12.77 0.13 & 10.61 0.06 & 8.89 0.04 & 4.90 99.9 & ... ... & 11 & T2/3 &\
052157.0-675700.1 & 77 & 13.43 0.01 & 14.36 0.01 & 14.42 0.01 & 14.32 0.01 & 14.66 0.03 & ... ... & 14.64 0.01 &&\
&& 14.29 0.12 & 13.50 0.27 & 11.90 0.15 & 9.96 0.10 & 3.52 99.9 & ... ... & 11 & T3: O8.5V+N &\
052157.5-675618.5 & 59 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.60 0.10 & 13.40 0.09 & 10.95 0.10 & 9.16 0.07 & 4.27 0.05 & ... ... & 0 & D &\
052157.9-675625.5 & 68 & 11.96 0.01 & 12.95 0.01 & 13.12 0.01 & ... ... & 13.32 0.03 & 13.28 0.04 & 13.16 0.04 &&\
&& 12.57 0.10 & 12.45 0.12 & 10.83 0.09 & 9.37 0.08 & ... ... & ... ... & 0 & S: O7III((f)) &\
052158.2-675554.2 & 44 & 17.62 0.03 & 18.42 0.02 & 18.32 0.02 & 17.82 0.02 & 17.18 0.11 & ... ... & 15.12 0.01 &&\
&& 12.34 0.07 & 11.79 0.07 & 10.27 0.05 & 8.81 0.04 & 5.12 99.9 & ... ... & 11 & T2/3 &\
052159.0-674437.2 & 76 & 19.57 0.15 & 19.55 0.07 & 18.81 0.10 & 18.55 0.01 & 17.57 0.14 & ... ... & 16.36 0.08 &&\
&& 14.13 0.02 & 13.79 0.01 & 11.43 0.01 & 9.71 0.01 & 6.68 0.04 & ... ... & 11 & T2, mul? &\
052159.6-675721.7 & 79 & 12.82 0.01 & 13.78 0.01 & 14.03 0.01 & 14.78 0.02 & 14.35 0.03 & ... ... & 14.23 0.03 &&\
&& 13.43 0.10 & 12.86 0.10 & ... ... & 10.09 0.24 & ... ... & ... ... & 11 & T3: O7.5V+N &\
052159.6-675715.6 & 78 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 14.32 0.36 & 13.90 0.30 & 11.89 0.27 & 10.05 0.26 & ... ... & ... ... & 0 & D &\
052200.4-675745.0 & 25 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.01 0.13 & 13.49 0.33 & 10.23 0.11 & 8.37 0.11 & ... ... & ... ... & 0 & D &\
052201.9-675732.5 & 56 & ... ... & 17.37 0.11 & 17.00 0.23 & 19.48 0.12 & 18.05 0.23 & ... ... & 16.17 0.10 &&\
&& 13.59 0.33 & 13.39 0.36 & 10.57 0.21 & 9.07 0.25 & ... ... & ... ... & 11 & YSO, mul? & in DC\
052202.0-675758.2 & 13 & ... ... & ... ... & ... ... & 22.52 0.39 & 21.06 99.9 & ... ... & 17.66 0.15 &&\
&& 12.69 0.12 & 11.50 0.08 & 9.36 0.05 & 7.77 0.03 & ... ... & ... ... & 11 & T2 & in DC\
052202.2-675753.6 & 11 & ... ... & ... ... & ... ... & 18.89 0.02 & 17.16 0.11 & ... ... & 14.93 0.04 &&\
&& 12.36 0.11 & 11.95 0.12 & 9.40 0.04 & 7.67 0.03 & 1.75 0.02 & ... ... & 11 & T2 &\
052202.3-674657.5 & 33 & ... ... & ... ... & ... ... & 20.81 0.10 & 18.28 0.19 & ... ... & 15.96 0.04 &&\
&& 13.18 0.17 & 12.86 0.03 & 10.35 0.06 & 8.51 0.03 & 2.41 0.01 & ... ... & 11 & YSO & in DC\
052202.8-674701.9 & 14 & 19.31 0.14 & 20.14 0.13 & 18.74 0.14 & 18.53 0.01 & 17.11 99.9 & ... ... & 15.18 0.01 &&\
&& 12.12 0.06 & 11.75 0.07 & 9.54 0.03 & 7.77 0.02 & 2.01 99.9 & -2.39 99.9 & 11 & T2 & at tip of DC\
052203.1-674703.5 & 16 & 17.26 0.10 & 17.95 0.05 & 17.66 0.11 & 17.28 0.01 & 16.42 0.07 & ... ... & 15.41 0.03 &&\
&& 12.10 0.02 & 11.78 0.02 & 9.64 0.03 & 7.81 0.03 & ... ... & ... ... & 11 & T2 &\
052203.4-675746.9 & 24 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & 16.58 0.09 &&\
&& 12.76 0.14 & 11.78 0.08 & 10.01 0.07 & 8.28 0.04 & 1.74 99.9 & ... ... & 11 & T1/2? & in DC\
052203.9-675743.7 & 52 & ... ... & ... ... & ... ... & 18.78 0.02 & 18.25 0.19 & ... ... & 16.19 0.07 &&\
&& 12.95 0.20 & 11.82 0.10 & 10.77 0.17 & 8.98 0.10 & ... ... & ... ... & 11 & T2 & in DC\
052204.1-674709.7 & 73 & 17.74 0.06 & 18.39 0.05 & 18.48 0.06 & 18.14 0.12 & ... ... & ... ... & ... ... &&\
&& 13.92 0.04 & 13.94 0.04 & 11.24 0.04 & 9.52 0.05 & ... ... & ... ... & 0 & S & on IR D\
052204.8-675744.6 & 42 & ... ... & ... ... & ... ... & 19.40 0.03 & 17.88 0.14 & ... ... & 15.62 0.05 &&\
&& 13.29 0.27 & 13.00 0.31 & 10.87 0.33 & 8.76 0.20 & ... ... & ... ... & 11 & T2/3 & at edge of DC\
052204.9-675720.6 & 62 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.45 0.12 & 13.42 0.12 & 11.16 0.23 & 9.22 0.20 & ... ... & ... ... & 0 & D &\
052204.9-675801.6 & 36 & 20.70 0.23 & 19.87 0.05 & 19.00 0.03 & ... ... & ... ... & ... ... & ... ... &&\
&& 13.08 0.16 & 12.61 0.11 & 10.25 0.14 & 8.54 0.06 & 2.08 0.02 & ... ... & 0 & D &\
052205.2-675741.6 & 38 & ... ... & ... ... & ... ... & 22.18 0.37 & 20.08 0.64 & ... ... & 16.25 0.08 &&\
&& 13.39 0.25 & 12.79 0.15 & ... ... & 8.61 0.23 & ... ... & ... ... & 11 & mul YSOs & in DC\
052205.3-675748.5 & 21 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & 17.22 0.13 &&\
&& 13.12 0.26 & 12.97 0.35 & 9.88 0.06 & 8.13 0.04 & 2.06 99.9 & ... ... & 11 & T2? & in DC\
052206.4-675659.2 & 28 & ... ... & ... ... & ... ... & 19.34 0.36 & 17.74 0.15 & ... ... & 15.16 0.04 &&\
&& 12.28 0.07 & 11.64 0.06 & 9.98 0.04 & 8.41 0.02 & 4.37 99.9 & ... ... & 11 & T2 &\
052207.3-675819.9 & 19 & 17.73 0.01 & 18.35 0.01 & 18.16 0.01 & 17.51 0.01 & 16.71 0.09 & ... ... & 15.38 0.05 &&\
&& 12.30 0.07 & 11.87 0.07 & 9.71 0.03 & 8.01 0.02 & 2.79 99.9 & ... ... & 11 & T3 & in HII\
052207.3-675826.8 & 26 & ... ... & ... ... & ... ... & 21.20 99.9 & 19.66 99.9 & ... ... & 15.21 0.04 &&\
&& 11.99 0.07 & 10.54 0.02 & 9.43 0.03 & 8.38 0.04 & 3.05 99.9 & ... ... & 11 & T2 & in DC\
052207.7-675649.4 & 85 & 18.68 0.08 & 18.91 0.04 & 19.03 0.04 & ... ... & ... ... & ... ... & ... ... &&\
&& 14.19 0.07 & 13.94 0.10 & 12.50 0.07 & 10.69 0.07 & ... ... & ... ... & 0 & S & in DR\
052208.5-675821.3 & 30 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 12.93 0.04 & 12.46 0.04 & 10.16 0.06 & 8.45 0.03 & 2.90 0.04 & ... ... & 0 & D &\
052208.6-675805.5 & 51 & ... ... & ... ... & ... ... & 19.35 0.05 & 17.83 0.15 & ... ... & 15.24 0.04 &&\
&& 12.32 0.08 & 11.68 0.08 & 10.42 0.10 & 9.01 0.09 & 4.44 0.25 & ... ... & 11 & mul YSOs &\
052208.6-675921.9 & 70 & 18.00 0.02 & 18.75 0.01 & 18.68 0.02 & 18.57 0.01 & 17.61 0.14 & ... ... & 16.43 0.09 &&\
&& 13.79 0.04 & 13.83 0.03 & 11.10 0.08 & 9.38 0.05 & 5.92 0.14 & ... ... & 11 & T2, mul? &\
052208.8-675325.2 & 60 & 18.61 0.06 & 18.98 0.03 & 18.68 0.04 & 18.05 0.01 & 17.37 0.11 & ... ... & 16.23 0.04 &&\
&& 13.35 0.11 & 13.13 0.15 & 10.79 0.04 & 9.19 0.02 & 5.63 0.02 & ... ... & 11 & T2 + T3/S &\
052208.9-674703.4 & 93 & 17.39 0.08 & 16.74 0.02 & 15.82 0.04 & 14.81 0.04 & 14.11 0.03 & 13.61 0.04 & 13.52 0.04 &&\
&& 13.33 0.01 & 13.38 0.18 & 12.64 0.09 & 11.37 0.14 & ... ... & ... ... & 0 & blue S &\
052211.9-675818.1 & 48 & ... ... & ... ... & ... ... & 22.79 0.27 & 19.55 0.31 & ... ... & 16.72 0.09 &&\
&& 14.01 0.13 & 12.69 0.17 & 10.92 0.10 & 8.97 0.05 & 2.04 99.9 & ... ... & 11 & T1 & in DC\
052212.0-674713.9 & 80 & 21.47 0.53 & 21.01 0.11 & 20.55 0.13 & 20.05 0.02 & 18.00 0.15 & ... ... & 15.89 0.03 &&\
&& 13.63 0.17 & 12.74 0.10 & 11.40 0.09 & 10.10 0.05 & 6.43 0.01 & ... ... & 11 & T2 &\
052212.3-675813.4 & 49 & ... ... & ... ... & ... ... & 18.63 0.01 & 17.35 0.11 & ... ... & 15.37 0.04 &&\
&& 12.99 0.17 & 12.24 0.12 & 10.61 0.10 & 8.97 0.07 & ... ... & ... ... & 11 & T2/3 & in DC\
052212.6-675832.2 & 2 & ... ... & ... ... & ... ... & 20.34 99.9 & 17.24 0.12 & ... ... & 12.87 0.01 &&\
&& 9.44 0.01 & 7.92 0.01 & 6.51 0.01 & 5.10 0.01 & sat. ... & -3.85 0.08 & 11 & T1 &\
052216.7-675837.7 & 55 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.92 0.14 & 13.57 0.15 & 10.85 0.10 & 9.07 0.09 & ... ... & ... ... & 0 & D &\
052216.8-680428.3 & 66 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.91 0.06 & 13.92 0.08 & 11.12 0.04 & 9.32 0.05 & 4.96 0.04 & ... ... & 0 & D peak &\
052216.9-680403.6 & 34 & ... ... & ... ... & ... ... & 21.22 0.08 & 18.98 0.27 & ... ... & 15.73 0.06 &&\
&& 12.81 0.09 & 11.59 0.04 & 9.96 0.03 & 8.52 0.03 & 4.01 0.01 & ... ... & 11 & T1 &\
052217.8-680432.9 & 69 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.80 0.06 & 13.68 0.08 & 11.07 0.06 & 9.36 0.06 & ... ... & ... ... & 0 & D &\
052218.9-675813.8 & 64 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& ... ... & 13.84 0.12 & 11.21 0.15 & 9.31 0.13 & ... ... & ... ... & 0 & D &\
052219.8-680436.8 & 4 & ... ... & 18.24 0.08 & 18.16 0.13 & 18.93 0.14 & 16.51 0.09 & ... ... & 14.93 0.04 &&\
&& 11.20 0.04 & 10.61 0.03 & 8.60 0.02 & 6.87 0.01 & 1.93 0.01 & -2.62 0.07 & 11 & T1/2 & in DC\
052221.0-680515.3 & 94 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 15.35 0.02 & 13.85 0.28 & 12.81 0.26 & 11.38 0.11 & 8.32 0.21 & ... ... & 0 & ES? &\
052227.7-675412.8 & 37 & ... ... & ... ... & ... ... & 21.85 0.22 & 18.85 0.24 & ... ... & 15.92 0.04 &&\
&& 13.05 0.13 & 12.95 0.15 & 10.32 0.06 & 8.55 0.03 & 4.20 0.04 & ... ... & 11 & T2? & in DC\
052229.1-675339.5 & 50 & 16.11 0.03 & 17.04 0.04 & 16.65 0.02 & 16.21 0.01 & 15.95 0.06 & ... ... & 15.12 0.02 &&\
&& 13.08 0.10 & 13.05 0.15 & 10.68 0.05 & 9.00 0.02 & 5.34 0.03 & ... ... & 11 & T3, mul &\
052230.4-675443.9 & 18 & 16.11 0.01 & 16.75 0.01 & 16.40 0.01 & ... ... & 14.84 0.08 & 14.42 0.10 & 13.62 0.11 &&\
&& 11.93 0.06 & 11.38 0.06 & ... ... & 7.88 0.09 & 2.38 0.01 & -2.69 0.09 & 0 & S & in HII\
052231.8-680319.2 & 27 & 18.75 0.10 & 19.65 0.06 & 19.59 0.06 & 19.44 0.03 & 17.73 0.14 & ... ... & 15.70 0.06 &&\
&& 12.75 0.10 & 12.33 0.09 & 10.11 0.05 & 8.40 0.03 & 3.85 0.01 & ... ... & 11 & T2 & in DC?\
052232.7-680301.7 & 53 & ... ... & ... ... & ... ... & 19.45 0.01 & 17.26 0.11 & ... ... & 15.06 0.04 &&\
&& 13.06 0.10 & 12.69 0.09 & 10.90 0.07 & 9.02 0.03 & 4.42 0.01 & ... ... & 11 & T2, mul & in DC\
052241.4-675508.2 & 84 & 16.15 0.01 & 16.19 0.01 & 15.55 0.01 & ... ... & 14.12 0.03 & 13.62 0.04 & 13.57 0.05 &&\
&& 13.37 0.12 & 13.33 0.17 & 12.02 0.16 & 10.52 0.06 & ... ... & ... ... & 0 & S &\
052242.0-675500.5 & 81 & ... ... & ... ... & ... ... & 20.86 0.03 & 18.51 0.22 & ... ... & 16.12 0.01 &&\
&& 14.04 0.04 & 13.21 0.02 & 11.66 0.13 & 10.24 0.10 & 5.50 0.02 & ... ... & 11 & T2 &\
052249.0-680111.1 & 67 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.93 0.16 & 13.76 0.11 & 11.20 0.16 & 9.36 0.09 & ... ... & ... ... & 0 & D &\
052249.2-680129.0 & 23 & 17.66 0.02 & 18.41 0.02 & 17.98 0.02 & 15.84 0.03 & 15.88 0.02 & ... ... & 14.77 0.03 &&\
&& 12.51 0.07 & 12.34 0.09 & 9.97 0.01 & 8.26 0.02 & 4.27 0.01 & ... ... & 11 & T2, mul &\
052251.0-680401.0 & 35 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.16 0.08 & 13.21 0.06 & 10.32 0.07 & 8.52 0.04 & ... ... & ... ... & 0 & D &\
052251.7-680436.7 & 39 & 18.35 0.06 & 19.07 0.02 & 18.76 0.03 & ... ... & ... ... & ... ... & ... ... &&\
&& 12.98 0.11 & 12.80 0.11 & 10.41 0.09 & 8.63 0.06 & ... ... & ... ... & 0 & S & near DC\
052253.7-680434.8 & 20 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 12.85 0.14 & 12.57 0.16 & 9.82 0.06 & 8.12 0.04 & 2.01 0.02 & ... ... & 0 & D &\
052254.6-680424.3 & 10 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 12.04 0.06 & 11.90 0.08 & 9.52 0.10 & 7.65 0.03 & ... ... & ... ... & 0 & D &\
052254.9-680422.2 & 6 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 11.93 0.05 & 11.67 0.07 & 9.32 0.04 & 7.49 0.07 & 1.64 0.01 & ... ... & 0 & D &\
052255.2-680409.5 & 8 & 17.16 0.06 & 18.31 0.04 & 18.00 0.04 & 17.18 0.01 & 15.63 0.05 & ... ... & 14.28 0.03 &&\
&& 12.08 0.07 & 11.70 0.07 & 9.37 0.03 & 7.63 0.02 & 2.35 0.01 & ... ... & 11 & T2 &\
052255.4-680431.6 & 22 & ... ... & 20.69 0.03 & 20.15 0.03 & ... ... & ... ... & ... ... & ... ... &&\
&& 13.15 0.12 & 13.05 0.14 & ... ... & 8.25 0.12 & ... ... & ... ... & 0 & D &\
052256.8-680406.9 & 12 & 19.23 0.07 & 20.00 0.04 & 19.52 0.05 & 18.61 0.01 & 16.75 0.09 & ... ... & 14.99 0.04 &&\
&& 12.25 0.09 & 12.26 0.12 & 9.46 0.04 & 7.76 0.02 & 3.29 0.01 & ... ... & 11 & T2 &\
052257.6-680414.1 & 40 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & 18.97 0.28 &&\
&& 13.52 0.24 & 12.14 0.09 & 10.28 0.06 & 8.70 0.04 & 3.65 99.9 & ... ... & 11 & T1/2 &\
052259.0-680346.3 & 83 & ... ... & ... ... & ... ... & 18.97 0.24 & 17.84 0.14 & ... ... & 15.61 0.05 &&\
&& 13.16 99.9 & 12.49 99.9 & 11.61 0.16 & 10.46 0.14 & 5.49 0.05 & ... ... & 11 & T2 &\
052308.7-680006.8 & 43 & ... ... & ... ... & ... ... & 20.50 0.03 & 18.69 0.18 & ... ... & 16.28 0.04 &&\
&& 13.53 0.17 & 12.94 0.15 & 10.61 0.07 & 8.78 0.04 & 3.51 0.01 & ... ... & 11 & YSC & in DC\
052309.4-680205.5 & 58 & 18.48 0.01 & 18.91 0.01 & 18.63 0.01 & ... ... & ... ... & ... ... & ... ... &&\
&& 13.69 0.04 & 13.78 0.04 & 10.88 0.05 & 9.15 0.05 & 6.04 0.02 & ... ... & 0 & D peak &\
052311.4-680040.9 & 54 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.52 0.06 & 13.41 0.04 & 10.81 0.06 & 9.07 0.05 & 4.62 0.03 & ... ... & 0 & D peak &\
052315.1-680017.0 & 57 & 13.40 0.01 & 14.19 0.01 & 14.10 0.01 & 13.86 0.01 & 13.81 0.04 & ... ... & 13.69 0.05 &&\
&& 13.02 0.17 & 12.62 0.17 & 11.20 0.12 & 9.10 0.06 & 2.32 0.01 & ... ... & 11 & T3 & in HII\
052318.0-675938.6 & 32 & ... ... & ... ... & ... ... & ... ... & 16.34 0.14 & 15.94 0.20 & 14.70 99.9 &&\
&& 13.15 0.10 & 12.73 0.07 & 10.25 0.06 & 8.51 0.04 & ... ... & ... ... & 0 & D &\
052318.0-675942.8 & 47 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 13.49 0.16 & 12.94 0.24 & 10.59 0.09 & 8.94 0.06 & ... ... & ... ... & 0 & D peak &\
052331.5-680107.9 & 29 & 17.64 0.03 & 18.27 0.02 & 17.59 0.02 & 16.85 0.01 & 16.12 0.06 & ... ... & 15.41 0.05 &&\
&& 12.91 0.12 & 12.83 0.16 & 10.19 0.06 & 8.42 0.04 & 4.44 0.01 & ... ... & 11 & T2 &\
052335.6-675235.4 & 7 & 16.58 0.02 & 17.19 0.03 & 16.95 0.02 & 16.66 0.01 & 15.86 0.06 & ... ... & 13.69 0.02 &&\
&& 11.33 0.02 & 10.60 0.02 & 9.03 0.01 & 7.59 0.01 & 2.22 0.01 & -1.43 0.08 & 11 & T2 &\
052340.6-680528.5 & 87 & 14.94 0.03 & 15.73 0.07 & 15.73 0.02 & 15.52 0.01 & 15.56 0.07 & ... ... & 15.20 0.18 &&\
&& 14.16 0.04 & 13.45 0.04 & 12.67 0.05 & 10.95 0.04 & 4.90 0.01 & 0.58 0.23 & 11 & T3 or B\[e\] S &\
052343.6-680034.2 & 3 & ... ... & ... ... & ... ... & 23.65 99.9 & 20.63 0.62 & ... ... & 15.76 0.05 &&\
&& 11.50 0.04 & 10.59 0.03 & 8.54 0.01 & 6.75 0.01 & 1.12 0.01 & -2.55 0.07 & 11 & T1/2 &\
052343.9-680056.1 & 88 & ... ... & ... ... & ... ... & 21.66 0.08 & 18.69 0.24 & ... ... & 16.97 0.11 &&\
&& 14.16 0.05 & 13.71 0.03 & 12.40 0.09 & 11.00 0.17 & ... ... & ... ... & 11 & mul YSOs &\
052350.1-675719.7 & 9 & ... ... & ... ... & ... ... & 20.35 0.04 & 18.81 0.21 & ... ... & 16.43 0.07 &&\
&& 12.62 0.09 & 11.46 0.04 & 9.48 0.03 & 7.63 0.01 & 1.35 0.01 & -2.54 0.06 & 11 & T1 &\
052351.1-675326.6 & 95 & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... & ... ... &&\
&& 14.54 0.01 & 13.60 0.22 & 12.52 0.02 & 11.54 0.12 & 9.07 0.07 & ... ... & 0 & ES? &\
[rcrrrll]{}
1 & 530 & 052042.1-675255.0 & $-$0.41 & 0.21 & YSO & G cut\
2 & 546 & 052116.2-674511.7 & $-$0.09 & $-$0.28 & YSO & E cut\
3 & 552 & 052129.7-675106.9 & $-$0.07 & 0.27 & PN & III? E?\
4 & 557 & 052147.1-675656.7 & $-$0.34 & 0.07 & YSO\_hp & III\
5 & 563 & 052155.3-674730.2 & $-$0.52 & $-$0.13 & YSO\_hp & II\
6 & 565 & 052159.0-674437.2 & $-$0.12 & 0.12 & YSO & II\
7 & 566 & 052202.0-675758.2 & $-$0.25 & 0.19 & YSO & II\
8 & 572 & 052207.3-675826.8 & 0.05 & $-$0.03 & YSO & II\
9 & 574 & 052211.5-675401.9 & $-$0.23 & 0.35 & YSO & G cut\
10 & 576 & 052212.0-674713.9 & $-$0.44 & 0.01 & YSO\_hp & II\
11 & 579 & 052212.6-675832.2 & $-$0.10 & $-$0.06 & YSO & I\
12 & 581 & 052216.9-680403.6 & $-$0.28 & $-$0.22 & YSO\_hp & I\
13 & 598 & 052249.2-680129.0 & $-$0.61 & 0.10 & YSO & II\
14 & 604 & 052259.0-680346.3 & $-$0.37 & 0.23 & YSO & II\
15 & 606 & 052302.2-680400.1 & $-$0.47 & 0.34 & YSO & G cut\
16 & 609 & 052308.7-680006.8 & $-$0.67 & 0.23 & YSO & YSC\
17 & 613 & 052318.5-680045.5 & $-$0.52 & $-$0.19 & YSO & E cut\
18 & 616 & 052335.6-675235.4 & $-$0.19 & $-$0.27 & YSO\_hp & II\
19 & 622 & 052351.3-680712.3 & $-$0.81 & 0.10 & YSO\_hp & E cut\
[lrcccrrrrrrrrrrl]{}
052212.6-675832.2 & 5.10 & I & 20-50 & I & 26 & 16000 & 37.7 & 8E+03 & 7.6E$-$03 & 0.0E+00 & 0.0E+00 & 18 & 92000 & 6.1 & DC, UCHII\
052350.1-675719.7 & 7.63 & I & 17-25 & I & 21 & 33000 & 7.5 & 3E+04 & 1.2E$-$03 & 0.0E+00 & 0.0E+00 & 56 & 62000 & 1.1 & DC\
052216.9-680403.6 & 8.52 & I & 8-15 & I,II & 8 & 16000 & 8.8 & 2E+05 & 1.0E$-$05 & 1.8E$-$07 & 6.0E$-$02 & 41 & 4300 & 7.2 &\
052211.9-675818.1 & 8.97 & I & 8-28 & I,II & 28 & 4200 & 300.0 & 1E+03 & 5.9E$-$04 & 0.0E+00 & 0.0E+00 & 56 & 26000 & 2.1 & DC\
052343.6-680034.2 & 6.75 & I/II & 17-37 & I,II & 20 & 18000 & 23.3 & 2E+04 & 5.9E$-$03 & 0.0E+00 & 0.0E+00 & 18 & 53000 & 27.5 & DC, UCHII\
052219.8-680436.8 & 6.87 & I/II & 17-24 & I & 21 & 8000 & 93.7 & 5E+03 & 3.9E$-$04 & 0.0E+00 & 0.0E+00 & 49 & 33000 & 1.9 & DC, UCHII\
052257.6-680414.1 & 8.70 & I/II & 10-16 & I,II & 16 & 8400 & 56.6 & 1E+04 & 6.5E$-$03 & 0.0E+00 & 0.0E+00 & 18 & 15000 & 29.7 & CO peak\
052335.6-675235.4 & 7.59 & II & 14-17 & II,III & 17 & 33000 & 5.3 & 9E+05 & 1.7E$-$08 & 6.7E$-$11 & 1.9E$-$06 & 18 & 30000 & 0.9 &\
052255.2-680409.5 & 7.63 & II & 13-18 & I,II & 13 & 30000 & 4.6 & 3E+05 & 4.0E$-$06 & 7.8E$-$07 & 3.6E$-$01 & 81 & 15000 & 0.0 & UCHII\
052202.2-675753.6 & 7.67 & II & 17-36 & I,II & 18 & 6400 & 125.0 & 6E+03 & 1.4E$-$04 & 8.8E$-$06 & 9.4E$-$02 & 87 & 23000 & 0.2 & DC?\
052256.8-680406.9 & 7.76 & II & 11-12 & I & 12 & 29000 & 4.3 & 2E+05 & 1.7E$-$05 & 1.8E$-$08 & 2.3E$-$02 & 87 & 11000 & 0.7 &\
052202.0-675758.2 & 7.77 & II & 9 & II & 9 & 24000 & 3.7 & 2E+06 & 0.0E+00 & 6.1E$-$07 & 6.8E$-$02 & 56 & 4100 & 9.9 & DC\
052202.8-674701.9 & 7.77 & II & 8-15 & I,II,III & 15 & 8900 & 47.3 & 1E+04 & 2.0E$-$04 & 3.0E$-$06 & 6.9E$-$02 & 49 & 13000 & 0.1 & DC tip\
052046.6-675255.1 & 7.81 & II & 17 & II & 17 & 33000 & 5.3 & 1E+06 & 0.0E+00 & 8.7E$-$07 & 1.5E$-$01 & 87 & 32000 & 0.0 &\
052203.1-674703.5 & 7.81 & II & 8-14 & I,III & 8 & 16000 & 8.8 & 2E+05 & 1.0E$-$05 & 1.8E$-$07 & 6.0E$-$02 & 41 & 4300 & 0.4 &\
052207.3-675826.8 & 8.38 & II & 7-25 & I,II & 11 & 4400 & 76.8 & 5E+03 & 1.8E$-$04 & 1.8E$-$04 & 4.0E$-$01 & 18 & 2700 & 7.8 & DC\
052231.8-680319.2 & 8.40 & II & 8-15 & I,II & 8 & 12000 & 13.9 & 1E+05 & 8.9E$-$05 & 5.2E$-$10 & 8.0E$-$04 & 18 & 3200 & 2.9 & DC?\
052206.4-675659.2 & 8.41 & II & 8-15 & I,II & 13 & 5900 & 78.2 & 1E+04 & 4.2E$-$03 & 4.0E$-$06 & 2.2E$-$01 & 18 & 6600 & 1.2 &\
052331.5-680107.9 & 8.42 & II & 8 & I & 8 & 16000 & 8.8 & 2E+05 & 1.0E$-$05 & 1.8E$-$07 & 6.0E$-$02 & 69 & 4300 & 0.9 &\
052203.9-675743.7 & 8.98 & II & 8-15 & I,II & 8 & 22000 & 3.4 & 8E+05 & 5.0E$-$07 & 5.5E$-$06 & 1.2E$-$01 & 31 & 2700 & 2.0 & DC\
052138.0-674630.3 & 9.52 & II & 9-10 & I & 9 & 7000 & 31.4 & 4E+04 & 4.5E$-$04 & 1.2E$-$05 & 2.4E$-$02 & 18 & 2200 & 3.5 &\
052212.0-674713.9 & 10.10 & II & 2-9 & I & 9 & 5100 & 48.7 & 3E+04 & 1.3E$-$03 & 5.1E$-$07 & 8.7E$-$02 & 18 & 1400 & 1.9 &\
052242.0-675500.5 & 10.24 & II & 5-10 & I & 9 & 6200 & 31.4 & 5E+04 & 6.3E$-$05 & 4.8E$-$07 & 6.1E$-$02 & 41 & 1300 & 3.6 &\
052259.0-680346.3 & 10.46 & II & 4-11 & I,II & 6 & 18000 & 3.6 & 5E+05 & 1.0E$-$05 & 2.9E$-$08 & 1.7E$-$03 & 18 & 1200 & 1.9 &\
052204.8-675744.6 & 8.76 & II/III & 8-33 & I,II,III & 9 & 6000 & 43.2 & 3E+04 & 6.1E$-$05 & 3.9E$-$09 & 2.4E$-$03 & 31 & 2200 & 0.7 & DC edge\
052158.2-675554.2 & 8.81 & II/III & 5-14 & I,III & 12 & 7100 & 51.8 & 2E+04 & 3.8E$-$03 & 1.7E$-$07 & 2.5E$-$03 & 18 & 6100 & 1.0 &\
052155.3-675634.9 & 8.89 & II/III & 8-9 & I,II & 8 & 22000 & 3.4 & 4E+05 & 6.2E$-$05 & 8.2E$-$08 & 1.6E$-$02 & 63 & 2300 & 0.5 &\
052212.3-675813.4 & 8.97 & II/III & 5-28 & I,II,III & 12 & 7100 & 51.8 & 2E+04 & 3.8E$-$03 & 1.7E$-$07 & 2.5E$-$03 & 18 & 6100 & 1.0 & superposed on DC\
052136.0-675443.4 & 9.32 & II/III & 6-11 & II,III & 6 & 13000 & 5.8 & 5E+05 & 2.3E$-$07 & 1.9E$-$09 & 3.2E$-$03 & 41 & 890 & 0.1 & dust pillar tip\
052129.7-675106.9 & 4.12 & III & 34 & II & 34 & 42000 & 7.8 & 5E+05 & 1.2E$-$06 & 1.1E$-$10 & 4.6E$-$04 & 41 & 160000 & 0.0 & $\sim$O9I, this study\
052147.1-675656.7 & 7.84 & III & 17 & III & 17 & 33000 & 5.3 & 1E+06 & 0.0E+00 & 8.7E$-$13 & 1.0E$-$06 & 81 & 32000 & 0.1 &\
052207.3-675819.9 & 8.01 & III & 9-15 & II & 9 & 25000 & 3.7 & 3E+05 & 3.0E$-$06 & 1.8E$-$08 & 1.1E$-$03 & 41 & 4600 & 0.4 &\
052315.1-680017.0 & 9.10 & III & 20-21 & III & 21 & 36000 & 5.9 & 1E+06 & 0.0E+00 & 4.7E$-$12 & 2.4E$-$06 & 75 & 50000 & 0.0 & B0-O5V\
052157.0-675700.1 & 9.96 & III & 9 & I & 9 & 9200 & 20.1 & 6E+04 & 9.6E$-$05 & 9.9E$-$06 & 5.5E$-$01 & 18 & 2800 & 0.0 & O8.5V\
052159.6-675721.7 & 10.09 & III & 16 & III & 16 & 32000 & 5.1 & 1E+06 & 0.0E+00 & 1.5E$-$11 & 4.7E$-$07 & 31 & 25000 & 0.0 & O7.5V\
052340.6-680528.5 & 10.95 & III & 12 & III & 12 & 28000 & 4.4 & 2E+06 & 0.0E+00 & 9.3E$-$14 & 7.4E$-$07 & 87 & 11000 & 0.0 & B\[e\]\
[llcccllcc]{}
B0522$-$5800 & 052212.6$-$675832.2 & I & I & 26 & O7-8 V & O7 V & 7.6E$-$03 & 4.1E$-$05\
B0523$-$6806(NE) & 052343.6$-$680034.2 & I/II & I,II & 20 & O8-9 V & O9 V & 5.9E$-$03 & 1.8E$-$05\
B0523$-$6806(SW) & 052219.8$-$680436.8 & I/II & I & 21 & O8-9 V & O8.5 V & 3.9E$-$04 & 2.2E$-$05\
B0523$-$6806 & 052255.2$-$680409.5 & II & I,II & 13 & B1-2 V & B0 V & 4.0E$-$06 & 1.0E$-$05\
[^1]: The Online SED Model Fitter is available at http://caravan.astro.wisc.edu/protostars/fitter/index.php.
[^2]: The code is available at http://caravan.astro.wisc.edu/protostars/codes/index.php.
[^3]: Owing to the $\sim300$ [km s$^{-1}$]{} redshift of the LMC, the filter transmission of the red-shifted [H$\alpha$]{} line is $\sim93$% of the peak transmission, thus the extracted [H$\alpha$]{} surface brightness and flux are multiplied by a correction factor of 1.07.
[^4]: Note that for the same spectral type, @Pa73 used a lower temperature scale than @SK82. We have adopted Schmidt-Kaler’s (1982) $UBV$ photometric calibrations for spectral types; therefore, we use Panagia’s (1973) calibration to convert ionizing power to stellar effective temperature and use Schmidt-Kaler’s (1982) calibration to convert from stellar effective temperature to spectral type.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Evolution of a Bose-Einstein condensate subject to a time-dependent external perturbation can be described by a time-dependent Bogoliubov theory: a condensate initially in its ground state Bogoliubov vacuum evolves into a time-dependent excited state which can be formally written as a time-dependent Bogoliubov vacuum annihilated by time-dependent quasiparticle annihilation operators. We prove that any Bogoliubov vacuum can be brought to a diagonal form in a time-dependent orthonormal basis. This diagonal form is tailored for simulations of quantum measurements on an excited condensate. As an example we work out phase imprinting of a dark soliton followed by a density measurement.'
address: ' Institute of Physics and Centre for Complex Systems, Jagiellonian University, Reymonta 4, 30-059 Kraków, Poland '
author:
- Jacek Dziarmaga and Krzysztof Sacha
date: ' March 4, 2005 '
title: ' Images of a Bose-Einstein condensate: diagonal dynamical Bogoliubov vacuum '
---
Quantum measurements on Bose-condensed systems can give quite unexpected results. For example, in the classic paper by Javanainen and Yoo [@Fock] a density measurement on a Fock state $|N/2,N/2\rangle$ with $N$ particles equally divided between two counter-propagating plane waves $e^{\pm ix}$ reveals an interference pattern $\rho(x|\varphi)\sim \cos^2(x-\varphi)$ with a phase $\varphi$ chosen randomly in every realization of the experiment. The Fock state has a uniform single particle density distribution, but its measurement unexpectedly reveals interference between the two counter-propagating condensates. The Fock state is a quantum superposition over $N$-particle condensates with different relative phases $\varphi$ in their wave functions [@phase], $|N/2,N/2\rangle\sim\int
d\varphi~|N:e^{+i(x-\varphi)}+e^{-i(x-\varphi)}\rangle$, but every single realization of the experiment reveals such a density distribution as if the state before the density measurement were one of the condensates $|N:e^{+i(x-\varphi)}+e^{-i(x-\varphi)}\rangle$ with a randomly chosen phase $\varphi$. This effect is best explained [@phase] when the density measurement, which is a destructive measurement of all particle positions at the same time, is replaced by an equivalent sequential measurement of one position after another. With an increasing number $n$ of measured positions a quantum state of the remaining $N-n$ particles gradually “collapses” from the initial uniform superposition over all phases to a state with a more and more localized phase $\varphi$. For a large $N$ a measurement of only a small fraction $\frac{n}{N}\ll 1$ of all particles practically collapses the state of remaining $N-n$ particles to a condensate with phase $\varphi$. The lesson from this instructive example [@Fock; @phase] is that when we want to predict possible outcomes of measurements on a Bose-condensed state, then it is helpful to rewrite the state as a superposition over condensates $\int {\cal D}\phi~\psi(\phi)~|N:\phi\rangle$ with different condensate wave functions $\phi(\vec r)$ in a hope that $P(\phi)\sim|\psi(\phi)|^2$ will give an approximate probability for different density measurement outcomes $\rho(\vec
r|\phi)=N|\phi(\vec r)|^2$. However, as the set of all condensates $|N:\phi\rangle$ is non-orthogonal and overcomplete, the “amplitude” $\psi(\phi)$ is not unique but depends on coordinates used to parameterize the space of condensate wave functions $\phi(\vec r)$. In general a special care should be taken to choose the right parameterization where the simple “Born rule” $P(\phi)\sim|\psi(\phi)|^2$ gives the correct probability.
The existing literature about measurements on Bose-condensed systems [@Fock; @phase] concentrates on simple most beautiful examples of Bose-condensed states. In this letter we construct for the first time a measurement theory able to predict probability of different density measurement outcomes on a realistic Bose-Einstein condensate of interacting particles evolving under influence of external time-dependent potentials. This problem is important because in many experiments, like phase imprinting of dark solitons [@Hannover] or condensate splitting in atom interferometers [@interferometer], manipulation of the condensate generates substantial dynamical depletion which can qualitatively affect measured density patterns. In addition to being realistic our theory also reveals a beautiful diagonal structure hidden in the quantum state of a condensate excited from its ground state by a time-dependent potential.
A generic Bose-Einstein condensate consisting of $N$ weakly interacting atoms can be described by a number-conserving version of the Bogoliubov theory [@CastinDum] which assumes that most particles occupy the condensate wave function. Time-dependent problems can be treated with the time-dependent version of the theory where the condensate wave function $\phi_0(t)$ evolves according to the time-dependent Gross-Pitaevskii equation (GPE) and the Bogoliubov modes $u_m(t)$ and $v_m(t)$ solve the time-dependent Bogoliubov-de Gennes equations (BdGE) [@CastinDum]. In this framework, starting with a condensate in the ground state, time-dependent perturbation excites the condensate to a state which is formally a time-dependent Bogoliubov vacuum $|0_b\rangle$ annihilated by the time-dependent quasiparticle annihilation operators b\_m(t) = . Here the operator $\hat a_0=\langle\phi_0(t)|\hat\psi\rangle$ annihilates an atom in a condensate wave function $\phi_0(t)$. The modes $u_m(t)$ and $v_m^*(t)$ are solutions of the time-dependent BdGE projected on the subspace orthogonal to $\phi_0$, i.e. $\langle u_m|\phi_0\rangle=\langle v_m^*|\phi_0\rangle=0$, and normalized so that $\langle u_m|u_n\rangle-\langle v_m|v_n\rangle=\delta_{mn}$. In the time-dependent vacuum $|0_b\rangle$ the reduced single particle density matrix (r,r’)= N\_0\^\*(r)\_0(r ’)+ \_[m=1]{}\^v\_m(r) v\_m\^\*(r ’), \[sing\] has common eigenstates with the operator =\_[m=1]{}\^|v\_mv\_m |= \_N\_|\_\^\*\_\^\*| . \[drho\] A mode $|\phi_\alpha^*\rangle$ is a non-condensate eigenmode of the density matrix (\[sing\]) occupied on average by ${\rm d}N_\alpha$ particles. Thanks to the completeness relation \_m |u\_mu\_m|-\_m |v\_m\^\*v\_m\^\*|=1, \[complet\] valid in the subspace orthogonal to $|\phi_0\rangle$, the operator ${\rm d}\hat\rho^*=\sum_m|v_m^*\rangle\langle v_m^*|$ has common eigenstates with the \_[m=1]{}\^|u\_mu\_m|= 1+[d]{}\^\*= \_(1+[d]{}N\_) |\_\_|. Furthermore $|\phi_\alpha\rangle$’s allow us to get a diagonal form of =\_m|v\_mu\_m|= \_\_|\_\^\*\_|, \[hatDelta\] which is a map between the subspace orthogonal to $|\phi_0\rangle$ and the subspace orthogonal to $|\phi_0^*\rangle$. Indeed, the property $\langle u_m|u_n\rangle-\langle v_m |v_n\rangle=\delta_{nm}$ of the Bogoliubov modes and Eq.(\[complet\]) imply a quasi-commutator ${\rm d}\hat\rho\hat\Delta=\hat\Delta{\rm d}\hat\rho^*$ from which follows the diagonal form in Eq.(\[hatDelta\]). What is more, owing to ${\rm d}\hat\rho^*(1+{\rm d}\hat\rho^*)=\hat\Delta^*\hat\Delta$, one gets $
|\Delta_\alpha|^2={\rm d}N_\alpha(1+{\rm d}N_\alpha).
$ In Ref. [@stanz] we have shown that in the stationary case the Bogoliubov vacuum has the following form |0\_b \~ ( a\_0\^a\_0\^+ \_[,=1]{}\^Z\_[,]{} \_\^\_\^ )\^[N/2]{} |0 , \[nz\] where $\hat{\tilde a}_\alpha$’s annihilate atoms in single particle states $\tilde\phi_\alpha$ orthogonal to the condensate wave function $\phi_0$. Matrix $Z_{\alpha,\beta}$ fulfills the equation v\_m|\_\^\*=u\_m| \_ Z\_[,]{}, \[Z\] which is equivalent to the set of the conditions $\hat b_m|0_b\rangle=0$. In the stationary and time reversal invariant case all modes can be chosen as real functions so that the matrix $Z_{\beta,\alpha}$ is real symmetric and can be diagonalized by an orthogonal transformation [@stanz; @leggett]. In a generic time-dependent case, or when a stationary $\phi_0$ breaks the $T$-invariance like in e.g. vortex state, the matrix $Z$ is complex and symmetric. However, using only the general properties of ${\rm d}\hat\rho$ and $\hat\Delta$ it is easy to show that the choice of $|\phi_\alpha\rangle$’s for the modes $|\tilde\phi_\alpha\rangle$ in Eq.(\[nz\]) makes the time-dependent Bogoliubov vacuum diagonal |0\_b \~ ( a\_0\^a\_0\^+ \_[=1]{}\^\_a\_\^a\_\^)\^[N/2]{} |0 , \[0lambda\] where the eigenvalues of the matrix $Z$ are \_=. \[lam\] Indeed, an action of $\sum_m\langle \phi_\alpha^*|v_m\rangle$ on both sides of Eq. (\[Z\]) yields $
\langle \phi_\alpha^*|{\rm d}\hat\rho|\phi^*_{\beta}\rangle=
\langle \phi_\alpha^*|\hat\Delta|\phi_{\alpha'}\rangle
Z_{\alpha',\beta}~,
$ equivalent to Eq.(\[lam\]).
Hence the procedure of time evolution of the Bogoliubov vacuum state in the particle representation becomes extremely simple: i) First one has to evolve Bogoliubov modes $(u_m,v_m)$ together with the condensate wave function $\phi_0$ which is easy because different modes evolve independently from each other. ii) Next one has to diagonalize the operator ${\rm d}\hat\rho=\sum_m|v_m\rangle\langle v_m|$ (diagonalization of $1+{\rm d}\hat\rho^*=\sum_m|u_m\rangle\langle u_m|$ turns out to be better convergent numerically) in order to get eigenvalues ${\rm d}N_\alpha$ and eigenvectors $|\phi_\alpha\rangle$. iii) Finally one calculates $\Delta_\alpha=\langle\phi_\alpha^*|\hat\Delta|\phi_{\alpha}\rangle$ and from Eq. (\[lam\]) obtains values of $\lambda_\alpha$. Note that there is a one to one correspondence between moduli of $\lambda_\alpha$ and the eigenvalues of the single particle density matrix ${\rm d}N_\alpha$: |\_|==. However, the phases of $\lambda$’s cannot be determined from the single particle matrix only, but they also depend on the eigenvalues $\Delta_\alpha=\langle\phi_\alpha^*|\hat\Delta|\phi_\alpha\rangle$, see Eq.(\[lam\]). Finally, once the phases are known it is convenient to make the transformation ’\_= |\_| , ’\_= \_e\^[12i[Arg]{}(\_)]{} . In the following we skip the primes.
For large $N$ the diagonal state (\[0lambda\]) can be rewritten as a superposition over $N$-particle condensates $$\begin{aligned}
|0_b\rangle~\sim~
\int dq~
e^{-\sum_{\alpha}
\frac{1-\lambda_\alpha}{2\lambda_\alpha}
q_\alpha^2 }~
|N:\phi(\vec r|q)\rangle~.
\label{0q}\end{aligned}$$ with the normalized condensate wave functions $$\phi(\vec r|q)~=~
\frac{ \phi_0(\vec r)+\frac{1}{\sqrt{N}}\sum_\alpha
q_\alpha\phi_\alpha(\vec r) }
{ \sqrt{1+\frac{1}{N}\sum_\alpha q_\alpha^2} }~.
\label{phiq}$$ The gaussian amplitude gives an accurate vacuum state when $\sum_\alpha {\rm d}N_\alpha \ll \sqrt{N}$. The gaussian square of the amplitude P(q)\~ \_e\^[-q\_\^2]{} \[Pq\] is a candidate for the probability distribution for different $q_\alpha$. If this choice is correct, then simulation of the density measurement amounts to choosing a random set of $q_\alpha$’s which determines the measured density (r|q)= N|(r|q)|\^2 . \[rhoq\] To show the correctness of Eq.(\[Pq\]), we average $\rho(\vec r|q)$ over $P(q)$ and for $\sum_\alpha {\rm d}N_\alpha\ll N$ we get the density (r)= N|\_0(r)|\^2+ \_ |\_(r)|\^2 \[rhoav\] averaged over different realizations of $q$. This average density should be equal to the expectation value of the density operator in the state (\[0q\]) 0\_b|\^(r)(r)|0\_b= N|\_0(r)|\^2+ \_N\_|\_(r)|\^2 \[rhoexp\] with ${\rm d}N_\alpha=\lambda_\alpha^2/(1-\lambda_\alpha^2)$. These two densities are the same for the highly occupied modes with ${\rm d}N_\alpha\gg 1$ or $\lambda_\alpha\approx 1^-$. There are discrepancies for poorly occupied modes but those modes give negligible contribution to the total density. The probability (\[Pq\]) accurately reproduces the single particle density matrix. This accuracy can be best explained when we look at the overlap between condensates N:(r|q)|N:(r|q’)= e\^[-\_(q\_-q’\_)\^2]{} . Condensates become orthogonal on the length scale of $1$. For the highly occupied modes the fluctuations of $q_\alpha\sim\sqrt{{\rm d}N_\alpha}$ are much greater than the width of the overlap and for these modes the overlap can be replaced by a delta function $\delta(q_\alpha-q'_\alpha)$. With the delta overlap the condensates are orthogonal and the probability $P(q)$ is simply given by the Born rule in Eq.(\[Pq\]).
To summarize, a density measurement, which is a measurement of all atomic positions, can be approximately simulated in two steps. In the first step a condensate wave function (\[phiq\]) is chosen randomly from the gaussian distribution $P(q)\sim\prod_\alpha \exp(-q_\alpha^2/2{\rm d}N_\alpha)$. In the second step atomic positions are measured in the chosen condensate. The first step already gives a smoothed density distribution $N|\phi(\vec r|q)|^2$ on top of which the second step only adds statistical fluctuations. In most applications one is not interested in the statistical fluctuations but in the smoothed density obtained after filtering out the statistical noise, compare the histogram and the solid line in Fig.2. Thus for most applications the density measurement can be very efficiently and accurately simulated with only the first step of the procedure which immediately gives a smoothed density $\rho(\vec r|q)=N|\phi(\vec r|q)|^2$ with randomly chosen $q$’s. In the following we give an example of this procedure: density measurement after phase imprinting of a dark soliton.
To describe evolution of the condensate wave function we solve the time-dependent GPE that in harmonic oscillator trap units reads i\_t\_0= -12\_x\^2\_0+ \_0+ g|\_0|\^2\_0+ V(t,x)\_0 . \[GPE\] Here $V(t,x)$ is an external potential created by the laser beam. As the effective 1D interaction strength we choose $g=7500$ corresponding to parameters of the Hannover experiment [@Hannover]. The dark soliton is a kink $\phi_0\sim\tanh(x/\xi)$ with the wave function winding a phase of $\pi$ as $x$ goes from negative to positive. At the same time the density drops to zero at $x=0$ explaining the name dark soliton. Experimentally the soliton is excited by a short laser pulse which imprints on the wave function a phase that changes by $\pi$ when one goes from negative to positive $x$. To simulate the phase imprinting we started with the ground state of the stationary GPE for a condensate in the harmonic trap, $\phi_0(x,0)$, and applied (similarly as in Ref.[@Law] where they use rather small $g=200$) the perturbation $V(t,x)=V_0[1+\tanh(x/l_0)]$ (where $V_0=140$, $l_0=0.05$) that lasted for $0.034$. Then $\phi_0(x,t)$ was evolved up to $t=0.15$ (or $1.7$ms). The density of the resulting wave function is plotted in Fig \[one\](a) where the dark soliton is clearly visible in the center of the trap.
The Bogoliubov modes solve the time-dependent BdGE i\_t u\_m&=&(H\_[GP]{}+g|\_0(x,t)|\^2) u\_m + g\_0\^2 v\_m, i\_t v\_m&=&-(H\_[GP]{}+g|\_0(x,t)|\^2)v\_m - g[\_0\^\*]{}\^2 u\_m, where $H_{\rm GP}=-\partial_x^2/2 +x^2/2+ V(x,t)+g|\phi_0(x,t)|^2$. We have evolved the Bogoliubov modes, starting with the eigenmodes of the stationary BdGE for a condensate in the ground state, up to $t=0.15$. The single particle density $\rho(x,t=0.15)$ is plotted in Fig. \[one\](b). In Fig. \[one\](c) the densities of condensate, noncondensate atoms and the total single particle density are plotted. At both values of $g$ (the realistic $g=7500$ here and the $g=200$ in Ref.[@Law]) the single particle density $\rho(x,t)$ shows that minimum at the soliton location is gradually filled with particles depleted from the condensate. These dynamical studies confirm earlier predictions that a static soliton is going to fill up after a few milliseconds [@greying].
After the evolution of the Bogoliubov modes was completed we calculated the eigenvalues $\lambda_\alpha$ and the non-condensate eigenmodes $\phi_\alpha(x)$ at the final time. It turns out that the soliton notch is filled up with atoms depleted to only one mode: the $\phi_1(x)$ with the largest ${\rm d}N_1=273$. In Fig.\[two\](d) the total depletion density $\sum_m|v_m(x)|^2$ (solid line) is plotted together with a barely visible dashed plot of the ${\rm d}N_1|\phi_1(x)|^2$. As the depletion in the soliton notch comes only from $\alpha=1$ and because we would like to know what will be seen in the soliton notch in a single experiment we truncate the final time-dependent Bogoliubov vacuum to $\left( \hat a_0^\dagger\hat a_0^\dagger +
\lambda_1
\hat a_1^\dagger\hat a_1^\dagger \right)^{N/2}~|0\rangle$. We made exact simulation of density measurement in the truncated vacuum using the algorithm of Ref.[@Fock] and Ref.[@greying]. A typical outcome is shown as the histogram in Fig.\[two\]. It turns out that every observed histogram can be very well fitted with the density $\rho(x|q_1)=N|\phi_0(x)+q_1\phi_1(x)/\sqrt{N}|^2$ where the only free parameter is $q_1$. The histogram in Fig.\[two\] is fitted by the density $\rho(x|14.5)$ shown in the figure as the solid line. This is in agreement with our measurement theory predicting the smoothed histograms to be $\rho(x|q_1)$ with a random $q_1$. In conclusion, we derived a convenient diagonal form of the time-dependent Bogoliubov vacuum which greatly facilitates simulations of quantum measurements on Bose-condensed systems.
Acknowledgements {#acknowledgements .unnumbered}
==================
We thank Zbyszek Karkuszewski for stimulating discussions. KS was supported by the KBN grant PBZ-MIN-008/P03/2003. The work of JD was supported by Polish Goverment scientific funds (2005-2008) as a research project.
[99]{}
J. Javanainen and S.Mi. Yoo, Phys. Rev. Lett. [**76**]{}, g161 (1996).
C. J. Pethick and H. Smith, [*Bose-Einstein Condensation in Dilute Gases*]{}, (Cambridge Univ. Press, 2002); Y. Castin and J. Dalibard, Phys. Rev. A [**55**]{}, 4330 (1997); K. Mølmer, Phys. Rev. A [**65**]{}, 021607 (2002); S. Ashhab and A.J. Leggett, Phys. Rev. A [**65**]{}, 023604 (2002).
S. Burger [*et al.*]{}, Phys. Rev. Lett. [**83**]{}, 5198 (1999).
Y. Shin [*et al.*]{}, Phys. Rev. Lett. [**92**]{}, 050405 (2004).
M. D. Girardeau and R. Arnowitt, Phys. Rev. [**113**]{}, 755 (1959); C. W. Gardiner, Phys. Rev. A [**56**]{}, 1414 (1997); M. D. Girardeau Phys. Rev. A [**58**]{}, 775 (1998); Y. Castin and R. Dum, Phys. Rev. A [**57**]{}, 3008 (1998);
J. Dziarmaga and K. Sacha, Phys. Rev. A [**67**]{} 033608 (2003).
A.J. Leggett, Rev. Mod. Phys.[**73**]{}, 307 (2001).
C.K. Law, Phys. Rev. A [**68**]{}, 015602 (2003).
J. Dziarmaga, Z.P. Karkuszewski and K. Sacha, J. Phys. B [**36**]{}, 1217 (2003).
![ Panel (a) shows condensate density, i.e. $N|\phi_0(x,t)|^2$, at the time $t=0.15$. In panel (b) single particle density $\rho(x,t)=N|\phi_0(x,t)|^2+\sum_m |v_m(x,t)|^2$ is plotted. Panel (c) shows single particle density $\rho(x,t)$ (solid line), condensate density $N|\phi_0(x,t)|^2$ (dashed line) and density of noncondensate atoms $\sum_m |v_m(x,t)|^2$ (dotted line) in the vicinity of the trap center. In panel (d) $\sum_m |v_m(x,t)|^2$ (solid line) is plotted together with ${\rm d}N_1|\phi_1(x,t)|^2$ (dashed line barely visible behind the solid line) where $\phi_1(x,t)$ is the most occupied mode (${\rm d}N_1\approx273$) orthogonal to $\phi_0(x,t)$. The parameters correspond to the Hannover experiment where $N=1.5\cdot 10^5$ (time $t=0.15$ corresponds to 1.7 ms). []{data-label="one"}](fig1.eps){width="8.6cm"}
![ The histogram is the normalized density distribution measured in a single exact simulation of the density measurement. The solid line is a fit to the histogram with one of the predicted $\rho(x|q_1)=N|\phi_0(x)+q_1\phi_1(x)/\sqrt{N}|^2$ with $q_1=14.5$. The dashed line is the total average density $\rho(x)$ which is equal to an average over densities measured in many experiments. []{data-label="two"}](fig2.eps){width="8.6cm"}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The dynamics of the quantum discord for two identical qubits in both two independent single-mode cavities and a common single-mode cavity are discussed. For the initial Bell state with correlated spins, while the entanglement sudden death can occur, the quantum discord vanishes only at discrete moments in the independent cavities and never vanishes in the common cavity. Interestingly, quantum discord and entanglement show opposite behaviors in the common cavity, unlike in the independent cavities. For the initial Bell state with anti-correlated spins, quantum discord and entanglement behave in the same way for both independent cavities and a common cavity. It is found that the detunnings always stabilize the quantum discord.'
address: |
$^{1}$Department of Physics, Zhejiang University, Hangzhou 310027, P. R. China\
$^{2}$Center for Statistical and Theoretical Condensed Matter Physics, Zhejiang Normal University, Jinhua 321004, P. R. China
author:
- 'Chen Wang$^{1}$'
- 'Qing-Hu Chen$^{1,2,}$'
title: 'Quantum discord dynamics of two qubits in the single-mode cavities'
---
Introduction
============
Quantum entanglement, originated from nonlocal quantum correlation, is fundamental in quantum physics both for understanding the nonlocality of quantum mechanics [@Einstein] and plays an important role in almost all efficient protocols for quantum computations and communications [@Nielsen]. Due to the unavoidable interaction with the environment, an initially entangled two-qubit system becomes totally disentangled after evolving for a finite time. This phenomena is called entanglement sudden death (ESD) [@yu] and has been recently demonstrated experimentally [@Almeida]. However, the entanglement may fail to capture the existence of the quantum correlation in some mixed separate states, in which the entanglement is considered not a good measure [@ollivier1; @oppenheim1].
Recently, a new kind of the quantum correlation, quantum discord (QD) has attracted a lot of attentions [@modi1]. It provides the alternative route for measurement, which is present even under separable states [@ollivier1]. The definition of the QD can be interpreted as the difference of the total quantum information of the two sub-systems $A$ and $B$ before and after the local operation on the one of them. The QD has been proved as a good measure of the non-classical correlations beyond entanglement. Furthermore, the QD has been indicated as the source to speed up the quantum computations [@datta1; @lanyon1].
Some works have been devoted to the QD dynamics of two qubits coupled to Markovian [@Werlang; @Altintas] and non-Markovian [@Fanchini] environments. The comparisons with entanglement dynamics have been also performed. However, the relevant study on two two-level atoms (qubits) coupled to independent or common single-mode cavities without dissipations has not been found in the literature, to the best of our knowledge. We believe that the QD dynamics in these qubit systems is also of fundamental interest. In addition, some essential pictures can be clearly described and unfolded in the framework of the simple model where the exact solutions are available. Actually, the entanglement dynamics for two independent Jaynes-Cummings (JC) atoms has been well studied previously [@Eberly1; @Yonac; @Ficek; @Sainz; @chen; @Agarwal]. The ESD was observed obviously from the initial Bell states with correlated spins. This feature would prevent the application of the entanglement as basic resource for quantum information processing. What is the consequence for the QD in this kind of the qubit system?. It is just the main topic of the present study.
In the present paper, we will study the QD dynamics for two identical qubits in both two independent identical single-mode cavities and one common single-mode cavity. Comparisons with the corresponding pairwise entanglement, i.e. concurrence, are also given. The paper is organized as follows. In Sec. II and III, we derive the time dependent QD in these two systems if initiated from two typical Bell states. In Sec. IV, the results are given and discussions are made. The conclusion is presented in the last section.
QD in two identical Jaynes-Cummings atoms
=========================================
The Hamiltonian of two identical Jaynes-Cummings atoms is shown as $$\begin{aligned}
\label{ham:1}
H_{\text{JC}}&=&\frac{\Delta}{2}({\sigma}^{A}_{z}+{\sigma}^{B}_{z}) +{\omega}%
(a^{\dag}a+b^{\dag}b) \\
&&+g(a^{\dag}{\sigma}^{A}_{-}+{\sigma}^{A}_{+}a)+ g(b^{\dag}{\sigma}^{B}_{-}+%
{\sigma}^{B}_{+}b). \nonumber\end{aligned}$$ where $\sigma^{A(B)}_{k} (k=x,y,z)$ is the Pauli operator of the atom A(B), shown as $\sigma_z=|\uparrow{\rangle}{\langle}\uparrow|-|\downarrow{\rangle}{%
\langle}\downarrow|$ and $\sigma_x=|\uparrow{\rangle}{\langle}%
\downarrow|+|\downarrow{\rangle}{\langle}\uparrow|$, with $|\uparrow{\rangle}
(|\downarrow{\rangle})$ the excited (ground) state of the two-level atom, $%
a^{\dag} (b^{\dag})$ and $a (b)$ are the creator and annihilator of the cavity A (B), respectively, $\Delta$ and $\omega$ are the frequencies of the atom and the cavity, $g$ is the atom-cavity coupling strength. Here we set $%
\hbar=1$ and the detunning $\delta=\Delta-\omega$.
We first study the evolution of the QD initiated from the Bell state with anti-correlated spins, which has the following form $$\begin{aligned}
|\Psi^{(1)}_{\text{Bell}}{\rangle}=\sin{\alpha}|{\downarrow}{\uparrow}{%
\rangle} +\cos{\alpha}|{\uparrow}{\downarrow}{\rangle}.\end{aligned}$$ Initially, the vacuum state of the cavity is considered, so the initial state of the whole system can be written as $$\begin{aligned}
~ |\Psi(0){\rangle}&=&(\sin{\alpha}|{\downarrow}{\uparrow}{\rangle} +\cos{%
\alpha}|{\uparrow}{\downarrow}{\rangle}){\otimes}|00{\rangle} \label{f:1} \\
&=&\sin{\alpha}|{\downarrow}{\uparrow}00{\rangle}+ \cos{\alpha}|{\uparrow}{%
\downarrow}00{\rangle}. \nonumber\end{aligned}$$ The time dependent wave function can be generally expressed as [@Eberly1] $$\begin{aligned}
~ |\Psi(t){\rangle}&=&x_1|{\uparrow}{\downarrow}00{\rangle} +x_2|{\downarrow}%
{\uparrow}00{\rangle} \label{f:2} \\
&&+x_3|{\downarrow}{\downarrow}10{\rangle} +x_4|{\downarrow}{\downarrow}01{%
\rangle}, \nonumber\end{aligned}$$ where the coefficients are $$\begin{aligned}
x_1&=&(Ae^{-i\lambda_{+}t}+Be^{-i\lambda_{-}t})\cos{\alpha} \\
x_2&=&(Ae^{-i\lambda_{+}t}+Be^{-i\lambda_{-}t})\sin{\alpha} \nonumber \\
x_3&=&C(e^{-i\lambda_{+}t}-e^{-i\lambda_{-}t})\cos{\alpha} \nonumber \\
x_4&=&C(e^{-i\lambda_{+}t}-e^{-i\lambda_{-}t})\sin{\alpha}. \nonumber\end{aligned}$$ with the eigenfrequencies as $$\begin{aligned}
~ \lambda_{\pm}=\omega+\frac{\delta}{2}{\pm}\frac{\sqrt{\delta^2+G^2}}{2},
\label{eigen:1}\end{aligned}$$ here $G=2g$. The auxiliary parameters are shown as $$\begin{aligned}
~ A&=&\frac{1}{2}(1+\frac{\delta}{\sqrt{\delta^2+G^2}}) \label{aux:1} \\
B&=&\frac{1}{2}(1-\frac{\delta}{\sqrt{\delta^2+G^2}}) \nonumber \\
C&=&\frac{G}{2\sqrt{\delta^2+G^2}}. \nonumber\end{aligned}$$
The pairwise density matrix from Eq. (\[f:2\]) under the standard basis $\{|{\downarrow}{\downarrow}{\rangle},|{\downarrow}{\uparrow}{\rangle}
, |{\uparrow}{\downarrow}{\rangle},|{\uparrow}{\uparrow}{\rangle}\}$ is thus expressed by tracing the freedoms of the cavities $\rho_{AB}(t)=\text{Tr}
_{cav}\{\rho(t)\}=\text{Tr}_{cav}\{|\Psi(t){\rangle}{\langle}\Psi(t)|\}$, $$\begin{aligned}
~ {\rho}_{AB}(t)&=&\frac{1}{4} \left(
\begin{array}{llll}
|x_3|^2+|x_4|^2 & 0 & 0 & 0 \\
0 & |x_2|^2 & x^{*}_1x_2 & 0 \\
0 & x^{*}_2x_1 & |x_1|^2 & 0 \\
0 & 0 & 0 & 0
\end{array}
\right). \label{rho:1}\end{aligned}$$
With this density matrix, the routine to derive the QD is formally given in the Appendix A. The von Neumann entropy for two atoms in Eq. (\[appsab\]) is given by $$\begin{aligned}
S(A,B)&=&-(|x_3|^2+|x_4|^2)\log(|x_3|^2+|x_4|^2) \\
&&-(|x_1|^2+|x_2|^2)\log(|x_1|^2+|x_2|^2), \nonumber\end{aligned}$$ and the sub-system entropies in Eq. (\[apps:2\]) are shown as $$\begin{aligned}
S(A)&=&-(1-|x_1|^2)\log(1-|x_1|^2) \\
&&-|x_1|^2\log|x_1|^2, \nonumber \\
S(B)&=&-(1-|x_2|^2)\log(1-|x_2|^2) \\
&&-|x_2|^2\log|x_2|^2. \nonumber\end{aligned}$$
From the Appendix A, one can find that the expressions of the elements in Eq. (\[apprho\]) are $$\begin{aligned}
v_+&=&|x_3|^2+|x_4|^2,~v_-=0,~w=|x_2|^2, \nonumber \\
x&=&|x_1|^2,~y=x_1x^{*}_2,~u=0.\end{aligned}$$ Moreover, $$\begin{aligned}
X_{1,+}&=&(|x_3|^2+|x_4|^2)\cos^2{\theta}+|x_2|^2\sin^2{\theta}, \nonumber
\\
X_{1,-}&=&|x_1|^2\cos^2{\theta}, \nonumber \\
|Y_1|^2&=&\frac{\sin^2{\theta}}{4}|x_1|^2|x_2|^2. \nonumber\end{aligned}$$ and $$\begin{aligned}
X_{2,+}&=&(|x_3|^2+|x_4|^2)\sin^2{\theta}+|x_2|^2\cos^2{\theta}, \nonumber
\\
X_{2,-}&=&|x_1|^2\sin^2{\theta}, \nonumber \\
|Y_2|^2&=&\frac{\sin^2{\theta}}{4}|x_1|^2|x_2|^2. \nonumber\end{aligned}$$ Therefore these parameters are independent of $\phi$. It follows that we can search the minimum of the conditional von Neumann entropy by only varying $ \theta$ in the regime $[0,\pi/2]$. Following the procedures outlined in Appendix A, we can finally derive the quantum discord. Since $\alpha$ is limited to $(0,\pi/2)$ , it can be numerically checked that $ \theta=\pi/4$ corresponds to the minimum of the conditional entropy in the following calculations. The minimum of the conditional von Neumann entropy reads $$\begin{aligned}
S(A|\Pi^{B})=-\sum_{\epsilon=\pm}{\eta}_{\epsilon}\log{\eta}_{\epsilon},\end{aligned}$$ where $$\begin{aligned}
{\eta}_{\pm}=\{1{\pm}[(1-2|x_1|^2)^2+4|x_1x_2|^2]^{1/2}\}/2.\end{aligned}$$ As a result, the quantum discord is finally given by $$\begin{aligned}
~ \mathcal{D}&=&-(1-|x_2|^2)\log(1-|x_2|^2)-|x_2|^2\log|x_2|^2 \nonumber
\label{qd:1} \\
&&+(|x_3|^2+|x_4|^2)\log(|x_3|^2+|x_4|^2) \nonumber \\
&&+(|x_1|^2+|x_2|^2)\log(|x_1|^2+|x_2|^2) \nonumber \\
&&-\sum_{\epsilon=\pm}{\eta}_{\epsilon}\log{\eta}_{\epsilon}.\end{aligned}$$ For later use, we also list the expression for concurrence derived in Ref. [@Eberly1] as $$\begin{aligned}
C_{AB}(t)=|\sin2\alpha|[1-4C^2\sin^{2}(\sqrt{\delta^2+G^2}t/2)].\end{aligned}$$
Next, we consider the Bell state with correlated spin as the initial atomic state, which is $$\begin{aligned}
|\Psi^{(2)}_{\text{Bell}}{\rangle}=\sin\alpha|{\downarrow}{\downarrow}{%
\rangle} +\cos\alpha|{\uparrow}{\uparrow}{\rangle}.\end{aligned}$$ Including the initial vacuum cavities, the wave function of the whole system can be expressed as $$\begin{aligned}
|\Psi{\rangle}(t)&=&x_1|{\uparrow}{\uparrow}00{\rangle} +x_2|{\downarrow}{%
\downarrow}11{\rangle}+x_3|{\uparrow}{\downarrow}01{\rangle} \nonumber \\
&&+x_4|{\downarrow}{\uparrow}10{\rangle}+x_5|{\downarrow}{\downarrow}00{%
\rangle},\end{aligned}$$ where the coefficients are $$\begin{aligned}
x_1&=&(Ae^{-i\lambda_+t}+Be^{-i\lambda_-t})^2\cos\alpha, \\
x_2&=&AB(e^{-i\lambda_+t}-e^{-i\lambda_-t})^2\cos\alpha, \nonumber \\
x_3&=&C(e^{-i\lambda_+t}-e^{-i\lambda_-t})(Ae^{-i\lambda_+t}+Be^{-i%
\lambda_-t})\cos\alpha, \nonumber \\
x_4&=&C(e^{-i\lambda_+t}-e^{-i\lambda_-t})(Ae^{-i\lambda_+t}+Be^{-i%
\lambda_-t})\cos\alpha, \nonumber \\
x_5&=&\sin\alpha, \nonumber\end{aligned}$$ The eigenfrequencies and the auxiliary parameters are the same as those in Eqs. (\[eigen:1\]) and (\[aux:1\]). Then under the standard basis $\{|{\downarrow}{\downarrow}{\rangle},|{\downarrow}{\uparrow}{\rangle}%
, |{\uparrow}{\downarrow}{\rangle},|{\uparrow}{\uparrow}{\rangle}\}$, the pairwise density matrix is shown as $$\begin{aligned}
~ {\rho}_{AB}(t)&=&\left(
\begin{array}{llll}
|x_2|^2+|x_5|^2 & 0 & 0 & x^{*}_1x_5 \\
0 & |x_4|^2 & 0 & 0 \\
0 & 0 & |x_3|^2 & 0 \\
x_1x^{*}_5 & 0 & 0 & |x_1|^2
\end{array}
\right). \label{rho:2}\end{aligned}$$ Hence, the joint von Neumann entropy is derive as $$\begin{aligned}
S(A,B)&=&-|x_3|^2\log|x_3|^2-|x_4|^2\log|x_4|^2 \\
&&-\sum_{\epsilon=\pm}\Omega_{\epsilon}\log\Omega_{\epsilon}, \nonumber\end{aligned}$$ where $$\begin{aligned}
\Omega_{\pm}&=&\{(|x_1|^2+|x_2|^2+|x_5|^2) \\
&&{\pm}\sqrt{(|x_2|^2+|x_5|^2-|x_1|^2)^2+4|x_1|^2|x_5|^2}\}/2. \nonumber\end{aligned}$$ And the sub-system entropy can be derive as $$\begin{aligned}
S(A)&=&-(|x_2|^2+|x_4|^2+|x_5|^2)\log(|x_2|^2+|x_4|^2+|x_5|^2) \nonumber \\
&&-(|x_1|^2+|x_3|^2)\log(|x_1|^2+|x_3|^2),~ \label{s:1} \\
S(B)&=&-(|x_2|^2+|x_3|^2+|x_5|^2)\log(|x_2|^2+|x_3|^2+|x_5|^2) \nonumber \\
&&-(|x_1|^2+|x_4|^2)\log(|x_1|^2+|x_4|^2).~ \label{s:2}\end{aligned}$$ Similar to the above Bell state with anti-correlated spins, if we focus on $ \alpha{\in}(0,\pi/2)$, $\theta=\pi/4$ also corresponds to the minimum of the conditional von Neumann entropy. Hence, the minimum of the conditional entropy is given by $$\begin{aligned}
~ S(A|\Pi^B)=-\sum_{\epsilon=\pm}\eta_{\epsilon}\log{\eta_{\epsilon}},
\label{s:3}\end{aligned}$$ where $$\begin{aligned}
\eta_{\pm}=\{1{\pm}\sqrt{(1-2|x_1|^2-2|x_3|^2)^2+4|x_1|^2|x_5|^2}\}/2.\end{aligned}$$ As a result, the quantum discord can be derived from Eqs. (\[s:1\]), (\[s:2\]), and (\[s:3\]) as $$\begin{aligned}
\mathcal{D}=S(B)-S(A,B)+S(A|\Pi^B).\end{aligned}$$
The concurrence in this case has been also derived previously [@Eberly1], and is also collected here $$C_{AB}(t)=\max\{0,f(t)\}$$ $$\begin{aligned}
f(t)&=&[1-4C^2\sin^{2}(\sqrt{\delta^2+G^2}t/2)]\\
&&[|\sin2\alpha|-8C^2\sin^{2}(\sqrt{\delta^2+G^2}t/2)\cos^{2}\alpha].\nonumber\end{aligned}$$ Specially at resonance ($\delta=0$), the entanglement sudden transition occurs only for $\alpha<\alpha_c$, where $\alpha_c=\pi/4$.
QD in two identical qubits in one common single-mode cavity
============================================================
The Hamiltonian of two identical qubits interacting with one common single-mode cavity reads $$H_{\textrm{DN2}}=\frac \Delta 2({\sigma }_z^A+{\sigma }_z^B)+{\omega
}(a^{\dag }a+\frac 12)+g\sum_{k=A,B}(a{\sigma }_k^{+}+{\sigma
}_ka^{\dag }). \nonumber$$ where $\ $ $a^{\dag }\ $ and $a\ $are the creator and annihilator of the common cavity. Actually, it is just the $N=2$ Dicke model [@dicke]. The detunning is also set as $\delta=\Delta-\omega$.
If the initial atom state is selected as the Bell state with anti-correlated spins, we can obtain the time dependent wavefunction as $$\begin{aligned}
~\label{atcof:1}
|\Psi(t){\rangle}=x_1|{\uparrow}{\downarrow}0{\rangle}+x_2|{\downarrow}{\uparrow}0{\rangle}+x_3|{\downarrow}{\downarrow}1{\rangle},\end{aligned}$$ with $$\begin{aligned}
x_1&=&{\langle}{\uparrow}{\downarrow}0|e^{-iH_{\textrm{DN2}}t}|\Psi^{(1)}_{\text{Bell}}{\rangle},\nonumber\\
x_2&=&{\langle}{\downarrow}{\uparrow}0|e^{-iH_{\textrm{DN2}}t}|\Psi^{(1)}_{\text{Bell}}{\rangle},\nonumber\\
x_3&=&{\langle}{\downarrow}{\downarrow}1|e^{-iH_{\textrm{DN2}}t}|\Psi^{(1)}_{\text{Bell}}{\rangle}.\nonumber\end{aligned}$$ Then the pairwise density matrix under standard basis $\{|{\downarrow}{\downarrow}{\rangle},|{\downarrow}{\uparrow}{\rangle},
|{\uparrow}{\downarrow}{\rangle},|{\uparrow}{\uparrow}{\rangle}\}$ is given by $$\begin{aligned}
~ {\rho}_{AB}(t)&=&\left(
\begin{array}{llll}
|x_3|^2 & 0 & 0 & 0 \\
0 & |x_2|^2 & x^{*}_1x_2 & 0 \\
0 & x_1x^{*}_2 & |x_1|^2 & 0 \\
0 & 0 & 0 & 0
\end{array}
\right). \label{rho:5}\end{aligned}$$ For the resonant case ($\delta=0$), we specify the coefficients of the wavefunction in Eq. (\[atcof:1\]) as $$\begin{aligned}
x_1&=&\frac{\cos\alpha}{2}(\cos{\sqrt{2}{\lambda}t}+1)
+\frac{\sin\alpha}{2}(\cos{\sqrt{2}{\lambda}t}-1),\nonumber\\
x_2&=&\frac{\cos\alpha}{2}(\cos{\sqrt{2}{\lambda}t}-1)
+\frac{\sin\alpha}{2}(\cos{\sqrt{2}{\lambda}t}+1),\nonumber\\
x_3&=&\frac{-i}{\sqrt{2}}(\cos\alpha+\sin\alpha)\sin{\sqrt{2}{\lambda}t}.\nonumber\end{aligned}$$ After the numerical checks, we find that $\theta=\pi/4$ corresponds to the minimum of the conditional entropy at arbitrary time. Hence the QD is described as $$\begin{aligned}
~ \mathcal{D}&=&-(1-|x_2|^2)\log(1-|x_2|^2)-|x_2|^2\log|x_2|^2 \nonumber
\label{qd:1} \\
&&+(|x_3|^2)\log(|x_3|^2)+(1-|x_3|^2)\log(1-|x_3|^2) \nonumber \\
&&-\sum_{\epsilon=\pm}{\eta}_{\epsilon}\log{\eta}_{\epsilon},\end{aligned}$$ with $$\begin{aligned}
{\eta}_{\pm}=\{1{\pm}[(1-2|x_1|^2)^2+4|x_1x_2|^2]^{1/2}\}/2.\end{aligned}$$ Besides, the concurrence of the two atoms can also be given as $$\begin{aligned}
C_{AB}(t)=2\max\{0,|x_1x_2|\}.\end{aligned}$$
![(Color online) Resonant dynamics of QD and concurrence of two identical JC atoms with the initial atomic Bell states with anti-correlated spins (a) and correlated spins (b). $\omega=1.0$. The blue solid line and the green dashed line are corresponding to QD and concurrence respectively. []{data-label="figure1"}](figure1.eps)
If starting from the initial Bell state with correlated spins, the wavefunction can be expressed as $$\begin{aligned}
|\Psi(t){\rangle}&=&x_1|{\uparrow}{\uparrow}0{\rangle}+x_2|{\downarrow}{\downarrow}2{\rangle}
+x_3|{\uparrow}{\downarrow}1{\rangle}+x_4|{\downarrow}{\uparrow}1{\rangle}\nonumber\\
&&+x_5|{\downarrow}{\downarrow}0{\rangle}.\end{aligned}$$ Hence the pairwise density matrix can be derived as $$\begin{aligned}
{\rho}_{AB}(t)&=&\left(
\begin{array}{llll}
|x_2|^2+|x_5|^2 & 0 & 0 & x_1^{*}x_5 \\
0 & |x_4|^2 & x^{*}_3x_4 & 0 \\
0 & x_3x^{*}_4 & |x_3|^2 & 0 \\
x_1x_5{*} & 0 & 0 & |x_1|^2
\end{array}
\right).~\label{rho:6}\end{aligned}$$ The coefficients at resonant condition are shown as $$\begin{aligned}
x_1&=&\frac{\cos\alpha}{3}e^{-i{\omega}t}(2+\cos{\sqrt{6}{\lambda}t}),\nonumber\\
x_2&=&\frac{\sqrt{2}\cos\alpha}{3}e^{-i{\omega}t}(\cos{\sqrt{6}{\lambda}t}-1),\nonumber\\
x_3&=&\frac{-i\cos\alpha}{\sqrt{6}}e^{-i{\omega}t}\sin{\sqrt{6}{\lambda}t},\nonumber\\
x_4&=&\frac{-i\cos\alpha}{\sqrt{6}}e^{-i{\omega}t}\sin{\sqrt{6}{\lambda}t},\nonumber\\
x_5&=&\sin\alpha.\nonumber\end{aligned}$$ Then we can obtain the QD numerically.
While for the concurrence at resonance, we know that the pairwise density matrix in Eq. (\[rho:6\]) has the form as $$\begin{aligned}
{\rho}_{AB}(t)&=&\left(
\begin{array}{llll}
v_+ & 0 & 0 & u^{*} \\
0 & y & y & 0 \\
0 & y & y & 0 \\
u & 0 & 0 & v_-
\end{array}
\right).~\label{rho:7}\end{aligned}$$ Then we can derive it as $$\begin{aligned}
C_{AB}(t)&=&2\max\{0, |x_1x_5|-|x_3|^2,\\
&&|x_3|^2-|x_1|\sqrt{|x_2|^2+|x_5|^2}\}.\nonumber\end{aligned}$$ From the definition, we find that there exists a critical bound for $\alpha$. The ESD happens only for $\alpha<\alpha_c$. The $\alpha_c$ is determined by $$\begin{aligned}
\tan^{2}\alpha_c-4\tan{\alpha_c}+1=0,\end{aligned}$$ resulting in $\alpha_c=\arctan(2-\sqrt{3})=\pi/12$.
![(Color online) Off-resonant dynamics of QD for two identical JC atoms with the initial atomic Bell states with anti-correlated spins (a) and correlated spins (b) for different detunnings $\delta=g, 2g$, and $5g$. $\omega=1.0$.[]{data-label="figure2"}](figure2.eps)
Results and discussions
=======================
First, we compare the QD with the concurrence in the two identical JC atoms with two initial atomic states, i. e. the Bell states with anti-correlated spins and correlated spins, for zero detunnings. The results are collected in Fig. 1. The evolution of both QD and concurrence for the initial Bell state with anti-correlated spins display similar behavior, as seen in Fig. 1(a). Yonac *et al.* [@Eberly1] has shown that the ESD only occurs in the initial atomic Bell states with correlated spin, where the entanglement can fall abruptly to zero and vanish for a period of time before revival. It is interesting to note from Fig. 1(b) that during the period of ESD, QD becomes small but is always finite, except vanish at discrete moments $t=(2k+1)\pi/G, (k=0,1,2,...$).
Then, we show the effects of the detunnings on the QD in independent cavities in Fig. 2, starting from these two Bell states. Interestingly, the amplitude of oscillation of the QD as a function of time is suppressed monotonically by the detunnings for both initial Bell stats. More importantly, the zeros of the QD at discrete instants shown in Fig. 1(b) disappear with the finite detunnings.
![(Color online) Resonant dynamics of QD and concurrence of two qubits coupled to a common cavity with the initial atomic Bell states with anti-correlated spins (a) and correlated spins (b). $\omega=1.0$. The blue solid line and the green dashed line are corresponding to QD and concurrence respectively. []{data-label="figure3"}](figure3.eps)
Next, we compare the QD with the concurrence in the two identical qubits coupled to the common cavity with two initial atomic Bell states. The results for zero detunning are presented in Fig. 3. For the initial atomic Bell state with anti-correlated spins, similar behaviors for both QD and concurrence are observed. For the initial atomic Bell state with correlated spins, one can find from Fig. 3(b) that the ESD can occur, but the QD never vanishes. Interestingly, QD and entanglement show opposite behaviors. Especially, during the period of ESD, the QD always becomes larger, in sharp contrast with that observed in the independent cavities (c.f. Fig. 1(b)).
![(Color online) Off-resonant dynamics of QD for two qubits coupled to a common cavity with the initial atomic Bell states with anti-correlated spins (a) and correlated spins (b) for different detunnings $\delta=g, 2g$, and $5g$. $\omega=1.0$.[]{data-label="figure4"}](figure4.eps)
Note that the critical parameter $\alpha_c=\pi/12$ below which the ESD can occurs in the common cavity is smaller than $\alpha_c=\pi/4$ in independent cavities. The instant vanish of QD is absent in the common cavity, implying that the common cavity enhance the QD. So, it is suggested that the quantum correlation in the common cavity is stronger than that in independent cavities in some sense.
The effect of the detunnings on the QD of two qubits coupled to a common cavity is also studied. As shown in Fig. 4(a), for the initial Bell state with anti-correlated spins, the amplitude of oscillation of the QD as a function of time is also suppressed monotonically by the detunnings and little bit larger than that in two cavities (c.f. Fig. 2(a)). While for the initial atomic Bell state with correlated spins with $\alpha=\pi/24$ where ESD can occur, the oscillation of the QD for two qubits in the common cavity is suppressed considerably with detunnings, as shown in Fig. 4(b). Especially, for large detunnings $\delta=5g$, the QD remains almost unchanged. In this case, we find that the components of the Bell stats show slightly variation with time for large detunning, due to the fact that large detunnings prevent the hopping for photons between different atomic levels to certain degree.
Conclusions
===========
In this paper, the QD dynamics of two qubits in both independent and common cavities are investigated. The comparisons with the entanglement evolution are also performed. For the initial atomic Bell state with anti-correlated spins, the QD and entanglement show the similar behaviors for both cavities. But for the initial atomic Bell state with correlated spins, the QD and entanglement behave in a remarkably different way. The ESD may occur for both cavities, but the QD never vanishes suddenly. For the independent cavities, the QD vanishes only at discrete instants and can be lifted with finite detunnings. In the common cavity the QD is always finite. Especially, the QD and entanglement display an opposite behavior in the common cavity, different from those in independent cavities. The detunnings play important role on the QD dynamics. It always stabilizes the QD, which could be helpful in the real applications of the QD as the better resource in quantum information science and quantum computing.
Acknowledgement
===============
This work was supported by National Natural Science Foundation of China under Grant Nos. 11174254 and 11104363, National Basic Research Program of China (Grants No. 2011CBA00103 and No. 2009CB929104).
Derivation of the quantum correlation
=====================================
In the present paper, the general pairwise density matrix under the standard basis $\{|{\downarrow }{\uparrow }{\rangle },|{\downarrow }{\uparrow }{%
\rangle },|{\uparrow }{\downarrow }{\rangle },|{\uparrow }{\uparrow }\}$, is shown as $$\begin{aligned}
~\label{apprho}
~{\rho }_{AB}(t)=\left(
\begin{array}{llll}
v_{+} & 0 & 0 & u^{*} \\
0 & w & y^{*} & 0 \\
0 & y & x & 0 \\
u & 0 & 0 & v_{-}
\end{array}
\right) .\end{aligned}$$ The von Neumann entropy of the two atoms is $$\begin{aligned}
~\label{appsab}
~S(A,B)=\sum_{\epsilon ={\pm },k=1,2}\Omega _{k,\epsilon }\log \Omega
_{k,\epsilon },\end{aligned}$$ with $$\begin{aligned}
\Omega _{1,\pm } &=&\frac{(v_{+}+v_{-}){\pm }\sqrt{(v_{+}-v_{-})^2+4|u|^2}}2,
\nonumber \\
\Omega _{2,\pm } &=&\frac{(w+x){\pm }\sqrt{(w-x)^2+4|y|^2}}2. \nonumber\end{aligned}$$ The reduced sub-system density matrices for A and B are obtained as $$\begin{aligned}
~\rho _A(t) &=&(v_{+}+w)|{\downarrow }{\rangle }_A{\langle }{\downarrow }%
|+(x+v_{-})|{\uparrow }{\rangle }_A{\langle }{\uparrow }|, \nonumber
\\
\rho _B(t) &=&(v_{+}+x)|{\downarrow }{\rangle }_B{\langle }{\downarrow }%
|+(w+v_{-})|{\uparrow }{\rangle }_B{\langle }{\uparrow }|. \nonumber\end{aligned}$$ Hence, we derive the corresponding von Neumann entropies as $$\begin{aligned}
~\label{apps:2}
S(A) &=&-(v_{+}+w)\log (v_{+}+w) \nonumber \\
&&-(x+v_{-})\log (x+v_{-}),\\
S(B) &=&-(v_{+}+x)\log (v_{+}+x) \nonumber \\
&&-(w+v_{-})\log (w+v_{-}),\nonumber\end{aligned}$$ While for the conditional density matrix $\rho _{A|\Pi ^B}$, projection basis are considered as $$\begin{aligned}
~|\Phi _1{\rangle }_B &=&\cos {\theta }|\downarrow {\rangle }_B+e^{i\phi
}\sin {\theta }|\uparrow {\rangle }_B, \nonumber \label{appf:1} \\
|\Phi _2{\rangle }_B &=&e^{-i\phi }\sin {\theta }|\downarrow {\rangle }%
_B-\cos {\theta }|\uparrow {\rangle }_B. \nonumber\end{aligned}$$ The conditional density operator is expressed as $$~\rho _{A|\Pi _k^B}=\Pi _k^B\rho _{AB}\Pi _k^B/p_k,$$ where $\Pi _k^B=\text{I}_A{\otimes }|\Phi _k{\rangle }_B{\langle }\Phi _k|$ and $p_k=\text{Tr}_{AB}\{\rho _{A|\Pi _k^B}\}$. Specifically under the projections in Eq. (\[appf:1\]), $$\begin{aligned}
\rho _{A|\Pi _k^B} &=&|\Phi _k{\rangle }_B{\langle }\Phi _k|{\otimes }\{|{%
\downarrow }{\rangle }_A{\langle }\downarrow |X_{k,+}+|{\uparrow }{\rangle }%
_A{\langle }\uparrow |X_{k,-} \nonumber \\
&&+|{\downarrow }{\rangle }_A{\langle }\uparrow |Y_k+|{\uparrow }{\rangle }_A%
{\langle }\uparrow |Y_k^{*}\}/p_k. \nonumber\end{aligned}$$ For $k=1$, we show $$\begin{aligned}
X_{1,+} &=&v_{+}\cos ^2{\theta }+w\sin ^2{\theta }, \nonumber \\
X_{1,-} &=&x\cos ^\theta +v_{-}\sin ^2{\theta }, \nonumber \\
Y_1 &=&(y^{*}e^{-i\phi }+u^{*}e^{i\phi })\sin {\theta }\cos {\theta }.
\nonumber\end{aligned}$$ For $k=2$, $$\begin{aligned}
X_{2,+} &=&v_{+}\sin ^2{\theta }+w\cos ^2{\theta }, \nonumber \\
X_{2,-} &=&x\sin ^2{\theta }+v_{-}\cos ^2{\theta }, \nonumber \\
Y_2 &=&-(y^{*}e^{-i\phi }+u^{*}e^{i\phi })\sin {\theta }\cos {\theta }.
\nonumber\end{aligned}$$ Then the eigenvalues of the conditional density matrix reads $$\begin{aligned}
\eta _{k,\pm } &=&\frac 1{2p_k}\{(X_{k,+}+X_{k,-}){\pm }[(X_{k,+}-X_{k,-})^2
\nonumber \\
&&+4|Y_k|^2]^{1/2}\}.\end{aligned}$$ The conditional von Neumann entropy is described as $$\begin{aligned}
S(A|\Pi ^B) &=&\sum_{k=1,2}-p_k\text{Tr}_A\{{\rho _{A|\Pi _k^B}}\log {\rho
_{A|\Pi _k^B}}\} \\
&=&-\sum_{\epsilon =\pm }\sum_{k=1,2}p_k{\eta }_{k,\epsilon }(\theta ,\phi
)\log {{\eta }_{k,\epsilon }(\theta ,\phi )}. \nonumber\end{aligned}$$ As a result, the quantum discord can be obtained by $$\mathcal{D}=\min \{S(B)-S(A,B)+S(A|\Pi ^B)\},$$ and the classical correlation is given as $$\mathcal{C}=\max \{S(A)-S(A|\Pi ^B)\}.$$
A. Einstein, B. Podolsky, and N. Rosen, Phys. Rev. **47**, 777 (1935); J. S. Bell, Physics **1**, 195 (1964).
M. A. Nielsen and I. L. Chuan, *Quantum Computational and Quantum information* (Cambridge University Press, Cambridge, UK 2000).
T. Yu and J. H. Eberly, Phys. Rev. Lett. **93**, 140404 (2004); T. Yu and J. H. Eberly, Phys. Rev. Lett. **97**, 140403 (2006).
M. P. Almeida, F. de Melo, M. Hor-Meyll, A. Salles, S. P. Walborn, P. H. Souto Ribeiro, and L. Davidovich, Science **316**, 579 (2007).
H. Ollivier and W. H. Zurek, Phys. Rev. Lett. **88**, 017901 (2001).
J. Oppenheim, M. Horodecki, P. Horodecki, and R. Horodecki, Phys. Rev. Lett. **89**, 180402 (2002).
K. Modi, A. Brodutch, H. Cable, T. Paterek, and V. Vedral, arXiv:1112.6238 (2011).
B. P. Lanyon, M. Barbieri, M. P. Almeida, and A. G. White, Phys. Rev. Lett. **101**, 200501 (2008).
A. Datta, A. Shaji, and C. M. Caves, Phys. Rev. Lett. **100**, 050502 (2008).
T. Werlang, S. Souza, F. F. Fanchini, and C. J. Villas Boas, Phys. Rev. A **80**, 024103 (2009).
F. Altintas and R. Eryigit, J. Phys. B: At. Mol. Opt. Phys. **44**, 125501(2011).
F. F. Fanchini, T. Werlang, C. A. Brasil, L. G. E. Arruda, and A. O. Caldeira, Phys. Rev. A **81**, 052107 (2010).
M. Yönac, T. Yu and J. H. Eberly, J. Phys. B: At. Mol. Opt. Phys. **39**, S621 (2006). M. Yönac, T. Yu and J. H. Eberly, J. Phys. B: At. Mol. Opt. Phys. **40**, S45 (2007).
S. Chan, M. D. Reid, and Z. Ficek, J. Phys. B: At. Mol. Opt. Phys. **42**, 065507 (2009).
I. Sainz and G. Bjork, Phys. Rev. A **76**, 042313 (2007).
Q. H. Chen, Y. Yang, T. Liu, and K. L. Wang, Phys. Rev. A **82**, 052306 (2010). S. Agarwal, S.M. Hashemi Rafsanjani and J.H. Eberly, arXiv:1201.2928 (2012).
R. H. Dicke, Phys. Rev. **93**, 99 (1954).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Estimation of the coin parameter(s) is an important part of the problem of implementing more robust schemes for quantum simulation using quantum walks. We present the estimation of the quantum coin parameter used for one-dimensional discrete-time quantum walk evolution using machine learning algorithms on their probability distributions. We show that the models we have implemented are able to estimate these evolution parameters to a good accuracy level. We also implement a deep learning model that is able to predict multiple parameters simultaneously. Since discrete-time quantum walks can be used as quantum simulators, these models become important when extrapolating the quantum walk parameters from the probability distributions of the quantum system that is being simulated.'
author:
- Parth Rajauria
- Prateek Chawla
- 'C. M. Chandrashekar'
title: 'Estimation of one-dimensional discrete-time quantum walk parameters by using machine learning algorithms'
---
Introduction {#sec:intro}
============
Quantum walks [@R58; @F86] are the quantum counterparts of classical random walks, and are therefore used as the basis to generate relevant models for controlled quantum dynamics of a particle. Much like a classical random walk, the formalism for quantum walks has also developed in two forms - the discrete-time quantum walk (DTQW) and the continuous-time quantum walk (CTQW). Both the formalisms exhibit features that enable them to effectively realise quantum computational tasks [@DW08; @C09; @A07; @MSS07; @BS06; @FGG07; @K08]. Quantum superposition and interference have the effect of allowing the quantum walker to have a quadratically faster spread in position space in comparison to a classical walker [@P88; @ADZ93; @M96; @K03; @CCD03; @CG04; @VA12]. This has applications in modelling dynamics of several quantum systems, like quantum percolation [@CB14; @KKNJ12; @CAC19], energy transport in photosynthesis [@ECR07; @MRLA08], graph isomorphism [@DW08], quantum algorithms [@CMC20; @PMCM13], and even in generating a scheme for universal quantum computation [@SCSC19; @SCASC20].
Much like a classical walk, the dynamics of a walker undergoing CTQW can be described only by a position Hilbert space, whereas a walker performing DTQW requires an additional degree of freedom to express its controlled dynamics. This is realised by the coin Hilbert space, which is an internal state of the walker, and provides the relevant additional degree of freedom for the walker. Tuning the parameters and evolution operators of a DTQW enables the walker to simulate several quantum mechanical phenomena, such as topological phases [@COB15; @SRFL08; @KRBD10; @AO13], relativistic quantum dynamics [@S06; @CBS10; @MC16; @C13; @MBD13; @MBD14; @AFF16; @P16; @RLBL05], localization [@J12; @C12; @CB15], and neutrino oscillations [@MMC17; @MP16]. Quantum walks have been experimentally implemented in various physical systems, such as photonic systems [@SCP10; @BFL10; @P10], nuclear magnetic resonance systems [@RLBL05], trapped ion systems [@SRS09; @ZKG10], and cold atoms [@KFCSWMW09].
It is known that the coin parameters of a quantum walk play a significant role in determining the overall dynamics of the system. For instance, the coin parameters of a split-step quantum walk determine topological phases [@COB15; @SRFL08; @KRBD10], neutrino oscillation [@MMC17; @MP16], and the mass of a Dirac particle [@MC16]. Thus the coin parameter is an important piece of the puzzle while using a quantum walk as a quantum simulation scheme. It thus becomes a crucial problem to be able to estimate the parameters of a quantum walk in order to facilitate better quantum simulations and also for further research into modelling realistic quantum dynamics.
In this regard, the problem partially becomes one of finding a pattern hidden in this complex data, and an effective approach is to use an algorithm that automates its learning process. In this context, the term ’learning’ aptly described in [@Mitchell97] : A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E. The task T in this particular case is defined as, to output a function $f(\mathbb{R}^N \rightarrow \mathbb{R})$, such that the input vector corresponds to the known probability distribution after $N$ steps and the output corresponds to the parameter $\theta$. It may, however, be noted that due to the No Free Lunch theorem [@W96], the machine learning strategies for different types of quantum walks are not likely to be the same as for this particular case.
It is known that the Quantum Fisher Information (QFI) [@P11; @TP13; @S13; @TA14] $H_w(\theta)$ of a quantum walker’s position quantifies a bound to the amount of information that can be extracted from the probability distribution [@P12; @SCP19]. It has also been shown that it is indeed possible to estimate the coin parameter of a DTQW with a single-parameter coin. However, the approaches that rely on QFI to predict the parameters in a DTQW share a similar constraint, viz, even though $H_w(\theta)$ is a continuous, single-valued function over $\theta$, the variations in the plot are small, and thus contribute to the error in determining $\theta$ [@SCP19]. In this work, we train various machine learning models and demonstrate that such models can indeed estimate the quantum walk parameters to a good accuracy. We also estimate the coin parameter and the number of steps for a DTQW simultaneously with a multilayer perceptron model, and demonstrate that it performs much better than a baseline model that does not learn with experience.
This paper is structured as follows. In section \[sec:dtqw\] we introduce a standard DTQW, an SSQW, and describe the evolution operators that determine their dynamics. We give a brief overview of the machine learning models used and their parameters in Section \[sec:ml\]. Section \[sec:res\] details our results of training the machine learning algorithms and their performance. We wrap up the paper in section \[sec:conc\] with a small discussion on our results and conclusions drawn.
Discrete-time quantum walk {#sec:dtqw}
==========================
The evolution of a walker executing DTQW is defined in a Hilbert space $\mathcal{H} = \mathcal{H}_C \otimes \mathcal{H}_P$, where $\mathcal{H}_C$ and $\mathcal{H}_P$ are the walker’s coin and position Hilbert spaces, respectively. The coin Hilbert space has the basis states $\left\{ \ket{\uparrow}, \ket{\downarrow} \right\}$, and the position Hilbert space is defined by the basis states $\ket{x}$, where $x \in \mathbb{Z}$. The evolution is described by two unitary operators, known as the coin and shift operations. The coin operator affects the coin Hilbert space, and the shift operator evolves the walker in a superposition of position states, the amplitudes of which are determined by the coin operation. In case of the one-dimensional DTQW, the most general unitary coin operator is an $SU(2)$ matrix, which has three independent parameters. That said, even one- and two-parameter coins are very useful while simulating various systems. As an example, the single-parameter split step DTQW is very effective to simulate neutrino oscillations [@MMC17], topological phases [@COB15], and the Dirac equation [@MC16]. In a one-dimensional DTQW, the coin operation is an $SU(2)$ matrix, defined as, $$\label{eq:eq2.1}
\hat{C}(\theta, \xi, \zeta)
= \begin{bmatrix}
e^{i\xi}\cos(\theta) & e^{i\zeta}\sin(\theta) \\
-e^{-i\zeta}sin(\theta) & e^{-i\xi}cos(\theta) \end{bmatrix} \otimes \sum_{x\in \mathbb{Z}} \ket{x} \bra{x}.$$
The shift operation is defined as, $$\label{eq:eq2.2}
\hat{S} = \sum_{x} \bigg[\ket{\uparrow}\bra{\uparrow} \otimes \ket{x-1}\bra{x} + \ket{\downarrow}\bra{\downarrow} \otimes \ket{x+1}\bra{x}\bigg].$$ The initial internal state of the walker is defined as, $$\label{eq:eq2.3}
\ket{\psi_0} = \left(\alpha \ket{\uparrow} + \beta \ket{\downarrow}\right) \otimes \ket{x=0}$$ and the complete evolution equation will be in the form, $$\label{eq:eq2.4}
\ket{\psi_N} = \big[ \hat{S}\hat{C} \big]^N \ket{\psi_0} = \hat{W}^N \ket{\psi_0},$$ where $N$ is the number of steps taken by the walker.
To obtain the the one-parameter coin form from Eq. (\[eq:eq2.1\]), we fix the values $\xi=0$ and $\zeta = -\frac{\pi}{2}$. This is the convention adopted in the remainder of this text. The one-parameter coin operator is thus given as, $$\label{eq:eq2.5}
\hat{C}(\theta) = \begin{bmatrix}
\cos(\theta) & -i\sin(\theta) \\
-i\sin(\theta) & \cos(\theta)
\end{bmatrix} \otimes \sum_{x\in \mathbb{Z}}\ket{x}\bra{x}.$$
Fig. \[fig:fig2.1\] shows the probability distributions in position space of a walker performing a DTQW with different values of $\theta$ as in Eq. (\[eq:eq2.5\]).
--------------------------------------- ---------------------------------------
{width="45.00000%"} {width="45.00000%"}
(a) (b)
{width="45.00000%"} {width="45.00000%"}
(c) (d)
--------------------------------------- ---------------------------------------
A special case of a DTQW is the split-step quantum walk (SSQW), in which case the evolution operator $\hat{W}$ is given by a composition of two half-steps, $$\label{eq:eq2.6}
\hat{W} = \hat{S}_+\hat{C}_{\theta_2}\hat{S}_-\hat{C}_{\theta_1},$$ Where $\hat{C}_{\theta_j}$ represent the two coin operations, which are defined in the same form as in Eq. (\[eq:eq2.5\]), and $\hat{S}_\pm$ are directed shift operations, defined as, $$\label{eq:eq2.7}
\begin{split}
\hat{S}_+ &= \sum_{x} \bigg[\ket{\uparrow}\bra{\uparrow} \otimes \ket{x}\bra{x} + \ket{\downarrow}\bra{\downarrow} \otimes \ket{x+1}\bra{x}\bigg] \\
\hat{S}_- &= \sum_{x} \bigg[\ket{\uparrow}\bra{\uparrow} \otimes \ket{x-1}\bra{x} + \ket{\downarrow}\bra{\downarrow} \otimes \ket{x}\bra{x}\bigg]
\end{split}.$$ In a one-dimensional visualisation, an SSQW implements the case when the components that are in the direction of $\ket{\uparrow}$ experience a different coin ($\hat{C}_{\theta_1}$) than the components in the $\ket{\downarrow}$ direction, which experience the coin $\hat{C}_{\theta_2}$. The evolution of the walk is still in the form as shown in Eq. (\[eq:eq2.4\]), with the appropriate walk operator $\hat{W}$ substituted from Eq. (\[eq:eq2.6\]).
Fig. \[fig:fig2.2\] shows the probability distributions obtained by a walker executing a SSQW with multiple combinations of $\theta_1$ and $\theta_2$. Unlike the two peak probability distribution from the single coin DTQW, SSQW can results in four peak probability distribution.
----------------------------------------- -----------------------------------------
{width="45.00000%"} {width="45.00000%"}
(a) (b)
{width="45.00000%"} {width="45.00000%"}
(c) (d)
----------------------------------------- -----------------------------------------
A brief note on machine learning {#sec:ml}
================================
Machine learning was originally created to try and create solutions to problems which were hard to define formally, such as recognizing faces in images, or cognition of spoken words. This necessitated the creation of algorithms that could learn from experience of examples supplied to them, so the programmer was absolved of the requirement to formally specify all knowledge needed to solve the problem beforehand.
While the hardcoded knowledge methods worked very well for small, relatively sterile environments such as the world of chess [@CHH02], such programs face difficulties in subjective and intuitive tasks such as understanding speech or object recognition. Inference-based methods have been suggested and implemented as a possible solution to this type of problem, but have not had much success [@LG89; @D15]. Modern approaches circumvent this difficulty by implementing algorithms that have the capability of recognising patterns in raw data on their own [@GBC16] - a trait now known as machine learning.
### Linear regression {#linear-regression .unnumbered}
Linear regression takes an input vector of features $\mathbf{x}\in\mathbb{R}^N$ and outputs a vector $\hat{\mathbf{y}} \in \mathbb{R}^k$, which is to be interpreted as a prediction for the actual vector $\mathbf{y} \in \mathbb{R}^k$. The model is named linear regression as it attempts to find a linear function from the input vector to the output. More mathematically, linear regression models calculate the prediction $\hat{\mathbf{y}}$ as $$\label{eq:eq2.8}
\hat{\mathbf{y}} = \mathbf{w}^T \mathbf{x} + \mathbf{b},$$ where $\mathbf{w}$ is a matrix of weights and $b$ is known as the bias vector. The weight $w_{ij}$ essentially determines how the feature $x_{i}$ affects the prediction $\hat{y}_{j}$. The bias vector is the value that this function tends to take in the absence of an input, and thus ensures that the mapping of features to predictions is an affine function.
### Ridge regression {#ridge-regression .unnumbered}
Due to the large sizes of the input vectors, a case may arise where the input variables may have near-linear relationships, a phenomenon known as multicollinearity [@B07]. Multicollinearity leads to unbiased linear regression estimates, but with high variances. Ridge regression is a method to improve estimation by adding a small amount of bias [@M98].
Assume that one requires the estimates of a vector $\mathbf{y}$, as in Eq. (\[eq:eq2.8\]). A traditional linear regression-based method would seek to minimize the sum of squared residuals $|| \mathbf{w}^T \mathbf{x} + \mathbf{b} - \mathbf{y} ||_2^2$, where $|| \cdot ||_2$ represents the Euclidean norm. Ridge regression introduces a regularization term in this in order to guide the solution towards preferred properties. Thus, the final minimization looks like,
$$\label{eq:eq2.9}
|| \mathbf{w}^T \mathbf{x} + \mathbf{b} - \mathbf{y} ||_2^2 + || \mathbf{\Gamma_x} ||_2^2$$
where $\mathbf{\Gamma_x}$ is known as the Tikhonov Matrix [@T63; @TGSY95]. In the case when this matrix is a multiple of the identity matrix (i.e. $\mathbf{\Gamma_x} = \alpha \mathbb{I}$), this process is the same as $L_2$ regularization. In this paper, we have chosen $\alpha=0.01$. In the method of lasso regression [@R96], the $L_1$ regularization is used instead.
### Nearest-neighbour regression {#nearest-neighbour-regression .unnumbered}
The $k$-Nearest Neighbour algorithm was proposed as a nonparametric method for pattern recognition, and is used for both classification and regression analyses [@CH67; @A92]. In our case, the task is that of regression, and therefore the output is the average of k closest examples in the feature space of the training set. The algorithm proceeds as follows:
1. Load the training data and initialize $k$ to the chosen number of neighbours.
2. For each example in the training data,
1. Calculate the distance between the query and the current example from the data.
2. Add the distance and index of this example to an ordered collection
3. Sort the ordered collection in ascending order of distances.
4. Pick the labels of first $k$ entries in this collection, and return the mean of the these labels.
This algorithm defers all computation until function evaluation, and only locally approximates the said function. Thus it is also a valid example of instance-based learning [@PE20; @H01]. As it uses the training examples in the local neighbourhood of the query to generate a prediction, it is extremely sensitive to the local structure of the data [@BGRS99]. Since only the $k$ closest training examples nearest to a query are considered to generate the prediction for it, this algorithm does not require an explicit training step.
### Multilayer Perceptron Models {#multilayer-perceptron-models .unnumbered}
Multilayer perceptron (MLP) models are a class of artificial neural networks, which are algorithms modelled after biological neural networks in animal brains [@MP43; @K56]. These algorithms learn by example, and consist of units known as neurons, which are loosely modelled after their biological counterparts. Fig. \[fig:fig2.3\] shows a basic MLP model with 3 neurons in the input layer, 2 hidden layers of four neurons each, and an output layer of 3 neurons. In this paper, the MLP model used has 200 neurons in the input layer, 200 in the hidden layer, and 2 in the output layer. The input and hidden layer neurons use a rectified linear unit (ReLU) as their activation function, and the output layer uses an exponential activation function.
--------------------------------------------- -------------------------------------------
{width="45.00000%"} {width="45.00000%"}
(a) (b)
--------------------------------------------- -------------------------------------------
The goal of this model is to approximate a function $\mathbf{y} = f (\mathbf{x})$ by creating a mapping $\mathbf{y} = f^* (\mathbf{x} ; \mathbf{\theta}) $ and then learning the parameters $\mathbf{\theta}$ so that $f^*$ approximates $f$ as closely as possible. This is a type of feedforward model [@GBC16], as there are no inputs that are fed back to the input from the output of the model. The model is provided a certain amount of labelled examples, which consist of both the input $\mathbf{x}$ and the output $\mathbf{y}$ (also known as the [*training set*]{}). The model uses these to tweak the parameters $\mathbf{\theta}$ the best it can, and then predicts $\hat{\mathbf{y}}$ corresponding to unlabelled values of $\mathbf{x}$ (also known as the [*testing set*]{}). This technique is known as supervised learning.
Our MLP model uses the cross-entropy between training data and the predictions made by the model as the cost function, and attempts to minimize it via an optimized gradient descent method. In our model, the optimizer used is Nadam. The Nadam (Nesterov-accelerated Adaptive Moment estimation) computes adaptive learning rates for each parameter by combining the approaches of Nesterov accelerated gradients and adaptive moment estimation algorithms [@D16; @N83; @KB15; @R17].
### Performance metric {#performance-metric .unnumbered}
We measure the performance of our model by computing the mean square error of the model on the testing set, defined as, $$\label{eq:eq2.10}
\text{MSE} = \frac{1}{n} \sum_{j = 1}^{k} \left[ \hat{y}_j - y_j \right]^2 ,$$
where $n$ is the number of elements in the testing set, and $k$ is the length of the output vector. Eq. (\[eq:eq2.10\]) may also be written in terms of Euclidean distance, $$\label{eq:eq2.11}
\text{MSE} = \frac{1}{n} || \hat{\mathbf{y}}^{(test)} - \mathbf{y}^{(test)} ||_2^2$$ of the predicted values and the actual values of $\mathbf{y}$. It is important to note here that the error is considered over the testing set and not the training set.
Results {#sec:res}
=======
In this work, we have used machine learning algorithms and trained an MLP model in an attempt to estimate the parameters of a one-dimensional DTQW. The models were trained on DTQW with $N = 500$, and $\theta$ varying from $\frac{\pi}{180}$ to $\frac{89\pi}{180}$, at an interval of $\frac{\pi}{1800}$. The dataset consists of DTQW distributions that resulted after using these parameters in conjunction with a symmetric initial state as, $$\label{eq:eq3.1}
\ket{\psi_0} = \frac{1}{\sqrt{2}}\left(\ket{\uparrow} + \ket{\downarrow}\right) \otimes \ket{x=0}.$$
We have chosen the ratio of training to testing data as $3:1$ ($75\% : 25\%$) for this dataset. As a result, $668$ randomly selected probability distributions are used to train our algorithm, and the rest are used to test its performance. Three regression models, namely, K-nearest neighbours, linear regression and ridge regression, were trained on this data, and their performance was evaluated by the mean square error (MSE) of their predictions. The results are shown as in Table \[tab:tab3.1\]. It is observed that while all the models return a very low MSE, but the performance of the linear regression model is a few orders of magnitude better than the next best performing model (k-nearest neighbours).
**Mean-squared error**
------------------------------------ ------------------------
Linear Regression 2.501 x $10^{-09}$
K-Nearest Neighbours (k = 5) 2.058 x $10^{-06}$
Ridge Regression ($\alpha$ = 0.01) 2.783 x $10^{-05}$
: Table showing MSE on test data for various machine learning models tasked with learning the coin parameter of a one-dimensional quantum walk. All the models shown here display a very low MSE, which is indicative of the fact that each of the models has successfully recognized, and is able to reproduce the patterns in the data.[]{data-label="tab:tab3.1"}
Using the same training data as specified above, the model was tasked to predict the value of $\theta$, and given the testing data corresponding to $\theta=\frac{\pi}{2}$, which is outside the range used to generate the training data. Table \[tab:tab3.2\] details the predictions by each of these models.\
**Expected $\theta$** **Predicted $\theta$** **Error (%)**
------------------------------------ ----------------------- ------------------------ ---------------
Linear Regression $90^{\circ}$ $89.956^{\circ}$ 0.048%
K-Nearest Neighbours (k = 5) $90^{\circ}$ $89.639^{\circ}$ 0.046%
Ridge Regression ($\alpha$ = 0.01) $90^{\circ}$ $89.958^{\circ}$ 0.401%
We also implemented an MLP model (detailed in section \[sec:ml\]) to try and predict two parameters ($\theta$ and $N$) of a 1D-DTQW simultaneously via deep learning. The model contains $3$ layers, out of which there is $1$ hidden layer. The model uses rectified linear units as the activation function for the first two layers (i.e. the input and hidden layers, respectively), and an exponential activation for the output layer. The neural net uses the [*Nadam*]{} optimizer. Other possible optimizers that can be used here are AdaGrad, AdaDelta and Adam. AdaGrad offers the benefit of eliminating the need of manually tuning the learning rate, but has the weakness that it accumulates a sum of squared gradients in its denominator over time, causing the learning rate to shrink and eventually become infinitesimally small, and the algorithm stop learning at this point. AdaDelta aims to reduce the aggressive decay of the learning rate in AdaGrad by fixing a window size of the accumulated past gradients. Adam tries to improve on both AdaGrad and AdaDelta by considering an exponentially decaying average of past gradients to adjust its learning rates. Nadam, however, improves on the performance of Adam by using Nesterov-accelerated gradients, and was chosen as the best fit for this case.
To train this network, we generated a new dataset, which contained all possible combinations of the parameters $\theta$ and $N$, varying between $[\frac{\pi}{180}, \frac{\pi}{2}]$ and $[1,499]$ respectively, with a step size of $1$ for $N$, and $\frac{\pi}{180}$ for $\theta$. The dataset thus comprised of $44,910$ different probability distributions. The training-testing split was chosen to be the same as the earlier case ($75\% : 25\%$), and the training set thus consisted of a randomly selected $33,682$ probability distributions of the total.
In order to judge the effectiveness of this model, we also designed a baseline model, which would always predict the mean of all supplied training values of $\theta$ and $N$ ($\frac{89\pi}{360}$ and $249$ respectively, for this case), for any input distribution. The MSE from this baseline model was found to be ${26077.4287}$.
Our neural network gave an output MSE of $916.0458$, which is a reduction of $96.488\%$ on the baseline error. This proves that the neural network is able to effectively learn and reproduce the patterns found in the DTQW probability distributions in order to reasonably estimate both the parameters of a 1D-DTQW.
We also tried to estimate selected parameters of an SSQW. A one-dimensional SSQW with two single-parameter coins as defined in Eq. (\[eq:eq2.5\]) which typically has 3 parameters, $\theta_1$, $\theta_2$, and $N$. For the purposes of this task, we varied $\theta_1$ in the range $[\frac{\pi}{1800}, \frac{\pi}{2}]$ at an interval of $\frac{0.04\pi}{180}$, while keeping $\theta_2 = \frac{\pi}{4}$, and $N=100$ fixed. We trained a linear regression model on this data, as well as a K-nearest neighbour model. The parameter was estimated with a good accuracy, as is seen from the contents of Table \[tab:tab3.3\].
**Mean-squared error**
----------------------------- -------------------------
K-Nearest Neighbour (k = 5) $3.112 \times 10^{-07}$
Linear Regression $1.265 \times 10^{-08}$
: Mean squared error on test data for ML models in predicting a single parameter in a split-step quantum walk. Both the models are able to learn the parameters well, but it can be seen that the linear regression model still yields an MSE that is an order of magnitude lower than the K-nearest neighbours model.[]{data-label="tab:tab3.3"}
Conclusions {#sec:conc}
===========
In this work, we have demonstrated the effectiveness of machine learning models while estimating the coin parameter in a DTQW and an SSQW. We have applied these models on the one-dimensional DTQW and SSQW, and attempted to estimate the parameters from these walks.
In the case of DTQW, we applied three different models in order to estimate the coin parameter $\theta$, and conclude that a machine learning-based approach is indeed able to estimate the parameter very well. We also attempt to use a neural network in order to try and predict the two parameters at once, and show that its prediction error reduces by a significant amount from the baseline error, implying that the network is able to distinguish and replicate patterns found in this data. We also use two different models in order to estimate one of the coin parameters of a SSQW.
It is important to keep in mind that the accuracy of the predictions will vary with the amount of training data it is given. Typically, larger datasets improve the accuracy. The accuracy is also very dependent on the model itself under consideration. There is no single known machine learning algorithm that may outperform all others, and it is thus important to choose the algorithm for implementation with care.
0.2in [**Acknowledgment:**]{} CMC would like to thank Department of Science and Technology, Government of India for the Ramanujan Fellowship grant No.:SB/S2/RJN-192/2014. We also acknowledge the support from Interdisciplinary Cyber Physical Systems (ICPS) programme of the Department of Science and Technology, India, Grant No.:DST/ICPS/QuST/Theme-1/2019/1 and US Army ITC-PAC contract no. FA520919PA139.
[99]{} G. V. Riazanov, “The Feynman path integral for the Dirac equation", [Zh. Eksp. Teor. Fiz. [**33**]{}, 1437 (1958), \[Sov. Phys. JETP **6**, 1107-1113 (1958)\].](http://www.jetp.ac.ru/cgi-bin/r/index/r/33/6/p1437?a=list)
R. P. Feynman,“Quantum mechanical computers", [Found. Phys. **16**, 507-531 (1986)](https://link.springer.com/article/10.1007/BF01886518).
B. L. Douglas and J. B. Wang, “A classical approach to the graph isomorphism problem using quantum walks," [J. Phys. A: Math. Theor. **41**, 7, 075303 (2008)](https://iopscience.iop.org/article/10.1088/1751-8113/41/7/075303)
A. M. Childs, “Universal computation by quantum walk," [Phys. Rev. Lett. **102**, 18, 180501 (2009).](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.102.180501)
A. Ambainis, “Quantum walk algorithm for element distinctness," [SIAM J. Comput. **37** (1), pp. 210–239 (2007).](https://doi.org/10.1137/S0097539705447311)
F. Magniez, M. Santha, and M. Szegedy, “Quantum algorithms for the triangle problem," [SIAM J. Comput. [**37**]{}, no. 2, pp. 413–424 (2007).](https://doi.org/10.1137/050643684)
H. Buhrman and R. Spalek, “Quantum verification of matrix products," [Proceedings of the $17^{th}$ annual ACM-SIAM SODA, pp. 880–889 (2006).](https://dl.acm.org/doi/abs/10.5555/1109557.1109654)
E. Farhi, J. Goldstone, and S. Gutmann, “A quantum algorithm for the Hamiltonian NAND tree," [arXiv:quant-ph/0702144 (2007)](https://arxiv.org/abs/quant-ph/0702144).
N. Konno, “Quantum walks," [Quantum potential theory, pp. 309–452, Springer (2008)](https://link.springer.com/chapter/10.1007/978-3-540-69365-9_7).
K. R. Parthasarathy, “The passage from random walk to diffusion in quantum probability," [J. App. Prob. [**25**]{}, pp. 151–166, 1988.](https://www.jstor.org/stable/3214153)
Y. Aharonov, L. Davidovich, and N. Zagury, “Quantum random walks," [Phys. Rev. A **48**, 1687-1690 (1993).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.48.1687)
D. A. Meyer, “From quantum cellular automata to quantum lattice gases," [J. Stat. Phys. [**85**]{}, no. 5-6, pp. 551–574 (1996).](https://link.springer.com/article/10.1007%2FBF02199356)
J. Kempe, “Quantum random walks: an introductory overview", [Contemp. Phys **44.4**, 307-327 (2003).](https://www.tandfonline.com/doi/abs/10.1080/00107151031000110776)
S. E. Venegas-Andraca, “Quantum walks: a comprehensive review", [Quantum. Info. Process **11**, 1015 (2012).](https://link.springer.com/article/10.1007/s11128-012-0432-5)
A. M. Childs, R. Cleve, E. Deotto, E. Farhi, S. Gutmann, and D. A. Spielman, “Exponential algorithmic speedup by a quantum walk," [Proceedings of the $35^{th}$ annual ACM symposium on Theory of computing, pp. 59–68, ACM (2003).](https://dl.acm.org/doi/10.1145/780542.780552)
A. M. Childs and J. Goldstone, “Spatial search by quantum walk," [Phys. Rev. A [**70**]{} 2, 022314 (2004).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.70.022314)
C. Chandrashekar and T. Busch, “Quantum percolation and transition point of a directed discrete-time quantum walk," [Sci. rep., [**4**]{}, 6583 (2014).](https://www.nature.com/articles/srep06583)
B. Kollár, T. Kiss, J. Novotny, and I. Jex, “Asymptotic dynamics of coined quantum walks on percolation graphs," [Phys. Rev. Let. [**108**]{} 23, 230505 (2012)](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.108.230505)
P. Chawla, C. V. Ambarish, C. M. Chandrashekar, “Quantum percolation in quasicrystals using continuous-time quantum walk," [J.Phys. Comm. [**3**]{}, 125004, (2019)](https://iopscience.iop.org/article/10.1088/2399-6528/ab5ce0).
G. S. Engel, T. R. Calhoun, E. L. Read, T.-K. Ahn, T. Mancal, Y.-C. Cheng, R. E. Blankenship, and G. R. Fleming, “Evidence for wavelike energy transfer through quantum coherence in photosynthetic systems," [Nat., [**446**]{}, 7137, (2007).](https://www.nature.com/articles/nature05678)
M. Mohseni, P. Rebentrost, S. Lloyd, and A. Aspuru Guzik, “Environment-assisted quantum walks in photosynthetic energy transfer," [J. Chem. Phys. **129**, 174106 (2008).](https://aip.scitation.org/doi/10.1063/1.3002335)
P. Chawla, R. Mangal, C. M. Chandrashekar, “Discrete-time quantum walk algorithm for ranking nodes on a network," [Quantum Inf. Proc. [**19**]{}, 158 (2020).](https://link.springer.com/article/10.1007/s11128-020-02650-4)
G. Paparo, M. Müller, F. Comellas, M. A. Martin-Delgado, “Quantum Google in a Complex Network," [Sci. Rep. **3**, 2773 (2013).](https://www.nature.com/articles/srep02773)
S. Singh, P. Chawla, A. Sarkar, C. M. Chandrashekar, “Computational power of single qubit discrete-time quantum walk," [arXiv:1907.04084. (2019).](https://arxiv.org/abs/1907.04084)
S. Singh, P. Chawla, A. Agarwal, S. Srinivasan, C. M. Chandrashekar, “Universal quantum computation using single qubit discrete-time quantum walk," [arXiv:2004.05956 (2020).](https://arxiv.org/pdf/2004.05956.pdf).
C. M. Chandrashekar, H. Obuse, T. Busch, “Entanglement properties of localized states in 1D topological quantum walks," [arXiv:1502.00436 \[quant-ph\] (2015).](https://arxiv.org/abs/1502.00436).
A. P. Schnyder, S. Ryu, A. Furusaki, A. W. W. Ludwig, “Classification of topological insulators and superconductors in three spatial dimensions," [Phys. Rev. B **78**, 195125 (2008).](https://journals.aps.org/prb/abstract/10.1103/PhysRevB.78.195125)
T. Kitagawa, M. Rudner, E. Berg, E. Demler, “Exploring Topological Phases with quantum walks," [Phys. Rev. A **82**, 033429 (2010).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.82.033429)
J. K. Asbóth , H. Obuse, “Bulk-boundary correspondence for chiral symmetric quantum walks," [Phys. Rev. B **88**, 121406(R) (2013).](https://journals.aps.org/prb/abstract/10.1103/PhysRevB.88.121406)
F. W. Strauch, “Relativistic quantum walks," [Phys. Rev. A [**73**]{} 5, 054302, (2006).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.73.054302)
C. Chandrashekar, S. Banerjee, and R. Srikanth, “Relationship between quantum walks and relativistic quantum mechanics," [Phys. Rev. A [**81**]{}, 6, 062340, (2010).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.81.062340)
A. Mallick, C. M. Chandrashekar, “Dirac cellular automaton from split-step quantum walk," [Sci. Rep. **6**, 25779 (2016).](https://www.nature.com/articles/srep25779)
C. M. Chandrashekar, “Two-component dirac-like hamiltonian for generating quantum walk on one-, two-and three-dimensional lattices," [Sci. Rep. **3**, 2829 (2013).](https://www.nature.com/articles/srep02829)
G. Di Molfetta, M. Brachet, and F. Debbasch, “Quantum walks as massless dirac fermions in curved space-time," [Phys. Rev. A [**88**]{}, 4, 042301 (2013).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.88.042301)
G. Di Molfetta, M. Brachet, and F. Debbasch, “Quantum walks in artificial electric and gravitational fields," [Phys. A: Stat. Mech. App [**397**]{}, pp. 157–168 (2014).](https://www.sciencedirect.com/science/article/abs/pii/S0378437113011059)
P. Arrighi, S. Facchini, and M. Forets, “Quantum walking in curved spacetime," [Quantum Inf. Proc. [**15**]{}, 8, pp. 3467–3486 (2016).](https://link.springer.com/article/10.1007%2Fs11128-016-1335-7)
A. Pèrez, “Asymptotic properties of the dirac quantum cellular automaton," [Phys. Rev. A [**93**]{}, 1, 012328 (2016).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.93.012328)
C. A. Ryan, M. Laforest, J.-C. Boileau, and R. Laflamme, “Experimental implementation of a discrete-time quantum random walk on an NMR quantum-information processor," [Phys. Rev. A **72** , 062317 (2005).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.72.062317)
A. Joye, “Dynamical localization for d-dimensional random quantum walks," [Quantum Inf. Proc. [**11**]{}, 5, 1251–1269 (2012).](https://link.springer.com/article/10.1007/s11128-012-0406-7)
C. M. Chandrashekar, “Disorder induced localization and enhancement of entanglement in one-and two-dimensional quantum walks," [arXiv:1212.5984 (2012)](https://arxiv.org/abs/1212.5984).
C. Chandrashekar and T. Busch, “Localized quantum walks as secured quantum memory," [EuroPhys. Lett., [**110**]{}, 1, 10005 (2015).](https://iopscience.iop.org/article/10.1209/0295-5075/110/10005)
A. Mallick, S. Mandal, and C. Chandrashekar, “Neutrino oscillations in discrete-time quantum walk framework," [J. Eur. Phys. C [**77**]{}, 2, 85 (2017).](https://link.springer.com/article/10.1140/epjc/s10052-017-4636-9)
G. Di Molfetta and A. Pèrez, “Quantum walks as simulators of neutrino oscillations in a vacuum and matter," [N. J. Phys. [**18**]{}, 10, 103038 (2016).](https://iopscience.iop.org/article/10.1088/1367-2630/18/10/103038)
A. Schreiber, K. N. Cassemiro, V. Potocek, A. Gabris, P. J. Mosley, E. Andersson, I. Jex, Ch. Silberhorn, “Photons Walking the Line: A quantum walk with adjustable coin operations," [Phys. Rev. Lett. **104**, 050502 (2010).](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.104.050502)
M. A. Broome, A. Fedrizzi, B. P. Lanyon, I. Kassal, A. Aspuru-Guzik, & A. G. White, “Discrete Single-Photon Quantum Walks with Tunable Decoherence," [Phys. Rev. Lett. **104**, 153602 (2010).](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.104.153602)
A. Peruzzo et. al , “Quantum Walks of Correlated Photons," [Science **329**, 1500 (2010).](https://science.sciencemag.org/content/329/5998/1500)
H. Schmitz, R. Matjeschk, Ch. Schneider, J. Glueckert, M. Enderlein, T. Huber, & T. Schaetz,“ Quantum walk of a trapped ion in phase space," [Phys. Rev. Lett **103**, 090504 (2009).](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.103.090504)
F. Zahringer, G. Kirchmair, R. Gerritsma, E. Solano, R. Blatt, C. F. & Roos, “Realization of a Quantum Walk with One and Two Trapped Ions," [Phys. Rev. Lett **104**, 100503 (2010).](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.104.100503)
M. Karski, L. F\`‘[o]{}rster, J-M. Choi, A. Steffen, W. Alt, D. Meschede, A. & Widera, “Quantum Walk in Position Space with Single Optically Trapped Atoms," [Science **325**, 5937, 174-177 (2009).](https://science.sciencemag.org/content/325/5937/174)
T. Mitchell, “Machine Learning." McGraw-Hill, New York, New York, 1997.
D. H. Wolpert, “The Lack of A Priori Distinctions Between Learning Algorithms," [Neur. Comput. [**8**]{} (7), 1341-1390 (1996).](https://dl.acm.org/doi/10.1162/neco.1996.8.7.1341)
M. G. A. Paris, “Quantum Estimation for Quantum Technology“, [Int. J. Quant. Inf. [**07**]{} (supp01), 125–137 (2009).](https://www.worldscientific.com/doi/10.1142/S0219749909004839)
G. Tóth, D. Petz, “Extremal properties of the variance and the quantum Fisher information“. [Phys. Rev. A [**87**]{}, 3, 032324 (2013).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.87.032324)
Yu, Sixia, “Quantum Fisher Information as the Convex Roof of Variance“. [arXiv:1302.5311 (2013](https://arxiv.org/abs/1302.5311)).
G. Tóth, I. Apellaniz,“Quantum metrology from a quantum information science perspective“. [J. Phys. A: Math. Theor. [**47**]{}, 42 424006 (2014).](https://iopscience.iop.org/article/10.1088/1751-8113/47/42/424006)
P. Hyllus, “Fisher information and multiparticle entanglement“. [Phys. Rev. A [**85**]{}, 2, 022321 (2012).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.85.022321)
S. Singh, C. M. Chandrashekar, M. G. A. Paris, “Quantum walker as a probe for its coin parameter" [Phys. Rev. A [**99**]{}, 052117 (2019).](https://journals.aps.org/pra/abstract/10.1103/PhysRevA.99.052117)
M. Campbell, A.J. Hoane Jr., F. Hsu. “Deep Blue," [Art. Int. [**134**]{} 1-2, pp 54-83 (2002).](https://www.sciencedirect.com/science/article/pii/S0004370201001291)
D. B. Lenat, R. V. Guha, “Building Large Knowledge-Based Systems; Representation and Inference in the Cyc Project" (1st ed.). Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc. (1989).
P. Domingos, “The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World," New York, New York, Basic Books (2015)
I. Goodfellow, Y. Bengio, A. Courville, “Deep Learning", MIT Press, MA, USA (2016).
R. M. O‘Brien, “A Caution Regarding Rules of Thumb for Variance Inflation Factors“. [Qual. Quant. [**41**]{}, 673–690. (2007).](https://link.springer.com/article/10.1007/s11135-006-9018-6)
M. Gruber, “Improving Efficiency by Shrinkage: The James–Stein and Ridge Regression Estimators." Boca Raton: CRC Press. (1998).
A. N. Tikhonov, “Solution of incorrectly formulated problems and the regularization method." [Dokl. Akad. Nauk SSSR [**151**]{}, 3, 501-504 (1963).](http://mi.mathnet.ru/eng/dan28329)
A. N. Tikhonov, A. Goncharsky, V. V. Stepanov, A. G. Yagola, “Numerical Methods for the Solution of Ill-Posed Problems". Netherlands: Springer Netherlands (1995).
R. Tibshirani, “Regression Shrinkage and Selection via the lasso," [J. R. Stat. Soc. Ser. B (meth.) [**58**]{} , 1 267–88 (1996)](https://rss.onlinelibrary.wiley.com/doi/10.1111/j.2517-6161.1996.tb02080.x).
T. M. Cover, P. E. Hart, “Nearest neighbor pattern classification," [IEEE Trans. Inf. Theor. [**13**]{}, 1, 21–27 (1967).](https://ieeexplore.ieee.org/document/1053964)
N. S. Altman, “An introduction to kernel and nearest-neighbor nonparametric regression," [Amer. Stat. [**46**]{}, 3, 175–185 (1992).](https://www.tandfonline.com/doi/abs/10.1080/00031305.1992.10475879)
S. M. Piryonesi, T. E. El-Diraby, “Role of Data Analytics in Infrastructure Asset Management: Overcoming Data Size and Quality Problems," [J. Trans. Engg. Part B: Pavements [**146**]{}, 2, 04020022 (2020).](https://ascelibrary.org/doi/10.1061/JPEODX.0000175)
T. Hastie, R. Tibshirani, J. H. Friedman, “The elements of statistical learning : data mining, inference, and prediction," [Springer, New York (2001).](https://www.springer.com/gp/book/9780387848570)
K. Beyer, J. Goldstein, R. Ramakrishnan, U. Shaft, “When Is ’Nearest Neighbor‘ Meaningful?" [ICDT 1999, Lec. Notes in Comp. Sc. [**1540**]{} Springer, Berlin, Heidelberg (1999).](https://rd.springer.com/chapter/10.1007/3-540-49257-7_15)
W. McCulloch, W. Pitts, “A Logical Calculus of Ideas Immanent in Nervous Activity," [Bull. Math. Biophy. [**5**]{}, 4, 115–133 (1943)](https://link.springer.com/article/10.1007/BF02478259).
S. C. Kleene, “Representation of Events in Nerve Nets and Finite Automata," [Ann. Math. St. [**34**]{} (Automata Studies), Princeton University Press, pp. 3–42 (1956).](https://doi.org/10.1515/9781400882618-002)
T. Dozat, “Incorporating Nesterov Momentum into Adam," [Proceedings of the ICLR Workshop, (2016).](cs229.stanford.edu/proj2015/054_report.pdf)
Yurii Nesterov, “A method for unconstrained convex minimization problem with the rate of convergence $\mathcal{O}$(1/$k^2$)," [Dok. ANSSSR (trans. Soviet. Math. Docl.) [**269**]{} pp. 543–547 (1983).](http://mpawankumar.info/teaching/cdt-big-data/nesterov83.pdf)
D. P. Kingma, J. L. Ba, “Adam: A Method for Stochastic Optimization." [$3^{rd}$ ICLR Conf. Proc., pp. 1–13 (2015).](https://arxiv.org/abs/1412.6980)
S. Ruder, “An overview of gradient descent optimization algorithms", [arXiv:1609.04747 \[cs.LG\] (2016)](https://arxiv.org/abs/1609.04747)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '[ Previous work has shown that the standard supergravity approximation can break down when using AdS/CFT duality to study certain top-down formulations of the jet stopping problem in strongly-coupled ${\cal N}{=}4$ super-Yang-Mills (SYM) plasmas, depending on the virtuality of the source of the “jet.” In this paper, we identify the nature of this breakdown: High-momentum gravitons in the gravitational dual get stretched into relatively large classical string loops by tidal forces associated with the black brane. These stringy excitations of the graviton are not contained in the supergravity approximation, but we show that the jet stopping problem can nonetheless still be solved by drawing on various string-theory methods (the eikonal approximation, the Penrose limit, string quantization in pp-wave backgrounds) to obtain a probability distribution for the late-time classical string loops. In extreme cases, we find that the gravitons are stretched into very long folded strings which are qualitatively similar to the folded classical strings originally used by Gubser, Gulotta, Pufu and Rocha to model the jet stopping problem. This makes a connection in certain cases between the different methods that have been used to study jet stopping with AdS/CFT and gives a specific example of a precise ${\cal N}{=}4$ SYM problem that generates such strings in the gravity description. ]{}'
author:
- 'Peter Arnold, Phillip Szepietowski, Diana Vaman, and Gabriel Wong'
title: ' Tidal stretching of gravitons into classical strings: application to jet quenching with AdS/CFT '
---
Introduction {#sec:intro}
============
Background
----------
Inspired by the observation of (and rapidly growing body of experimental information on) jet quenching in relativistic heavy ion collisions, there has for many years been an interest in the theory of jet quenching and what can be learned about that theory by studying interesting limiting cases. One of the simplest-to-pose thought experiments is this: How far does a very-high momentum excitation (the potential precursor of a would-be jet) travel in a thermal QCD medium before it loses energy, stops, and thermalizes in the medium? And how does the answer to that question depend on the effective strength $\alphas$ of the strong coupling?
This question can be addressed from first principles in various theoretical limits. One such limit is that of weak coupling, which in principle applies to asymptotically large temperatures $T$ and jet energies $E$, for which the relevant running values of $\alphas$ are small. In that limit, the stopping distance $\lstop$ for a high-energy parton ($E \gg T$) scales with energy as $E^{1/2}$, up to logarithms.[^1] A contrasting limit of interest occurs when the running values of $\alphas$ relevant to jet stopping are all large.[^2] This problem is not very tractable from first principles in QCD itself, but, through gauge-gravity duality, progress can be made for QCD-like plasmas with gravity duals, such as ${\cal N}{=}4$ super Yang Mills (SYM) theory. For some years, people have considered various ways to study analogs of jet stopping in such plasmas, namely the stopping distance for various types of localized, high-momentum excitations. The exact stopping distance depends on details of exactly how the “jet” is prepared, but universally these studies have found that the maximum possible stopping distance $\lmax$ scales with energy as $E^{1/3}$ [@GubserGluon; @HIM; @CheslerQuark; @adsjet; @adsjet2; @CHR], in contrast to the weak-coupling scaling of $E^{1/2}$. This is an interesting theoretical result because it teaches us that the scaling of jet stopping with energy depends on the strength of the coupling. It remains an open question (which we will not answer here) how $E^{1/3}$ starts to move toward $E^{1/2}$ as one lowers the coupling, and vice versa.[^3]
The stopping distance of high-momentum, localized excitations traveling through the plasma depends on more than just the energy of the excitation. Depending on exactly how one creates the excitation (the “jet”), one may get stopping distances $\lstop$ significantly smaller than the maximum $\lmax$. As an example from weak coupling, imagine that we spread out the total energy and momentum $E$ of the jet among 10 partons, each having energy $E/10$, rather than putting it all into a single parton of energy $E$. Each of the 10 partons has lower energy than the single one and so will stop sooner; so the stopping distance for the high-momentum excitation depends on how many high-energy partons we use in the initial state. In the weak-coupling case, the maximum stopping distance $\ell_{\rm max}$ corresponds to the particular initial state where all the energy is concentrated into a single initial parton.
In the strong-coupling case, we cannot speak of individual partons, but the stopping distance again depends on how we prepare the initial high-momentum excitation. In our work [@adsjet; @adsjet2; @R4], we create the initial excitation in a way that is analogous to what you would get if a high-momentum, slightly virtual photon (or graviton or other massless particle) decayed hadronically in the quark-gluon plasma, as depicted in fig. \[fig:Wdecay\]. Alternatively, one could consider the decay of a high-momentum on-shell W boson (also depicted). For these methods of creating “jets,” one finds that the maximum possible stopping distance scales as $$\lmax \sim \frac{E^{1/3}}{T^{4/3}} \,.
\label {eq:lmax}$$ As we will review later, it turns out that the stopping distance may be made smaller than (\[eq:lmax\]) by varying the virtuality $-q^2 \equiv -q_\mu q^\mu$ of the virtual photon (or equivalently the mass-squared $M_{\rm w}^2$ of the on-shell W boson) [@HIM; @adsjet2]. The important point is that there is a range of stopping distances $\lstop \lesssim \lmax$ for our “jets,” depending on the details of how those excitations are created.
![ \[fig:Wdecay\] Examples of the decay of a very high-energy (a) slightly virtual photon, (b) slightly virtual graviton, or (c) on-shell $W^+$ boson, inside a standard-model quark-gluon plasma, producing high-momentum partons moving to the right. In the context of ${\cal N}{=}4$ super Yang Mills, the $q$, $u$, and $\bar d$ above represent adjoint-color fermions or scalars carrying R charge. For strong coupling, of course, one should not picture perturbatively, as in this figure, the high-momentum excitation created in the plasma by the decay. ](Wdecay.eps)
Most top-down studies of jet stopping using gauge-gravity duality have studied the infinite color and infinite coupling limit, $N_{\rm c}{=}\infty$ and $\lambda{=}\infty$, where $\lambda \equiv N_{\rm c} g_{\rm YM}^2$ is the ’t Hooft coupling. To understand the true high-energy behavior, however, it is important to study the corrections to these limits. As an example, fig. \[fig:stopvE\] shows two different scenarios one might imagine for the maximum stopping distance $\lmax$ for strongly-coupled ${\cal N}{=}4$ SYM. One is that $\lmax$ grows like $E^{1/3}$ at high energy, up to arbitrarily high energies. The other is that it starts growing like $E^{1/3}$ at high energy $E \gg T$, but then crosses over to some different power-law behavior once $E$ exceeds some positive power of $\lambda$ times $T$ (e.g. $\lambda^2 T$ in the figure). In the latter case, $E^{1/3}$ would not be the true behavior for arbitrarily large $E$ and large but finite $\lambda$. But there is no way to tell the difference between these two scenarios if one only has $\lambda{=}\infty$ calculations! For this reason, three of us analyzed the parametric size of finite-$\lambda$ corrections to jet stopping distances in ref. [@R4]. We found that the formal expansion in $1/\sqrt\lambda$ (which corresponds to an expansion in the string parameter $\alpha'$ on the gravity side) breaks down for some jets and is safe for others, depending on the stopping distance $\lstop$ of the jet (and therefore on the virtuality $-q^2$), as depicted in fig. \[fig:R4\]. The stopping distance and virtuality parametrize the horizontal axis in this figure. The vertical axis is a relative measure of the importance of a given correction compared to the $\lambda{=}\infty$ result (see ref. [@R4] for details). The curves are labeled by the sequence of higher-curvature terms in the gravitational dual theory action that correspond, via the AdS/CFT correspondence, to a sequence of corrections in powers of $1/\sqrt\lambda$ in the 3+1 dimensional ${\cal N}{=}4$ SYM theory. Throughout, $N_{\rm c}$ is taken to be infinite. The result of this study was that, for $\lambda \gg 1$, corrections to the $\lambda{=}\infty$ result are parametrically small for $\lambda^{-1/6}\lmax \ll \lstop \lesssim \lmax$. In particular, corrections to the maximum stopping distance $\lmax \propto E^{1/3}$ are small. But the interesting case is when jets are created in such a way that
\[eq:fun\] $$T^{-1} \ll \lstop \lesssim \lambda^{-1/6} \lmax ,
\label {eq:lstopwindow}$$ which is $$T^{-1} \ll
\lstop \lesssim \frac{\bigl(E/\sqrt\lambda\>\bigr)^{1/3}}{T^{4/3}} \,.$$
In this case, the fate of $\lambda{=}\infty$ results for the stopping distance was unclear. For $\lstop \sim \lambda^{-1/6} \lmax$, all the corrections are the same size, and so the formal expansion in powers of $1/\sqrt\lambda$ has broken down. Yet the individual corrections are all small (of relative importance $\lambda^{-1/2}$) for that $\lstop$. From fig. \[fig:R4\], we cannot tell whether the [*sum*]{} of the corrections to $\lambda{=}\infty$ will remain small for $\lstop \lesssim \lambda^{-1/6}\lmax$ or whether, instead, the $\lambda{=}\infty$ calculation becomes useless there.
![ \[fig:stopvE\] Examples of two different scenarios for the high-energy ($E \gg T$) behavior of the maximum jet stopping distance $\lmax(E)$ which are indistinguishable with $\lambda{=}\infty$ calculations. ](stopvE.eps)
![ \[fig:R4\] A parametric picture of the relative importance of higher-derivative corrections to the low-energy supergravity action as a function of the stopping distance $\ell_{\rm stop}$ (using the $\lambda{=}\infty$ result for $\ell_{\rm stop}$). The axis are both logarithmic, and an importance of $1$ indicates that the individual correction would, by itself, significantly modify the $\lambda{=}\infty$ analysis. The measure of “importance” is explained in ref. [@R4]. Also shown, as an alternative horizontal axis, is the 4-dimensional virtuality $-q^2$ of the source that created the jet, where ${\hat E}\equiv E/T$. ](corrections.ps)
The purpose of the present paper is to understand the physics (on the gravity side) of what is going on in the region (\[eq:fun\]) where the naive expansion in powers of $1/\sqrt\lambda$ (powers of $\alpha'$) breaks down, and to figure out how to account for the effect of this physics on the jet stopping distance. Note that the interesting window (\[eq:lstopwindow\]) of stopping distances exists only if the energy is large enough that $T^{-1} \ll \lambda^{-1/6} \lmax$. By (\[eq:lmax\]), this requires $E \gg \lambda^{1/2} T$, which we will assume throughout the rest of this paper.
Before outlining what we have done, it will be useful to first explain one other qualitative feature of the $\lambda{=}\infty$ calculation. Excitations created in the field theory correspond to excitations created on the boundary of AdS$_5$-Schwarzschild, which then fall towards the black brane over time, such as depicted in fig. \[fig:fall\]. The 3-space distance that this excitation travels before falling into the horizon matches the stopping distance of the corresponding excitation in ${\cal N}{=}4$ SYM.[^4] For $\lstop \ll \lmax$, which includes the region (\[eq:fun\]) of interest, there is a nice simplification. On the gravity side, the excitation falling in fig. \[fig:fall\] turns out to be a spatially [*small*]{} wavepacket which can be treated in the geometric optics approximation. The wavepacket’s motion is the same (up to parametrically small corrections) as that of a 5-dimensional “particle” traveling in the AdS$_5$-Schwarzschild geometry, and so it follows a geodesic whose trajectory is easily calculated in terms of the 4-momentum $q_\mu$ of the excitation. (See section \[sec:geodesic\] for more detail.)
![ \[fig:fall\] Qualitative sketch of the motion though AdS$_5$-Schwarzschild of a wave packet with high 3-momentum in the $x^\three$ direction. As measured by $x^\zero$, the particle takes infinitely long to reach the horizon. Of special importance is the parametric scale $x^\five_\star$ in the fifth dimension, where the trajectory turns over and beyond which progress in $x^\three$ rapidly slows to a stop. ](fall.eps)
It will be important for what follows to remember that the AdS/CFT correspondence is really a correspondence between field theory and 10-dimensional string theory. In the strong-coupling limit $\lambda{=}\infty$ of the field theory, the general correspondence reduces to one between the field theory and the infrared limit of the string theory, which is a supergravity theory. The quanta of the supergravity fields correspond to string states that are massless in flat 10-dimensional spacetime, such as the graviton. For $\lambda{=}\infty$, the well-known gravitational dual of finite-temperature ${\cal N}=4$ SYM is Type IIB supergravity in an (AdS$_5$-Schwarzschild)$\times S^5$ background.
What we find
------------
The classical wave packet falling in fig. \[fig:fall\] is a localized, classical excitation of the supergravity fields. For the sake of specificity, consider the case where it is an excitation of the background gravitational field. In general, a classical wave is a coherent superposition of the corresponding quanta, and so our high-frequency classical wavepacket is a coherent superposition of high-frequency gravitons. Since our wavepacket behaves like a particle (in the geometric optics approximation appropriate for $\lstop \ll \lmax$), let’s follow just one of these gravitons. So think of the trajectory in fig. \[fig:fall\] as that of a single high-momentum graviton with a localized wave function. As we will discuss later, interactions between the gravitons that make up the wavepacket are very small, and so it is adequate to think about the evolution of individual high-momentum gravitons.
A graviton is really a tiny loop of string whose internal degrees of freedom are in their ground state. Because of the gravitational field from the black brane, this closed string will feel tidal forces as it falls, which will try to stretch the string in some directions and squeeze it in others. As the graviton gets further from the boundary (and so closer to the black brane), the tidal forces will increase, and eventually they will become large enough to excite the internal string degrees of freedom of the graviton. We will argue that it is the excitation of these string degrees of freedom that is responsible for the breakdown of the expansion in fig. \[fig:R4\] in the problem region (\[eq:fun\]).
We will find that, in the problematic case (\[eq:fun\]) where $\lstop \ll \lambda^{-1/6}\lmax$, the tidal forces are strong enough to stretch that loop of string to become classically large before the stopping distance is reached. This is why stringy corrections cannot be ignored in that case, explaining the breakdown of the expansion in fig. \[fig:R4\]. (In contrast, the tidal forces are not strong enough to excite the graviton’s internal degrees of freedom soon enough when $\lstop \gg \lambda^{-1/6}\lmax$.) Though the resulting classical string loop will be large compared to the size of a graviton, we should ask how its size compares to the stopping distance $\lstop$. We will find that the ratio of (i) the stretched, classical string’s size in the direction of motion $x^\three$ to (ii) the stopping distance $\lstop$ is parametrically of order $$\frac{(\delta x^\three)_{\rm string}}{\lstop}
\sim
\lambda^{-1/4} \, \ln^{1/2}\biggl(\frac{\lambda^{-1/6} \lmax}{\lstop}\biggr)
.
\label {eq:ratio}$$ Because of the $\lambda^{-1/4}$, this ratio is typically parametrically small for large but finite $\lambda$, and we will argue that the stretching of the graviton into a string (and the accompanying breakdown of the formal expansion in $1/\sqrt{\lambda}$ in fig. \[fig:R4\]) then has sub-leading impact on $\lambda{=}\infty$ results for the stopping distance. But (\[eq:ratio\]) also includes cases where the stretching of the string may play an important role: If one considers a situation where the argument of the logarithm in (\[eq:ratio\]) is exponentially large, then the logarithm can be large enough to compensate for the factor of $\lambda^{-1/4}$. We discuss this situation further in our conclusions, where we make contact with folded classical string configurations that were originally used by Gubser et al. [@GubserGluon] to model jet stopping.
Since the tidal forces stretch a quantum string (the graviton) into a larger classical string, one may wonder whether or not it is possible to do a real, detailed calculation of the transition between the two. Having restricted attention to a single graviton, our problem reduces to following the evolution of a single closed string in the AdS$_5$-Schwarzschild background. In general, it is not known (for all practical purposes) how to quantize a string in an AdS$_5$-Schwarzschild background. But remember that our graviton is localized and so only probes a region of space-time near the geodesic depicted in fig. \[fig:fall\]. It is enough to consider only a narrow region of the space-time that lies near a null geodesic, as depicted in fig. \[fig:fallpp\], and so we may treat the full background metric in an approximation (known as a Penrose limit) that treats displacements from the null geodesic as small. The resulting approximation to the background metric is an example of what is known as a pp-wave background, and it [*is*]{} known how to quantize a string in a pp-wave background. In particular, it will be possible to calculate the probability distribution of the shape of the classical string loop. The methods we use are similar to previous works by other authors on the excitation of string modes in scattering processes and/or in pp-wave backgrounds [@Veneziano; @PRT; @HS; @GPZS; @dVS].
![ \[fig:fallpp\] The (pink) shaded area represents a narrow region of space-time around the null geodesic of fig. \[fig:fall\]. The AdS$_5$-Schwarzschild metric in this region may be approximated as a pp-wave background for the purpose of quantizing a small, falling loop of string that describes a graviton (or other particle) in the initial falling wavepacket. ](fallpp.eps)
In the next section, we give a simple, back-of-the-envelope argument that yields the result $(\delta x^\three)/\ell_{\rm stop} \sim \lambda^{-1/4}$ but is not precise enough to explain the logarithmic factor in (\[eq:ratio\]). Back-of-the-envelope estimates are sometimes enlightening and sometimes frustratingly unconvincing, and the rest of the paper is devoted to a more formal calculation along the lines just described. We set up our notational conventions and review $\lambda{=}\infty$ calculations in section \[sec:setup\]. Next we review in section \[sec:R4review\] the source of the breakdown of the $\alpha'$ expansion for $\lstop \ll \lambda^{-1/6} \lmax$, depicted in fig. \[fig:R4\], and use it to motivate why these problems should be overcome by following the quantum evolution of single closed strings in the AdS$_5$-Schwarzschild background. In section \[sec:Penrose\], we take the relevant Penrose limit of the AdS$_5$-Schwarzschild metric, crudely depicted in fig. \[fig:fallpp\]. We are then ready to quantize the string in section \[sec:quantize\] and solve for the evolution of the graviton into a classical loop of string. We then compute the average size of the resulting classical loops of string, which parametrically gives (\[eq:ratio\]), including the logarithmic factor, for the interesting case $\lstop \ll \lambda^{-1/6} \lmax$. Some readers may wonder how a quantum treatment of the stretching of gravitons can underlie the analysis, given that the gravitational/string dual theory is supposed to be classical in the limit $\Nc{=}\infty$ taken in this paper. We discuss this in section \[sec:gravitons\], where we also give a more detailed justification for treating the gravitons in the wavepacket as independent. In section \[sec:PenroseCheck\] (supplemented by Appendix \[app:PenroseCheck\]), we then revisit the Penrose limit used in our analysis and verify that it is justified, provided that (\[eq:ratio\]) is small. Finally, we offer our conclusions in section \[sec:conclusion\].
A back-of-the-envelope estimate {#sec:envelope}
===============================
In this section, we will make a parametric estimate of the amount of tidal stretching of the string compared to the size of the stopping distance $\lstop$. In a later section, we will review the $\lambda{=}\infty$ results for how the stopping distance depends on the energy and 4-virtuality of our jet source, but here the only thing we will need to know is that the stopping distance given by following a null geodesic as in fig. \[fig:fall\] is proportional to a power of the slope $dx^\three/dx^\five$ of that geodesic where it starts, at the boundary. The more downward-directed one starts the trajectory in fig. \[fig:fall\], the less distance it will travel in $x^\three$ before reaching the horizon.
Now interpret the trajectory of fig. \[fig:fall\] as a trajectory for the center of mass of a tiny, falling loop of string. Once the string gets far enough from the boundary that the tidal forces dominate over the string tension, then the string tension becomes ignorable, and different pieces of the string will fall independently along their own geodesics, the string stretching accordingly. Imagine plotting two such geodesics, for the two bits of the string loop that are most separated. The separation of those geodesics is a measure of the extent of the tidally-stretched loop of string as it falls towards the horizon. The [*proper*]{} size of the string should start out of order the quantum mechanical size $\lgrav$ of the graviton, which is roughly set by dimensional analysis in terms of the string tension $\tension$ as $$\lgrav_{\rm graviton} \sim \tension^{-1/2} \sim \sqrt{\alpha'} ,
\label {eq:lgrav}$$ where $\alpha' = 1/2\pi\tension$ is the string slope parameter.
Very close to the boundary, the tidal forces due to the black hole are negligible, and the closed loop of string is in its ground state. We can set up our two geodesics above so that, correspondingly, they maintain constant proper separation $\lgrav_{\rm graviton}$ near the boundary, where the AdS$_5$-Schwarzschild metric approaches a purely AdS$_5$ metric. To see how this works, imagine making a 4-dimensional boost from (i) the plasma rest frame, in which we create an excitation with large 4-momentum $q^\mu = (\omega,0,0,q_\three) \simeq (E,0,0,E)$ and relatively small 4-virtuality $-q^2 \ll E^2$, to (ii) the excitation’s initial rest frame, where the 4-momentum is instead $(\sqrt{-q^2},0,0,0)$. The Lorentz boost factor for this transformation is $$\gamma = \sqrt{\frac{\omega^2}{-q^2}}
\simeq \sqrt{\frac{E^2}{-q^2}}
\gg 1 .
\label {eq:gamma}$$ In AdS$_5$, the trajectory in the new frame will drop straight down away from the boundary, as depicted by the dashed line in fig. \[fig:boost\]a.
![ \[fig:boost\] (a) Parallel geodesics in AdS$_5$ in the Lorentz frame where the excitation is at rest in 3-space. These geodesics maintain a constant proper separation as they fall into the bulk, and this separation should be thought of as of order the characteristic size ($\sim \sqrt{\alpha'}$) of the closed quantum string loop describing the graviton (or other massless string mode). The narrow red loops are meant to be suggestive of the closed string loop. (b) The same picture boosted to the original Lorentz frame. (c) A picture of how those geodesics evolve in AdS$_5$-Schwarzschild rather than AdS$_5$. The early-time behavior is the same as (b). \[For classical oscillating string solutions, the strings depicted in (a) may be thought of as snapshots at moments when the string’s proper extent in $x^\three$ is at, say, maximum (or half-maximum or whatever). Such solutions would similarly oscillate in the $z \ll z_\star$ part of (b) but not in the $z \gg z_\star$ part, where tidal forces dominate over tension.\] ](boost.eps)
Now consider the graviton as an extended object with proper size $\lgrav$. The two straight solid null lines in fig. \[fig:boost\]a depict the extent of the graviton in AdS$_5$ in the excitation’s rest frame at early times. In Poincare coordinates for pure AdS$_5$, $$(ds)^2 =
\frac{\Rads^2}{z^2} \,\bigl (dx^\mu \eta_{\mu\nu} dx^\nu + (dz)^2 \bigr) ,$$ null geodesics are straight lines. Here $\Rads$ is the radius of the 5-sphere $S^5$. We parametrize the two solid lines of fig. \[fig:boost\]a as $$x^I = \bigl(\gamma_+,0,0,\pm\beta_+\gamma_+,1\bigr) \, z
\label {eq:xline}$$ with $\beta_+ \ll 1$ and $\gamma_+ \equiv (1-\beta_+^2)^{-1/2} \simeq 1$. Because of the warp factor in the metric, these two lines are parallel and maintain constant proper separation $$\sqrt{\Delta x^\three \, g_{\three\three} \, \Delta x^\three}
= 2\beta_+ \gamma_+ \Rads
\simeq 2\beta_+ \Rads
\label {eq:separation}$$ as a function of the rest-frame time. Setting this proper separation to be of order the graviton size $\lgrav$ given by (\[eq:lgrav\]) then gives $$\beta_+ \sim
\frac{\lgrav_{\rm graviton}}{\Rads} \sim \frac{\sqrt{\alpha'}}{\Rads} \,.$$ The AdS/CFT dictionary relates $\alpha'$ in the string theory with the ’t Hooft coupling of ${\cal N}{=}4$ SYM by [@MAGOO] $$\frac{\alpha'}{\Rads^2} = \lambda^{-1/2}
\equiv (g_{\rm YM}^2 N_{\rm c})^{-1/2} .
\label {eq:lambda}$$ So $$\beta_+ \sim \lambda^{-1/4} ,$$ and (\[eq:xline\]) gives $$x^I \sim \bigl(1,0,0,\pm\lambda^{-1/4},1\bigr) \, z .$$
Now boost back to the original plasma frame using (\[eq:gamma\]) to get the early-time trajectories depicted by solid lines in fig. \[fig:boost\]b: $$x^I \sim
\Bigl(\gamma(1 \pm \lambda^{-1/4}),0,0,\gamma(1\pm\lambda^{-1/4}),1\Bigr) \, z
,$$ where we have used $\gamma \gg 1$ (\[eq:gamma\]). Then $$\frac{dx^\three}{dz} \Bigg|_{\rm initial} \simeq \gamma(1\pm\lambda^{-1/4}) ,$$ which represents a small relative variation $$\frac{\Delta(dx^\three/dz)}{dx^\three/dz} \Bigg|_{\rm initial}
\sim \lambda^{-1/4}
\label {eq:relative}$$ in the initial slope $dx^\three/dz$ of the trajectory. As discussed before, the stopping distance (which requires a calculation in the full AdS$_5$-Schwarzschild metric) covered by a null geodesic is power-law related to this initial slope, and so the difference $\Delta \lstop$ in how far the two bits of string travel also has the same small size (\[eq:relative\]) relative to $\lstop$: $$\frac{\Delta\lstop}{\lstop} \sim \lambda^{-1/4} .$$ This is just our parametric result (\[eq:ratio\]) quoted in the introduction but without the logarithmic factor. The logarithmic factor requires a more detailed analysis.
Setup {#sec:setup}
=====
Notation
--------
In this paper, we will use Greek letters for 4-dimensional space-time indices ($\mu,\nu={\mathsf{0,1,2,3}}$) and upper-case roman letters for indices that run over all 5 dimensions of AdS$_5$-Schwarzschild ($I,J={\mathsf{0,1,2,3,5}}$). One form of the metric we use for AdS$_5$-Schwarzschild is $$(ds)^2 = \frac{\Rads^2}{z^2} \bigl[-f (dt)^2 + (d\x)^2 + f^{-1} (dz)^2\bigr]
,
\label {eq:metric}$$ where $z$ is the coordinate $x^\five$ of the fifth dimension, $\Rads$ is the radius of the 5-sphere (which will drop out of final results), and $$f \equiv 1 - \frac{z^4}{\zh^4} \,.$$ The boundary is at $z{=}0$, and the horizon is at $$\zh = \frac{1}{\pi T} \,.
\label {eq:zh}$$ We will not need to worry about the details of regularizing the location of the boundary in this work.
Review of $\lambda{=}\infty$ calculation {#sec:geodesic}
----------------------------------------
Fig. \[fig:Wdecay\] gave a cartoon picture of how we create our “jets” in the plasma. Readers may find a precise description of the field theory problem in any of the previous papers [@adsjet; @adsjet2; @jj; @R4] utilizing this method, but we will not need those details here. Suffice it to say that, in the gravity description, the boundary is perturbed in a localized region of space-time in such a way as to create an excitation with large 4-momentum $q^\mu \simeq (E,0,0,E)$ and a relatively small amount of time-like 4-virtuality $-q^2 \ll E^2$. The response of the system is then tracked at late times to see where the excitation comes to a stop.
The supergravity field that is excited depends on the nature of the source for the jet and is the supergravity field dual to the vertex operator in fig. \[fig:Wdecay\]. For example, if we study the decay of a slightly off-shell graviton in the 4-dimensional quantum field theory, then the relevant supergravity excitation is in the 5-dimensional gravitational field; if we were to study the decay of a gauge boson weakly coupled to R charge in the 4-dimensional quantum field theory, then the supergravity excitation would be in a corresponding 5-dimensional gauge field; and so forth.
As mentioned in the introduction, there is a nice simplification to the $\lambda{=}\infty$ analysis of this problem [@adsjet2] when $\lstop \ll \lmax$. On the gravity side, the wavepacket’s motion is then the same (up to parametrically small corrections) as that of a 5-dimensional “particle” traveling in the AdS$_5$-Schwarzschild geometry, and so it follows a geodesic whose trajectory is easily calculated in terms of the 4-momentum $q_\mu$ of the excitation. The exact geodesic depends on the mass of the 5-dimensional “particle” and so on the mass $m$ of the 5-dimensional supergravity field that we have excited, but this mass (if any) may be ignored in the high energy limit.[^5]
As a result, attention may be restricted to null geodesics, which are given by $$x^\mu(x^\five) = \int \sqrt{g_{\five\five}} \, dx^\five \>
\frac{ g^{\mu\nu} q_\nu }
{ (-q_\alpha g^{\alpha\beta} q_\beta)^{1/2} }
\,.
\label {eq:geodesic}$$ Using the metric (\[eq:metric\]), the $\lambda{=}\infty$ stopping distance is then found to be [@adsjet2] $$\ell_{\rm stop} \simeq
\int_0^{\zh} dz \> \frac{|\q|}{\sqrt{-q^2 + \frac{z^4}{\zh^4} |\q|^2}}
\simeq
\frac{\Gamma^2(\frac14)}{(4\pi)^{1/2}} \left( \frac{E^2}{-q^2} \right)^{1/4}
\frac{1}{2\pi T}
\,.
\label {eq:stop}$$ An important feature of the integral in (\[eq:stop\]) is that it is dominated by small values of $z$, of order $$z_\star
\equiv \zh \left( \frac{-q^2}{|\q|^2} \right)^{1/4}
\simeq \zh \left( \frac{-q^2}{E^2} \right)^{1/4}
\ll \zh .
\label {eq:zstar}$$ $z_\star$ corresponds to the parametric scale $x^\five_\star$ in fig. \[fig:fall\], where the trajectory turns over and beyond which 3-space motion rapidly slows to a stop. The stopping distance is determined by the behavior of the trajectory at $z \sim z_\star$.
Discussion of $1/\sqrt{\lambda}$ expansion {#sec:R4review}
==========================================
We now want to consider corrections to the $\lambda{=}\infty$ results for the stopping distance in the case $\Nc = \infty$. First, we take a moment to review the generic story of $1/\sqrt\lambda$ corrections in the AdS/CFT correspondence, which relates [@MAGOO] $$g_{\rm string} = \frac{\lambda}{4\pi\Nc} = \frac{g_{\rm YM}^2}{4\pi} \,,$$ $$\frac{1}{2\pi\tension \Rads^2} = \frac{\alpha'}{\Rads^2} = \lambda^{-1/2}
\equiv (g_{\rm YM}^2 N_{\rm c})^{-1/2} ,$$ where $g_{\rm string}$ is the string loop expansion parameter. The string tension $\tension$ sets the mass scale for massive string excitations, and so $\alpha'\to 0$ corresponds to taking the scale for massive string excitations to infinity. For $\lambda{=}\infty$, the strongly-coupled 4-dimensional quantum field theory is therefore dual to the infrared limit of the 10-dimensional string theory, namely supergravity, in the appropriate background. For large but finite $\lambda$, massive string modes are not completely ignorable, and the effective supergravity theory of the massless modes gets corrections, in the form of higher-dimensional terms in its action, from integrating out the effects of the massive modes. Schematically, the effective supergravity Lagrangian becomes[^6] $$\begin{gathered}
{\cal L} ~~\sim~~
R
~~+~~
\bigr[ \alpha'^3 R^4 + \alpha'^5 D^4 R^4 + \alpha'^6 D^6 R^4 + \cdots \bigl]
\\
~~+~~
\bigr[ \alpha'^5 D^2 R^5 + \alpha'^6 D^4 R^5 + \alpha'^7 D^6 R^5
+ \cdots \bigl]
~~+~~ \cdots ,
\label {eq:Leff}\end{gathered}$$ where we have focused just on the gravitational fields for simplicity. $R$ represents factors of the Riemann tensor, and we have not shown numerical coefficients or how the indices contract. For $\Nc{=}\infty$, there are no loop effects ($g_{\rm string}{=}0$), and accounting for the massive string modes in the effective theory is analogous to replacing the effects of the W boson by the Fermi 4-point interaction in electroweak theory. So, for example, the $R^4$ terms in (\[eq:Leff\]) are calculated from string amplitudes for $2\to 2$ graviton scattering and, crudely speaking, they correspond to processes which involve intermediate massive string states, as depicted schematically in fig. \[fig:string4\]. (A more accurate statement would be that they correspond to the full string amplitude for graviton-graviton scattering minus the sum of the $s$, $t$, and $u$-channel graviton exchange diagrams that one would calculate in the $\alpha'{=}0$ supergravity theory.) The $R^5$ terms similarly account for corrections to the 5-point graviton interaction, and so forth.
![ \[fig:string4\] A picture of massive string mode corrections to graviton-graviton scattering that are accounted for by the $D^m R^4$ corrections to the effective supergravity action. ](string4.eps)
In our application, we are interested in the evolution of a high-energy excitation propagating through the soft AdS$_5$-Schwarzschild background. For simplicity of presentation, we will focus on the case where the excitation is in the 5-dimensional gravitational fields, though our conclusions will not be sensitive to this assumption. A classical gravitational excitation can also be thought of as a coherent configuration of gravitons, and the relevant string scattering amplitudes are those where two of the external lines are the incoming and outgoing high-energy gravitons and the others are the soft background field. So, for a 4-point scattering amplitude such as fig. \[fig:string4\], the relevant kinematic limit is that depicted in fig. \[fig:string4kin\]. With the notation used in that figure, the high-energy limit corresponds to potentially large $s=-(p_1+p_2)_I (p_1+p_2)^I$ but small $t=-(p_1-p_3)_I (p_1-p_3)^I$. \[Here and throughout we may think of the $p$’s as 5-dimensional momenta in AdS$_5$-Schwarzschild rather than 10-dimensional momenta in (AdS$_5$-Schwarzschild)$\times S^5$ because in our problem there is no interesting dynamics associated with the 5-sphere $S^5$.\] As discussed in ref. [@R4], the $D^m R^4$ terms in (\[eq:Leff\]) all become equally important in the jet stopping problem when this 5-dimensional $\sqrt{s}$ becomes large enough at $z \sim z_\star$ to excite massive string modes in fig. \[fig:string4kin\]. The string mass scale is of order $1/\sqrt{\alpha'}$, and this condition $$\sqrt{s_{\mbox{\scriptsize (5-dim)}^{\vphantom{2}}}} \gtrsim \frac{1}{\sqrt{\alpha'}}
\label {eq:Mthreshold}$$ is shown in ref. [@R4] to be the same, in the context of the jet stopping problem, as the condition $$\lstop \lesssim \lambda^{-1/6} \lmax ,$$ which is the problematic case (\[eq:fun\]) highlighted in our introduction. In this region, massive string states in the intermediate state in fig. \[fig:string4kin\] are kinematically accessible and cannot be ignored.
![ \[fig:string4kin\] A high-energy graviton, depicted as a string loop, interacting twice with the AdS$_5$-Schwarzschild background gravitational field. ](string4kin.eps)
As the high-energy excitation falls from the boundary to the horizon, as in fig. \[fig:fall\], it does not just interact with the background field once or twice but does so over and over again, as depicted in fig. \[fig:stringmany\]. If the massive string states are kinematically accessible as in (\[eq:Mthreshold\]), then they cannot be neglected in any of the internal lines, which means in the effective theory language of (\[eq:Leff\]) that all $D^m R^n$ terms will also become important. This is just what happens at the $\lstop \sim \lambda^{-1/6} \lmax$ point in fig. \[fig:R4\], where all the corrections become the same size, corresponding to the threshold $\sqrt{s_{\mbox{\scriptsize (5-dim)}^{\vphantom{2}}}} \sim
1/\sqrt{\alpha'}$. Also, note that if massive string modes are kinematically accessible for intermediate states, then they are also accessible as final states. (In our problem, however, there is never really an ultimate “final” asymptotic state of the excited graviton because the excited graviton falls into the black brane.)
![ \[fig:stringmany\] Like fig. \[fig:string4kin\] but with many interactions with the soft background field. ](stringmany.eps)
Interaction with a pure AdS$_5$ background does not produce these massive string mode excitations. On a technical level, the $R^4$ interactions really involve the Weyl tensor (the traceless part of the Riemann tensor), which vanishes for AdS$_5$ but not for AdS$_5$-Schwarzschild. So it is only the gravitational effect due to the presence of the black brane that contributes to massive string mode excitation. As a result, the effects of excited string modes are negligible at the boundary and become stronger as one moves away from it (and so closer to the black brane). At some distance from the boundary which we will review later, the gravitational effects of the black brane become strong enough that (\[eq:Mthreshold\]) is satisfied, which is when string modes may first be excited.
From the point of view of an effective theory (\[eq:Leff\]) of gravitons, having all the correction terms become the same size (or worse), seems like an hopeless disaster for the purpose of computations. However, the picture of fig. \[fig:stringmany\] suggests a different tack. What is happening is that the 10-dimensional gravitons which make up the classical excitation are really tiny (quantum) loops of string which are getting their internal string degrees of freedom excited as they fall in the background gravitational field. Specifically, internal degrees of freedom of a small object are affected by gravitational [*tidal*]{} forces, which try to compress the object in some directions and stretch it in others. In any case, consider the fate of a single graviton as depicted by fig. \[fig:stringmany\]: a high-momentum object moves through a soft background field. Various authors have previously studied applications of the eikonal approximation to string scattering [@eikonal; @Veneziano]. The upshot, as reviewed below, is that fig. \[fig:stringmany\] may be replaced by the evolution of a single string quantized in the classical background field.
One might simply assert that the right thing to do in the eikonal limit is to quantize a single string in the classical background field, but there is a very nice paper by D’Appollonio, Di Vecchia, Russo, and Veneziano [@Veneziano] that explicitly checks this in a closely related context. As the source of their gravitational field, they take a stack of $N$ coincident Dp-branes in otherwise-flat 10-dimensional space-time. They then probe this gravitational field by scattering a high momentum, massless closed string (such as a graviton) from it. The geometry of the situation is depicted in fig. \[fig:Dpscatt\], which shows the probe particle moving in two of the asymptotically-flat spatial directions perpendicular to the Dp-branes. The particle is deflected by the gravitational field of the Dp-branes. (If these were D0-branes in 4-dimensional space-time, this could be a picture of a particle deflected by the gravitational field of the Sun.) Their problem is slightly different from our problem in that their particle eventually escapes back to infinity, given the geometry of their setup, but never mind that. They further assume that the impact parameter $b$ in fig. \[fig:Dpscatt\] is large enough that $bE \gg 1$. This is the assumption that the background field experienced by the particle is soft, its momentum components $\sim 1/b$ small compared to the particle’s energy. After discussing the eikonal approximation more generally, they then make the following check. Consider the elastic amplitude for the string loop to interact exactly twice with the Dp-branes as it flies by and to emerge in the same massless state at the end. They perform this calculation in two different ways. One way is to do the full string calculation in the presence of the D-branes, as depicted in fig. \[fig:Dpscatt2\]. The other way is to simply quantize a single string loop in the classical gravitational background caused by the D-branes by taking the Penrose limit of the metric near the string trajectory and quantizing the string in the resulting pp-wave background. Then they calculate to second order in that background. In the eikonal limit, they verify that they get exactly the same result with either method. The calculated probability for the string to remain in its massless state drops rapidly below 1 once the kinematic threshold for exciting internal string modes is exceeded.
![ \[fig:Dpscatt\] A high energy massless string mode, such as a graviton, deflected by the gravitational field sourced by a stack of Dp-branes. The plane of the figure is a plane orthogonal to the Dp-branes. (So, for instance, a D1-brane could be visualized as a line extending out of the page.) ](Dpscatt.eps)
![ \[fig:Dpscatt2\] The 2nd-order scattering amplitude for a graviton to elastically scatter from a stack of Dp-branes, calculated as a string scattering amplitude with two connections to the Dp-branes. This is only a topological picture: As described in fig. \[fig:Dpscatt\], the motion of the graviton is orthogonal to the Dp-branes in the problem studied by D’Appollonio et al. [@Veneziano]. ](Dpscatt2.eps)
With this reassurance, we now turn to taking the Penrose limit and quantizing the string in our own problem.
The Penrose Limit {#sec:Penrose}
=================
We begin by taking the Penrose limit to describe a narrow region around the null geodesic (\[eq:geodesic\]), depicted earlier in fig. \[fig:fallpp\].[^7] For an overview of taking Penrose limits, see Refs. [@Blau; @BlauLecture; @BlauFermi].
The null geodesic (\[eq:geodesic\]) can be written as
\[eq:geodesic2\] $$dx^\mu = \frac{g^{\mu\nu} q_\nu}{\omega} \> du ,$$ where $u$ is an affine parameter for the geodesic determined by[^8] $$du = \omega \,
\frac{ \sqrt{g_{\five\five}} \, dx^\five}
{ (-q_\alpha g^{\alpha\beta} q_\beta)^{1/2} }
\,,
\label {eq:du}$$
which can be integrated to give $u$ as a function of $x^\five$. The normalization of $u$ is just convention, and the canceling factors of $\omega=-q_\zero$ in (\[eq:geodesic2\]) have just been chosen to give $u$ dimensions of length. In the metric (\[eq:metric\]), $u$ is given by $$u = \Rads^2 \int \frac{dz}{z^2(1 - f |\q|^2/\omega^2)^{1/2}}
\,,
\label {eq:uz0}$$ which is linearly divergent as $z \to 0$. Our convention will be to define $u$ as running from $u(z{=}0) = -\infty$ at the boundary to $u(z{=}\infty) = 0$ at the black brane singularity, and so $$u(z) = - \Rads^2 \int_z^{\infty} \frac{dz}{z^2(1 - f |\q|^2/\omega^2)^{1/2}}
\,.
\label {eq:uformula}$$ The important thing to remember in what follows is that [*late*]{} times correspond to [*small*]{} negative values of $u$ (which we will later also call $\tau$). Also, though it will be convenient to have defined $u=0$ to be at the singularity $z = \infty$, all of the physics we discuss will only depend on what happens outside of the horizon, $z < \zh$.
We’ll refer to our reference geodesic as $\bar x^\mu(x^\five)$, which we choose to start at the origin $\bar x^\mu = 0$ on the boundary. Now measure 4-positions in AdS$_5$-Schwarzschild relative to this geodesic by defining $$\Delta x^\mu \equiv x^\mu - \bar x^\mu(x^\five) .
\label{eq:deltax}$$ Take $\q$ to be in the $x^\three$ direction. Changing coordinates from $x^\five$ and $\Delta x^\zero$ to $u{=}u(x^\five)$ and $$v \equiv \frac{q_\mu \> \Delta x^\mu}{\omega}
= - \Delta x^\zero + \frac{|\q|}{\omega} \, \Delta x^\three
\label {eq:v}$$ puts the AdS$_5$-Schwarzschild metric (\[eq:metric\]) into the form $$\begin{gathered}
(ds)^2 =
2 \, du \, dv
+ \frac{\Rads^2}{z^2} \Biggl[
(dx^\one)^2 + (dx^\two)^2
+ \frac{(\omega^2 - f |\q|^2)}{\omega^2} \, (d\Delta x^\three)^2
\\
+ 2 f \, \frac{|\q|}{\omega} \, dv \, d\Delta x^\three
- f \, (dv)^2
\Biggr] ,
\label {eq:metricuv}\end{gathered}$$ where $f$ is now implicitly a function of $u$. The Penrose limit consists of keeping only the terms in the metric that would dominate after a scaling of coordinates $$u \to u,
\quad
v \to \gamma^{-2} \, v
\quad
x^i \to \gamma^{-1} x^i
\label{eq:scaling}$$ for very large $\gamma$. This is analogous to making what would be a very large boost ($u \to \gamma u$, $v \to \gamma^{-1} v$) in flat space, and so looking at physics close to the light cone, followed by rescaling all coordinates by a factor of $\gamma^{-1}$. For (\[eq:metricuv\]), the resulting limit is $$(ds)^2_{\rm pp} =
2 \, du \, dv
+ \frac{\Rads^2}{z^2} \Biggl[
(dx^\one)^2 + (dx^\two)^2
+ \frac{(\omega^2 - f |\q|^2)}{\omega^2} \, (d\Delta x^\three)^2
\Biggr] ,
\label {eq:metricpp0}$$ which is a particular example of a pp-wave metric. When we are done with our analysis, we will go back and check that this approximation is justified. The coordinates used in (\[eq:metricpp0\]) are known as Rosen coordinates.
The metric (\[eq:metricpp0\]) has the schematic form $$(ds)^2 = 2 \, du \, dv + \sum_i \kappa_i(u) \, (dy_i)^2 .
\label {eq:metricyhat}$$ It is useful to normalize the last term by switching to Brinkmann coordinates $$\hat y_i \equiv y_i \sqrt{\kappa_i(u)}
\label {eq:yhat}$$ and $$\hat v \equiv
v - \tfrac12 \sum_i \partial_u(\ln \sqrt{\kappa_i}) \, \hat y_i^2
\label {eq:vhat}$$ to give $$(ds)^2 =
2 \, du \, d\hat v
+ \sum_i (d\hat y_i)^2
+ \Bigl(\sum_i \frac{\partial_u^2 \sqrt{\kappa_i}}{\sqrt{\kappa_i}}
\, \hat y_i^2 \Bigr) (du)^2
\,.$$ In our case, using (\[eq:uz0\]) to rewrite $$\partial_u =
\frac{z^2}{\Rads^2} \, \Bigl(1-f\frac{|\q|^2}{\omega^2}\Bigr)^{1/2} \partial_z ,
\label {eq:partialu}$$ the metric in Brinkmann coordinates is $$ds^2_{\rm pp} = 2 \, du \, d\hat v
+ (d\hat x^\one)^2 + (d\hat x^\two)^2 + (d\Delta\hat x^\three)^2
+ {\cal G}(u,\hat x^\one, \hat x^\two, \Delta\hat x^\three) \,
(du)^2
\label {eq:ppmetrichat}$$ with
\[eq:Gdefs\] $${\cal G}(u,\hat x^\one,\hat x^\two,\Delta\hat x^\three)
= {\cal G}_1(u) \, \bigl[(\hat x^\one)^2 + (\hat x^\two)^2\bigr]
+ {\cal G}_3(u) \, (\Delta\hat x^\three)^2 ,$$ $${\cal G}_1(u) = {\cal G}_2(u)
= \frac{\partial_u^2(z^{-1})}{z^{-1}}
= \frac{z^3 f' |\q|^2}{2 \Rads^4\omega^2}
= - 2 \, \frac{z^6 |\q|^2}{\zh^4 \Rads^4\omega^2}
\simeq - 2 \, \frac{z^6}{\zh^4 \Rads^4}
\,,
\label {eq:calG12}$$ $${\cal G}_3(u)
= \frac{\partial_u^2\bigl[z^{-1}(\omega^2-f|\q|^2)^{1/2}\bigr]}
{z^{-1}(\omega^2-f|\q|^2\bigr)^{1/2}}
= \frac{z^3 (f'-zf'') |\q|^2}{2 \Rads^4 \omega^2}
= 4 \frac{z^6 |\q|^2}{\zh^4 \Rads^4 \omega^2}
\simeq 4 \frac{z^6}{\zh^4 \Rads^4}
\,.
\label {eq:calG3}$$
Here, primes denote derivatives with respect to $z$, and $z{=}z(u)$ is implicitly a function of $u$, determined by inverting (\[eq:uformula\]). The metric (\[eq:ppmetrichat\]) would be flat if not for the ${\cal G}_i$, which arise from tidal forces. Note that these tidal terms vanish for null geodesics in pure AdS ($f{=}1$, or equivalently $\zh{\to}\infty$), in agreement with general arguments that the Penrose limit of AdS is flat Minkowski space [@Blau]. They also vanish in AdS$_5$-Schwarzschild if $\q{=}0$ (i.e. if the excitation fell straight down in the $x^\five$ direction as a function of time $x^\zero$).
Quantizing our falling string loop {#sec:quantize}
==================================
Overview
--------
We may now follow other authors to quantize a string in a pp-wave background [@Veneziano; @PRT; @HS; @GPZS; @dVS]. The bosonic sector is described by a $\sigma$-model in the pp-wave background metric: $$S =
- \frac{1}{4\pi\alpha'} \int d\tau \> d\sigma
\sqrt{-\gamma} \, \gamma^{\alpha\beta}
(\partial_\alpha X^I)(\partial_\beta X^J) \,
g_{IJ}(X) ,$$ where $\gamma$ is the world-sheet metric and $X^I$ are the world-sheet fields corresponding to the coordinates. For the pp-wave space-time metric (\[eq:ppmetrichat\]), this takes the form $$S = S_0
- \frac{1}{4\pi\alpha'} \int d\tau \> d\sigma
\sqrt{-\gamma} \, \gamma^{\alpha\beta}
(\partial_\alpha U)(\partial_\beta U) \,
{\cal G}(U,\Delta \hat {\bm X}) ,
\label {eq:action}$$ where $S_0$ is the Minkowski string action. Identifying world-sheet time $\tau$ with the affine parameter $u$ for the pp-wave space-time metric (\[eq:ppmetrichat\]) then gives a constraint equation for $\partial_\alpha\hat V$ and gives the light-cone gauge Lagrangian $$\begin{aligned}
L &=
\frac{p^u}{2} \int_0^{2\pi} \frac{d\sigma}{2\pi} \sum_i
\left(
\partial_\tau \Delta \hat X^i \, \partial_\tau \Delta \hat X^i
- (\alpha'p^u)^{-2}
\partial_\sigma \Delta \hat X^i \,
\partial_\sigma \Delta \hat X^i
+ {\cal G}_i(\tau) \, \Delta \hat X^i \, \Delta\hat X^i
\right)
\nonumber\\
&=
\frac{p^u}{2} \sum_i \sum_{n=-\infty}^{\infty}
\left(
\partial_\tau \Delta \hat{X^i_n}^* \, \partial_\tau \Delta \hat X^i_n
- \omega_{i,n}^2(\tau) \,
\Delta \hat{X^i_n}^* \, \Delta\hat X^i_n
\right)
\label {eq:Lin}\end{aligned}$$ for the $\Delta\hat X^i$, where $p^u = p_{\hat v} \simeq E$ and $\Delta \hat X^i = \sum_n \Delta \hat X^i_n(\tau) \, e^{in\sigma}$ and $$\omega_{i,n}^2(\tau) \equiv
\frac{n^2}{(\alpha' p^u)^2} - {\cal G}_i(\tau) .
\label {eq:win}$$ (We have suppressed the fields corresponding to $S^5$ coordinates because they will play no role in our discussion.) Note that we have chosen a convention where $\tau$ has units of time and $\sigma$ is dimensionless.[^9]
Each mode $\Delta\hat X^i_n$ of the string is a time-dependent harmonic oscillator problem with classical frequency $\omega_{i,n}(\tau)$. The $\Delta\hat X^\one_n$ and $\Delta\hat X^\two_n$ modes are tidally compressed as the string moves away from the boundary since ${\cal G}_1 = {\cal G}_2$ is negative in (\[eq:calG12\]), so that the curvature $\omega_{1,n}(\tau)$ of the harmonic oscillator potential increases with time $\tau{=}u$. In contrast, the $\Delta\hat X^\three_n$ oscillators are tidally stretched, since ${\cal G}_3$ is positive in (\[eq:calG3\]). We will focus on these $\Delta\hat X^\three$ degrees of freedom and so will focus on $$\omega_{3,n}^2(\tau) =
\frac{n^2}{(\alpha' p^u)^2} - {\cal G}_3(\tau) .
\label {eq:w3n}$$ When the string gets far enough from the boundary (i.e. at late enough times $\tau$), the ${\cal G}_\three$ term in (\[eq:w3n\]) becomes dominant and $\Delta \hat X^\three_n$ oscillators become unstable. Physically, this is when tidal forces come to dominate over string tension. Using (\[eq:calG3\]), this instability occurs when $z \ge z_n$, where $$z_n
\simeq \left( \frac{n \zh^2 \Rads^2}{2 \alpha' E} \right)^{1/3}
= \lambda^{1/6} \left( \frac{n \zh^2}{2 E} \right)^{1/3}
= n^{1/3} z_1 .$$ The instability for the center-of-mass mode $n{=}0$ is not particularly interesting: it has nothing to do with exciting internal degrees of freedom of the string and just reflects the slight spread of the falling wavepacket in fig. \[fig:fall\] due to curvature effects. That is, it reflects physics already incorporated in earlier work [@adsjet2] on our jet stopping problem and has nothing to do with gravitons being tidally stretched into classical strings. So, in what follows, we will focus on the $n\not=0$ modes, and the first tidal instability kicks in at $z \simeq z_1$.
Recall that $z_\star$, defined in (\[eq:zstar\]), characterizes the scale where the $x^\three$ motion of the bulk excitation in fig. \[fig:fall\] is coming to a stop (no significant motion for $z \gg z_\star$). How does the instability scale $z_n$ above compare to $z_\star$? Using (\[eq:stop\]) and (\[eq:lmax\]), the ratio is $$\frac{z_n}{z_\star}
\sim \frac{n^{1/3} \lambda^{1/6} (\zh^2/E)^{1/3}}{\zh (-q^2/E^2)^{1/4}}
\sim \frac{n^{1/3} \lstop}{\lambda^{-1/6} \lmax} \,.$$ So the tidal instability kicks in before the stopping distance is reached ($z_1 \lesssim z_\star$) precisely when we are in the interesting regime $\lstop \lesssim \lambda^{-1/6} \lmax$ (\[eq:fun\]) identified in earlier work as the case where stringy corrections become important. This is the case we focus on. Correspondingly, the modes which become tidally unstable for $z \lesssim z_\star$ are $n \lesssim n_\star$ with $$n_\star \sim \left(\frac{\lstop}{\lambda^{-1/6} \lmax}\right)^{-3}
.
\label {eq:nstar}$$
Although the instability develops at $z = z_n$, we will see later that the modes $n \lesssim n_\star$ do not have time to stretch significantly until $z \sim z_\star$. This fact is suggested by the geodesic picture in fig. \[fig:boost\]: For $z\ll z_\star$, the impact of the black hole on the evolution in fig. \[fig:boost\]c is negligible, and so the evolution at those times is well approximated by the pure AdS case of figs.\[fig:boost\]a and b, for which the proper size of the string remains constant.
Once a given mode becomes unstable, the quantum mechanics of that mode will be somewhat analogous to a standard quantum mechanics thought experiment: What is the longest time that an idealized pencil can be balanced on its tip before it falls? Because of the uncertainty principle, the pencil cannot be started simultaneously at rest and perfectly vertical, and so it must fall. The pencil might be started in a Gaussian wavepacket chosen to maximize the average fall time. But it is clear that, once the top of the pencil has fallen a macroscopic distance, [*classical*]{} mechanics will suffice to describe its subsequent motion: at that time, its position and momentum may, to excellent approximation, be considered as simultaneously well defined. For late enough times $t_1$, the pencil’s motion for $t > t_1$ is approximately classical, with the only effect of its initial wave packet at $t{=}0$ being to determine a classical probability distribution for the pencil end’s position at $t_1$. The corresponding momentum at $t_1$ is given (to excellent approximation) by the momentum the pencil would have picked up falling classically to that position from vertical.
In our problem, we will see that, for the string modes $n \lesssim n_\star$ of interest, the analogous transformation from a quantum description to a probability distribution for classical configurations occurs when $z \gg z_\star$.
Solution of the time-dependent harmonic oscillators
---------------------------------------------------
### Basics
The distinction between $\Delta X^i$ (the difference between $X^i$ and the reference geodesic) and $X^i$ does not affect the $n \not= 0$ modes that are our focus. Similarly for the normalized coordinates $\Delta \hat X^i$. So we will make our notation a little less cumbersome and henceforth write $\Delta\hat X^i_n$ as simply $\hat X^i_n$ (for $n\not=0$).
Each of the real degrees of freedom $\sqrt2 \Re \hat X^i_n$ and $\sqrt2 \Im \hat X^i_n$ in (\[eq:Lin\]) have a harmonic oscillator Lagrangian of the form $$L = \tfrac12 \, m \bigl( \dot\dof^2 - \omega^2(\tau) \, \dof^2 \bigr) ,
\label {eq:Lho}$$ with the translation $m \to p^u \simeq E$ and $\omega^2(\tau) \to \omega_{i,n}^2(\tau)$. The squared frequency $\omega^2(\tau)$ starts at a non-zero value $\omega^2(-\infty)$ and then changes with time $\tau$. The quantum mechanical solution to such time-dependent harmonic oscillator problems has a long history. Useful explicit formulas for wave functions may be found in ref. [@KimLee], with applications to strings in pp-wave backgrounds in refs. [@PRT; @GPZS].[^10] In our case, the harmonic oscillators all start in their ground state (the string state describing a graviton) at early times ($\tau \to -\infty$) and so start with Gaussian wave functions. For a time-dependent harmonic oscillator that starts as a Gaussian $\psi(\dof) \propto \exp[-\dof^2/4\sigma^2]$ at some time $\tau_0$, one may check that the Schrödinger equation $$i \dot\psi
= \Bigl[ - \frac{1}{2m} \, \partial_\dof^2
+ \frac12 \, m \, \omega^2(\tau) \, \dof^2 \Bigr] \psi$$ is solved by $$\psi(\dof,\tau) \propto
\frac{1}{\sqrt{\chi(t)}} \,
\exp\left[
\frac{i}{2} \, \frac{\dot\chi(t)}{\chi(t)} \, m \dof^2
\right] ,
\label {eq:psi}$$ where the complex-valued function $\chi(\tau)$ satisfies the classical equation of motion $$\ddot \chi = - \omega^2(\tau) \, \chi
\label {eq:chi}$$ with initial conditions $$\chi(\tau_0) = 1,
\qquad
\dot\chi(\tau_0) = \frac{i}{2m\sigma^2} \,.$$ In our case, where we start in the early-time ground state, that’s $$\chi(\tau_0) = 1,
\qquad
\dot\chi(\tau_0) = i \, \omega(-\infty)
\label {eq:chii}$$ with $\tau_0 \to -\infty$.
It will be useful to have an expression for the corresponding probability distribution $|\psi(\dof,\tau)|^2$ for $\dof$. From (\[eq:psi\]), this is just a Gaussian distribution $$\operatorname{Prob}(\dof,\tau) =
\frac{e^{-\dof^2/2\dof_{\rm rms}^2(\tau)}}{(2\pi)^{1/2} \, \dof_{\rm rms}(\tau)}$$ with width $$\dof_{\rm rms} = \left[ 2m \Im \frac{\dot\chi}{\chi} \right]^{-1/2}
= \left[ \frac{m(\chi^*\dot\chi-\dot\chi^*\chi)}{i|\chi|^2} \right]^{-1/2}
.$$ But $\chi^*\dot\chi-\dot\chi^*\chi$ is a Wronskian of the two solutions $\chi$ and $\chi^*$ of (\[eq:chi\]) and so is time-independent and may be evaluated at $\tau{=}\tau_0$ using (\[eq:chii\]), giving $$\dof_{\rm rms}(\tau)
= \frac{|\chi(\tau)|}{\sqrt{2m\,\omega(-\infty)}}
= |\chi(\tau)| \, \dof_{\rm rms}(-\infty) .
\label {eq:xrms}$$ Using (\[eq:w3n\]), $x_{\rm rms}(-\infty)$ corresponds to $\sqrt{\alpha'/2n}$ in our application.
Our remaining task is to solve the classical equation of motion (\[eq:chi\]) for $\chi$. For the case of interest $\hat X^\three_n$, using (\[eq:zstar\]), (\[eq:partialu\]), (\[eq:calG3\]), (\[eq:w3n\]), and $\omega \simeq |\q| \simeq E$, the $\chi$ equation may be put into the form
\[eq:diffeqs\] $$\begin{aligned}
\frac{d^2\chi}{d\bar\tau^2}
&= -4(\xi^6 - \bar z^6) \chi ,
\label {eq:chixi}
\\
\frac{d\bar z}{d\bar\tau} \> &= \bar z^2 (1+\bar z^4)^{1/2} ,
\label {eq:zbar}\end{aligned}$$
where $$\begin{aligned}
\bar z &\equiv \frac{z}{z_\star} ,
\\
\bar\tau &\equiv \frac{z_\star^3}{\zh^2 \Rads^2} \, \tau ,
\\
\xi &= \xi_n \equiv n^{1/3} \xi_1 \equiv
\left(\frac{n}{2}\right)^{1/3}
\frac{\zh^2/z_\star}{\lambda^{-1/6} \zh^{4/3} E^{1/3}} .
\label {eq:xidef}\end{aligned}$$ In these variables, the initial conditions (\[eq:chii\]) on $\chi$ are $$\chi(\bar\tau_0) = 1,
\qquad
\frac{d\chi}{d\bar\tau}(\bar\tau_0) = 2i\xi^3 \,.
\label {eq:chiinit}$$ Note from (\[eq:lmax\]) and (\[eq:stop\]) that, parametrically, $$\xi_1 \sim \frac{\lstop}{\lambda^{-1/6}\lmax} ,$$ and so the smallness of $\xi_1$ is a specific measure of how far we are into the interesting regime (\[eq:fun\]) of $\lstop \lesssim \lambda^{-1/6}\lmax$. The modes $n \lesssim n_\star$ of interest to us correspond to $\xi_n \lesssim 1$.
Using the above equations, one may now check our earlier claim that the string does not stretch significantly in proper size at early times $z \ll z_\star$ ($\bar z \ll 1$).[^11] But we are more interested in what the string does at late times, on which we now focus.
### Late-time behavior {#sec:late}
For $z \gg z_\star$ (which is $\bar z \gg 1$), the $\bar z$ equation (\[eq:zbar\]) gives $$\bar z = (-3 \bar\tau)^{-1/3} ,
\label {eq:zbarlate}$$ remembering that our convention is that $\tau$ is negative and that $\tau(z{=}\infty)=0$. Note from (\[eq:zbarlate\]) that $-\bar\tau$ is very small at the horizon $z = \zh$, where $-\bar\tau \sim (\zh/z_\star)^{-3} \sim (-q^2/E^2)^{3/4} \sim (\lstop T)^{-3}$. Throughout this paper we assume that $-q^2 \ll E^2$ and so $\lstop \gg T^{-1}$.
Plugging (\[eq:zbarlate\]) into the $\chi$ equation (\[eq:chixi\]) yields late-time ($\bar z \gg 1$) solutions $\chi \propto (-\bar\tau)^{-1/3}$ and $\chi \propto (-\bar\tau)^{4/3}$. The dominant solution will be $$\chi \propto (-\bar\tau)^{-1/3} .
\label {eq:chilate}$$
Though $\chi$ is a complex-valued function whose purpose is to track the evolution of the wavepacket, exactly the same arguments as above give that a classical trajectory would have late-time behavior $\dof \propto (-\bar\tau)^{-1/3}$. That means that $\dot\dof \propto (-\bar\tau)^{-4/3}$, and so $\dof$ and $\dot\dof$ both become large at late times, justifying a classical description at late times.[^12] The classical relation between the two is determined by $\dof \propto (-\bar\tau)^{-1/3}$ to be $$\frac{\dot x}{x} \simeq \frac{1}{-3\bar\tau} \simeq \bar z^3 .
\label {eq:xdotclassical}$$
We may extract the proportionality constant in the late-time behavior (\[eq:chilate\]) by solving (\[eq:diffeqs\]) numerically with initial conditions (\[eq:chiinit\]) for $\chi$, matching the late time behavior of the numerical solution to (\[eq:chilate\]), and repeating the calculation for earlier and earlier values of $\tau_0$ in order to take the $\tau_0 \to -\infty$ limit. Our result is that the late-time behavior is $$|\chi(\bar\tau)| \to \frac{C(\xi)}{(-\bar\tau)^{1/3}}
\label {eq:Cdef}$$ with $C(\xi)$ given by fig. \[fig:Cxi\].[^13] We show in Appendix \[app:Cxibig\] that the limiting behavior for large $\xi$ is
\[eq:Cxilimits\] $$C(\xi) \simeq \frac{\Gamma(\tfrac56)}{\pi^{1/2}\xi}
\qquad
\mbox{for}~\xi \gg 1 ,
\label {eq:Clargexi}$$ shown as a dashed curve in the figure. In the opposite limit of $\xi \ll 1$, our numerical results approach a constant $$C(\xi) \simeq 0.6428
\qquad
\mbox{for}~\xi \ll 1.
\label {eq:C0}$$
![ \[fig:Cxi\] The proportionality constant $C(\xi)$ in (\[eq:Cdef\]), which determines the late-time width of the probability distribution for the amplitude of a string mode. The sloping dashed curve shows the large-$\xi$ approximation (\[eq:Clargexi\]). ](Cxi.eps)
From (\[eq:w3n\]), (\[eq:xrms\]) and (\[eq:Cdef\]), and remembering that the analogs of $x$ are $\sqrt{2} \Re X^i_n$ and $\sqrt{2} \Im X^i_n$, the late time probability distribution of mode amplitudes $\hat X^\three_n$ is given by a Gaussian with width $$\bigl|\hat X^\three_n\bigr|_{\rm rms}
\simeq
\frac{C(\xi_n)}{(-\bar\tau)^{1/3}}
\left(\frac{\alpha'}{2n}\right)^{1/2}
\qquad
\mbox{for}~{-}\bar\tau \ll 1.$$ Using (\[eq:zbarlate\]), that may be rewritten as $$\bigl|\hat X^\three_n\bigr|_{\rm rms}
\simeq
3^{1/3} \, C(\xi_n) \, \frac{z}{z_\star}
\left(\frac{\alpha'}{2n}\right)^{1/2}
\qquad
\mbox{for}~ z \gg z_*,\xi_n.
\label {eq:Xhatrms}$$ As in (\[eq:xdotclassical\]), the corresponding momenta in this classical regime are related to the mode amplitudes $\hat X^\three_n$ by $$\partial_{\bar\tau} \hat X^\three_n
\simeq \frac{\hat X^\three_n}{-3\bar\tau}
\simeq \frac{z^3}{z_\star^3} \, \hat X^\three_n \,.$$
Using (\[eq:metricpp0\]) and (\[eq:yhat\]), the conversion between the normalized coordinate $\Delta \hat x^\three$ and the displacement $\Delta x^\three$ from the reference geodesic in Poincare coordinates is $$\Delta x^\three =
% \frac{z}{\Rads} \left( \frac{-q^2}{|\q|^2} + \frac{z^4}{\zh^4} \right)^{-1/2}
\frac{z}{\Rads} \left( 1 - f \frac{|\q|^2}{\omega^2} \right)^{-1/2}
\simeq
\frac{z}{\Rads} \left( \frac{z_\star^4 + z^4}{\zh^4} \right)^{-1/2}
\Delta\hat x^\three ,$$ which is $$\Delta x^\three \simeq
\frac{\zh^2}{z \Rads} \, \Delta\hat x^\three
\qquad \mbox{for}~z \gg z_\star .$$ So, from (\[eq:Xhatrms\]), the amplitudes of the stretched modes in the Poincare coordinate system are $$\bigl|X^\three_n\bigr|_{\rm rms}
\simeq
3^{1/3} \, C(\xi_n) \, \frac{\zh^2}{z_\star \Rads}
\left(\frac{\alpha'}{2n}\right)^{1/2}
=
\frac{3^{1/3} \, C(\xi_n) \, \zh^2}{(2n)^{1/2} \lambda^{1/4} z_\star}$$ for fixed $\tau$ (and so fixed $z$) in the classical regime. Using (\[eq:stop\]) and (\[eq:zstar\]), this may be written as $$\bigl|X^\three_n\bigr|_{\rm rms}
\simeq
\frac{3^{1/3} (8\pi)^{1/2}\, C(\xi_n)}{n^{1/2}\,\Gamma^2(\tfrac14)}
\, \lambda^{-1/4} \lstop
\label {eq:Xrms}$$ for $z \gg z_\star,\xi_n$.
Note that fixed-$\tau$ (i.e. fixed-$z$) slices of the string worldsheet look different than fixed-$x^0$ slices of the string worldsheet, which is why our depiction of the string at various times $x^0$ in fig. \[fig:boost\]c were not horizontal.
The size of the string at late times
------------------------------------
Parametrically, the average size (\[eq:Xrms\]) of each stretched mode $X^\three_n$ in Poincare coordinates is just the $\lambda^{-1/4} \lstop$ that we found in our back-of-the-envelope argument in section \[sec:envelope\]. But we should now look at what happens if we sum up all the modes to get the total average size of the string. A convenient measure of the size scale of the string is the average rms deviation from the center of the string, $$(\delta X^\three)_{\rm rms}
\equiv \biggl\langle\>
\overline{\bigl(
X^\three - \overline{X^\three} \,
\bigr)^2 \,}
\>\biggr\rangle^{1/2} ,$$ where overlines indicate averaging over the string worldsheet position $\sigma$ and the angle brackets indicate averaging over the late-time classical probability distribution for each mode amplitude. This is given by $$(\delta X^\three)_{\rm rms}
= \biggl(
2 \sum_{n=1}^\infty \bigl\langle \bigl|X^\three_n\bigr|^2 \bigr\rangle
\biggr)^{1/2} .
\label {eq:size0}$$ $\bigl\langle \bigl|X^\three_n\bigr|^2 \bigr\rangle$ is just the square of what we called $\bigl|X^\three_n\bigr|_{\rm rms}$ in (\[eq:Xrms\]). Combining the limiting forms (\[eq:Cxilimits\]) with (\[eq:Xrms\]), and recalling from (\[eq:xidef\]) that $\xi_n = n^{1/3} \xi_1$, $$|X^\three_n\bigr|^2_{\rm rms} \simeq
\left( \frac{3^{1/3} (8\pi)^{1/2}}{\Gamma^2(\tfrac14)}
\, \lambda^{-1/4} \lstop \right)^2 \times
\begin {cases}
\frac{[C(0)]^2}{n} , & n \ll n_\star ; \\
\frac{\Gamma^2(\tfrac56)}{\pi \xi_1^2 n^{5/3}} , & n \gg n_\star ,
\end {cases}
\label {eq:Xcases}$$ where $C(0)$ is given by (\[eq:C0\]). The sum in (\[eq:size0\]) is therefore convergent at large $n$ (more on that in a moment) and is dominated by a logarithm coming from $n=1$ up to $n \sim n_\star$. At leading order in inverse powers of that logarithm, $$(\delta X^\three)_{\rm rms}
\simeq |X^\three_1|_{\rm rms} \sqrt{2 \ln n_\star}
\simeq
\frac{3^{1/3} 4\sqrt\pi \, C(0)}{\Gamma^2(\tfrac14)}
\, \lambda^{-1/4} \lstop \sqrt{\ln n_\star} .
\label {eq:result0}$$ Using (\[eq:nstar\]), this may be rewritten as[^14] $$(\delta X^\three)_{\rm rms}
\simeq
0.8660 \, \lambda^{-1/4} \lstop \,
\ln^{1/2}\biggl(\frac{\lambda^{-1/6} \lmax}{\lstop}\biggr) ,
\label {eq:result}$$ where a parametric expression for the argument of the logarithm is adequate if we are only keeping track of the logarithmic term.
Ignoring the numerical constant in front, (\[eq:result\]) is the parametric result (\[eq:ratio\]) that we presented in the introduction. One could go on to evaluate the non-logarithmic corrections to (\[eq:result\]), but (\[eq:result\]) is good enough for our present purposes. We are mostly interested in the parametric size of the answer, so that we can determine whether the extent of the string in $x^\three$ remains small compared to the stopping distance $\lstop$ in the case of large but finite $\lambda$.
Qualitatively, what is the origin of the logarithmic factor in (\[eq:result0\]) and (\[eq:result\])? Imagine doing the same calculation of $(\delta X^\three)_{\rm rms}$ for the graviton instead of for the stretched, classical string. In the ground state, all the modes $X^i_n$ have an rms size proportional to $n^{-1/2}$ and so the sum $\sum_n \langle |X^i_n|^2 \rangle$ is log divergent in the ultraviolet. This does not mean, however, that the energy or momentum carried by the graviton is smeared over infinite spatial distances:[^15] in the string’s ground state, the bosonic $n\not=0$ mode contributions to energy and momentum are canceled by the fermionic mode contributions, which we have ignored in our analysis. Only the bosonic mode amplitudes, however, can become “big” due to tidal stretching, and it is the modes whose amplitudes have grown big that we consider when we make the classical approximation at late times. Our expression (\[eq:Xcases\]) is only valid until $n$ gets so large that the bosonic mode has not been significantly excited ($\xi_n \sim \bar z$, and so $n \sim n_\star (z/z_\star)^3 \gg n_\star$ when $z \gg z_\star$). For yet larger $n$ the modes will be in their ground state and the canceling contribution of fermionic modes to physical quantities of interest will come into play. The contribution of such high modes is sub-leading and is simply discarded when we approximate the string as a classical string at late times. In the resulting treatment of the classical string, there is no ambiguity in what the size of the string means, as highlighted by the convergence of the sum of (\[eq:Xcases\]).[^16]
The upshot is that the logarithm in (\[eq:result0\]) arises because of (i) the logarithmic UV divergence associated with the bosonic modes in the ground state, combined with (ii) the fact that the bosonic mode amplitudes with $n \ll n_\star$ all grow by an equal large factor from tidal stretching (and no longer cancel against fermionic modes in their physical consequences), while those with $n \gg n_\star$ do not grow significantly in comparison.
Discussion of the single graviton approximation {#sec:gravitons}
===============================================
We have followed the evolution of a single graviton as it is stretched into a classical string loop. We will now take a moment to discuss in more detail the premise that we may follow a single graviton.
First, a single high-momentum string loop might split into two, but such splitting is $1/\Nc$ suppressed and so may be ignored in the $\Nc{=}\infty$ limit that we take in this paper.
The same argument might be given against the possibility of two gravitons merging if not for the fact that the localized gravitational wavepacket describing our excitation in the bulk is classical, which means it contains a correspondingly [*large*]{} number of gravitons. We need a different argument for why interactions between the gravitons that make up the classical wavepacket may be ignored. For this, we need to review in slightly more detail the formalism of refs. [@adsjet; @adsjet2] for creating the excitation in the first place. In that formalism, the initial photon or W boson or whatever of fig. \[fig:Wdecay\] is replaced by a localized, external classical field. Specifically, we add a source term to the Lagrangian, $${\cal L} \to
{\cal L} + {\cal N} \, O(x) \, e^{i\bar k\cdot x} \, \Lambda_L(x) ,
\label {eq:source}$$ where ${\cal N}$ is an arbitrarily small source amplitude, $O(x)$ is a source operator (corresponding to the vertex in fig. \[fig:Wdecay\]), $$\bar k^\mu \simeq (E,0,0,E)$$ is the large 4-momentum of the desired excitation, and $\Lambda_L(x)$ is a slowly varying envelope function that localizes the source near the origin in both $x^\three$ and time. The amplitude ${\cal N}$ is chosen to be small so that we can treat the external source as a small perturbation to the strongly-interacting gauge theory, so that the source will never create more than one jet with energy $E$ at a time. On an event-by-event basis, the source will usually do nothing at all, but on rare occasions ($\propto |{\cal N}|^2$), it will create an excitation with energy $\simeq E$. For small enough ${\cal N}$, it will essentially never create excitations with energies $2E$, $3E$, etc. [@adsjet2]. But taking ${\cal N}$ small also means that the bulk excitation created by the source on the boundary can be treated in linear response: the self-interactions of the bulk excitation with itself are ignorable. For this reason, we may ignore interactions between the gravitons (or other type of particles) that make up the bulk excitation.
Some readers may wonder how we can focus on gravitons, which are quantum mechanical objects, when the dual theory for $\Nc{=}\infty$ is supposed to be a classical theory. As an analogy, consider a classical electromagnetic wave with polarization $({\bm e}_x+{\bm e}_y)/\sqrt2$. If we choose to, we may think of this classical wave as a coherent superposition of photons which are in the quantum state $\bigl(|x\rangle + |y\rangle\bigr)/\sqrt2$, where $|x\rangle$ and $|y\rangle$ are states corresponding to polarization in the $x$ and $y$ directions. Now measure the polarization by putting the wave through a linear polarizer oriented in the $x$ direction. As we all know, we may view this classically as picking out the ${\bm e}_x/\sqrt2$ component of the wave, or alternatively we may view it in terms of photons as saying that each photon has probability $1/2$ of being $x$-polarized. A discussion of classical physics in one description is equivalent to a quantum-mechanical discussion of probabilities for the behavior of individual quanta in the other description.
The same reasoning applies to our description of the classical bulk wavepacket in terms of individual gravitons. An important conceptual lesson from this is that our wavepacket is not a single graviton that evolves into a single classical string loop with probability distribution given by a Gaussian with size (\[eq:result\]) for each mode $n$. Instead, the wavepacket is a large number of gravitons that independently evolve into a large number of classical string loops that independently have that probability distribution. Following the analogy with photon polarizations further, one could presumably replace the quantum description involving probability amplitudes for the degrees of freedom of individual string loops by a classical description in terms of a classical string field theory, promoting the Schrödinger wave functional $\psi[{X_n^i}]$ for a single string to a classical wave functional $\psi[{X_n^i}]$ that could be used to describe all the physics discussed in this paper. We have not pursued this latter approach because we thought that the description in terms of gravitons was simpler, more intuitive, and more directly related to the previous literature on tidal excitation.
Checking the Penrose limit {#sec:PenroseCheck}
==========================
Now that we have our final answer (\[eq:result\]) for the size of the classical string that is produced by stretching, we should go back and verify that the string is not so big, or so far away from the reference geodesic, that the Penrose limit taken in section \[sec:Penrose\] breaks down. We need to check that the $dv\>d\Delta x^\three$ and $(dv)^2$ terms in the AdS$_5$-Schwarzschild metric (\[eq:metricuv\]), which were dropped in the Penrose limit, are parametrically small compared to the $du\>dv$ term, $$\frac{f\Rads^2}{z^2} \frac{|\q|}{\omega} \> |dv \> d\Delta x^\three|
\quad \mbox{and} \quad
\frac{f\Rads^2}{z^2} \> (dv)^2
\quad \ll \quad
|du\>dv| ,$$ for the string motions that we have found. Dividing both sides by $|du\>dv|$ and using $\omega \simeq |\q|$, we may rewrite these conditions as $$\frac{f\Rads^2}{z^2} \left|\frac{d\Delta x^\three}{du}\right|
\quad \mbox{and} \quad
\frac{f\Rads^2}{z^2} \left|\frac{dv}{du} \right|
\quad \ll \quad
1 .
\label {eq:PenroseConstraints}$$ We check these conditions on the string motion in appendix \[app:PenroseCheck\], where we find that the condition on $|d\Delta x^\three/du|$ is the strongest and requires $$\lambda^{-1/4} \sqrt{\ln n_\star} \ll 1
\label {eq:PenroseFinal0}$$ in order for our earlier analysis to be valid. Using our result (\[eq:result0\]), this condition may be written as $$(\delta X^\three)_{\rm rms} \ll \lstop .
\label {eq:PenroseFinal}$$ That is, the Penrose limit only breaks down if one considers the extreme case (to be discussed in a moment) where the string becomes as large as the stopping distance itself.
Conclusion {#sec:conclusion}
==========
In our scheme for creating “jets,” we have seen different behaviors in the dual theory depending on the virtuality (and so the stopping distance) of the jet. For $\lstop \gg \lambda^{-1/6}\lmax$, the gravitons (or other massless string modes) composing the excitation in the gravity description remain gravitons until after the excitation has stopped moving in the $x^\three$ direction, and there is no difficulty in using the supergravity approximation for the calculation. For $\lstop \ll \lambda^{-1/6}\lmax$, each graviton is instead stretched into a classical string loop. However, provided that $$\lambda^{-1/4} \, \ln^{1/2}\biggl(\frac{\lambda^{-1/6} \lmax}{\lstop}\biggr)
\ll 1 ,$$ the size of that string remains small compared to the stopping distance $\lstop$. The string remains close to its reference geodesic, and so corrections to the $\lambda{=}\infty$ description of the jet stopping are parametrically small (if one only attempts to resolve details on size scales large compared to the size of the string). However, if instead $$\lambda^{-1/4} \, \ln^{1/2}\biggl(\frac{\lambda^{-1/6} \lmax}{\lstop}\biggr)
\gtrsim 1 ,
\label {eq:biglog}$$ then the string loop will stretch out to be parametrically as large as the stopping distance itself. Our quantum analysis of the string breaks down in this case (because of the failure of the Penrose limit), but we can see what happens qualitatively by tracking what happens to the $\lambda^{-1/4} \log^{1/2} \ll 1$ results as we increase the logarithm towards $\lambda^{-1/4} \log^{1/2} \sim 1$.
In particular, a nice way to visualize what happens is to follow the classical evolution of a closed string that initially starts with a proper size $\Sigma$ of order $\sqrt{\alpha' \ln n_\star}$, which is roughly the initial rms size from the modes $n \lesssim n_\star$ which become classically excited. Increasing the logarithm towards $\lambda^{-1/4} \sqrt{\ln n\star} \sim 1$ is equivalent to increasing $\Sigma$ towards $\sim \Rads$. Fig. \[fig:string\] compares examples of such evolution for the cases (a) $\lambda^{-1/4} \log^{1/2} \ll 1$ and (b) $\lambda^{-1/4} \log^{1/2} \sim 1$. (More details of exactly how we initialized our classical string may be found in appendix \[app:string\].) The interesting feature of fig. \[fig:string\]b is that, at late times, the string looks like the original picture advocated by Gubser et al. [@GubserGluon] of gluon jets as dual to the evolution of a trailing, folded classical string falling in AdS$_5$-Schwarzschild. Our string is a folded closed string, as depicted in the cartoon of fig. \[fig:fold\]a, whereas the one studied by Gubser et al. was a folded infinite open string, as depicted by fig. \[fig:fold\]b.[^17] However, the left end of the string in these figures, which is very close to the horizon, does not play a significant role in the effect on the boundary theory, and so the physics of these two situations is approximately the same.
![ \[fig:string\] Examples of numerical solutions of the evolution of a falling classical string loop that starts near the boundary with proper size (a) $\Sigma \ll \Rads$ and (b) $\Sigma \sim \Rads$. These are snapshots of the string at fixed $x^\zero$. See appendix \[app:string\] for details of the initial condition. ](string1.eps "fig:") ![ \[fig:string\] Examples of numerical solutions of the evolution of a falling classical string loop that starts near the boundary with proper size (a) $\Sigma \ll \Rads$ and (b) $\Sigma \sim \Rads$. These are snapshots of the string at fixed $x^\zero$. See appendix \[app:string\] for details of the initial condition. ](string2.eps "fig:")
![ \[fig:fold\] Schematic pictures of classical folded strings. (a) A closed folded string produced by extreme tidal stretching of a graviton in our method of generating “jets” in the case of $\lambda^{-1/4} \log^{1/2} \gtrsim 1$ (\[eq:biglog\]), and (b) the infinite, folded open string considered by Gubser et al. [@GubserGluon]. In the latter case, the trailing string continues to get closer and closer to the horizon as $x^\three \to -\infty$. ](fold.eps)
Historically, the original motivation of our own method for posing “jet” stopping problems [@adsjet], outlined in the introduction to this paper and motivated by fig. \[fig:Wdecay\], was to give a precise field theory problem in ${\cal N}{=}4$ SYM that could be solved, beginning to end, using gauge-gravity duality. It has not previously been know how to precisely set up a problem in ${\cal N}{=}4$ SYM that corresponds to earlier studies of jets [@GubserGluon; @HIM; @CheslerQuark] that made use of classical strings in the gravity description. It is interesting to now make contact between our approach and Gubser et al.’s classical string approach, in the particular limit (\[eq:biglog\]), which can be roughly rewritten as $$T^{-1} \ll \lstop
\lesssim
T^{-4/3} \left( \frac{E}{\sqrt\lambda} \right)^{1/3} e^{-O(\lambda^{1/2})} ,
\label {eq:overlap}$$ where the $T^{-1} \ll \lstop$ is thrown in to emphasize that we’ve always been assuming $-q^2 \ll E^2$ and so $T^{-1} \ll \lstop$ throughout, and where $O$ means “of order.” Alternatively, in terms of the virtuality $-q^2$ of the source of our “jet,” (\[eq:overlap\]) is $$E^2 \gg -q^2 \gtrsim
T^{4/3} E^{2/3} \lambda^{2/3} e^{+O(\lambda^{1/2})}
.$$ This window of stopping lengths only appears once the jet energy is large enough that $$E \gg T \sqrt{\lambda} \> e^{+O(\lambda^{1/2})} .$$
Even though there is a region of overlap (\[eq:overlap\]) of our results with strings that look similar to those of Gubser et al., there are still important differences once we get out of this range. Gubser et al. found a [*maximum*]{} stopping distance of order $T^{-4/3} (E/\sqrt{\lambda})^{1/3}$, as do other methods that also model excitations with semi-infinite classical strings in the gravity dual [@CheslerQuark]. In contrast, the types of excitations that we create, through processes like fig. \[fig:Wdecay\], have a parametrically larger maximum stopping distance of order $\lmax \sim T^{-4/3} E^{1/3}$.
In this paper, we have studied the case of large but finite $\lambda$, but we have kept $\Nc{=}\infty$. It is important to mention, following Gubser et al. [@GubserGluon], that the case of finite $\Nc$ may be qualitatively different. For finite $\Nc$, the folded classical string may break, creating many string loops. It would be interesting to understand whether or not such breaking would impact the description of jet stopping in ${\cal N}{=}4$ SYM.
Our basic argument in this paper has been that the breakdown of the $1/\sqrt{\lambda}$ expansion seen in the earlier supergravity calculation of ref. [@R4] is explained by the tidal stretching of gravitons into classical string loops. Logically, however, to show that tidal stretching was the only source of difficulty, we should now do a systematic analysis of sub-leading corrections to the string evolution that we have presented here and show that those corrections are indeed parametrically small. We expect that they will be, but we leave that for someone to study another day.
Finally, we mention that an alternative to our method for creating “jets” is to produce gluon jets as synchrotron radiation from heavy quarks that are forced into circular motion. The corresponding jet stopping problem has been investigated for strongly-coupled ${\cal N}{=}4$ SYM by Chesler, Ho, and Rajagopal [@CHR]. We suspect that problems similar to those treated here could be set up in that context as well, with appropriate choices of the heavy quark velocity and synchrotron radius for a given temperature $T$. It would be interesting to investigate.
This work was supported, in part, by the U.S. Department of Energy under Grant No. DE-SC0007984.
Large $\xi$ behavior of $C(\xi)$ {#app:Cxibig}
================================
For large $\xi$, the $\bar z^6$ term in the differential equation (\[eq:chixi\]) for $\xi$ can be ignored until $\bar z \gg 1$. At that point, however, we may use the simple large-$\bar z$ result (\[eq:zbarlate\]) for $\bar z$. Substituting this into (\[eq:chixi\]) gives $$\frac{d^2\chi}{d\bar\tau^2}
= -4\left[\xi^6 - \frac{1}{(-3\bar\tau)^2}\right] \chi ,$$ whose solution is $$\chi =
(-\pi \xi^3 \bar\tau)^{1/2} H_{5/6}^{(2)}\bigl(-2 \xi^3 \bar\tau \bigr) .$$ The late-time behavior $\tau\to 0$ is $$|\chi| \simeq
\frac{\Gamma(\tfrac56)}{\pi^{1/2} \xi (-\bar\tau)^{1/3}}
\,.$$
Checking the Penrose limit: details {#app:PenroseCheck}
===================================
In this appendix, we will check whether the evolution of our strings satisfy the conditions (\[eq:PenroseConstraints\]) required for taking the Penrose limit. It is illuminating to do this in two different ways. First, we will make a rough argument based on following divergent geodesics, similar to the style of argument that we used in section \[sec:envelope\]. Then we will give alternative (but physically equivalent) arguments based on string-based results and formulas.
Tracking diverging geodesics
----------------------------
As in section (\[sec:envelope\]), let us characterize the string by following null geodesics that roughly trace different bits of string and which deviate slightly from our reference geodesic. As we’ll discuss again later, this approximation amounts to ignoring the tension in the string. From the null geodesic formula (\[eq:geodesic\]) and the metric (\[eq:metric\]), the $x^\three$ coordinate for such geodesics is given by $$\frac{dx^\three}{dz} = \frac{\hat q_\three}{\sqrt{1-f\hat\q^2}} ,
\label {eq:dDx3a}$$ where $$\hat q_\mu \equiv \frac{q_\mu}{\omega} = (-1,\hat\q) .$$ Remembering that $\Delta x^\mu \equiv x^\mu - \bar x^\mu(z)$ is the deviation relative to the reference geodesic, we have $$\frac{d\Delta x^\three}{dz} =
\frac{\hat q_\three}{\sqrt{1-f\hat\q^2}}
- \frac{\bar{\hat q}_\three}{\sqrt{1-f\bar{\hat\q}^2}}
.$$ Expand to first order in $\Delta\hat q_3 \equiv \hat q_3 - \bar{\hat q}_3$: $$\frac{d\Delta x^\three}{dz} \simeq
\frac{\Delta\hat q_\three}{(1-f\bar{\hat\q}^2)^{3/2}}
.
\label {eq:dv1}$$ Then using (\[eq:partialu\]) (and defining $u$ with respect to the reference geodesic $\bar x$), $$\frac{f \Rads^2}{z^2} \, \frac{d\Delta x^\three}{du} \simeq
\frac{f \, \Delta\hat q_\three}{1-f\hat\q^2}
.
\label {eq:combo1}$$ Since $1-f\hat\q^2 \simeq (z_\star^4+z^4)/\zh^4$, the combination (\[eq:combo1\]) is largest for $z \lesssim z_\star$, and the $d\Delta x^\three/du$ condition in (\[eq:PenroseConstraints\]) requires $$\Delta \hat q_\three \ll \frac{z_\star^4}{\zh^4}
\label {eq:cond1a}$$ for the Penrose limit. Use (\[eq:stop\]) to relate this to the stopping distance: $$\lstop \sim \frac{\zh^2}{z_\star}
\sim \zh \left( \frac{E^2}{-q^2} \right)^{1/4}
\sim \frac{\zh}{(1-\hat q_\three)^{1/4}} ,$$ so that $$\Delta\lstop
\sim \frac{\zh\,\Delta\hat q_\three}{(1-\hat q_\three)^{5/4}}
\sim \frac{\Delta\hat q_\three \, \lstop}{1-\hat q_\three}
\sim \Delta\hat q_\three \, \lstop \, \frac{\zh^4}{z_\star^4} .
\label {eq:cond1b}$$ Combining (\[eq:cond1a\]) and (\[eq:cond1b\]) gives the condition $$\Delta\lstop \ll \lstop
\label {eq:condition}$$ quoted in (\[eq:PenroseFinal\]).
Now turn to the condition on $dv/du$ in (\[eq:PenroseConstraints\]). The definition (\[eq:v\]) of $v$ gives $$dv = \bar{\hat q}_\mu \, d(\Delta x^\mu)
= - d(\Delta x^\zero) + \bar{\hat q}_\three \, d(\Delta x^\three),
\label {eq:dv2}$$ and so we need a formula for $d(\Delta x^\zero)$. The analog of (\[eq:dDx3a\]) is $$\frac{dx^\zero}{dz} = \frac{- f^{-1} \hat q_\zero}{\sqrt{1-f\hat\q^2}} ,$$ with expansion $$\frac{d\Delta x^\zero}{dz} \simeq
\frac{\hat q_\three \, \Delta\hat q_\three}{(1-f\bar{\hat q}^2)^{3/2}}
.
\label {eq:dv3}$$ Combining (\[eq:dv1\]), (\[eq:dv2\]), and (\[eq:dv3\]), gives $dv/du \simeq 0$. We therefore have to go back and make our expansions to second-order in $\Delta\hat\q$. The result is $$\frac{dv}{dz} \simeq
- \frac{(\Delta\hat\q_\perp)^2}{2(1-f\bar{\hat q}_3^2)^{1/2}}
- \frac{(\Delta \hat q_\three)^2}{2(1-f \bar{\hat q}_3^2)^{3/2}}
\,,$$ and so $$\frac{f\Rads^2}{z^2} \left|\frac{dv}{du}\right| \simeq
\frac{f(\Delta\hat\q_\perp)^2}{2}
+ \frac{f(\Delta \hat q_\three)^2}{2(1-f \bar{\hat q}_3^2)}
\,.$$ The corresponding condition on $dv/du$ in (\[eq:PenroseConstraints\]) is then $$f(\Delta\hat\q_\perp)^2
\quad \mbox{and} \quad
\frac{f(\Delta \hat q_\three)^2}{(1-f \bar{\hat q}_3^2)}
\quad \ll \quad
1 .
\label {eq:dvdugeo}$$ The first condition is strongest for $z \ll \zh$ and the second for $z \lesssim z_\star$, giving $$|\Delta\hat\q_\perp|
\quad \mbox{and} \quad
|\Delta \hat q_\three| \, \frac{\zh^2}{z_\star^2}
\quad \ll \quad
1 .
\label {eq:cond2}$$ Using (\[eq:cond1b\]), the condition involving $\Delta\hat q_\three$ becomes $$\Delta\lstop \ll \lstop \frac{\zh^2}{z_\star^2} \,.
\label {eq:condition9}$$ Since $z_\star \ll \zh$, this is weaker than the previous condition (\[eq:condition\]).
Now turn to the other condition, $|\Delta\q_\perp| \ll 1$ in (\[eq:cond2\]). To estimate $|\Delta\q_\perp|$, return to the arguments of section \[sec:envelope\], but now, in the rest frame, include an initial proper displacement of the two geodesics in $\x^\perp$ of the same parametric size as the initial proper displacement in $x^\three$. Following through the argument, one finds $$x^I \simeq
\Bigl(\gamma(1 + \beta\beta_+),{\bm\beta}_\perp,
\gamma(\beta + \beta_+),1\Bigr) \, z$$ with $\beta_\perp \sim \beta_+$. Then $$\Delta \hat q_\three
= \Delta \frac{q_\three}{q_\zero}
= \Delta \frac{dx^\three/dz}{dx^\zero/dz}
\simeq \Delta \frac{\gamma(\beta + \beta_+)}{\gamma(1+\beta\beta_+)}
\simeq \frac{\beta_+}{\gamma^2}$$ and $$\Delta \hat\q_\perp
= \Delta \frac{\q_\perp}{q_\zero}
= \Delta \frac{d\x^\perp/dz}{dx^\zero/dz}
\simeq \Delta \frac{{\bm\beta}_\perp}{\gamma(1+\beta\beta_+)}
\simeq \frac{{\bm\beta}_\perp}{\gamma} ,$$ so that $$\frac{|\Delta\hat\q_\perp|}{|\Delta\hat q_\three|}
\sim \gamma \sim \sqrt{\frac{E^2}{-q^2}} \sim \frac{\zh^2}{z_\star^2}
\,.$$ So, using (\[eq:cond1b\]), $$|\Delta\hat\q_\perp| \sim |\Delta\hat q_\three| \, \frac{z^2}{z_\star^2}
\sim \frac{\Delta\lstop}{\lstop} \, \frac{z_\star^2}{\zh^2} \,.
\label {eq:qperpidentify}$$ The condition $|\Delta\hat\q_\perp| \ll 1$ is therefore the same as the previous condition (\[eq:condition9\]) and so is also weaker than (\[eq:condition\]).
String-based formulas {#app:PenroseCheckString}
---------------------
Another way to get to the same result is to start from formulas based on our string analysis in section \[sec:quantize\]. We will also need the constraint equation for $\hat V$ generated by (\[eq:action\]) in $\tau=U$ gauge, which is $$\partial_\tau \hat V =
-\tfrac12 \sum_i \left[
(\partial_\tau \Delta\hat X^i)^2
+ \frac{1}{(\alpha' p^u)^2} (\partial_\sigma \Delta\hat X^i)^2
+ {\cal G}_i \, (\Delta\hat X^i)^2
\right] .$$ Using the definition of (\[eq:vhat\]) of $\hat v$, the corresponding constraint for $V$ is $$\partial_\tau V =
-\tfrac12 \sum_i \left[
\kappa_i \Bigl(\partial_\tau \frac{\Delta\hat X^i}{\sqrt{\kappa_i}}\Bigr)^2
+ \frac{1}{(\alpha' p^u)^2} (\partial_\sigma \Delta\hat X^i)^2
\right] ,
\label {eq:dvstring}$$ where $\kappa_i = \kappa_i(\tau)$ is the notation introduced in eq.(\[eq:metricyhat\]) with $$\kappa_1 = \kappa_2 = \frac{\Rads^2}{z^2} \,,
\qquad
\kappa_3 = \frac{\Rads^2}{z^2} \frac{(\omega^2-f|\q|^2)}{\omega^2}
\simeq \frac{\Rads^2}{z^2} \, \frac{(z_\star^4+z^4)}{\zh^4} .
\label {eq:kappa}$$ The relationship (\[eq:yhat\]) between $\Delta X^i$ and $\Delta\hat X^i$ is $$\Delta\hat X^i \equiv \sqrt{\kappa_i} \, \Delta X^i .
\label {eq:Xhat}$$ Alternatively, (\[eq:dvstring\]) is just the constraint equation $T_{\tau\tau}=0$ that one would derive in Rosen coordinates (\[eq:metricpp0\]).
Here and throughout, we will take results derived in the Penrose limit but then use them to check, after the fact, whether the Penrose limit was valid. We will focus on the evolution of the string modes after they become unstable ($z \gg z_n$). The string equation of motion from the Lagrangian (\[eq:Lin\]) is $$\partial_\tau^2 \Delta\hat X_n^i =
- \omega_{i,n}^2(\tau) \, \Delta\hat X_n^i ,$$ where $\omega_{i,n}^2(\tau) = n^2(\alpha'p^u)^{-2} - {\cal G}_i(\tau)$. At late times $z \gg z_n$, the tidal terms ${\cal G}_i$ dominate over the string tension terms $n^2(\alpha' p^u)^{-2}$, and the equation of motion becomes $$\partial_\tau^2 \Delta\hat X_n^i \simeq
{\cal G}_i \, \Delta\hat X_n^i ,$$ which it is convenient to rewrite as $$\partial_\tau^2 \Delta\hat X_n^i \simeq
\frac{\partial_\tau^2 \sqrt{\kappa_i}}{\sqrt{\kappa_i}} \, \Delta\hat X_n^i .
\label {eq:EOMtensionless}$$ Two independent solutions for $\Delta\hat X$ in the tensionless approximation (\[eq:EOMtensionless\]) are $$A^i(\tau) \propto \sqrt{\kappa_i(\tau)} ,
\qquad
B^i(\tau) \propto
\sqrt{\kappa_i(\tau)} \int_{-\infty}^\tau \frac{d\tau'}{\kappa_i(\tau')} \,.
\label {eq:AB}$$ We will take $A^i_n$ and $B^i_n$ to be normalized to be $\sim 1$ at $z \sim z_n$. The late-time ($z \gg z_n)$ string solutions will then be given by some probabilistic superposition $$\Delta\hat X^i_n(\tau) = a^i_n \, A^i_n(\tau) + b^i_n \, B^i_n(\tau)$$ where $a^i_n$ and $b^i_n$ are of order the size scale $\sqrt{\alpha'/n}$ of the proper size of the ground state of the $n$-th mode, which characterizes the string for $z \lesssim z_n$.
From (\[eq:partialu\]) or (\[eq:zbar\]), the relationship between $\tau$ and $z$ is $$-\tau = \zh^2 \Rads^2 \int_z^\infty \frac{dz}{z^2(z_\star^4 + z^4)^{1/2}}
\sim
\zh^2\Rads^2
\begin {cases}
\frac{1}{z_\star^2 z}\,, & z \ll z_\star ; \\[4pt]
\frac{1}{z^3}\,, & z \gg z_\star .
\end {cases}$$ Given the expressions (\[eq:kappa\]) for $\kappa_i$, the resulting behavior of the solution (\[eq:AB\]) is then $$\begin{aligned}
A_n^\one = A_n^\two &\sim \frac{z_n}{z} \,
\\
B_n^\one = B_n^\two &\sim
\begin {cases}
1, & z \ll z_\star ; \\[4pt]
\frac{z_\star}{z} \,, & z \gg z_\star .
\end {cases}
\\
A_n^\three &\sim
\begin {cases}
\frac{z_n}{z}, & z \ll z_\star ; \\[4pt]
\frac{z_n z}{z_\star^2}\,, & z \gg z_\star .
\end {cases}
\\
B_n^\three &\sim
\begin {cases}
1, & z \ll z_\star ; \\[4pt]
\frac{z}{z_\star}\,, & z \gg z_\star .
\end {cases}\end{aligned}$$ Note that $B_n^i$ dominates over $A_n^i$ for $z \gg z_n$. For $z \gg z_\star$, we have $B_n^\three \propto A_n^\three \propto z$, and this behavior corresponds to the dominant $(-\bar\tau)^{-1/3}$ late-time behavior discussed in section \[sec:late\]. At late times, $B^\three$ differs from a multiple of $A^\three$ by corrections of absolute size of order $z_\star^4/z^4$, which is the sub-leading $(-\bar\tau)^{4/3}$ late-time solution discussed in the main text just before (\[eq:chilate\]). A simple way to see this is to rewrite the definition of $B^i$ in (\[eq:AB\]) as $$B^i(\tau) \propto
\sqrt{\kappa_i(\tau)} \int_{-\infty}^0 \frac{d\tau'}{\kappa_i(\tau')}
- \sqrt{\kappa_i(\tau)} \int_{\tau}^0 \frac{d\tau'}{\kappa_i(\tau')} \,,$$ where the first term is proportional to $A^i(\tau)$ and the second is (at late times) the sub-leading solution. Similarly, at late times ($z \gg z_\star$), $B^\one$ differs from a multiple of $A^\one$ by sub-leading corrections of order $z_\star^2/z^2$.
### Condition on $d(\Delta x^\three)/du$
Now let us investigate the Penrose limit condition (\[eq:PenroseConstraints\]) on $d(\Delta x^\three)/du$, which can be rewritten as $$\frac{f\Rads^2}{z^2}
\left| \partial_\tau
\Bigl(\frac{\Delta\hat X^\three}{\sqrt{\kappa_3}}\Bigr) \right|
\ll 1 .
\label {eq:Cond1}$$ The $A \propto \sqrt{\kappa}$ solutions do not give any contribution to the $\tau$ derivative above, and so $$\partial_\tau
\Bigl(\frac{\Delta\hat X_n^\three}{\sqrt{\kappa_3}}\Bigr)
= b_n^\three \partial_\tau
\Bigl(\frac{B_n^\three}{\sqrt{\kappa_3}}\Bigr)
\sim \sqrt{\frac{\alpha'}{n}} \, \frac{z_\star^4/\Rads\zh^4}{\kappa_3}
= \frac{\lambda^{-1/4}}{\sqrt{n}} \,
\frac{z^2 z_\star^4}{\Rads^2 (z_\star^4+z^4)} \,.
\label {eq:dX3}$$ The condition (\[eq:Cond1\]) has to be satisfied for every point $\sigma$ on the string. But we can get a sufficient parametric condition by taking the rms average over the modes $n \lesssim n_\star$ which are excited: $$\frac{f\Rads^2}{z^2}
\left| \partial_\tau
\Bigl(\frac{\Delta\hat X^\three}{\sqrt{\kappa_3}}\Bigr)
\right|_{\rm rms}
\ll 1 ,
\label {eq:cond1}$$ which then gives $$\lambda^{-1/4} \sqrt{\ln n_\star} \,
\frac{f z_\star^4}{(z_\star^4+z^4)}
\ll 1 .$$ This condition is strongest for $z \lesssim z_\star$ and then gives the condition (\[eq:PenroseFinal0\]) quoted in the main text.
To relate this to the earlier geodesic analysis, recall that $\Delta\hat X^i/\sqrt{\kappa_i} = \Delta X^i$ and so note that (\[eq:dX3\]) gave $$\frac{f\Rads^2}{z^2}
\left| \partial_\tau \Delta X^\three \right|_{\rm rms}
\sim
\lambda^{-1/4} \sqrt{\ln n_\star} \,
\frac{f z_\star^4}{(z_\star^4+z^4)} .$$ Comparing to (\[eq:combo1\]) then identifies $\Delta \hat q_\three$ as $$\Delta \hat q_\three
\sim \frac{z_\star^4}{\zh^4} \, \lambda^{-1/4} \sqrt{\ln n_\star}
\sim \frac{z_\star^4}{\zh^4} \, \frac{\Delta\lstop}{\lstop} \,,
\label {eq:dq3identify}$$ consistent with (\[eq:cond1b\]).
### Condition on $dv/du$
Now turn to the condition on $dv/du$. The first term in the constraint formula (\[eq:dvstring\]), which is what remains if one ignores the tension terms in that equation, simply reproduce the conditions (\[eq:dvdugeo\]) found in the earlier analysis based on diverging geodesics. To see this, consider the $i=3$ contribution of the first term in (\[eq:dvstring\]) to the $dv/du$ constraint in (\[eq:PenroseConstraints\]): $$\frac{f \Rads^2}{z^2} \, \kappa_3
\Bigl(\partial_\tau \frac{\Delta\hat X^\three}{\sqrt{\kappa_3}}\Bigr)^2
\ll 1 .
\label {eq:Cond2}$$ Using (\[eq:kappa\]) and (\[eq:dX3\]), the contribution from mode $n$ to the left-hand side of (\[eq:Cond2\]) is of order $$\frac{\lambda^{-1/2}}{n} \,
\frac{f z_\star^8}{\zh^4(z_\star^4+z^4)}
\,.$$ Summing over the excited modes gives $$\lambda^{-1/2} \ln(n_\star) \, \frac{f z_\star^8}{\zh^4(z_\star^4+z^4)}
\ll 1 .
\label {eq:condition999}$$ This is the same as the second condition in (\[eq:dvdugeo\]), with the identification (\[eq:dq3identify\]) noted before. Similarly, using (\[eq:kappa\]) and (\[eq:AB\]), $$\frac{f \Rads^2}{z^2} \, \kappa_\perp
\Bigl(\partial_\tau \frac{\Delta\hat X^\perp}{\sqrt{\kappa_\perp}}\Bigr)^2
\ll 1$$ generates $$\lambda^{-1/2} \ln(n_\star) \, \frac{f z_\star^4}{\zh^4} \ll 1 ,$$ which is the same as the first condition in (\[eq:dvdugeo\]), using the identifications of (\[eq:qperpidentify\]) and (\[eq:dq3identify\]).
Finally, there are the tension terms in the constraint formula (\[eq:dvstring\]) for $dv/du$. One may check that these also do not generate any constraints stronger than (\[eq:condition\]).
Simulation details for fig. \[fig:string\] {#app:string}
==========================================
Before we explain the initial conditions used to simulate the string evolution in fig. \[fig:string\], we will take a short detour and explain how to generate a string evolution as depicted in fig. \[fig:boost\]a.
First, let us notice that near the boundary, the metric is approximately AdS. So the string evolves at very early times in a geometry which is AdS. Moreover, in the rest frame of the string considered in fig. \[fig:boost\]a, if the string has a sufficiently large energy, the string excitations probe a geometry which is the Penrose limit of an in-falling massless excitation in AdS. In other words, these excitations probe flat space. To show this explicitly, we start by expanding around the null geodesic $\bar x^\mu = (z,{\bm 0})$. Defining, as in equations (\[eq:uformula\]–\[eq:v\]), $\Delta x^\mu = x^\mu -\bar x^\mu(z), v=-\Delta x^\zero, u=-\Rads^2/z$, yields $$ds^2
= 2 \> du \> dv
-\frac{dv^2}{z^2}
+ \sum_{i=\one,\two,\three}\frac{(d\Delta x^i)^2}{z^2}
\,.$$ Taking the Penrose limit by scaling the coordinates as in (\[eq:scaling\]) brings the metric into the Rosen form $$ds^2
= 2 \> du \> dv
+ \sum_{i=\one,\two,\three}\frac{(d\Delta x^i)^2}{z^2}
\,,$$ which is just (\[eq:metricpp0\]) with $\zh \to \infty$ (i.e. $f \to 1$). While this is not immediately recognizable as a flat space metric, a change of coordinates to bring the metric into Brinkmann form, $\hat x^i=-u\,\Delta x^i/\Rads, \hat v=v-\sum_i (\Delta x^i)^2/(2u)$, yields $$ds^2= 2 \> du \> d\hat v + \sum_{i=\one,\two,\three}(d\hat x^i)^2 ,$$ which is the $\zh \to \infty$ limit of (\[eq:ppmetrichat\]). In this geometry we consider the evolution of an initial string configuration with the string shrunk to a point and with $\partial_\tau \MT(0,\sigma) = A$, $\partial_\tau\hat X^\three(0,\sigma) = B\cos\sigma$, $\partial_\tau \MZ(0,\sigma) =
\sqrt{(\partial_\tau \MT)^2-(\partial_\tau \hat X^\three)^2}$, where we have defined $\MT$ and $\MZ$ by $u=(\MT+\MZ)/2$ and $\hat v=\MT-\MZ$. The other string coordinates remain zero throughout this section. Since the string is point-like at the initial time, the stress-tensor constraint $T_{\tau\sigma}=0$ is trivially satisfied, and the second stress tensor constraint $T_{\tau\tau}=0$ was used to solve for the initial condition on $\partial_\tau\MZ$. If we choose $A$ and $B$ such that $B \ll A$, the string has an approximate analytic solution of the form $\MT=A\tau$, $\hat X^\three=B\cos\sigma\sin\tau$, $\MZ \approx A\tau$, or, in the Brinkmann coordinates, $U\approx A\tau$, $\hat V\approx 0$, $\hat X^\three=B\cos\sigma\sin\tau$.
Next we switch to Rosen coordinates, and we require that the string is again point-like at the initial time $\tau=\tau_c$: $\Delta X^\zero(\tau_c,\sigma)=0$, $\Delta X^\three(\tau_c,\sigma)=0$ and $Z(\tau_c,\sigma)=z_c$, where we introduced a boundary regulator $z_c\approx 0$. In order to satisfy this set of initial conditions we only have to perform a shift in $U$ in the previous Brinkmann solution, and use the mapping from Brinkmann to Rosen coordinates: $U=-\Rads^2/Z \approx A(\tau-\tau_c)-\Rads^2/z_c$, $V\approx 0$, $X^\three = B(Z/\Rads)\cos\sigma\sin(\tau-\tau_c)$. This is a solution which oscillates, as depicted in fig. \[fig:oscillate\] for small $z$. The envelope of these oscillations grows linearly with $z$ and roughly corresponds to fig. \[fig:boost\]a.
![ \[fig:oscillate\] An (approximate) AdS string solution for small $z$ showing the type of string we create with our initial conditions, viewed in the 4-dimensional string rest frame. The snapshots of the string are equally spaced in $\tau$. $z_{\rm c}$ could be any point of the oscillation where the length of the string vanishes. The straight line geodesics of fig. \[fig:boost\]a may be roughly thought of as corresponding to the dashed lines above, which depict the envelope of the oscillation. ](oscillate.eps)
The next step is to perform a very large boost along $x^\three$ starting from the rest frame we have been discussing so far. This will give us the appropriate initial conditions used to generate fig. \[fig:string\]a. Again, the string starts as a point-like object very close to the boundary: $Z(\tau=0,\sigma)=z_c=0.01$, $X^\zero(\tau,\sigma)=0$, $X^\three(\tau,\sigma)=0$. We take $z_h=1$. The initial (worldsheet) time $\tau=0$ derivatives have the following expressions: $\partial_\tau X^\zero(0,\sigma)=1300+0.02\times 1299\cos\sigma$, $\partial_\tau X^\three(0,\sigma) =1299+0.02\times 1300\cos\sigma$, $\partial_\tau Z(0,\sigma)
=\sqrt{(\partial_\tau X^\zero)^2 f^2(z_c)- (\partial_\tau X^\three)^2 f(z_c)}$, where $f(z_c)=1-(z_c/z_h)^4$. Here $1300$ is the energy of the boosted string and $1299$ is the momentum along the $x^\three$ direction in the appropriate units.
Earlier in this appendix, we discussed Penrose limits of AdS in order to motivate our choice of initial conditions. However, we now simulate the evolution of our classical string from these initial conditions in the full AdS$_5$-Schwarzschild geometry, with no such approximation, to obtain fig. \[fig:string\]a. In order to generate a string evolution to sufficiently late times, we followed [@CheslerQuark; @Herzog] and introduced an appropriate stretching function $\stretch$, $$\stretch =
\bigg(1+\frac{(X^\three)^2}{ z_h^2}\bigg)^m
\bigg(\frac{z_h-z}{z_h-z_c}\bigg)^n
\bigg(\frac{z_c}{z}\bigg)^p ,
% , \qquad {\rm with}\;\; m=-0.07, n=1, p=4,
\label{eq:stretching}$$ with $m=-0.07$, $n=1$, $p=4$. Then redefine the worldsheet time $\tau$ so that the worldsheet metric is $$\gamma_{\alpha\beta}=\begin{pmatrix} -\stretch&0\\0&\stretch^{-1}
\end{pmatrix}.$$ The numerical evolution of the classical string was carried out using Mathematica.
Lastly, to generate fig. \[fig:string\]b, the initial conditions were modified as follows. The string begins its evolution once more as a point-like object close to the boundary $Z(\tau=0,\sigma)=z_c=0.01$, $X^\zero(\tau,\sigma)=0, X^\three(\tau,\sigma)=0$. But the initial time $\tau=0$ derivatives are now $\partial_\tau X^\zero(0,\sigma)=1300+0.6\times 1299 \cos\sigma,
\partial_\tau X^\three(0,\sigma)=1299+0.6\times 1300 \cos\sigma$, and, as before, the remaining derivative is obtained from the constraint $T_{\tau\tau}=0$.
The main difference between the two initial conditions is that we have increased the small parameter, which is a placeholder for $\sqrt{(\alpha'/\Rads^2) \ln n_{\star}}$ from 0.02 in the previous numerical simulation to 0.6 here. \[As a side comment, this parameter cannot be increased to arbitrarily large values or else the argument of the square root in $\partial_\tau Z$ becomes negative.\] The stretching function used here is of the same form as in (\[eq:stretching\]), but with different exponents $m=-0.08, n=1, p=4.5$. This string quickly stretches in $z$. Once one of the folding points reaches the horizon, it remains frozen at that particular value of $x^\three$. The net result is that we have now generated a string which is folded back, with one folding point at the horizon and trailing behind the rest of the string, while the other folding point is still close to the boundary. This string is moving with some momentum in the $x^\three$ direction.
P. B. Arnold, S. Cantrell and W. Xiao, “Stopping distance for high energy jets in weakly-coupled quark-gluon plasmas,” Phys. Rev. D [**81**]{}, 045017 (2010) \[arXiv:0912.3862 \[hep-ph\]\]. R. Baier, Y. L. Dokshitzer, A. H. Mueller, S. Peigné and D. Schiff, “The Landau-Pomeranchuk-Migdal effect in QED,” Nucl. Phys. B [**478**]{}, 577 (1996) \[arXiv:hep-ph/9604327\]; “Radiative energy loss of high energy quarks and gluons in a finite-volume quark-gluon plasma,” Nucl. Phys. B [**483**]{}, 291 (1997) \[arXiv:hep-ph/9607355\]; “Radiative energy loss and $p_\perp$-broadening of high energy partons in nuclei,” Nucl. Phys. B [**484**]{}, 265 (1997) \[arXiv:hep-ph/9608322\]. B. G. Zakharov, “Fully quantum treatment of the Landau-Pomeranchuk-Migdal effect in QED and QCD,” JETP Lett. [**63**]{}, 952 (1996) \[arXiv:hep-ph/9607440\]; “Radiative energy loss of high energy quarks in finite-size nuclear matter and quark-gluon plasma,” JETP Lett. [**65**]{}, 615 (1997) \[arXiv:hep-ph/9704255\]. H. Liu, K. Rajagopal, U. A. Wiedemann, “Calculating the jet quenching parameter from AdS/CFT,” Phys. Rev. Lett. [**97**]{}, 182301 (2006) \[hep-ph/0605178\].
S. S. Gubser, D. R. Gulotta, S. S. Pufu and F. D. Rocha, “Gluon energy loss in the gauge-string duality,” JHEP [**0810**]{}, 052 (2008) \[arXiv:0803.1470 \[hep-th\]\]. Y. Hatta, E. Iancu and A. H. Mueller, “Jet evolution in the ${\cal N}{=}4$ SYM plasma at strong coupling,” JHEP [**0805**]{}, 037 (2008) \[arXiv:0803.2481 \[hep-th\]\]. P. M. Chesler, K. Jensen, A. Karch and L. G. Yaffe, “Light quark energy loss in strongly-coupled ${\cal N}{=}4$ supersymmetric Yang-Mills plasma,” Phys. Rev. D [**79**]{}, 125015 (2009) \[arXiv:0810.1985 \[hep-th\]\]. P. Arnold, D. Vaman, “Jet quenching in hot strongly coupled gauge theories revisited: 3-point correlators with gauge-gravity duality,” JHEP [**1010**]{}, 099 (2010) \[arXiv:1008.4023 \[hep-th\]\]. P. Arnold and D. Vaman, “Jet quenching in hot strongly coupled gauge theories simplified,” JHEP [**1104**]{}, 027 (2011) \[arXiv:1101.2689 \[hep-th\]\]. P. M. Chesler, Y. -Y. Ho and K. Rajagopal, “Shining a Gluon Beam Through Quark-Gluon Plasma,” arXiv:1111.1691 \[hep-th\]. P. Arnold, P. Szepietowski and D. Vaman, “Coupling dependence of jet quenching in hot strongly-coupled gauge theories,” JHEP [**1207**]{}, 024 (2012) \[arXiv:1203.6658 \[hep-th\]\]. G. D’Appollonio, P. Di Vecchia, R. Russo and G. Veneziano, “High-energy string-brane scattering: Leading eikonal and beyond,” JHEP [**1011**]{}, 100 (2010) \[arXiv:1008.4773 \[hep-th\]\]. G. Papadopoulos, J. G. Russo and A. A. Tseytlin, “Solvable model of strings in a time dependent plane wave background,” Class. Quant. Grav. [**20**]{}, 969 (2003) \[hep-th/0211289\].
G. T. Horowitz and A. R. Steif, “Space-Time Singularities in String Theory,” Phys. Rev. Lett. [**64**]{}, 260 (1990); “Strings In Strong Gravitational Fields,” Phys. Rev. D [**42**]{}, 1950 (1990). E. G. Gimon, L. A. Pando Zayas and J. Sonnenschein, “Penrose limits and RG flows,” JHEP [**0209**]{}, 044 (2002) \[hep-th/0206033\]. H. J. de Vega and N. G. Sanchez, “Strings falling into space-time singularities,” Phys. Rev. D [**45**]{}, 2783 (1992); H. J. de Vega, M. Ramon Medrano and N. G. Sanchez, “Classical and quantum strings near space-time singularities: Gravitational plane waves with arbitrary polarization,” Class. Quant. Grav. [**10**]{}, 2007 (1993). P. Arnold and D. Vaman, “4-point correlators in finite-temperature AdS/CFT: Jet quenching correlations,” JHEP [**1111**]{}, 033 (2011) \[arXiv:1109.0040 \[hep-th\]\]. E. Witten, “Anti-de Sitter space and holography,” Adv. Theor. Math. Phys. [**2**]{}, 253 (1998) \[arXiv:hep-th/9802150\]. O. Aharony, S. S. Gubser, J. M. Maldacena, H. Ooguri and Y. Oz, “Large N field theories, string theory and gravity,” Phys. Rept. [**323**]{}, 183 (2000) \[hep-th/9905111\]. S. Stieberger, “Constraints on Tree-Level Higher Order Gravitational Couplings in Superstring Theory,” Phys. Rev. Lett. [**106**]{}, 111601 (2011) \[arXiv:0910.0180 \[hep-th\]\]. D. Amati, M. Ciafaloni and G. Veneziano, “Superstring Collisions at Planckian Energies,” Phys. Lett. B [**197**]{}, 81 (1987); “Higher Order Gravitational Deflection And Soft Bremsstrahlung In Planckian Energy Superstring Collisions,” Nucl. Phys. B [**347**]{}, 550 (1990). “Effective action and all order gravitational eikonal at Planckian energies,” Nucl. Phys. B [**403**]{}, 707 (1993); D. N. Kabat and M. Ortiz, “Eikonal quantum gravity and Planckian scattering,” Nucl. Phys. B [**388**]{}, 570 (1992) \[hep-th/9203082\]; L. Cornalba, M. S. Costa, J. Penedones and R. Schiappa, “Eikonal Approximation in AdS/CFT: From Shock Waves to Four-Point Functions,” JHEP [**0708**]{}, 019 (2007) \[hep-th/0611122\]; L. Cornalba, M. S. Costa and J. Penedones, “Eikonal approximation in AdS/CFT: Resumming the gravitational loop expansion,” JHEP [**0709**]{}, 037 (2007) \[arXiv:0707.0120 \[hep-th\]\]; S. B. Giddings and R. A. Porto, “The Gravitational S-matrix,” Phys. Rev. D [**81**]{}, 025002 (2010) \[arXiv:0908.0004 \[hep-th\]\]. L. A. Pando Zayas and J. Sonnenschein, “On Penrose limits and gauge theories,” JHEP [**0205**]{}, 010 (2002) \[hep-th/0202186\]. M. Blau, J. M. Figueroa-O’Farrill and G. Papadopoulos, “Penrose limits, supergravity and brane dynamics,” Class. Quant. Grav. [**19**]{}, 4753 (2002) \[hep-th/0202111\]. M. Blau, “Penrose Waves and Penrose Limits,” lecture given at 2004 Saalburg/Wolfesdorf Summer School, http://www.blau.itp.unibe.ch/Lecturenotes.html
M. Blau, D. Frank and S. Weiss, “Fermi coordinates and Penrose limits,” Class. Quant. Grav. [**23**]{}, 3993 (2006) \[hep-th/0603109\]. S. P. Kim and C. H. Lee, “Nonequilibrium quantum dynamics of second order phase transitions,” Phys. Rev. D [**62**]{}, 125020 (2000) \[hep-ph/0005224\]. S. P. Kim and D. N. Page, “Classical and quantum action phase variables for time dependent oscillators,” Phys. Rev. A [**64**]{}, 012104 (2001) \[quant-ph/0101076\]. M. Karliner, I. R. Klebanov and L. Susskind, “Size And Shape Of Strings,” Int. J. Mod. Phys. A [**3**]{}, 1981 (1988); D. Mitchell and B. Sundborg, “Measuring the size and shape of strings,” Nucl. Phys. B [**349**]{}, 159 (1991); S. Jain, “Absence of initial singularities in superstring cosmology,” gr-qc/9708018. C. P. Herzog, A. Karch, P. Kovtun, C. Kozcaz and L. G. Yaffe, “Energy loss of a heavy quark moving through N=4 supersymmetric Yang-Mills plasma,” JHEP [**0607**]{}, 013 (2006) \[hep-th/0605158\].
[^1]: A specific weak-coupling calculation of the stopping distance for QCD in the high-energy limit may be found in ref. [@stop]. However, the scaling of this result was implicit in the early pioneering work of refs. [@BDMPS; @Zakharov] on bremsstrahlung and energy loss rates in QCD plasmas. Introducing supersymmetry will not change the conclusion that the stopping distance scales as $E^{1/2}$ (up to logarithms) at weak coupling.
[^2]: In a weak-coupling analysis, the two running couplings relevant to jet stopping are, roughly, $\alphas(T)$ and $\alphas(Q_\perp)$, where $Q_\perp \sim (\hat q E)^{1/4}$ grows slowly with energy and is the scale of the typical relative momentum between two daughter partons when a high energy parton splits through hard bremsstrahlung or pair production. ($\hat q \sim \alphas^2 T^3$ is a scale characteristic of the plasma that parametrizes transverse momentum diffusion of high-energy partons.) A third limiting case of interest, not addressed in this paper, is where $\alphas(T)$ is large but $\alphas(Q_\perp)$ is small. See, for example, Liu, Rajagopal, and Wiedemann [@LRW].
[^3]: See the conclusion of ref. [@R4] for further discussion of this point.
[^4]: See ref. [@adsjet2] for a discussion in the context of the present paper, but this correspondence is implicit in the earlier work of refs. [@GubserGluon; @HIM; @CheslerQuark].
[^5]: The relevant 5-dimensional supergravity field is the one dual to the operator used to create the “jet” excitation in ${\cal N}{=}4$ SYM. The mass $m$ of the supergravity field is determined by the conformal dimension $\Delta$ of that operator, e.g. $(\Rads m)^2 = \Delta(\Delta-d)$ for scalar operators [@Witten], where $d{=}4$. We emphasize that $m$ is a mass in the 5-dimensional supergravity theory and has nothing to do with “mass” of a jet from the point of view of the 3+1 dimensional ${\cal N}{=}4$ SYM quantum field theory.
[^6]: The precise details of which terms $D^m R^n$ appear independently in (\[eq:Leff\]) and which do not will not matter to our discussion, but Table I of ref. [@Stieberger] gives a nice summary of what’s currently known at tree level (i.e. corresponding to $\Nc{=}\infty$).
[^7]: Penrose limits have previously been studied in AdS$_5$-Schwarzschild by Pando Zayas and Sonnenschein [@LeoPenrose], but the null geodesic studied was different. Their geodesic fell straight toward the horizon in AdS$_5$-Schwarzschild, corresponding to $\q{=}0$ in our problem rather than $|\q| \simeq E$. Their geodesics also have non-trivial motion on the 5-sphere $S^5$. In our application, no dynamical evolution of the $S^5$ degrees of freedom takes place: the $S^5$ internal modes of the string stay in their ground state, and the $S^5$ zero modes of the string simply remain in a quantum state given by the $S^5$ harmonic of the supergravity field of interest. Equivalently, our worldsheet vacuum has $S^5$ conserved charges that are equal to those of the $S^5$ harmonic of the supergravity field of interest.
[^8]: This $u$ should not be confused with the coordinate $u \equiv (z/\zh)^2$ used in earlier work by some of the authors [@adsjet; @adsjet2].
[^9]: A more conventional normalization might be to define a dimensionless $\tau$ as $\tau \equiv u/(\alpha' p^u)$ instead of our $\tau \equiv u$. One may convert to this convention by everywhere replacing our $\tau$ by $\alpha' p^u \tau$. In our convention, the world-sheet metric is $$\gamma_{\alpha\beta} \propto
\begin{pmatrix}
-(\alpha' p^u)^{-1} & 0 \\
0 & \alpha' p^u
\end {pmatrix} .$$
[^10]: The case (\[eq:psi\]) of interest to us corresponds to the ground state cases ($\ell{=}0$ or $n{=}0$) of refs. [@PRT; @GPZS; @KimLee], and our $\chi$ here corresponds to their $\chi^*$, $v^*$, or $u^*$ respectively. However, there are various normalization issues in the formulas in all of these references, which all differ from us and from each other by overall powers of our $\chi^*/\chi$ in their expressions for the final wave function, though most differences are probably typographic errors. $\chi^*/\chi$ is just a time-dependent but $\dof$-independent complex phase, and so such differences will not matter for the evaluation of an expectation value in one of these states. But, if one wants explicit solutions to the time-dependent Schrödinger equation, then the phase needs to be correct. A simple test of any such result is to consider the case where $\omega(\tau)$ is constant and verify that $\psi_n(\dof,t)$ reproduces the correct time dependence $\exp\bigl[-i(n+\tfrac12)\omega \tau\bigr]$. Accordingly, the $(\chi/\chi^*)^\ell$ in (5.37) of ref. [@PRT] should be $(\chi/\chi^*)^{\frac{\ell}{2}+\frac14}$, the $[v^i/{v^i}^*]$ in (2.16) \[(2.15) in the preprint\] of ref. [@GPZS] should be $[v^i/{v^i}^*]^{\frac{\ell}{2}+\frac14}$, and the $(u/u^*)^{n/2}$ in (3.9) of ref. [@KimLee] should be $(u/u^*)^{\frac{n}{2}+\frac14}$ if one wants to keep track of the overall time-dependent phases and have results that solve the Schrödinger equation. [*Note added:*]{} Formulas with the correct phase may be found in ref. [@KimPage].
[^11]: The argument is a distraction and so relegated to this footnote. For $z_n \ll z \ll z_\star$ ($\xi \ll \bar z \ll 1$), one may ignore the $\xi^6$ term in (\[eq:chixi\]) and treat $\bar z$ as small. In this case the two independent solutions are approximately ${\cal A} \simeq (\xi/\bar z)(1 + \tfrac12 \, \bar z^4)$ and ${\cal B} \simeq 1 + \tfrac15 \, \bar z^4$, with $\partial_{\bar\tau} \simeq \bar z^2 (1 + \tfrac12 \bar z^4) \partial_{\bar z}$. (See Appendix \[app:PenroseCheckString\] for a more general discussion of $z \gg z_n$ solutions, for which the ones here are the small $\bar z$ limit.) To match to the oscillating solution for $\chi$ for $\bar z \ll \xi$ determined by (\[eq:chiinit\]), the superposition of ${\cal A}$ and ${\cal B}$ for $\bar z \gg \xi$ must be $\chi \simeq O(1) \, {\cal A} + O(1) \, {\cal B}$, where $O(1)$ denotes coefficients with magnitude of order 1. We then see that $\chi$ is approximately constant over the time period $\xi \ll \bar z \ll 1$.
[^12]: For example, at late times the exponential in the wavepacket (\[eq:psi\]) becomes $\exp[i S]$, where $S \propto \dof^2/(-\tau)$. The WKB condition $|\partial_\dof^2 S| \ll (\partial_\dof S)^2$ is satisfied as $\tau \to 0$ for $\dof \propto (-\tau)^{-1/3}$. We will see shortly that the proportionality constant in (\[eq:chilate\]) is of order 1 for the modes $n \lesssim n_\star$ of interest. If one keeps track parametrically of all the proportionality constants in the exponential $\exp[i S]$, one finds more specifically that the WKB condition is satisfied when $-\bar\tau \ll 1$ (i.e. $\bar z \gg 1$).
[^13]: For numerical work, it is mildly convenient to eliminate $\bar\tau$ and express all of the relevant equations solely in terms of $\bar z$, giving $\bar z^4 (1+\bar z^4) \chi'' + 2 \bar z^3(1+2\bar z^4) \chi'
= -4(\xi^6-\bar z^6)\chi$ and $\chi'(\bar z_0) = 2 i \xi^3/\bar z_0^2$ (with $\bar z_0 \to 0$) and $|\chi(\bar z)| \to 3^{1/3} \, C(\xi) \, \bar z$ (as $\bar z \to \infty$).
[^14]: A comment on randomness: $|X_n^\three|_{\rm rms}$, determined by (\[eq:Xcases\]), is the width of a Gaussian distribution for $X_n^\three$ that is centered on $X_n^\three = 0$. In contrast, at leading log order, the $\sigma$-averaged extent ${\cal S} \equiv \Bigl[\,\overline{\bigl(
X^\three - \overline{X^\three} \,
\bigr)^2 \,}\,\Bigr]^{1/2}$ of the late-time string is [*always*]{} given by the $(\delta X^\three)_{\rm rms}$ of (\[eq:result\]). That’s because, when the logarithm is large, the calculation of the logarithm in (\[eq:result\]) involves a sum over a large number $n_\star$ of modes, all contributing to the coefficient of the leading log, and so the probability distribution of ${\cal S}^2$ is narrowly peaked about its probabilistic average $(\delta X^\three)_{\rm rms}^2 \equiv \langle {\cal S}^2 \rangle$ by the central limit theorem. The $n{=}1$ mode by itself is a substantial piece of the first correction beyond leading-log order, and so randomness will enter if one moves beyond a leading-log analysis of ${\cal S}^2$.
[^15]: For some discussion of the unobservability of the UV log divergence and and how the “size” of a low-mass string excitation should be interpreted, see refs. [@size].
[^16]: We point out in passing that this is not an obscure issue specific to relativistic strings. If one quantizes small transverse vibrations of an idealized violin string $\bigl(L = \int dz [ \tfrac12 \rho \, \dot\x_\perp^2
+ \tfrac12 \tension \, \x_\perp^{\prime 2}] \bigr)$, the same $\sum_n n^{-1}$ divergence arises in the calculation of the mean-square displacement of the string. But, if you took a snapshot in time of an actual vibrating violin string (with classically large amplitude), you would not have any real-world confusion about what the mean-square displacement of that string meant. You would only get confused if you perversely chose to resolve the (idealized) violin string on such small distance scales that you could see the quantum uncertainty of very high modes that had not been excited, and for such measurements the classical description of the violin string would be inadequate. If you’re not interested in such high resolution, then the classical result for the average displacement is an excellent approximation and is physically meaningful information.
[^17]: More precisely, Gubser et al. first considered a folded open string that stretched out from beyond the horizon, as in fig. 1 of ref. [@GubserGluon]. But in actual calculations, they focused on the trailing infinite folded string, as in fig. 2 of that reference.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Differential cross sections for the reaction $\gamma p \rightarrow
\eta{'} p$ have been measured with the CLAS spectrometer and a tagged photon beam with energies from 1.527 to 2.227 GeV. The results reported here possess much greater accuracy than previous measurements. Analyses of these data suggest for the first time the coupling of the $\eta{'}N$ channel to both the $S_{11}(1535)$ and $P_{11}(1710)$ resonances, known to couple strongly to the $\eta{N}$ channel in photoproduction on the proton, and the importance of $J=3/2$ resonances in the process.
author:
- 'M. Dugger'
- 'J.P. Ball'
- 'P. Collins'
- 'E. Pasyuk'
- 'B.G. Ritchie'
- 'G. Adams'
- 'P. Ambrozewicz'
- 'E. Anciant'
- 'M. Anghinolfi'
- 'B. Asavapibhop'
- 'G. Asryan'
- 'G. Audit'
- 'H. Avakian'
- 'H. Bagdasaryan'
- 'N. Baillie'
- 'N.A. Baltzell'
- 'S. Barrow'
- 'V. Batourine'
- 'M. Battaglieri'
- 'K. Beard'
- 'I. Bedlinskiy'
- 'M. Bektasoglu'
- 'M. Bellis'
- 'N. Benmouna'
- 'B.L. Berman'
- 'N. Bianchi'
- 'A.S. Biselli'
- 'B.E. Bonner'
- 'S. Bouchigny'
- 'S. Boiarinov'
- 'R. Bradford'
- 'D. Branford'
- 'W.J. Briscoe'
- 'W.K. Brooks'
- 'S. Bültmann'
- 'V.D. Burkert'
- 'C. Butuceanu'
- 'J.R. Calarco'
- 'S.L. Careccia'
- 'D.S. Carman'
- 'B. Carnahan'
- 'S. Chen'
- 'P.L. Cole'
- 'A. Coleman'
- 'P. Coltharp'
- 'D. Cords'
- 'P. Corvisiero'
- 'D. Crabb'
- 'H. Crannell'
- 'J.P. Cummings'
- 'E. De Sanctis'
- 'R. DeVita'
- 'P.V. Degtyarenko'
- 'H. Denizli'
- 'L. Dennis'
- 'A. Deur'
- 'K.V. Dharmawardane'
- 'K.S. Dhuga'
- 'C. Djalali'
- 'G.E. Dodge'
- 'J. Donnelly'
- 'D. Doughty'
- 'P. Dragovitsch'
- 'S. Dytman'
- 'O.P. Dzyubak'
- 'H. Egiyan'
- 'K.S. Egiyan'
- 'L. Elouadrhiri'
- 'A. Empl'
- 'P. Eugenio'
- 'R. Fatemi'
- 'G. Fedotov'
- 'G. Feldman'
- 'R.J. Feuerbach'
- 'T.A. Forest'
- 'H. Funsten'
- 'M. Garçon'
- 'G. Gavalian'
- 'G.P. Gilfoyle'
- 'K.L. Giovanetti'
- 'F.X. Girod'
- 'J.T. Goetz'
- 'R.W. Gothe'
- 'K.A. Griffioen'
- 'M. Guidal'
- 'M. Guillo'
- 'N. Guler'
- 'L. Guo'
- 'V. Gyurjyan'
- 'C. Hadjidakis'
- 'R.S. Hakobyan'
- 'J. Hardie'
- 'D. Heddle'
- 'F.W. Hersman'
- 'K. Hicks'
- 'I. Hleiqawi'
- 'M. Holtrop'
- 'J. Hu'
- 'M. Huertas'
- 'C.E. Hyde-Wright'
- 'Y. Ilieva'
- 'D.G. Ireland'
- 'B.S. Ishkhanov'
- 'M.M. Ito'
- 'D. Jenkins'
- 'H.S. Jo'
- 'K. Joo'
- 'H.G. Juengst'
- 'J.D. Kellie'
- 'M. Khandaker'
- 'K.Y. Kim'
- 'K. Kim'
- 'W. Kim'
- 'A. Klein'
- 'F.J. Klein'
- 'A.V. Klimenko'
- 'M. Klusman'
- 'M. Kossov'
- 'L.H. Kramer'
- 'V. Kubarovsky'
- 'J. Kuhn'
- 'S.E. Kuhn'
- 'J. Lachniet'
- 'J.M. Laget'
- 'J. Langheinrich'
- 'D. Lawrence'
- 'T. Lee'
- 'A.C.S. Lima'
- 'K. Livingston'
- 'K. Lukashin'
- 'J.J. Manak'
- 'C. Marchand'
- 'L.C. Maximon'
- 'S. McAleer'
- 'B. McKinnon'
- 'J.W.C. McNabb'
- 'B.A. Mecking'
- 'M.D. Mestayer'
- 'C.A. Meyer'
- 'T. Mibe'
- 'K. Mikhailov'
- 'R. Minehart'
- 'M. Mirazita'
- 'R. Miskimen'
- 'V. Mokeev'
- 'S.A. Morrow'
- 'V. Muccifora'
- 'J. Mueller'
- 'G.S. Mutchler'
- 'P. Nadel-Turonski'
- 'J. Napolitano'
- 'R. Nasseripour'
- 'S. Niccolai'
- 'G. Niculescu'
- 'B.B. Niczyporuk'
- 'R.A. Niyazov'
- 'M. Nozar'
- 'G.V. O’Rielly'
- 'M. Osipenko'
- 'A.I. Ostrovidov'
- 'K. Park'
- 'C. Paterson'
- 'S.A. Philips'
- 'J. Pierce'
- 'N. Pivnyuk'
- 'D. Pocanic'
- 'O. Pogorelko'
- 'S. Pozdniakov'
- 'B.M. Preedom'
- 'J.W. Price'
- 'Y. Prok'
- 'D. Protopopescu'
- 'L.M. Qin'
- 'B.A. Raue'
- 'G. Riccardi'
- 'G. Ricco'
- 'M. Ripani'
- 'F. Ronchetti'
- 'G. Rosner'
- 'P. Rossi'
- 'D. Rowntree'
- 'P.D. Rubin'
- 'F. Sabatié'
- 'C. Salgado'
- 'J.P. Santoro'
- 'V. Sapunenko'
- 'R.A. Schumacher'
- 'V.S. Serov'
- 'A. Shafi'
- 'Y.G. Sharabian'
- 'J. Shaw'
- 'S. Simionatto'
- 'A.V. Skabelin'
- 'E.S. Smith'
- 'L.C. Smith'
- 'D.I. Sober'
- 'M. Spraker'
- 'A. Stavinsky'
- 'S.S. Stepanyan'
- 'S. Stepanyan'
- 'B.E. Stokes'
- 'P. Stoler'
- 'I.I. Strakovsky'
- 'S. Strauch'
- 'M. Taiuti'
- 'S. Taylor'
- 'D.J. Tedeschi'
- 'U. Thoma'
- 'R. Thompson'
- 'A. Tkabladze'
- 'S. Tkachenko'
- 'C. Tur'
- 'M. Ungaro'
- 'M.F. Vineyard'
- 'A.V. Vlassov'
- 'K. Wang'
- 'L.B. Weinstein'
- 'H. Weller'
- 'D.P. Weygand'
- 'M. Williams'
- 'E. Wolin'
- 'M.H. Wood'
- 'A. Yegneswaran'
- 'J. Yun'
- 'L. Zana'
- 'J. Zhang'
title: '$\eta{''}$ photoproduction on the proton for photon energies from 1.527 to 2.227 GeV'
---
Understanding the structure of the proton is challenging due to the great complexity of this strongly interacting multi-quark system [@Cap00]. Of particular utility in investigating nucleon structure are those production mechanisms and observables that help isolate individual excited states of the nucleon and determine the importance of specific contributions. Since the electromagnetic interaction is well understood, photoproduction offers one of the more powerful methods for studying the nucleon. Since the $\eta$ and $\eta{'}$ mesons have isospin 0, $\eta N$ and $\eta{'} N$ final states can only originate (in one-step processes) from isospin $I = 1/2$ intermediate states. Therefore, the reactions $\gamma p \rightarrow \eta p$ and $\gamma p -> \eta{'} p$ isolate $I = 1/2$ resonances, thereby providing an “isospin filter” for the spectrum of broad, overlapping nucleon resonances, a useful simplification for theoretical efforts to predict the large number of excited nucleon states.
Thus, photoproduction of the $\eta{'}$ meson from the proton is an excellent tool for clarifying the details of the nucleon resonance spectrum. However, existing data for the $\gamma p \to
\eta{'} p$ reaction come from only a few exclusive or semi-exclusive measurements due to the limitations of experimental facilities. While previous experiments [@abbhhm; @ahhm; @saphir] detected fewer than 300 $\eta{'}$ events, in the measurements described here, over 2$\times 10^5$ $\eta{'}$ photoproduction events were detected and used to extract differential cross sections.
The differential cross sections for the reaction $\gamma p \rightarrow
\eta{'} p$ were measured with the CEBAF Large Acceptance Spectrometer (CLAS) [@CLAS] and the bremsstrahlung photon tagging facility [@tagger] at the Thomas Jefferson National Accelerator Facility. The cross sections were part of a program of meson production measurements using the same CLAS, tagger, and target configuration. Tagged photons, with energies $E_\gamma$ between 0.49 and 2.96 GeV, were incident on an 18-cm-long liquid hydrogen target placed at the center of CLAS. (The threshold for $\eta{'}$ photoproduction on the proton is $E_\gamma$ = 1.447 GeV.) The event trigger required the coincidence of a post-bremsstrahlung electron passing through the focal plane of the photon tagger and at least one charged particle detected in CLAS. Tracking of the charged particles through the magnetic field within CLAS by drift chambers provided determination of their charge, momentum and scattering angle. This information, together with the particle velocity measured by the time-of-flight scintillators, provided particle identification for each particle detected in CLAS and its corresponding momentum four-vector. Particle identification was generally unambiguous; in the case of proton identification, the fraction of particles misidentified as protons made up a background of less than $2\times10^{-3}$.
The $\gamma p \to p X$ missing mass was used to identify photoproduced mesons through detection of the proton recoiling into the CLAS from the cryogenic target. As seen in the missing mass spectrum in Fig. 1, the resolution obtained is sufficient for clear identification of the photoproduced $\pi^0, \eta, \rho + \omega, \eta{'}$, and $\varphi$ mesons, the latter four peaks situated atop a multi-pion background. The missing mass spectrum was binned in center-of-mass scattering angle and photon energy to extract meson yields for each angle/energy bin. The CLAS acceptance limited the measurement of the $\gamma p \to \eta{'}p $ reaction to photon energies above 1.527 GeV ($W$ = 1.94 GeV) and $\eta{'}$ center-of-mass scattering angles $\vartheta^{\eta{'}}_{\rm{c.m.}}$ in the range $-0.8 \leq \cos\vartheta^{\eta{'}}_{\rm{c.m.}} \leq 0.8$. For the $\eta{'}$ measurements reported here, a total of 15 non-overlapping bins in incident photon energy $E_\gamma$ were used, each about 50 MeV wide. (For convenience, the photon energy bins are labeled by the energy of the centroid of the bin.) The photon energies ranged from bins centered on $E_\gamma$ from 1.527 up to 2.227 GeV, corresponding to center-of-mass energies $W$ from 1.94 to 2.25 GeV. Above this energy range, the yield for $\eta{'}$ photoproduction was too low to permit the extraction of reliable cross sections. The background subtraction (as exemplified in the inset in Fig. \[mm\]) assumed a mixture of two-, three-, and four-pion contributions, along with contributions from the $\rho^0$.
\[scale=0.35\][mm.eps]{}
The proton detection efficiency for CLAS was measured empirically [@Dug01; @Aug99] using the reaction $\gamma p \rightarrow p \pi^+ \pi^-$. Both pions were required to be detected in the event and both must have been produced by the same photon from the bremsstrahlung beam. A missing mass reconstruction from the kinematical information for the two pions was performed to determine if a proton should have been seen in the CLAS in a particular phase-space volume. The presence or absence of a proton yielded an empirical measure of the momentum-dependent proton detection efficiency for that volume. Efficiency uncertainties for $\eta{'}$ photoproduction, dominated by the statistical uncertainty in the number of protons scattered and detected, were determined for each bin, and ranged from $\sim$1% at the lowest energies to $\sim$2% at the highest energies.
The results reported here represent the first measurements for $\eta{'}$ photoproduction utilizing an [*absolute*]{} measurement of the photon flux [@gflux]. The photon flux for the entire tagger photon energy range was determined by measuring the rate of scattered electrons detected in each counter of the focal plane of the bremsstrahlung photon tagger by sampling focal plane hits not in coincidence with CLAS. The detection rate for the scattered electrons was integrated over the livetime of the experiment and converted to the total number of photons on target for each counter of the tagger focal plane. The tagging efficiency was measured in dedicated runs with a Total Absorption Counter (TAC) [@tagger], which directly counted all photons in the beam [@gflux].
Ideally, one would use a well-known reaction in the energy range used for these measurements to confirm the validity of the photon flux measurement technique and to estimate the uncertainties in the photon flux normalization. However, no large database exists for any photoproduction reaction over the range of photon energies for which we report $\eta{'}$ cross sections here. As an alternative, the pion photoproduction database is quite extensive. The SAID parameterization [@SAID] provides a very good description of that database. The SAID analysis incorporates many observables for all channels of pion photoproduction. The existing $\pi^0$ photoproduction cross section database below 1.5 GeV is quite dense. (The data below 1.5 GeV make up 95% of the published measurements on $\pi^0$ photoproduction on the proton.) The SAID solution (SM02) is in very good agreement with those existing data. Thus, SAID can be assumed to provide the correct energy and angular dependence for the $\pi^0$ photoproduction cross section in that energy range within its estimated normalization uncertainty of 2%. The existing data above 1.5 GeV are much more scarce and have significantly larger uncertainties. Therefore, we have used that parametrization to ascertain the validity of the procedures used here by comparing that SAID parametrization to $\pi^0$ photoproduction cross sections (for $E_\gamma$ from 0.675 to 1.525 GeV) extracted from data taken simultaneously with the $\eta{'}$ measurements reported here (for $E_\gamma$ from 1.527 to 2.227 GeV), using the same absolute normalization techniques for both reactions [@pi0_soon].
In order to determine the $\pi^0$ cross sections for this experiment over the photon energy range from 0.675 to 1.525 GeV, the empirically measured proton detection efficiency for CLAS had to be supplemented by a Monte-Carlo estimate of the detection efficiency for protons from $\pi^0$ photoproduction because the phase space occupation of protons for the $\gamma p \rightarrow p \pi^+
\pi^-$ reaction becomes sparse at higher energies when rebinned for $\gamma p \rightarrow p \pi^0$ efficiencies. Agreement between the empirical and Monte-Carlo methods, where sufficient statistics made comparison possible, was within 3%.
For $E_\gamma$ from 0.675 to 1.525 GeV and the range of $\cos(\vartheta^{\pi^0}_{\rm{c.m.}})$ used here, our entire set of $\pi^0$ differential cross sections, comprised of 19 energy bins each with 12 bins in $\cos(\vartheta^{\pi^0}_{\rm{c.m.}})$ (228 points, in total) was easily fit by the SAID parametrization with a single overall constant factor $N$ = 1.02 ($\chi^2_{reduced} = 1.3$). [*[This overall agreement throughout the energy range implies that the absolute normalization technique is sound, and additionally indicates the detector acceptance also is well-determined.]{}*]{}
To estimate the uncertainty in the photon flux measurement, a more refined fit of our measured differential cross sections for $\pi^0$ photoproduction for [*[each]{}*]{} photon energy bin to the SAID parametrization was performed, determining a single overall constant factor $N_E$ for [*[each]{}*]{} photon energy bin. For $E_\gamma$ = 0.675 to 1.525 GeV, these $N_E(E_\gamma)$ values were produced, binned into a histogram, and fit with a simple Gaussian. The centroid of the fit to $N_E(E_\gamma)$ was 1.02, as before. The standard deviation $\sigma(N_E(E_\gamma))$ of the $N_E(E_\gamma)$ values was 4%. We conservatively estimate the absolute normalization systematic uncertainty to be about 5%.
![Differential cross sections for $\eta{'}$ photoproduction on the proton (black squares). Other results from SAPHIR [@saphir] (blue triangles) are shown for comparison. Error bars shown include systematic and statistical uncertainties. \[diff\_pic\]](exp2.eps)
![Differential cross sections for $\eta{'}$ photoproduction on the proton. Also shown are results from Nakayama and Haberzettl [@hk] (Red lines: $u$-channel contributions. Green lines: $t$-channel contributions. Blue lines: Sum of all $s$-, $t$-, and $u$-channel contributions), and a model (black lines) inspired by A. Sibirtsev *et al.* [@elster_and_co], as discussed in the text. Error bars shown include systematic and statistical uncertainties. \[diff\_comp\]](comp.eps)
The differential cross sections for $\eta{'}$ photoproduction obtained are shown in Figs. \[diff\_pic\] and \[diff\_comp\]. In general, the angular distributions, while flat at threshold, show a continuing increase in slope at forward angles with increasing photon energy. At the highest energies, growth at backward angles is also seen. These general features are suggestive of coupling to an s-channel resonance near threshold, with increasing contributions of t- and u-channel exchange as the energy above threshold increases. The SAPHIR measurements [@saphir] are shown for comparison in Fig. \[diff\_pic\]. The CLAS data, with much smaller error bars and smaller photon energy bins (SAPHIR has energy bins of 100 MeV for energies below 1.84 GeV and 200 MeV wide bins above), generally agree with the SAPHIR results within the very large error bars of the latter, but the CLAS values are nonetheless systematically lower. The excellent agreement noted above between the SAID parametrization and the $\pi^0$ photoproduction cross sections measured here, using the same normalization techniques as used for these $\eta{'}$ cross sections, strongly suggests the absolute normalization determined here is correct.
Included in Fig. \[diff\_comp\] are the results (shown as red, green, and blue lines) representing a consistent analysis of the reactions $\gamma p
\rightarrow p \eta{'}$ and $p p \rightarrow p p \eta{'}$ by Nakayama and Haberzettl (NH) [@hk]. The NH analysis is based upon a relativistic meson-exchange model of hadronic interactions including coupled-production mechanisms. We have also performed calculations (black lines) using a relativistic meson-exchange model by A. Sibirtsev *et al.* [@elster_and_co] as a recipe. For both models, allowed processes include $s$-, $t$-, and $u$-channel contributions. The intermediate mesons in the $t-$channel exchanges are the $\omega$ and $\rho^0$ in both cases. Both models here also included the $S_{11}(1535)$ and $P_{11}(1710)$ resonances ($J = 1/2$), which are known to decay strongly to the ${\eta}N$ channel [@PDG]. The NH model also includes two additional $S_{11}$ and two additional $P_{11}$ resonances, albeit with relatively small couplings. In contrast to the fit of the SAPHIR data in Ref. [@saphir], the present adaptation of the NH model to our data now also requires $J = 3/2$ resonances \[$P_{13}(1940)$, $D_{13}(1780)$, and $D_{13}(2090)$\]. Since the NH model fits the data better than our calculations, the inclusion of these additional $J = 3/2$ resonances appears to be beneficial.
A comparison of the predictions of these two different approaches can provide insight into which physical contributions are most successful at explaining features of the observed cross sections. The forward peaking of the cross sections at the highest energies is dominated by $t$-channel exchange. Addition of the $S_{11}(1535)$ state contributes mainly to the overall initial rise and fall of the total cross sections below 1.7 GeV. We note that this is the first time that $S_{11}(1535)$ and $P_{11}(1710)$ resonances, known to strongly couple to the $\eta{N}$ channel in photoproduction, have been used in fits as contributions to the $\eta{'}N$ photoproduction channel. The $J=3/2$ resonances included by NH are especially useful in obtaining the correct shape of the differential cross sections for the energies from 1.728 to 1.879 GeV. The $u$-channel exchange causes the backward-angle enhancement seen around 2 GeV and above. (The general behavior of individual $t-$ and $u-$channel contributions can be seen in Fig. \[diff\_comp\] and Ref. [@hk2].)
Since the $\eta{'}$ meson is the only flavor singlet of the fundamental pseudoscalar meson nonet, studies of the reaction can also help yield information on the role of glue states in excitations of the nucleon. The flavor-singlet axial charge of the nucleon ($G_A(0)$) is related to the $\eta{'}$-nucleon-nucleon and gluon-nucleon-nucleon coupling constants ($g_{\eta{'}NN}$ and $g_{GNN}(0)$, respectively) through the flavor-singlet Goldberger-Treiman relation [@ven]: $$2 m_N G_A(0) = F g_{\eta{'}NN} - \frac{F^2 m_{\eta{'}}^2}{N_F} g_{GNN}(0), \label{eq:ga0}$$ where $m_N$ is the mass of the nucleon, $m_{\eta{'}}$ is the $\eta{'}$ mass, $F$ is an invariant decay constant that reduces to $F_{\pi}$ (pion decay constant) if the $U(1)_A$ anomaly were turned off [@feld], and $N_F$ equals the number of flavors. When first measured [@emc], the singlet axial charge was found to have a value of $G_A(0) = 0.20 \pm 0.35$. (A more recent calculation [@hirai] gives $G_A(0) = 0.213 \pm 0.138$.) At that time, the importance of the second term in Eq. \[eq:ga0\] was unappreciated, and this low value of $G_A(0)$ was surprising: Since $g_{\eta{'}NN}$ is considered to be correlated with the fraction of the nucleon spin carried by its constituent quarks [@elster_and_co], that fraction would then be consistent with zero. Thus, neglecting the gluonic portion of Eq. \[eq:ga0\] was one of the causes of the so-called “spin crisis." However, by including the gluonic degrees of freedom in Eq. \[eq:ga0\], the value of $g_{\eta{'}NN}$ can be large, provided that it is nearly canceled by $g_{GNN}(0)$. This equation then can be used to indirectly determine the gluonic coupling to the nucleon given a value of $g_{\eta{'}NN}$.
The observed $u$-channel contribution seen here allows the $g_{\eta{'}NN}$ coupling to be extracted (albeit in a model-dependent way). The value of $g_{\eta{'}NN}$ found from the particular NH fit shown here is 1.33, whereas our results using the model of Ref. [@elster_and_co] provides 1.46. Since differential cross sections alone do not provide sufficient constraints to these models, these $g_{\eta{'}NN}$ values should be taken with caution. Nonetheless, both values of $g_{\eta{'}NN}$ are consistent with the analysis of Ref. [@feld] which gives 1.4 $\pm$ 1.1. Moreover, even though the uncertainty in $g_{\eta{'}NN}$ precludes a definitive statement about the value of $g_{GNN}(0)$, Eq. \[eq:ga0\] can be carried out taking $N_F = 3$, $F = F_{\pi} = 0.131$ GeV, $g_{\eta{'}NN} = 1.4$, and $G_A(0) = 2.13$, yielding the result that $g_{GNN}(0) = 41$ GeV$^{-3}$.
In conclusion, the differential cross sections presented here are the first high-quality data for the $\gamma p \rightarrow \eta{'} p$ reaction. An analysis of the data with two different models of the process suggests for the first time contributions from both the $S_{11}(1535)$ and $P_{11}(1710)$ nucleon resonances to the $\eta{'}N$ channel in photoproduction, the two resonances previously identified as strongly coupling to the $\eta{N}$ channel [@PDG]. Using two different theoretical descriptions of the data, these cross sections suggest a value for the $\eta{'}$-nucleon-nucleon coupling constant $g_{\eta{'}NN}$ of 1.3-1.5, consistent with previous theoretical estimates of this quantity. These data should continue to prove quite useful in guiding future experimental and theoretical investigations of the structure of the nucleon.
The authors gratefully acknowledge the Jefferson Lab Accelerator Division staff. We thank W. Kaufmann, H. Haberzettl and K. Nakayama for useful discussions and assistance. This work was supported by the National Science Foundation, the Department of Energy (DOE), the Deutsche Forschungsgemeinschaft (through an Emmy Noether grant to U.T.), the French Centre National de la Recherche Scientifique and Commissariat à l’Energie Atomique, the Italian Istituto Nazionale di Fisica Nucleare, and the Korean Science and Engineering Foundation. The Southeastern Universities Research Association operates Jefferson Lab for DOE under contract DE-AC05-84ER40150.
See, e.g., S. Capstick and W. Roberts, Prog. in Part. and Nucl. Phys [**45**]{}, S241 (2000).
ABBHHM Collaboration, Phys. Rev. [**175**]{}, 1669 (1968).
AHHM Collaboration, Nucl. Phys. B [**108**]{}, 45 (1976).
R. Plötzke [*et al.*]{}, Phys. Lett. B [**444**]{}, 555 (1998).
B. Mecking [*et al.*]{}, Nucl. Instr. Meth. [**A** 503]{}, 513 (2003).
D. Sober [*et al.*]{}, Nucl. Instr. Meth. [**A** 440]{}, 263 (2000).
M. Dugger, Ph.D. dissertation, Arizona State University (unpublished, 2001).
T. Auger, Ph.D. dissertation, Université de Paris (unpublished, 1999). J. Ball and E. Pasyuk, CLAS note 2005-002 (2005) http://www.jlab.org/Hall-B/notes/clas\_notes05/2005-002.pdf
R. A. Arndt, W. J. Briscoe, I. I. Strakovsky, and R. L. Workman, Phys. Rev. C [**66**]{}, 055213 (2002); R. A. Arndt, I. I. Strakovsky, R. L. Workman, and M. M. Pavan, Phys. Rev. C **52**, 2120 (1995).
M. Dugger, [*et al.*]{}, The results for $\pi^0$ photoproduction will be reported in a forthcoming publication.
K. Nakayama, and H. Haberzettl, private communication. The results shown here are based on an extended version of the model given in .
S. Eidelman [*et al.*]{}, Phys. Lett. B [**592**]{}, 1 (2004).
K. Nakayama and H. Haberzettl, Phys. Rev. C [**69**]{}, 065212 (2004).
G.M. Shore, G. Veneziano, Nucl. Phys. [**B381**]{}, 23 (1992)
T. Feldmann, Int. J. Mod. Phys. A [**15**]{} 159 (2000); also available as hep-ph/9907491.
J. Ashman [*et al.*]{}, Nucl. Phys. B [**328**]{}, 1 (1989). M. Hirai, S. Kumano, N. Saito, Phys. Rev. D [**69**]{} 054021 (2004); also available as hep-ph/0312112.
A. Sibirtsev, Ch. Elster, S. Krewald, and J. Speth, AIP Conf. Proc.[**717**]{}, 837 (2004); A. Sibirtsev, Ch. Elster, S. Krewald, and J. Speth, nucl-th/0303044.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We use the DECaLS DR3 survey photometry matched to the SDSS-III/BOSS DR12 spectroscopic catalog to investigate the morphology and stellar mass-size relation of luminous red galaxies (LRGs) within the CMASS and LOWZ galaxy samples in the redshift range $0.2<z<0.7$. The large majority of both samples is composed of early-type galaxies with De Vaucouleurs profiles, while only less than 20% are late-type exponentials. We calibrate DECaLS effective radii using the higher resolution CFHT/MegaCam observations and optimise the correction for each morphological type. By cross-matching the photometric properties of the early-type population with the Portsmouth stellar mass catalog, we are able to explore the high-mass end of the distribution using a large sample of 313,026 galaxies over 4380deg$^{2}$. We find a clear correlation between the sizes and the stellar masses of these galaxies, which appears flatter than previous estimates at lower masses. The sizes of these early-type galaxies do not exhibit significant evolution within the BOSS redshift range, but a slightly declining redshift trend is found when these results are combined with $z\sim0.1$ SDSS measurements at the high-mass end. The synergy between BOSS and DECaLS has important applications in other fields, including galaxy clustering and weak lensing.'
author:
- |
\
$^1$European Space Astronomy Centre (ESAC), 28692 Villanueva de la Cañada, Madrid, Spain\
$^2$Instituto de Física, Universidade de São Paulo, São Paulo, SP, Brazil\
$^3$Instituto de Astrofísica de Andalucía (IAA)/ CSIC, Granada, E-18008, Spain\
$^4$Instituto de Física Teórica (IFT) UAM/CSIC, Universidad Autónoma de Madrid, Cantoblanco, E-28049 Madrid, Spain\
$^5$Campus of International Excellence UAM/CSIC, Cantoblanco, E-28049 Madrid, Spain\
$^6$Departamento de Física Teórica M-8, Universidad Autónoma de Madrid, Cantoblanco, 28049 Madrid, Spain\
$^7$Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, CA, 94720, USA\
bibliography:
- './bibliography.bib'
date:
title: 'The mass-size relation of LRGs from BOSS and DECaLS'
---
galaxies: distances and redshifts galaxies: evolution galaxies: photometry galaxies: structure galaxies: statistics cosmology: observations cosmology: theory large-scale structure of Universe
Introduction {#sec:intro}
============
The SDSS-III/Baryon Oscillation Spectroscopic Survey [BOSS; @Eisenstein2011; @Dawson2013] provided unprecedented statistics at the high-mass end by measuring the spectra of about 1.5 million luminous red galaxies [LRGs; @Eisenstein2001] over 10,000deg$^2$ of sky down to magnitude $r\sim 22.2$ and within the redshift range $0.2<z<0.7$. This data set has been used not only to accurately measure the baryon acoustic oscillation feature [BAO; @Eisenstein2005; @Anderson2014; @2017MNRAS.470.2617A], but also to study the massive galaxy population at $z\sim0.55$. BOSS allowed us to characterise the red/blue color bimodality observed in LRGs [@2013MNRAS.432..359T; @2014MNRAS.437.1109R; @2016MNRAS.462.2218F; @AMD2016a], to constrain the high-mass end of the stellar mass and luminosity functions of these massive galaxies [@2013MNRAS.435.2764M; @2013MNRAS.436..697B; @2016MNRAS.457.4021L; @2016MNRAS.455.4122B; @2017MNRAS.467.2217B; @AMD2016a] and to measure the intrinsic relation between galaxy luminosity and velocity dispersion [@AMD2016b; @2017MNRAS.468...47M]. Despite these achievements, the morphological and structural properties of BOSS LRGs have been difficult to probe due to the poor SDSS image quality (median seeing of 2").
More recently, the Dark Energy Camera Legacy Survey[^1] (DECaLS) of the SDSS Equatorial Sky has been designed to obtain high-quality images that cover 6700$\,\rm{deg^2}$ in three optical bands $(g,\,r,\,z)$. With a limiting magnitude of $r\leq 23.4$ and a median seeing of $1.2"$, it allows a narrower and more efficient target selection for the DESI survey . DECaLS improves dramatically the quality of the SDSS data set, providing also deeper photometry.
Besides the classification of galaxies through their morphology and shape parameters, the stellar mass-size relation has been explored in a number of works as a powerful scaling law to connect fundamental galaxy properties. @2010MNRAS.404.2087B studied the distribution of stellar mass (M$_{\star}$), size, velocity dispersion, luminosity and color as a function of galaxy morphology and concentration index for SDSS massive early-type galaxies. They claimed that sample selections based on colour or concentration lead to significantly different scaling relations. @2011MNRAS.412L...6B investigated further these dependencies in a sample of SDSS early-type galaxies (ETGs) and found that there is a particular stellar mass scale (M$_{\star}\sim2\times10^{11}\,\rm{M_{\odot}}$) beyond which major mergers start to dominate the assembly histories of these massive galaxies. @2013ApJ...778L...2C identified the same mass scale as the transition point between two processes that regulate the mass-size distribution of galaxies in dense environments and in the field. From one side, spiral galaxies are replaced by bulge-dominated fast-rotator ETGs, with the same mass-size relation and mass distribution as in the field. On the other hand, the slow-rotator ETGs are segregated in mass from the fast ones, and their size increases proportionally to their mass. These evidences suggest that bulge growth (outside-in evolution) and bulge-related environmental quenching dominate in the low-mass end, while dry mergers (inside-out evolution) and halo-related quenching shape the mass and size growth at the high-mass end.
@2013ApJ...779...29H investigated the impact of different large-scale environments (i.e., field, group and clusters) on the size of massive ETGs at $z\sim0$. At fixed stellar mass, they did not find any significant dependence of the central and satellite ETG sizes on the environment. The mass-size relation of these galaxies is independent of the host halo mass and the galaxy position within the halo. This result is not sensitive to different galaxy selections based on morphology, star formation, or central density. @2011MNRAS.415.3903T studied the buildup of the mass-size relation of elliptical galaxies from $z\sim0$ up to $z\sim1$, using observations from SDSS and HST/GOODS. They did not find any evidence for age segregation at fixed stellar mass. This rules out the scenario of a present-day mass-size relation progressively established through a bottom-up sequence in which older galaxies populate its lower tail, remaining in place since their formation. Their result supports instead the hypothesis that the local mass-size relation is defined at $z\sim1$, with all galaxies occupying a region half of the size of the present-day distribution. @2003MNRAS.343..978S explored the connection between galaxy size and luminosity (or stellar mass) using $z\sim0.1$ SDSS data and found a trend which is significantly steeper for early- than for late-type galaxies.
Recently, @2017arXiv170704979Z analysed the dependence of the luminosity- or mass-size relation on galaxy concentration and morphology in the SDSS DR7 Main galaxy sample. They found a clear trend of smaller sizes and steeper slope for early-type elliptical galaxies. @2011MNRAS.418.1055M studied the morphology and size of BOSS luminous massive galaxies using HST/COSMOS photometry and found that about $74\%$ of them are early-type elliptical or lenticular, while the rest are late-type spirals. @2014ApJ...789...92B compared galaxy size measurements in SDSS, SDSS-III/BOSS and COSMOS data at $0.1\lesssim z\lesssim 0.7$ to derive accurate corrections for the galaxy effective radii (i.e. sizes). @2017ApJ...837..147H investigated the redshift-size relation in massive ETGs in the UltraVISTA and CANDELS surveys. They found evidence of a significant mass build up at $r<3\,$kpc beyond $z>4$, and a clear evolutionary change at $z\sim1.5$, when the galaxy progenitor stops growing in-situ through disk star formation and accretes minor mergers. @2017arXiv170103526S explored the ratio between galaxy size and dark matter halo virial radius at $z\lesssim3$ using data from GAMA and CANDELS. They found very little dependence on stellar mass and lower ratios at high redshift for more massive galaxies.
In this work, we aim to characterise the morphology and the stellar mass-size relation of the well-known SDSS-III/BOSS DR12 CMASS and LOWZ galaxy samples [@Anderson2012; @Bolton2012; @Anderson2014; @2015ApJS..219...12A] within the redshift range $0.2<z<0.7$. To this purpose, we match these BOSS spectroscopic samples to the DECaLS DR3 photometric catalog. We calibrate DECaLS sizes using the high-resolution (0.6" median seeing) CFHT/MegaCam observations and optimise the correction individually for each morphological type. By cross-matching our DECaLS selections with the Portsmouth [@2013MNRAS.435.2764M] stellar mass catalog at $0.2<z<0.7$, we are able to constrain the M$_{\star}$–size relation of very massive LRGs in a sample of unprecedented size at these redshifts. Our cross-matched BOSS-DECaLS galaxy samples with CFHT calibrated sizes are made publicly available for the community on the <span style="font-variant:small-caps;">Skies and Universes</span>[^2] database.
The paper is organised as follows: Section \[sec:data\] describes the data sets used in our analysis. In Section \[sec:calibration\] we explain how the DECaLS effective radii are calibrated using CFHT observations. In Section \[sec:results\] we present our results: the morphology of BOSS galaxies, their stellar mass-size relation and their size evolution. We compare with previous studies in Section\[sec:comparison\] and summarize our conclusions in Section \[sec:summary\]. For the analysis we adopt the cosmology: $h=0.6777,\,\Omega_m=0.3071,\,\Omega_{\Lambda}=0.6929,\,n=0.96,\,\sigma_8=0.8228$ [@Planck2014].
Data and galaxy selections {#sec:data}
==========================
![Footprint of the cross-matched DECaLS-BOSS galaxy sample (green area) versus the original SDSS-III/BOSS coverage (grey).[]{data-label="fig:completeness"}](plots/footprint.pdf){width="0.96\linewidth"}
![$(z-W1)$ vs. $(g-z)$ color distributions of the cross-matched DECaLS-BOSS LOWZ (top) and CMASS (bottom) samples. The contours denote the $1\,\sigma$ and $2\,\sigma$ uncertainty regions.[]{data-label="fig:selezioni"}](plots/DECaLS_LOWZ_distr.pdf "fig:"){width="1.04\linewidth"} ![$(z-W1)$ vs. $(g-z)$ color distributions of the cross-matched DECaLS-BOSS LOWZ (top) and CMASS (bottom) samples. The contours denote the $1\,\sigma$ and $2\,\sigma$ uncertainty regions.[]{data-label="fig:selezioni"}](plots/DECaLS_CMASS_distr.pdf "fig:"){width="1.04\linewidth"}
We use the DECam Legacy Survey (DECaLS) DR3 photometric catalog[^3] row-by-row-matched to the SDSS DR12 spectroscopic galaxy sample[^4]. DECaLS is an optical survey on the 4m Blanco telescope at Cerro Tololo Inter-American Observatory designed to complement the SDSS, SDSS-III, SDSS-IV and DESI surveys with high-quality images from 6700$\,\rm{deg^2}$ of extragalactic sky in the equatorial region in three optical bands $(g,\,r,\,z)$. The DECaLS DR3 photometric catalog also includes the infrared WISE[^5] bands (W1, W2, W3, W4). The sky coverage lies within $-18^{\circ} < \delta < +34^{\circ}$ in celestial and $|b| > 18^{\circ}$ in Galactic coordinates. DECaLS has improved dramatically the quality of the SDSS imaging data, providing a deeper photometry with limiting magnitude of $r\leq 23.4$ and a median seeing of $1.2"$.
In the cross-matched catalog introduced above, we select the BOSS CMASS and LOWZ galaxy samples of LRGs (hereafter our parent samples") using the SDSS spectroscopic flags[^6].
We further exclude point-like sources from the parent samples by imposing the DECaLS condition `TYPE!=PSF`. We recover 238,008 CMASS and 75,018 LOWZ galaxies, respectively, i.e., about 31% and 23% of the original BOSS samples. The missing galaxies are not observed by DECaLS DR3, which has an effective area of 4380deg$^{2}$, much smaller than the 9376deg$^{2}$ of the SDSS-III/BOSS, as shown in Figure \[fig:completeness\]. In Figure \[fig:selezioni\], we display our LOWZ (top panel) and CMASS (bottom) parent samples in the DECaLS color-color plane. We use the $g$ and $z$-band magnitudes from DECaLS and the W1 infrared magnitude from WISE to highlight the color properties of BOSS LRGs in DECaLS photometry.
Beside DECaLS magnitudes, for the analysis we adopt DECaLS effective radii, surface brightness profiles and galaxy morphologies. We perform galaxy size calibrations using data from two different surveys: the MegaPrime/MegaCam[^7] at CFHT and the Cosmic Evolution Survey (COSMOS)[^8]. The first one has a $1\,$deg$^2$ field-of-view with a resolution of 0.187" per pixel and a median seeing of $\sim 0.7"$. It provides much better imaging quality, which is key to precisely determine galaxy sizes and morphological types. The second survey was originally designed to probe galaxy formation and evolution over a 2deg$^2$ equatorial field with imaging by most of the major space-based telescopes and a number of large ground based telescopes.
We adopt @2013MNRAS.435.2764M stellar masses for the galaxies in our parent samples to study the mass-size relation of LRGs at $0.2<z<0.7$. These are estimated by fitting model spectral energy distributions to the BOSS observed magnitudes.
Galaxy size calibration {#sec:calibration}
=======================
{width="0.48\linewidth"}{width="0.48\linewidth"}{width="0.48\linewidth"} {width="0.48\linewidth"}
------------ ---------------------------------- ---------------------------------- -- -- --
$0.15\leq z<0.3$ $0.3\leq z<0.43$
$R_0$\[arcsec\]$\alpha$ $R_0$\[arcsec\]$\alpha$
CFHT DeV 1.226$\pm$0.020 -0.324$\pm$0.014 1.141$\pm$0.009 -0.395$\pm$0.011
CFHT Exp 1.370$\pm$0.168 -0.672$\pm$0.154 1.292$\pm$0.094 -0.652$\pm$0.143
COSMOS DeV 1.241$\pm$0.289 -0.079$\pm$0.166 1.556$\pm$0.168 -0.439$\pm$0.128
COSMOS Exp – –
$0.43< z\leq 0.55$ $0.55< z<0.7$
$R_0$\[arcsec\]$\alpha$ $R_0$\[arcsec\] $\alpha$
CFHT DeV 1.009$\pm$0.006 -0.469$\pm$0.009 0.952$\pm$0.006 -0.547$\pm$0.011
CFHT Exp 2.085$\pm$0.147-0.276$\pm$0.020 2.123$\pm$0.143 -0.247$\pm$0.018
COSMOS DeV 1.256 $\pm$0.126-0.186$\pm$0.083 1.847$\pm$0.356 -0.832$\pm$0.344
COSMOS Exp – –
------------ ---------------------------------- ---------------------------------- -- -- --
In order to correct our galaxy size measurements from seeing effects [@1993MNRAS.264..961S; @2003AJ....125.1882B; @2014ApJ...789...92B], we calibrate DECaLS effective radii with the latest CFHT (see Section \[sec:data\]) observations. We cross-match our CMASS and LOWZ samples with the data available in the four CFHT fields. Only galaxies with De Vaucouleurs and exponential profiles are employed. For those objects surviving the matching (4721 in CMASS and 2050 in LOWZ), we compare their radii measured in both surveys. We define the DECaLS circularized radius as $R_{\rm{DECaLS}}=R_{\rm{eff}}\sqrt{(b/a)}$, where $R_{\rm{eff}}$ is the DECaLS effective radius, while $a$ and $b$ are the semi-major and semi-minor ellipse axes, respectively. For the calibration we use the following functional form: $$R_{\rm{DECaLS}}^{\rm{calib}}=R_{\rm{DECaLS}}\times f(R_{\rm{DECaLS}}),
\label{eq:correctionfunc}$$ where $f(R_{\rm{DECaLS}})$ is the calibration function depending on DECaLS size defined as: $$f(R_{\rm{DECaLS}})=\left (\frac{R_{\rm{DECaLS}}}{R_0}\right )^\alpha.
\label{eq:calib}$$ We separately fit CMASS and LOWZ galaxies with De Vaucouleurs and exponential profiles to find the optimal parameters $\alpha$ and $R_0$. As part of the fitting procedure, we perform sigma-clipping, rejecting those objects located more than $2\,\sigma$ away from the mean of the $R_{\rm{CFHT}}/R_{\rm{DECaLS}}$ distribution. The excluded points are considered outliers in what follows. The best-fit parameters are reported in Table \[tab:corrtable\]. In the top panels of Figure \[fig:corr\], we display DECaLS versus CFHT effective radii of the LOWZ (left) and CMASS (right) samples, respectively. The grey points are DECaLS original radii before the CFHT calibration; the blue contours are the corrected sizes. The effect of the CFHT calibration lowers DECaLS effective radii by a $\sim$40% factor, fully consistent with the statistical correction made by @2011MNRAS.418.1055M using the Zurich Estimator of Structural Types [ZEST; @2007ApJS..172..406S] measurements. In what follows, we extrapolate and apply this calibration to the entire CMASS and LOWZ parent samples.
In order to test the CFHT calibration, we also derive an independent correction by cross-matching DECaLS with COSMOS data. Even though the overlap between the two data sets is very small – only 67 galaxies survive the matching for CMASS and 56 for LOWZ – the result is consistent with the CFHT analysis, as shown in the bottom panels of Figure \[fig:corr\]. Here we show DECaLS LOWZ on the left and CMASS on the right side. The grey points are the DECaLS radii before correction and the blue filled squares are the sizes calibrated using COSMOS data. The blue empty squares are the outliers, i.e. those objects located more than $2\,\sigma$ away from the mean of the corrected distribution.
Results {#sec:results}
=======
In this section, we present our main results: the morphology of the cross-matched BOSS-DECaLS CMASS and LOWZ samples and the stellar mass-size relation for their early-type galaxy population.
The morphology of BOSS LRGs {#sec:morfologia}
---------------------------
{width="0.45\linewidth"} {width="0.45\linewidth"}
{width="0.5\linewidth"}{width="0.5\linewidth"}
We use the DECaLS surface brightness profile classification as an indicator of the morphology of CMASS and LOWZ galaxies. In DECaLS, the following profiles have been fitted to individual objects:
- De Vaucouleurs: Sersic [@1968adga.book.....S] profile with $n=4$.
- Exponential: Sersic profile with $n=1$.
- Composite: linear combination of a De Vaucouleurs and an exponential profile with the same source center.
- Simple: exponential profile with a fixed 0.45" effective radius and circular shape.
We find that 64% (89%) of CMASS (LOWZ) galaxies have De Vaucouleurs profiles; 14% (4%) are exponentials; 17% (1%) are simple and 5% (6%) are composite. Galaxies with De Vaucouleurs profiles are typically early-type/ellipticals, while exponentials correspond to late-type/spirals . Composite profiles are a mixture of the two previous configurations. Simple profiles are used when any other profile with varying radius does not yield a significantly better $\chi^2$ (note that the number of parameters is penalized in the determination of the goodness of fit).
The CMASS selection allows for a fraction of bluer objects in the sample, which increases with redshift [@Eisenstein2001; @AMD2016a]. This explains the presence of galaxies with exponential profiles. Interestingly, the fraction of galaxies with De Vaucouleurs profiles increases significantly from the LOWZ to the CMASS sample, as the fraction of exponentials decreases. In Figure \[fig:giplot\], we show the $(g-z)$ color distributions in both the LOWZ (left) and CMASS (right) samples for De Vaucouleurs and elliptical galaxies separately. In the CMASS sample, galaxies with De Vaucouleurs profiles are significantly redder than those showing an exponential profile, as expected from the early-late type association. Interestingly, this separation is less obvious in the LOWZ sample, which might be due to the presence of more dusty spirals having an exponential profile. Note that the red/blue separation in the CMASS sample is more evident in the $(g-i)$ color distribution (i.e., $(g-i)=2.35$), as shown in @2011MNRAS.418.1055M, @Dawson2013, @2013MNRAS.435.2764M, @2014MNRAS.437.1109R, @2016MNRAS.462.2218F, and @2017ApJ...836...87L.
The fraction of late-type and early-type galaxies that we find in our samples is approximately consistent, given the uncertainties and differences between different methods, with results from @2011MNRAS.418.1055M, @2013MNRAS.435.2764M and @AMD2016a using the SDSS photometry.
In Figure \[fig:hlr\], we show the effective radius distribution of the LOWZ (left) and CMASS (right) samples, highlighting the contribution from the different morphologies. In both populations, the early-type De Vaucouleurs galaxy distribution peaks at $\rm{R_{DECaLS}}\sim7\,\rm{kpc}$, exponentials around $8\,\rm{kpc}$, composite at $12\,\rm{kpc}$ and simple below $5\,\rm{kpc}$. Most of the galaxies classified as “composite" have a companion nearby preventing to accurately measure their effective radius. Due to this configuration, composite galaxies have on average larger radii and wider size distributions compared to the other morphologies. The number of galaxies and the number density (per unit deg$^2$) of each sample are reported in Table \[tab:densities\].
--------------- ---------------- ------------------------------- --------------- -- ---------------- ------------------------------- --------------- --
N$_{\rm{gal}}$ $n_{\rm{dens}}$\[deg$^{-2}$\] fraction\[%\] N$_{\rm{gal}}$ $n_{\rm{dens}}$\[deg$^{-2}$\] fraction\[%\]
Total 84,986 19.4 100 239,431 54.7 100
DeVaucouleurs 75,441 17.2 89 154,004 35.2 64
Exponential 3464 0.8 4 33,681 7.7 14
Simple 1062 0.3 1 41,292 9.4 17
Composite 5019 1.1 6 10,454 2.4 5
--------------- ---------------- ------------------------------- --------------- -- ---------------- ------------------------------- --------------- --
In Figure \[fig:hlr\], the median seeing at the corresponding redshift of each sample is represented by a solid vertical line. The DECaLS PSF is dominated by seeing on scales of 1-1.2", which corresponds to a FWHM of about 2.8kpc at the mean redshift of LOWZ ($z\sim0.3$) and about 3.9kpc at the mean redshift of CMASS ($z\sim0.55$). This makes the effective radius distribution fall sharply at small radii. For LOWZ galaxies, however, this effect is less pronounced due to their larger angular size compared to CMASS objects. In what follows, we exclude from our samples those objects classified as “simple", which have effective radius significantly lower than these thresholds.
The mass-size relation of LRGs at $0.2<z<0.7$ {#sec:Mstar}
---------------------------------------------
Hereafter, we will focus only on LRGs with De Vaucouleurs profiles.
{width="0.48\linewidth"} {width="0.48\linewidth"} {width="0.48\linewidth"} {width="0.48\linewidth"}
------------------------- ------------------ ------------------ -------------------- ------------------ --
$0.2\leq z<0.3$ $0.3\leq z<0.4$ $0.43\leq z< 0.55$ $0.55\leq z<0.6$
$f$ 1.00 0.87 0.57 1.0
$\sigma$ 0.12 0.20 0.20 0.22
$\log{(M_1/M_{\odot})}$ 11.24 11.27 11.24 11.36
$A$ 0.238$\pm$0.044 0.219$\pm$0.022 0.202$\pm$0.021 0.172$\pm$0.015
$B$ -1.947$\pm$0.509 -1.706$\pm$0.263 -1.493$\pm$0.241 -1.141$\pm$0.178
------------------------- ------------------ ------------------ -------------------- ------------------ --
Figure \[fig:masssize\] displays the circularized effective radius as a function of stellar mass for the DECaLS LOWZ (upper row) and CMASS (lower row) samples, respectively, in four bins of redshift ($0.2\leq z<0.3$ and $0.3\leq z<0.4$ for LOWZ; $0.43\leq z<0.55$ and $0.55\leq z<0.6$ for CMASS). The density contours are approximately corrected from stellar-mass incompleteness using the analytic formula from @2016MNRAS.457.4021L: $$c=\frac{f}{2} \left[1+\rm{erf}{\left(\frac{\log{M_{\star}/M_1}}{\sigma}\right)} \right],
\label{eq:comp_leauthaud}$$ where the parameter values are chosen at the mean redshift of our samples, see Table\[tab:param\_mass\_size\]. As expected, we find a correlation, although mild, between effective radius and stellar mass in our cross-matched BOSS-DECaLS samples. The mean size estimates in bins of stellar mass are displayed on top of each distribution as blue points; the error bars correspond to the $\pm1\,\sigma$ dispersion around the mean. A linear fit of the form $\rm{\log{(R_{DECaLS}/kpc)}}=\,\textit{A}\,\rm{\log{(M_{\star}/M_{\odot})}}\,+\,\textit{B}$ is also shown in each panel of Figure \[fig:masssize\] as a blue solid line; the corresponding parameters are given in Table\[tab:param\_mass\_size\]. The slope of the mass-size relation increases mildly across our redshift range $0.2\leq z<0.7$, with values of $A\sim0.17-0.24$.\
\
BOSS provides unprecedented statistics at the high-mass end, as compared to previous surveys and samples at similar redshifts. Establishing a fair comparison at these stellar masses is therefore tricky. Instead, in Figure \[fig:masssize\], we show results from two relatively large lower-mass samples. The first one is a selection of quiescent galaxies observed in CFHT SDSS Stripe 82 [@2017MNRAS.469.4523C], with stellar masses from the S82 Massive Galaxy Catalog[^9] [S82-MGC; @2015ApJS..221...15B]. The second one is composed by early-type galaxies detected using COSMOS [@2013MNRAS.428.1715H]. When combined, the BOSS mass-size relation appears as a natural higher-mass continuation of those lower-mass relations, but displaying a significantly flatter slope (the typical slope at lower-masses is $A\sim0.47-0.61$).
The apparent flattening observed in the mass-size relation might be due to residual incompleteness and selection effects that we could not take into account in the analysis, and to the CFHT size calibration. In Figure\[fig:masssize\], we overplot the linear fit to the uncalibrated relation (black dotted line), which is flatter ($A\sim0.20-0.45$) than the lower-mass measurements, but steeper than the corrected relation, especially towards higher redshifts. By comparing these two fits, one can appreciate the effect of the CFHT calibration on the DECaLS size estimates, which are reduced by a factor $\sim0.5-0.25$dex. Note also that the size correction has a stronger effect on the higher redshift bins (i.e., CMASS), as expected from the right panel of Figure\[fig:corr\].
The possibility remains that the apparent flattening of the mass-size relation towards the high-mass end is related to the well-documented curvature of scaling relations for early-type galaxies [see e.g., @2007MNRAS.377..402D; @2009MNRAS.394.1978H; @2011MNRAS.412L...6B; @2013ApJ...769L...5K; @2013MNRAS.432.1862C; @2013MNRAS.432.1709C; @AMD2016a; @AMD2016b; @2017MNRAS.468...47M]. In BOSS, particularly, this phenomenon was reported by @AMD2016b when analysing the intrinsic $L-\sigma$ relation for the red sequence population. In Section\[sec:comparison\], we discuss possible interpretations of this result.
The redshift-size relation of LRGs at $0.2<z<0.7$ {#sec:z-size}
-------------------------------------------------
We have analysed the redshift evolution of the average size of massive LRGs from the BOSS-DECaLS cross-matched samples. This measurement, due to the mass-size relation itself, is very sensitive to the particular stellar mass range observed, so comparisons with previous results should be taken with caution.
Figure \[fig:zsize\] displays the mean effective radius of our LOWZ (blue point) and CMASS (red square) samples, in which only galaxies with De Vaucouleurs profile are considered; the error bars correspond to $\pm1\,\sigma$ scatter around the mean. Our results are obtained by integrating over the entire stellar mass range. The empty black triangles represent previous estimates from SDSS and SDSS-III/BOSS [@2014ApJ...789...92B] calibrated against HST/COSMOS data and selected in a narrow bin of stellar mass.
The redshift evolution of the DECaLS early-type galaxy sizes calibrated using CFHT data is overall consistent with a flat trend, i.e. no evolution. This is in good agreement with CFHT observations in Stripe 82 of quiescent ETGs [@2017MNRAS.469.4523C]. However, when we combine our nearly flat results with the SDSS measurements at $z\sim 0.1$ [@2003MNRAS.343..978S empty magenta point], the evolutionary trend mildly declines with redshift and reconciles with @2014ApJ...789...92B. The effective radius estimates presented by @2014ApJ...789...92B are systematically smaller than our results and their evolutionary trend is overall similarly flat.
Interestingly, when we limit our measurements to very high masses, $\log{(\rm{M_{\star}/ M_{\odot}})}>11.8$, we find a slope steeply declining with redshift. This is in line with current estimates for very massive ETGs in ULTRAVISTA and CANDELS/3D-HST [@2017ApJ...837..147H] and with the massive ETGs at $11.2< \log{(\rm{M_{\star}/M_{\odot}})} <12$ observed in COSMOS [@2013MNRAS.428.1715H].
![Redshift-size relation of our DECaLS CMASS (red filled square) and LOWZ (blue filled point) galaxies, compared to the SDSS-III/BOSS (black empty triangles) results from @2014ApJ...789...92B. We also show the $z\sim0.1$ SDSS Main galaxy sample measurement from @2003MNRAS.343..978S (magenta empty point) which, combined with our results, suggests a mildly declining redshift trend. The dot-dashed line is the fit to the COSMOS ETGs with $11.2< \log{(\rm{M_{\star}/M_{\odot}})} <12$ [@2013MNRAS.428.1715H].[]{data-label="fig:zsize"}](plots/DECaLS_CMASS_zsize_lesspoins.pdf){width="1.05\linewidth"}
Comparison with previous studies {#sec:comparison}
================================
We have measured the stellar mass-size relation for massive early-type galaxies within the redshift range $0.2< z < 0.7$. When compared with lower-mass results, our measurement shows a relative flattening of this relation, especially at higher redshift.
At face value, it seems that the observed flattening of the mass-size relation could be related to the well-documented curvature of the scaling relations towards the high-mass end, which has been extensively addressed in the literature for early-type galaxies [@2009MNRAS.394.1978H; @2007MNRAS.377..402D; @2011MNRAS.412L...6B; @2013ApJ...769L...5K; @2013MNRAS.432.1862C; @2013MNRAS.432.1709C; @AMD2016b]. In particular, @2009MNRAS.394.1978H studied the stellar mass-size relation in a sample of $\sim50,000$ SDSS ETGs at $z\sim0.1$ and found evidence for a deviation from the linear behaviour: galaxies with $\rm{\log{(M_{\star}/M_{\odot})}}\gtrsim11.5$ have larger sizes than expected. The slope of the regression line depends on the weighting scheme adopted to correct from survey incompleteness and ranges from $A\sim1$ (unitary weights) to $A\sim0.47$ ($1/V_{\rm{max}}(L)$ weights). @2011MNRAS.412L...6B demonstrated that different scaling relations for ETGs all point to two preferential mass scales, $3\times10^{10}$ and $2\times10^{11}\,\rm{M_{\odot}}$, as places where fundamental physical processes happen. @2013ApJ...769L...5K investigated the Faber-Jackson correlation between velocity dispersion $\sigma$ and total galaxy luminosity separately for elliptical galaxies with and without cores. Using the mass-to-light ratio, they related $\sigma$ to the stellar mass. They found that the velocity dispersion of core ellipticals increases much more slowly with luminosity and mass, compared to the coreless galaxies. They claimed that this is an evidence for dry major mergers as the dominant growth mode of the most massive elliptical galaxies. @AMD2016b found a steep slope and small scatter for the L-$\sigma$ relation of the massive red sequence population at $z\sim0.55$ using the CMASS sample.
Although our measurement, in combination with lower-mass results, seems generally consistent with the curvature of the scaling relations towards the high-mass end, it is noteworthy that this behaviour appears to go in the opposite direction to what is reported by @2009MNRAS.394.1978H at low redshift. As mentioned above, they find that SDSS ETGs at the high-mass end are progressively larger than expected (from a linear relation). Establishing a fair comparison is, however, hindered by sample differences. Besides focusing on a different redshift range, their conclusion is drawn mostly from an intermediate-mass sample (the high-mass end corresponds to the tail of the distribution), whereas our results are obtained from a larger sample covering exclusively the high-mass end (and after comparing with independent lower-mass measurements at the same redshift). Follow-up work will be specifically devoted to addressing this question.
We have also measured the redshift evolution of the average size of massive early-type galaxies from $z=0.7$. Our results are consistent with a non-evolving scenario. This conclusion is in agreement with results from @2017ApJ...851...34B, who detected no growth in the stellar mass of massive (i.e., $\rm{\log(M_{\star}/M_{\odot})>11.2}$) galaxies over $0.3<z<0.65$. @AMD2016b also found results generally consistent with no evolution of the high-mass end of the L-$\sigma$ relation all the way to $z=0$.
Summary and future work {#sec:summary}
=======================
We have studied the morphology, the stellar mass-size relation and the size evolution of the SDSS-III/BOSS DR12 CMASS and LOWZ spectroscopic galaxy samples cross-matched with the DECaLS DR3 ($g,\,r,\,z$) deeper and higher-quality image photometry. The resulting CMASS and LOWZ selections include about 31% and 23% of the original BOSS samples. We find that the large majority of both populations is composed of early-type galaxies with De Vaucouleurs profiles, while only less than 20% of them are late-type spirals with exponential profiles. The fraction of ETG clearly increases from LOWZ to CMASS. We calibrate the DECaLS sizes of these galaxies against the available observations from CFHT/Megacam and COSMOS with better image quality. We obtain an excellent agreement between these two independent corrections and our results are fully consistent with @2011MNRAS.418.1055M using ZEST [@2007ApJS..172..406S] data. By cross-matching our CMASS and LOWZ galaxies with De Vaucouleurs profiles with the Portsmouth [@2013MNRAS.435.2764M] stellar mass catalog for SDSS-III/BOSS LRGs at $0.2<z<0.7$, we are able to study the high-mass end of the distribution up to $\log{(\rm{M\star/M_{\odot}})}\sim\,12.2$ with unprecedented statistics for 313,026 galaxies over 4380deg$^{2}$. Our main results can be summarized as:
1. the BOSS-DECaLS mass-size relation for massive early-type galaxies exhibits a clear correlation with an apparent flattening in the slope compared to previous estimates from ETGs in CFHT SDSS Stripe 82 at lower masses [@2013MNRAS.428.1715H; @2017MNRAS.469.4523C]. Further analysis is needed to determine what causes this behaviour. The apparent flattening might be explained by the fact that scaling relations for the most massive early-type galaxies can be systematically different from the same relations at lower masses [e.g., @AMD2016b; @2009MNRAS.394.1978H; @2011MNRAS.412L...6B; @2013ApJ...769L...5K].\
2. we find no evolution in the BOSS-DECaLS ETG sizes over $0.2<z<0.7$. This result is consistent with the non-evolving scenario found by @AMD2016b in the high-mass end of the L-$\sigma$ relation all the way to $z=0$. In addition, it is consolidated by the no-growth detection in the stellar mass of Stripe 82 Massive galaxies within $0.3<z<0.65$ [@2017ApJ...851...34B]. If we focus only on the most massive galaxies at $\log{(\rm{M_{\star}/ M_{\odot}})}>11.8$, the slope of their evolution changes to steeply declining with redshift. This is in agreement with current estimates for very massive ETGs in ULTRAVISTA and CANDELS/3D-HST [@2017ApJ...837..147H] and in COSMOS [@2013MNRAS.428.1715H].\
3. combining our BOSS-DECaLS size measurements with the SDSS results at $z\sim 0.1$ [@2003MNRAS.343..978S], the evolutionary trend mildly declines with redshift and reconciles with @2014ApJ...789...92B. This is consistent with a passive evolution scenario for LRGs from $z\sim0.55$ [@2013MNRAS.435.2764M; @AMD2016a; @AMD2016b; @2017ApJ...851...34B].
This work provides a galaxy sample with unprecedented statistics that can be used to further investigate morphological and size-related aspects in the evolution of LRGs. In addition, this cross-matched sample can be used to study the dependence of clustering on morphological and size-related properties of LRGs. Our cross-matched BOSS-DECaLS CMASS and LOWZ samples with CFHT calibrated sizes are made publicly available for the community on the <span style="font-variant:small-caps;">Skies and Universes</span>[^10] database.
In a follow-up study, we will attempt to deconvolve the uncertainties on the effective radius and the residual incompleteness effects present in the mass-size relation using a similar forward-modeling Bayesian method as the one presented in @AMD2016a [@AMD2016b]. Within this framework, we will be able to measure the mass-size relation for the [*[intrinsic]{}*]{} red sequence population photometrically identified in @AMD2016a. We also plan to look at the dust properties and star formation history of these galaxies by cross-matching them with the available data from the infra-red Herschel[^11] ESA mission.
In the near future, the Subaru HSC-CCP[^12] Collaboration will provide ulta-deep multicolor images down to $r_{\rm{AB}}\sim28$ with 0.6" median seeing, which wil be key to improve the current constraints on galaxy size and morphology. New-generation spectroscopic surveys such as SDSS-IV/eBOSS [@2016AJ....151...44D], DESI [@2015AAS...22533607S] and Euclid [@2011arXiv1110.3193L; @2015arXiv150502165S] will produce enormous data sets with high-resolution out to redshift $z\sim2$. These observations will allow us to better understand the galaxy formation paradigm on small scales, and to coherently link it to the evolution of the large scale structure of our Universe.
Acknowledgments {#acknowledgments .unnumbered}
===============
GF is supported by a European Space Agency (ESA) Research Fellowship at the European Space Astronomy Centre (ESAC), in Madrid, Spain.
AMD acknowledges support from the Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP), through the grant 2016/23567–4.
GF and FP acknowledge financial support from MINECO grant AYA2014-60641-C2-1-P.
GF and FP are thankful to the Lawrence Berkeley National Laboratory for hosting the first phase of this work. GF and coauthors wish to thank D. Lang for insightful discussions on DECaLS morphology details.
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is <http://www.sdss3.org/>.
SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University.
The Legacy Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS; NOAO Proposal ID \# 2014B-0404; PIs: David Schlegel and Arjun Dey), the Beijing-Arizona Sky Survey (BASS; NOAO Proposal ID \# 2015A-0801; PIs: Zhou Xu and Xiaohui Fan), and the Mayall z-band Legacy Survey (MzLS; NOAO Proposal ID \# 2016A-0453; PI: Arjun Dey). DECaLS, BASS and MzLS together include data obtained, respectively, at the Blanco telescope, Cerro Tololo Inter-American Observatory, National Optical Astronomy Observatory (NOAO); the Bok telescope, Steward Observatory, University of Arizona; and the Mayall telescope, Kitt Peak National Observatory, NOAO. The Legacy Surveys project is honored to be permitted to conduct astronomical research on Iolkam Du’ag (Kitt Peak), a mountain with particular significance to the Tohono O’odham Nation.
NOAO is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.
This project used data obtained with the Dark Energy Camera (DECam), which was constructed by the Dark Energy Survey (DES) collaboration. Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science Foundation, the Ministry of Science and Education of Spain, the Science and Technology Facilities Council of the United Kingdom, the Higher Education Funding Council for England, the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign, the Kavli Institute of Cosmological Physics at the University of Chicago, Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A&M University, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo, Financiadora de Estudos e Projetos, Fundacao Carlos Chagas Filho de Amparo a Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Cientifico e Tecnologico and the Ministerio da Ciencia, Tecnologia e Inovacao, the Deutsche Forschungsgemeinschaft and the Collaborating Institutions in the Dark Energy Survey. The Collaborating Institutions are Argonne National Laboratory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones Energeticas, Medioambientales y Tecnologicas-Madrid, the University of Chicago, University College London, the DES-Brazil Consortium, the University of Edinburgh, the Eidgenossische Technische Hochschule (ETH) Zurich, Fermi National Accelerator Laboratory, the University of Illinois at Urbana-Champaign, the Institut de Ciencies de l’Espai (IEEC/CSIC), the Institut de Fisica d’Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig-Maximilians Universitat Munchen and the associated Excellence Cluster Universe, the University of Michigan, the National Optical Astronomy Observatory, the University of Nottingham, the Ohio State University, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, and Texas A&M University.
BASS is a key project of the Telescope Access Program (TAP), which has been funded by the National Astronomical Observatories of China, the Chinese Academy of Sciences (the Strategic Priority Research Program “The Emergence of Cosmological Structures” Grant \# XDB09000000), and the Special Fund for Astronomy from the Ministry of Finance. The BASS is also supported by the External Cooperation Program of Chinese Academy of Sciences (Grant \# 114A11KYSB20160057), and Chinese National Natural Science Foundation (Grant \# 11433005).
The Legacy Survey team makes use of data products from the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE), which is a project of the Jet Propulsion Laboratory/California Institute of Technology. NEOWISE is funded by the National Aeronautics and Space Administration.
The Legacy Surveys imaging of the DESI footprint is supported by the Director, Office of Science, Office of High Energy Physics of the U.S. Department of Energy under Contract No. DE-AC02-05CH1123, by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract; and by the U.S. National Science Foundation, Division of Astronomical Sciences under Contract No. AST-0950945 to NOAO.
[^1]: <http://legacysurvey.org/decamls/>
[^2]: <http://www.skiesanduniverses.org/>
[^3]: <http://legacysurvey.org/dr3/files/>
[^4]: <https://data.sdss.org/sas/dr12/sdss/spectro/redux/>
[^5]: <http://wise.ssl.berkeley.edu/index.html>
[^6]: <http://www.sdss.org/dr13/algorithms/boss_galaxy_ts/>
[^7]: <http://www.cfht.hawaii.edu/Instruments/Imaging/MegaPrime/>
[^8]: <http://cosmos.astro.caltech.edu/>
[^9]: <http://www.ucolick.org/~kbundy/massivegalaxies/>
[^10]: <http://www.skiesanduniverses.org/>
[^11]: <http://sci.esa.int/herschel/>
[^12]: <http://hsc.mtk.nao.ac.jp/ssp/>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Euclidean embeddings of data are fundamentally limited in their ability to capture latent semantic structures, which need not conform to Euclidean spatial assumptions. Here we consider an alternative, which embeds data as discrete probability distributions in a Wasserstein space, endowed with an optimal transport metric. Wasserstein spaces are much larger and more flexible than Euclidean spaces, in that they can successfully embed a wider variety of metric structures. We exploit this flexibility by learning an embedding that captures semantic information in the Wasserstein distance between embedded distributions. We examine empirically the representational capacity of our learned Wasserstein embeddings, showing that they can embed a wide variety of metric structures with smaller distortion than an equivalent Euclidean embedding. We also investigate an application to word embedding, demonstrating a unique advantage of Wasserstein embeddings: We can visualize the high-dimensional embedding directly, since it is a probability distribution on a low-dimensional space. This obviates the need for dimensionality reduction techniques like t-SNE for visualization.'
author:
- |
Charlie Frogner\
MIT CSAIL and MIT-IBM Watson AI Lab\
`[email protected]`\
Farzaneh Mirzazadeh\
MIT-IBM Watson AI Lab and IBM Research\
`[email protected]`\
Justin Solomon\
MIT CSAIL and MIT-IBM Watson AI Lab\
`[email protected]`
bibliography:
- 'w2embedding.bib'
title: |
Learning Embeddings into\
Entropic Wasserstein Spaces
---
Introduction
============
Learned embeddings form the basis for many state-of-the-art learning systems. Word embeddings like word2vec [@MikolovNIPS13], GloVe [@pennington2014glove], fastText [@BojanowskiGJM17], and ELMo [@PetersNIGCLZ18] are ubiquitous in natural language processing, where they are used for tasks like machine translation [@NeubigSWFMPQSAG18], while graph embeddings [@Nickel0TG16] like node2vec [@Grover16] are used to represent knowledge graphs and pre-trained image models [@SimonRD16] appear in many computer vision pipelines.
An effective embedding should capture the semantic structure of the data with high fidelity, in a way that is amenable to downstream tasks. This makes the choice of a target space for the embedding important, since different spaces can represent different types of semantic structure. The most common choice is to embed data into Euclidean space, where distances and angles between vectors encode their levels of association [@MikolovNIPS13; @WestonBU11; @KirosZS14; @MirzazadehGS14]. Euclidean spaces, however, are limited in their ability to represent complex relationships between inputs, since they make restrictive assumptions about neighborhood sizes and connectivity. This drawback has been documented recently for tree-structured data, for example, where spaces of negative curvature are required due to exponential scaling of neighborhood sizes [@NickelK17; @NickelK18].
In this paper, we embed input data as probability distributions in a Wasserstein space. Wasserstein spaces endow probability distributions with an *optimal transport* metric, which measures the distance traveled in transporting the mass in one distribution to match another. Recent theory has shown that Wasserstein spaces are quite flexible—more so than Euclidean spaces—allowing a variety of other metric spaces to be embedded within them while preserving their original distance metrics. As such, they make attractive targets for embeddings in machine learning, where this flexibility might capture complex relationships between objects when other embeddings fail to do so.
Unlike prior work on Wasserstein embeddings, which has focused on embedding into Gaussian distributions [@Muzellec18; @Zhu18], we embed input data as discrete distributions supported at a fixed number of points. In doing so, we attempt to access the full flexibility of Wasserstein spaces to represent a wide variety of structures.
Optimal transport metrics and their gradients are costly to compute, requiring the solution of a linear program. For efficiency, we use an approximation to the Wasserstein distance called the Sinkhorn divergence [@Cuturi13], in which the underlying transport problem is regularized to make it more tractable. While less well-characterized theoretically with respect to embedding capacity, the Sinkhorn divergence is computed efficiently by a fixed-point iteration. Moreover, recent work has shown that it is suitable for gradient-based optimization via automatic differentiation [@GenevayPC18]. To our knowledge, our work is the first to explore embedding properties of the Sinkhorn divergence.
We empirically investigate two settings for Wasserstein embeddings. First, we demonstrate their representational capacity by embedding a variety of complex networks, for which Wasserstein embeddings achieve higher fidelity than both Euclidean and hyperbolic embeddings. Second, we compute Wasserstein word embeddings, which show retrieval performance comparable to existing methods. One major benefit of our embedding is that the distributions can be visualized directly, unlike most embeddings, which require a dimensionality reduction step such as t-SNE before visualization. We demonstrate the power of this approach by visualizing the learned word embeddings.
Preliminaries
=============
Optimal transport and Wasserstein distance
------------------------------------------
The [**$p$-Wasserstein distance**]{} between probability distributions $\mu$ and $\nu$ [over a metric space $\cal X$]{} is $$\begin{aligned}
\label{eq:wasserstein-distance}
{\cal W}_p(\mu, \nu ) = \left(\inf_{\pi \in \Pi(\mu, \nu)} \int_{{\cal X} \times {\cal X}} d(x_1,x_2)^p \,d\pi(x_1, x_2)\right)^{\frac{1}{p}},
\end{aligned}$$ where the infimum is taken over [**transport plans**]{} $\pi$ that distribute the mass in $\mu$ to match that in $\nu$, with the $p$-th power of the [**ground metric**]{} $d(x_1, x_2)$ on $\mathcal{X}$ giving the cost of moving a unit of mass from support point $x_1\in\cal X$ underlying distribution $\mu$ to point $x_2\in\cal X$ underlying $\nu$. The Wasserstein distance is the cost of the optimal transport plan matching $\mu$ and $\nu$ [@villani2003topics].
In this paper, we are concerned with [**discrete distributions**]{} supported on finite sets of points in $\mathbb R^n$: $$\mu = \sum_{i=1}^M \mathbf{u}_i \delta_{\mathbf{x}^{(i)}}
\qquad\qquad\textrm{and}\qquad\qquad
\nu = \sum_{i=1}^N \mathbf{v}_i \delta_{\mathbf{y}^{(i)}}.$$ Here, $\mathbf u$ and $\mathbf v$ are vectors of nonnegative weights summing to $1$, and $\{\mathbf x^{(i)}\}_{i=1}^M,\{\mathbf y^{(i)}\}_{i=1}^N\subset\mathbb R^n$ are the **support points**. In this case, the transport plan $\pi$ matching $\mu$ and $\nu$ in becomes discrete as well, supported on the product of the two support sets. Define $D\in\mathbb R_+^{M\times N}$ to be the matrix of pairwise ground metric distances, with $D_{ij}=d(\mathbf x^{(i)}, \mathbf y^{(j)})$. Then, for discrete distributions, is equivalent to solving the following: $$\begin{aligned}
{\mathcal W}_p(\mu,\nu)^p = \min_{T\ge 0} \; {\text{tr}}(D^p T^\top) \quad \text{ subject to } \quad T {\boldsymbol}{1} = {\boldsymbol}{u}, \quad T^\top{\boldsymbol}{1} = {\boldsymbol}{v}, \label{eq:OT}
\end{aligned}$$ with $T_{ij}$ giving the transported mass between $\mathbf{x}_i$ and $\mathbf{y}_j$. The power $D^p$ is taken elementwise.
Sinkhorn divergence {#sec:prelim-sinkhorn}
-------------------
Equation \[eq:OT\] is a linear program that can be challenging to solve in practice. To improve efficiency, recent learning algorithms use an entropic regularizer proposed by @Cuturi13. The resulting **Sinkhorn divergence** solves a modified version of : $$\begin{aligned}
\mathcal W_p^\lambda (\mu,\nu)^p=
\min_{T\ge 0} \; {\text{tr}}(D^p T^\top) &+ \lambda {\text{tr}}\left(T (\log(T) - {\boldsymbol}{1}{\boldsymbol}{1}^\top)^\top\right)\; \text{ s.t. }\; T {\boldsymbol}{1} = {\boldsymbol}{u}, \; T^\top{\boldsymbol}{1} = {\boldsymbol}{v},\label{eq:Sinkhorn}
\end{aligned}$$ where $\log(\cdot)$ is applied elementwise and $\lambda\ge 0$ is the regularization parameter. For $\lambda>0$, the optimal solution takes the form $T^* = \Delta({\boldsymbol}{r}) \exp\left(\nicefrac{-D^p}{\lambda}\right) \Delta({\boldsymbol}{c})$, where $\Delta({\boldsymbol}{r})$ and $\Delta({\boldsymbol}{c})$ are diagonal matrices with diagonal elements ${\boldsymbol}{r}$ and ${\boldsymbol}{c}$, resp. Rather than optimizing over matrices $T$, one can optimize for ${\boldsymbol}{r}$ and ${\boldsymbol}{c}$, reducing the size of the problem to $M+N$. This can be solved via *matrix balancing*, starting from an initial matrix $K := \exp(\frac{-D^p}{\lambda})$ and alternately projecting onto the marginal constraints until convergence: $$\begin{aligned}
\label{eq:sinkhorniter}
{\boldsymbol}{r} \leftarrow {\boldsymbol}{u} ./K{\boldsymbol}{c} \qquad\qquad
{\boldsymbol}{c} \leftarrow {\boldsymbol}{v}./ K^\top{\boldsymbol}{r}.\end{aligned}$$ Here, $./$ denotes elementwise division for vectors.
Beyond simplicity of implementation, has an additional advantage for machine learning: The steps of this algorithm are differentiable. With this observation in mind, @GenevayPC18 incorporate entropic transport into learning pipelines by applying automatic differentiation (back propagation) to a fixed number of Sinkhorn iterations.
What can we embed in theory? {#sec:what-is-embeddable}
----------------------------
Given two metric spaces $\mathcal{A}$ and $\mathcal{B}$, an embedding of $\mathcal{A}$ into $\mathcal{B}$ is a map $\phi : \mathcal{A} \rightarrow \mathcal{B}$ that approximately preserves distances, in the sense that the distortion is small: $$\label{eq:distortion}
L d_{\mathcal{A}}(u, v) \leq d_{\mathcal{B}}(\phi(u), \phi(v)) \leq C L d_{\mathcal{A}}(u, v), \quad \forall u, v \in \mathcal{A},$$ for some uniform constants $L > 0$ and $C \geq 1$. The distortion of the embedding $\phi$ is the smallest $C$ such that holds.
One can characterize how “large” a space is (its *representational capacity*) by the spaces that embed into it with low distortion. In practical terms, this capacity determines the types of data (and relationships between them) that can be well-represented in the embedding space. $\mathbb{R}^n$ with the Euclidean metric, for example, embeds into the $L^1$ metric with low distortion, while the reverse is not true [@Deza09]. We do not expect Manhattan-structured data to be well-represented in Euclidean space, no matter how clever the mapping.
Wasserstein spaces are very large: Many spaces can embed into Wasserstein spaces with low distortion, even when the converse is not true. $\mathcal{W}_p(\mathcal{A})$, for $\mathcal{A}$ an arbitrary metric space, embeds any product space $\mathcal{A}^n$, for example [@kloeckner2010geometric], via discrete distributions supported at $n$ points. Even more generally, certain Wasserstein spaces are [**universal**]{}, in the sense that they can embed arbitrary metrics on finite spaces. $\mathcal{W}_1(\ell^1)$ is one such space [@bourgain1986metrical], and it is still an open problem to determine if $\mathcal{W}_1(\mathbb{R}^k)$ is universal for any $k < +\infty$. Recently it has been shown that every finite metric space embeds the $\frac{1}{p}$ power of its metric into $\mathcal{W}_p(\mathbb{R}^3)$, $p > 1$, with vanishing distortion [@andoni2015snowflake]. A hopeful interpretation suggests that $\mathcal{W}_1(\mathbb{R}^3)$ may be a plausible target space for arbitrary metrics on symbolic data, with a finite set of symbols; we are unaware of similar universality results for $L^p$ or hyperbolic spaces, for example.
The reverse direction, embedding Wasserstein spaces into others, is well-studied in the case of discrete distributions. Theoretical results in this domain are motivated by interest in efficient algorithms for approximating Wasserstein distances by embedding into spaces with easily-computed metrics. In this direction, low-distortion embeddings are difficult to find. $\mathcal{W}_2(\mathbb{R}^3)$, for example, is known not to embed into $L^1$ [@andoni2016impossibility]. Some positive results exist, nevertheless. For a Euclidean ground metric, for example, the $1$-Wasserstein distance can be approximated in a wavelet domain [@Shirdhonkar:2008approximate] or by high-dimensional embedding into $L^1$ [@Indyk:2003fast].
In §\[sec:empirical\], we empirically investigate the embedding capacity of Wasserstein spaces, by attempting to learn low-distortion embeddings for a variety of input spaces. For efficiency, we replace the Wasserstein distance by its entropically-regularized counterpart, the Sinkhorn divergence (§\[sec:prelim-sinkhorn\]). The embedding capacity of Sinkhorn divergences is previously unstudied, to our knowledge, except in the weak sense that the approximation error with respect to the Wasserstein distance vanishes with the regularizer taken to zero [@carlier2017convergence; @genevay2018sample].
Related work
------------
While learned vector space embeddings have a long history, there is a recent trend in the representation learning community towards more complex target spaces, such as spaces of probability distributions [@VilnisM14; @athiwaratkun2018on], Euclidean norm balls [@MirzazadehRDS15; @Farza17], Poincaré balls [@NickelK17], and Riemannian manifolds [@NickelK18]. From a modeling perspective, these more complex structures assist in representing uncertainty about the objects being embedded [@VilnisM14; @bojchevski2018deep] as well as complex relations such as inclusion, exclusion, hierarchy, and ordering [@MirzazadehRDS15; @Vendrov15; @athiwaratkun2018on]. In the same vein, our work takes probability distributions in a Wasserstein space as our embedding targets.
The distance or discrepancy measure between target structures is a major defining factor for a representation learning model. $L_p$ distances as well as angle-based discrepancies are fairly common [@MikolovNIPS13], as is the KL divergence [@kullback1951information], when embedding into probability distributions. For distributions, however, the KL divergence and $L_p$ distances are problematic, in the sense that they ignore the geometry of the domain of the distributions being compared. For distributions with disjoint support, for example, these divergences do not depend on the separation between the supports. Optimal transport distances [@villani2008optimal; @Cuturi13; @peyre2017computational; @Solomon18], on the other hand, explicitly account for the geometry of the domain. Hence, models based on optimal transport are gaining popularity in machine learning; see [@rubner1998metric; @courty2014domain; @FrognerZMAP15; @kusner2015word; @arjovsky2017wasserstein; @GenevayPC18; @ClaiciCS18; @singh2018context] for some examples.
Learned embeddings into Wasserstein spaces are relatively unexplored. Recent research proposes embedding into Gaussian distributions [@Muzellec18; @Zhu18]. Restricting to parametric distributions enables closed-form expressions for transport distances, but the resulting representation space may lose expressiveness. We note that @courty2018learning study embedding in the opposite direction, from Wasserstein into Euclidean space. In contrast, we learn to embed into the space of discrete probability distributions endowed with the Wasserstein distance. Discrete distributions are dense in $\mathcal W_2$ [@kloeckner2012approximation; @brancolini2009long].
Learning Wasserstein embeddings
===============================
The learning problem
--------------------
We consider the task of recovering a pairwise distance or similarity relationship that may be only partially observed. We are given a collection of objects $\mathcal{C}$—these can be words, symbols, images, or any other data—as well as samples $\left\{\left(u^{(i)}, v^{(i)}, r(u^{(i)}, v^{(i)})\right)\right\}$ of a target relationship $r : \mathcal{C} \times \mathcal{C} \rightarrow \mathbb{R}$ that tells us the degree to which pairs of objects are related.
Our objective is to find a map $\phi : \mathcal{C} \rightarrow \mathcal{W}_p(\mathcal{X})$ such that the relationship $r(u, v)$ can be recovered from the Wasserstein distance between $\phi(u)$ and $\phi(v)$, for any $u, v \in \mathcal{C}$. Examples include:
1. <span style="font-variant:small-caps;">Metric embedding:</span> $r$ is a distance metric, and we want $\mathcal{W}_p(\phi(u), \phi(v)) \approx r(u, v)$ for all $u, v \in \mathcal{C}$.
2. <span style="font-variant:small-caps;">Graph embedding:</span> $\mathcal{C}$ contains the vertices of a graph and $r : \mathcal{C} \times \mathcal{C} \rightarrow \{0, 1\}$ is the adjacency relation; we would like the neighborhood of each $\phi(u)$ in $\mathcal{W}_p$ to coincide with graph adjacency.
3. <span style="font-variant:small-caps;">Word embedding:</span> $\mathcal{C}$ contains individual words and $r$ is a semantic similarity between words. We want distances in $\mathcal{W}_p$ to predict this semantic similarity.
Although the details of each task require some adjustment to the learning architecture, our basic representation and training procedure detailed below applies to all three examples.
Optimization {#sec:learning-optimization}
------------
Given a set of training samples $\mathcal{S} = \left\{\left(u^{(i)}, v^{(i)}, r^{(i)}\right)\right\}_{i=1}^N \subset \mathcal{C} \times \mathcal{C} \times \mathbb{R}$, we want to learn a map $\phi : \mathcal{C} \rightarrow \mathcal{W}_p(\mathcal{X})$. We must address two issues.
First we must define the range of our map $\phi$. The whole of $\mathcal{W}_p(\mathcal{X})$ is infinite-dimensional, and for a tractable problem we need a finite-dimensional output. We restrict ourselves to discrete distributions with an *a priori* fixed number of support points $M$, reducing optimal transport to the linear program in . Such a distribution is parameterized by the locations of its support points $\{\mathbf{x}^{(j)}\}_{j=1}^M$, forming a point cloud in the ground metric space $\mathcal{X}$. For simplicity, we restrict to uniform weights $\mathbf u,\mathbf v\propto\mathbf 1$, although it is possible to optimize simultaneously over weights and locations. As noted in [@brancolini2009long; @kloeckner2012approximation; @ClaiciCS18], however, when constructing a discrete $M$-point approximation to a fixed target distribution, allowing non-uniform weights does not improve the asymptotic approximation error.[^1]
The second issue is that, as noted in §\[sec:prelim-sinkhorn\], exact computation of $\mathcal{W}_p$ in general is costly, requiring the solution of a linear program. As in [@GenevayPC18], we replace $\mathcal{W}_p$ with the Sinkhorn divergence $\mathcal{W}_p^{\lambda}$, which is solvable by a the fixed-point iteration in . Learning then takes the form of empirical loss minimization: $$\label{eq:optimization-loss-minimization}
\phi_{\ast} = \operatorname*{arg\,min}_{\phi \in \mathcal{H}} \frac{1}{N} \sum_{i=1}^N \mathcal{L}\left(\mathcal{W}_p^{\lambda}\left(\phi(u^{(i)}), \phi(v^{(i)})\right), r^{(i)}\right),$$ over a hypothesis space of maps $\mathcal{H}$. The loss $\mathcal{L}$ is problem-specific and scores the similarity between the regularized Wasserstein distance $\mathcal{W}_p^\lambda$ and the target relationship $r$ at $\left(u^{(i)}, v^{(i)}\right)$. As mentioned in §\[sec:prelim-sinkhorn\], gradients are available from automatic differentiation of the Sinkhorn procedure, and hence with a suitable loss function the learning objective can be optimized by standard gradient-based methods. In our experiments, we use the Adam optimizer [@KingmaB14].
Empirical study {#sec:empirical}
===============
Representational capacity: embedding complex networks
-----------------------------------------------------
We first demonstrate the representational power of learned Wasserstein embeddings. As discussed in §\[sec:what-is-embeddable\], theory suggests that Wasserstein spaces are quite flexible, in that they can embed a wide variety of metrics with low distortion. We show that this is true in practice as well.
To generate a variety of metrics to embed, we take networks with various patterns of connectivity and compute the shortest-path distances between vertices. The collection of vertices for each network serves as the input space $\mathcal{C}$ for our embedding, and our goal is to learn a map $\phi : \mathcal{C} \rightarrow \mathcal{W}_p(\mathbb{R}^k)$ such that the $1$-Wasserstein distance $\mathcal{W}_1(\phi(u), \phi(v))$ matches as closely as possible the shortest path distance between vertices $u$ and $v$, for all pairs of vertices. We learn a [**minimum-distortion embedding**]{}: Given a fully-observed distance metric $d_{\mathcal{C}} : \mathcal{C} \times \mathcal{C} \rightarrow \mathbb{R}$ in the input space, we minimize the mean distortion: $$\phi_{\ast} = \operatorname*{arg\,min}_{\phi} \frac{1}{{n \choose 2}} \sum_{j > i} \frac{|\mathcal{W}_1^{\lambda}(\phi(v_i), \phi(v_j)) - d_{\mathcal{C}}(v_i, v_j)|}{d_{\mathcal{C}}(v_i, v_j)}.$$ $\phi$ is parameterized as in §\[sec:learning-optimization\], directly specifying the support points of the output distribution.
We examine the performance of Wasserstein embedding using both random networks and real networks. The random networks in particular allow us systematically to test robustness of the Wasserstein embedding to particular properties of the metric we are attempting to embed. Note that these experiments do not explore generalization performance: We are purely concerned with the representational capacity of the learned Wasserstein embeddings.
For random networks, we use three standard generative models: Barabási–Albert [@albert2002statistical], Watts–Strogatz [@watts1998collective], and the stochastic block model [@holland1983stochastic]. [**Random scale-free networks**]{} are generated from the Barabási–Albert model, and possess the property that distances are on average much shorter than in a Euclidean spatial graph, scaling like the logarithm of the number of vertices. [**Random small-world networks**]{} are generated from the Watts–Strogatz model; in addition to log-scaling of the average path length, the vertices of Watts–Strogatz graphs cluster into distinct neighborhoods. [**Random community-structured networks**]{} are generated from the stochastic block model, which places vertices within densely-connected communities, with sparse connections between the different communities. We additionally generate [**random trees**]{} by choosing a random number of children[^2] for each node, progressing in breadth-first order until a specified total number of nodes is reached. In all cases, we generate networks with $128$ vertices.
We compare against two baselines, trained using the same distortion criterion and optimization method: [**Euclidean embeddings**]{}, and [**hyperbolic embeddings**]{}. Euclidean embeddings we expect to perform poorly on all of the chosen graph types, since they are limited to spatial relationships with zero curvature. Hyperbolic embeddings model tree-structured metrics, capturing the exponential scaling of graph neighborhoods; they have been suggested for a variety of other graph families as well [@zhao2011fast].
[0.49]{} ![Random networks: Learned Wasserstein embeddings achieve lower distortion than Euclidean/hyperbolic embeddings. Hyperbolic embeddings outperform specifically on random trees.[]{data-label="fig:empirical-capacity-random-networks"}](capacity_barabasi_distortion "fig:"){width="\textwidth"}
[0.49]{} ![Random networks: Learned Wasserstein embeddings achieve lower distortion than Euclidean/hyperbolic embeddings. Hyperbolic embeddings outperform specifically on random trees.[]{data-label="fig:empirical-capacity-random-networks"}](capacity_watts_distortion "fig:"){width="\textwidth"}
[0.49]{} ![Random networks: Learned Wasserstein embeddings achieve lower distortion than Euclidean/hyperbolic embeddings. Hyperbolic embeddings outperform specifically on random trees.[]{data-label="fig:empirical-capacity-random-networks"}](capacity_block_distortion "fig:"){width="\textwidth"}
[0.49]{} ![Random networks: Learned Wasserstein embeddings achieve lower distortion than Euclidean/hyperbolic embeddings. Hyperbolic embeddings outperform specifically on random trees.[]{data-label="fig:empirical-capacity-random-networks"}](capacity_tree_distortion "fig:"){width="\textwidth"}
Figure \[fig:empirical-capacity-random-networks\] shows the result of embedding random networks.[^3] As the total embedding dimension increases, the distortion decreases for all methods. Importantly, Wasserstein embeddings achieve lower distortion than Euclidean and hyperbolic embeddings, establishing their flexibility under the varying conditions represented by the different network models. In some cases, the Wasserstein distortion continues to decrease long after the other embeddings have saturated their capacity. As expected, hyperbolic space significantly outperforms both Euclidean and Wasserstein specifically on tree-structured metrics.
We test $\mathbb{R}^2$, $\mathbb{R}^3$, and $\mathbb{R}^4$ as ground metric spaces. For all of the random networks we examined, the performance between $\mathbb{R}^3$ and $\mathbb{R}^4$ is nearly indistinguishable. This observation is consistent with theoretical results (§\[sec:what-is-embeddable\]) suggesting that $\mathbb{R}^3$ is sufficient to embed a wide variety of metrics.
We also examine fragments of real networks: an ArXiv co-authorship network, an Amazon product co-purchasing network, and a Google web graph [@snapnets]. For each graph fragment, we choose uniformly at random a starting vertex and then extract the subgraph on $128$ vertices taken in breadth-first order from that starting vertex. Distortion results are shown in Figure \[fig:empirical-capacity-real-networks\]. Again the Wasserstein embeddings achieve lower distortion than Euclidean or hyperbolic embeddings.
[0.32]{} ![Real networks: Learned Wasserstein embeddings achieve lower distortion than Euclidean and hyperbolic embeddings of real network fragments.[]{data-label="fig:empirical-capacity-real-networks"}](capacity_condmat_distortion "fig:"){width="\textwidth"}
[0.32]{} ![Real networks: Learned Wasserstein embeddings achieve lower distortion than Euclidean and hyperbolic embeddings of real network fragments.[]{data-label="fig:empirical-capacity-real-networks"}](capacity_amazon_distortion "fig:"){width="\textwidth"}
[0.32]{} ![Real networks: Learned Wasserstein embeddings achieve lower distortion than Euclidean and hyperbolic embeddings of real network fragments.[]{data-label="fig:empirical-capacity-real-networks"}](capacity_webgoogle_distortion "fig:"){width="\textwidth"}
word2cloud: Wasserstein word embeddings
---------------------------------------
In this section, we embed words as point clouds. In a sentence $s = (\mathbf{x}_0, \dots, \mathbf{x}_n)$, a word $\mathbf{x}_i$ is associated with word $\mathbf{x}_j$ if $\mathbf{x}_j$ is in the *context* of $\mathbf{x}_i$, which is a symmetric window around $\mathbf{x}_i$. This association is encoded as a label $r$; $r_{\mathbf{x}_i,\mathbf{x}_j}=1$ if and only if $|i-j|\le l$ where $l$ is the window size. For word embedding, we use a contrastive loss function [@HadsellCL06] $$\phi_{\ast} = \operatorname*{arg\,min}_{\phi} \sum_s \sum_{\mathbf{x}_i,\mathbf{x}_j \in s}r_{\mathbf{x}_i,\mathbf{\mathbf{x}}_j} \Big({\mathcal W}_1^\lambda \big( \phi(\mathbf{x}_i), \phi({\mathbf x}_j)\big)\Big)^2 + (1- r_{\mathbf{x}_i,\mathbf{x}_j}) \Big(\Big[m - {\mathcal W}_1^\lambda\big(\phi (\mathbf{x}_i), \phi(\mathbf{x}_j)\big)\Big]_+\Big)^2,$$ which tries to embed words $\mathbf{x}_i, \mathbf{x}_j$ near each other in terms of $1$-Wasserstein distance (here ${\mathcal W}_1^\lambda$) if they co-occur in the context; otherwise, it prefers moving them at least distance $m$ away from one another. This approach is similar to that suggested by @MikolovNIPS13, up to the loss and distance functions.
We use a Siamese architecture [@BromleyBBGLMSS93] for our model, with negative sampling (as in [@MikolovNIPS13]) for selecting words outside the context. The network architecture in each branch consists of a linear layer with $64$ nodes followed by our point cloud embedding layer. The two branches of the Siamese network connect via the Wasserstein distance, computed as in §\[sec:prelim-sinkhorn\]. The training dataset is [Text8]{}[^4], which consists of a corpus with $17M$ tokens from Wikipedia and is commonly used as a language modeling benchmark. We choose a vocabulary of $8000$ words and a context window size of $l = 2$ (i.e., 2 words on each side), $\lambda = 0.05$, number of epochs of $3$, negative sampling rate of $1$ per positive and Adam [@KingmaB14] for optimization.
We first study the effect of dimensionality of the point cloud on the quality of the semantic neighborhood captured by the embedding. We fix the total number of output parameters, being the product of the number of support points and the dimension of the support space, to $63$ or $64$ parameters. Table \[tbl:word2cloud5NN\] shows the 5 nearest neighbors in the embedding space. Notably, increasing the dimensionality directly improves the quality of the learned representation. Interestingly, it is more effective to use a budget of $64$ parameters in a $16$-point, $4$-dimensional cloud than in a $32$-point, $2$-dimensional cloud.
--------------------------------------- ---------- ------------------------------------------------------
one: f, two, i, after, four
${\mathcal W}_1^\lambda(\mathbb R^2)$ united: series, professional, team, east, central
algebra: skin, specified, equation, hilbert, reducing
one: two, three, s, four, after
${\mathcal W}_1^\lambda(\mathbb R^3)$ united: kingdom, australia, official, justice, officially
algebra: binary, distributions, reviews, ear, combination
one: six, eight, zero, two, three
${\mathcal W}_1^\lambda(\mathbb R^4)$ united: army, union, era, treaty, federal
algebra: tables, transform, equations, infinite, differential
--------------------------------------- ---------- ------------------------------------------------------
: Change in the 5-nearest neighbors when increasing dimensionality of each point cloud with fixed total length of representation.[]{data-label="tbl:word2cloud5NN"}
------------ ------- --------------------------------------- ----------------------------------------- ----------------------------------------- ------ ------- ------ ------ ------
$\#$ ${\mathcal W}_1^\lambda(\mathbb R^2)$ ${{\mathcal W}_1^\lambda(\mathbb R^3)}$ ${{\mathcal W}_1^\lambda(\mathbb R^4)}$ R M S G W
Task Name Pairs 17M 17M 17M — 63M 631M 900M 100B
RG-65 65 0.32 0.67 0.81 0.27 -0.02 0.50 0.66 0.54
WS-353 353 0.15 0.27 0.33 0.24 0.10 0.49 0.62 0.64
WS-353-S 203 0.23 0.40 0.44 0.36 0.15 0.61 0.70 0.70
WS-353-R 252 0.05 0.19 0.21 0.18 0.09 0.40 0.56 0.61
MC-30 30 0.04 0.45 0.54 0.47 -0.14 0.57 0.66 0.63
Rare-Word 2034 0.06 0.22 0.10 0.29 0.11 0.39 0.06 0.39
MEN 3000 0.25 0.28 0.26 0.24 0.09 0.57 0.31 0.65
MTurk-287 287 0.40 0.38 0.49 0.33 0.09 0.59 0.36 0.67
MTurk-771 771 0.11 0.23 0.25 0.26 0.10 0.50 0.32 0.57
SimLex-999 999 0.09 0.05 0.07 0.23 0.01 0.27 0.10 0.31
Verb-143 144 0.03 0.03 0.16 0.29 0.06 0.36 0.44 0.27
------------ ------- --------------------------------------- ----------------------------------------- ----------------------------------------- ------ ------- ------ ------ ------
: Performance on a number of similarity benchmarks when dimensionality of point clouds increase given a fixed total number of parameters. The middle block shows the performance of the proposed models. The right block shows the performance of baselines. The training corpus size when known appears below each model name. []{data-label="tbl:sim_benchmarks"}
Next we evaluate these models on a number of benchmark retrieval tasks from [@faruqui-2014], which score a method by the correlation of its output similarity scores with human similarity judgments, for various pairs of words. Results are shown in Table \[tbl:sim\_benchmarks\]. The results of our method, which use Sinkhorn distance to compute the point cloud (dis)similarities, appear in the middle block of Table \[tbl:sim\_benchmarks\]. Again, we mainly see gradual improvement with increasing dimensionality of the point clouds. The right block in Table \[tbl:sim\_benchmarks\] shows baselines: Respectively, RNN(80D) [@KombrinkMKB11], Metaoptimize (50D) [@TurianRB10], SENNA (50D) [@Collobert11] Global Context (50D) [@huang2012improving] and word2vec (80D) [@MikolovNIPS13]. In the right block, as in [@faruqui-2014], the cosine similarity is used for point embeddings. The reported performance measure is the correlation with ground-truth rankings, computed as in [@faruqui-2014]. Note there are many ways to improve the performance: increasing the vocabulary/window size/number of epochs/negative sampling rate, using larger texts, and accelerating performance. We defer this tuning to future work focused specifically on NLP.
### Direct, interpretable visualization of high-dimensional embeddings
Wasserstein embeddings over low-dimensional ground metric spaces have a unique property: We can [**directly visualize the embedding**]{}, which is a point cloud in the low-dimensional ground space. This is not true for most existing embedding methods, which rely on dimensionality reduction techniques such as t-SNE for visualization. Whereas dimensionality reduction only approximately captures proximity of points in the embedding space, with Wasserstein embeddings we can display the exact embedding of each input, by visualizing the point cloud.
We demonstrate this property by visualizing the learned word representations. Importantly, each point cloud is strongly clustered, which leads to apparent, distinct modes in its density. We therefore use kernel density estimation to visualize the densities. In Figure \[fig:interpret-0-0\], we visualize three distinct words, thresholding each density at a low value and showing its upper level set to reveal the modes. These level sets are overlaid, with each color in the figure corresponding to a distinct embedded word. The density for each word is depicted by the opacity of the color within each level set.
It is easy to visualize multiple sets of words in aggregate, by assigning all words in a set a single color. This immediately reveals how well-separated the sets are, as shown in Figure \[fig:interpret-0-1\]: As expected, military and political terms overlap, while names of sports are more distant.
Examining the embeddings in more detail, we can dissect relationships (and confusion) between different sets of words. We observe that each word tends to concentrate its mass in two or more distinct regions. This multimodal shape allows for multifaceted relationships between words, since a word can partially overlap with many distinct groups of words simultaneously. Figure \[fig:interpret-0-2\] shows the embedding for a word that has multiple distinct meanings ([kind]{}), alongside synonyms for both senses of the word ([nice, friendly, type]{}). We see that [kind]{} has two primary modes, which overlap separately with [friendly]{} and [type]{}. [nice]{} is included to show a failure of the embedding to capture the full semantics: Figure \[fig:interpret-0-3\] shows that the network has learned that [nice]{} is a city in France, ignoring its interpretation as an adjective. This demonstrates the potential of this visualization for debugging, helping identify and attribute an error.
[0.49]{} ![Directly visualizing high-dimensional word embeddings.[]{data-label="fig:interpret-0"}](interpret_three_words "fig:"){width="\textwidth"}
[0.49]{} ![Directly visualizing high-dimensional word embeddings.[]{data-label="fig:interpret-0"}](interpret_separation_sports_politics_military "fig:"){width="\textwidth"}
[0.49]{} ![Directly visualizing high-dimensional word embeddings.[]{data-label="fig:interpret-0"}](interpret_multimeaning_kind "fig:"){width="\textwidth"}
[0.49]{} ![Directly visualizing high-dimensional word embeddings.[]{data-label="fig:interpret-0"}](interpret_multimeaning_nice "fig:"){width="\textwidth"}
Discussion and conclusion
=========================
Several characteristics determine the value and effectiveness of an embedding space for representation learning. The space must be large enough to embed a variety of metrics, while admitting a mathematical description compatible with learning algorithms; additional features, including direct interpretability, make it easier to understand, analyze, and potentially debug the output of a representation learning procedure. Based on their theoretical properties, Wasserstein spaces are strong candidates for representing complex semantic structures, when the capacity of Euclidean space does not suffice. Empirically, entropy-regularized Wasserstein distances are effective for embedding a wide variety of semantic structures, while enabling direct visualization of the embedding.
Our work suggests several directions for additional research. Beyond simple extensions like weighting points in the point cloud, one observation is that we can lift nearly *any* representation space $\mathcal X$ to distributions over that space $\mathcal W(\mathcal X)$ represented as point clouds; in this paper we focused on the case $\mathcal X =\mathbb R^n$. Since $\mathcal X$ embeds within $\mathcal W(\mathcal X)$ using $\delta$-functions, this might be viewed as a general “lifting” procedure increasing the capacity of a representation. We can also consider other tasks, such as co-embedding of different modalities into the same transport space. Additionally, our empirical results suggest that theoretical study of the embedding capacity of Sinkhorn divergences may be profitable. Finally, following recent work on computing geodesics in Wasserstein space [@seguy2015principal], it may be interesting to invert the learned mappings and use them for interpolation.
### Acknowledgements {#acknowledgements .unnumbered}
J. Solomon acknowledges the support of Army Research Office grant W911NF-12-R-0011 (“Smooth Modeling of Flows on Graphs”), of National Science Foundation grant IIS-1838071 (“BIGDATA:F: Statistical and Computational Optimal Transport for Geometric Data Analysis”), from an Amazon Research Award, from the MIT–IBM Watson AI Laboratory, and from the Skoltech–MIT Next Generation Program.
[^1]: In both the non-uniform and uniform cases, the order of convergence in $\mathcal{W}_p$ of the nearest weighted point cloud to the target measure, as we add more points, is $\mathcal{O}(M^{-1/d})$, for a $d$-dimensional ground metric space. This assumes the underlying measure is absolutely continuous and compactly-supported.
[^2]: Each non-leaf node has a number of children drawn uniformly from $\{2, 3, 4\}$.
[^3]: The solid line is the median over $20$ randomly-generated inputs, while shaded is the middle $95\%$.
[^4]: From <http://mattmahoney.net/dc/text8.zip>
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.